X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/5d5c5d0d5b79ade9a973d55186ffda2638ba2b6e..4d15aeb193b2c68f1d38666c317f8d3734f5f083:/bsd/vfs/vfs_cluster.c?ds=sidebyside diff --git a/bsd/vfs/vfs_cluster.c b/bsd/vfs/vfs_cluster.c index ddcb1d373..70eecc5ff 100644 --- a/bsd/vfs/vfs_cluster.c +++ b/bsd/vfs/vfs_cluster.c @@ -1,31 +1,29 @@ /* - * Copyright (c) 2000-2004 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2014 Apple Inc. All rights reserved. * - * @APPLE_LICENSE_OSREFERENCE_HEADER_START@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * - * This file contains Original Code and/or Modifications of Original Code - * as defined in and that are subject to the Apple Public Source License - * Version 2.0 (the 'License'). You may not use this file except in - * compliance with the License. The rights granted to you under the - * License may not be used to create, or enable the creation or - * redistribution of, unlawful or unlicensed copies of an Apple operating - * system, or to circumvent, violate, or enable the circumvention or - * violation of, any terms of an Apple operating system software license - * agreement. - * - * Please obtain a copy of the License at - * http://www.opensource.apple.com/apsl/ and read it before using this - * file. - * - * The Original Code and all software distributed under the License are - * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER - * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, - * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. - * Please see the License for the specific language governing rights and + * This file contains Original Code and/or Modifications of Original Code + * as defined in and that are subject to the Apple Public Source License + * Version 2.0 (the 'License'). You may not use this file except in + * compliance with the License. The rights granted to you under the License + * may not be used to create, or enable the creation or redistribution of, + * unlawful or unlicensed copies of an Apple operating system, or to + * circumvent, violate, or enable the circumvention or violation of, any + * terms of an Apple operating system software license agreement. + * + * Please obtain a copy of the License at + * http://www.opensource.apple.com/apsl/ and read it before using this file. + * + * The Original Code and all software distributed under the License are + * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER + * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, + * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. + * Please see the License for the specific language governing rights and * limitations under the License. - * - * @APPLE_LICENSE_OSREFERENCE_HEADER_END@ + * + * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ /* Copyright (c) 1995 NeXT Computer, Inc. All Rights Reserved */ /* @@ -73,101 +71,234 @@ #include #include #include +#include #include #include #include #include +#include #include #include #include #include +#include +#include #include #include #include +#include #include +#include +#include -#define CL_READ 0x01 -#define CL_ASYNC 0x02 -#define CL_COMMIT 0x04 -#define CL_PAGEOUT 0x10 -#define CL_AGE 0x20 -#define CL_DUMP 0x40 -#define CL_NOZERO 0x80 -#define CL_PAGEIN 0x100 -#define CL_DEV_MEMORY 0x200 -#define CL_PRESERVE 0x400 -#define CL_THROTTLE 0x800 -#define CL_KEEPCACHED 0x1000 +#include +#if 0 +#undef KERNEL_DEBUG +#define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT +#endif + + +#define CL_READ 0x01 +#define CL_WRITE 0x02 +#define CL_ASYNC 0x04 +#define CL_COMMIT 0x08 +#define CL_PAGEOUT 0x10 +#define CL_AGE 0x20 +#define CL_NOZERO 0x40 +#define CL_PAGEIN 0x80 +#define CL_DEV_MEMORY 0x100 +#define CL_PRESERVE 0x200 +#define CL_THROTTLE 0x400 +#define CL_KEEPCACHED 0x800 +#define CL_DIRECT_IO 0x1000 +#define CL_PASSIVE 0x2000 +#define CL_IOSTREAMING 0x4000 +#define CL_CLOSE 0x8000 +#define CL_ENCRYPTED 0x10000 +#define CL_RAW_ENCRYPTED 0x20000 +#define CL_NOCACHE 0x40000 + +#define MAX_VECTOR_UPL_ELEMENTS 8 +#define MAX_VECTOR_UPL_SIZE (2 * MAX_UPL_SIZE_BYTES) + +#define CLUSTER_IO_WAITING ((buf_t)1) + +extern upl_t vector_upl_create(vm_offset_t); +extern boolean_t vector_upl_is_valid(upl_t); +extern boolean_t vector_upl_set_subupl(upl_t,upl_t, u_int32_t); +extern void vector_upl_set_pagelist(upl_t); +extern void vector_upl_set_iostate(upl_t, upl_t, vm_offset_t, u_int32_t); struct clios { + lck_mtx_t io_mtxp; u_int io_completed; /* amount of io that has currently completed */ u_int io_issued; /* amount of io that was successfully issued */ int io_error; /* error code of first error encountered */ int io_wanted; /* someone is sleeping waiting for a change in state */ }; +struct cl_direct_read_lock { + LIST_ENTRY(cl_direct_read_lock) chain; + int32_t ref_count; + vnode_t vp; + lck_rw_t rw_lock; +}; + +#define CL_DIRECT_READ_LOCK_BUCKETS 61 + +static LIST_HEAD(cl_direct_read_locks, cl_direct_read_lock) + cl_direct_read_locks[CL_DIRECT_READ_LOCK_BUCKETS]; + +static lck_spin_t cl_direct_read_spin_lock; + static lck_grp_t *cl_mtx_grp; static lck_attr_t *cl_mtx_attr; static lck_grp_attr_t *cl_mtx_grp_attr; -static lck_mtx_t *cl_mtxp; +static lck_mtx_t *cl_transaction_mtxp; + +#define IO_UNKNOWN 0 +#define IO_DIRECT 1 +#define IO_CONTIG 2 +#define IO_COPY 3 + +#define PUSH_DELAY 0x01 +#define PUSH_ALL 0x02 +#define PUSH_SYNC 0x04 + + +static void cluster_EOT(buf_t cbp_head, buf_t cbp_tail, int zero_offset); +static void cluster_wait_IO(buf_t cbp_head, int async); +static void cluster_complete_transaction(buf_t *cbp_head, void *callback_arg, int *retval, int flags, int needwait); +static int cluster_io_type(struct uio *uio, int *io_type, u_int32_t *io_length, u_int32_t min_length); static int cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int non_rounded_size, - int flags, buf_t real_bp, struct clios *iostate); -static int cluster_iodone(buf_t bp, void *dummy); -static int cluster_rd_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize); -static int cluster_hard_throttle_on(vnode_t vp); + int flags, buf_t real_bp, struct clios *iostate, int (*)(buf_t, void *), void *callback_arg); +static int cluster_iodone(buf_t bp, void *callback_arg); +static int cluster_ioerror(upl_t upl, int upl_offset, int abort_size, int error, int io_flags, vnode_t vp); +static int cluster_is_throttled(vnode_t vp); -static int cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags); -static int cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, - off_t headOff, off_t tailOff, int flags); -static int cluster_nocopy_read(vnode_t vp, struct uio *uio, off_t filesize); -static int cluster_nocopy_write(vnode_t vp, struct uio *uio, off_t newEOF); -static int cluster_phys_read(vnode_t vp, struct uio *uio, off_t filesize); -static int cluster_phys_write(vnode_t vp, struct uio *uio, off_t newEOF); -static int cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, int xsize, int flags); +static void cluster_iostate_wait(struct clios *iostate, u_int target, const char *wait_name); -static void cluster_rd_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl_readahead *ra); +static void cluster_syncup(vnode_t vp, off_t newEOF, int (*)(buf_t, void *), void *callback_arg, int flags); -static int cluster_push_x(vnode_t vp, struct cl_extent *, off_t EOF, int flags); -static void cluster_push_EOF(vnode_t vp, off_t EOF); +static void cluster_read_upl_release(upl_t upl, int start_pg, int last_pg, int take_reference); +static int cluster_copy_ubc_data_internal(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty, int take_reference); -static int cluster_try_push(struct cl_writebehind *, vnode_t vp, off_t EOF, int can_delay, int push_all); +static int cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t filesize, int flags, + int (*)(buf_t, void *), void *callback_arg); +static int cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length, + int flags, int (*)(buf_t, void *), void *callback_arg); +static int cluster_read_contig(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length, + int (*)(buf_t, void *), void *callback_arg, int flags); -static void sparse_cluster_switch(struct cl_writebehind *, vnode_t vp, off_t EOF); -static void sparse_cluster_push(struct cl_writebehind *, vnode_t vp, off_t EOF, int push_all); -static void sparse_cluster_add(struct cl_writebehind *, vnode_t vp, struct cl_extent *, off_t EOF); +static int cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t oldEOF, off_t newEOF, + off_t headOff, off_t tailOff, int flags, int (*)(buf_t, void *), void *callback_arg); +static int cluster_write_direct(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, + int *write_type, u_int32_t *write_length, int flags, int (*)(buf_t, void *), void *callback_arg); +static int cluster_write_contig(vnode_t vp, struct uio *uio, off_t newEOF, + int *write_type, u_int32_t *write_length, int (*)(buf_t, void *), void *callback_arg, int bflag); -static kern_return_t vfs_drt_mark_pages(void **cmapp, off_t offset, u_int length, int *setcountp); +static int cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, u_int32_t xsize, int flags, int (*)(buf_t, void *), void *callback_arg); + +static int cluster_read_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize, int (*callback)(buf_t, void *), void *callback_arg, int bflag); +static void cluster_read_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl_readahead *ra, int (*callback)(buf_t, void *), void *callback_arg, int bflag); + +static int cluster_push_now(vnode_t vp, struct cl_extent *, off_t EOF, int flags, int (*)(buf_t, void *), void *callback_arg); + +static int cluster_try_push(struct cl_writebehind *, vnode_t vp, off_t EOF, int push_flag, int flags, int (*)(buf_t, void *), void *callback_arg, int *err); + +static void sparse_cluster_switch(struct cl_writebehind *, vnode_t vp, off_t EOF, int (*)(buf_t, void *), void *callback_arg); +static int sparse_cluster_push(void **cmapp, vnode_t vp, off_t EOF, int push_flag, int io_flags, int (*)(buf_t, void *), void *callback_arg); +static void sparse_cluster_add(void **cmapp, vnode_t vp, struct cl_extent *, off_t EOF, int (*)(buf_t, void *), void *callback_arg); + +static kern_return_t vfs_drt_mark_pages(void **cmapp, off_t offset, u_int length, u_int *setcountp); static kern_return_t vfs_drt_get_cluster(void **cmapp, off_t *offsetp, u_int *lengthp); static kern_return_t vfs_drt_control(void **cmapp, int op_type); -int is_file_clean(vnode_t, off_t); + +/* + * For throttled IO to check whether + * a block is cached by the boot cache + * and thus it can avoid delaying the IO. + * + * bootcache_contains_block is initially + * NULL. The BootCache will set it while + * the cache is active and clear it when + * the cache is jettisoned. + * + * Returns 0 if the block is not + * contained in the cache, 1 if it is + * contained. + * + * The function pointer remains valid + * after the cache has been evicted even + * if bootcache_contains_block has been + * cleared. + * + * See rdar://9974130 The new throttling mechanism breaks the boot cache for throttled IOs + */ +int (*bootcache_contains_block)(dev_t device, u_int64_t blkno) = NULL; + + +/* + * limit the internal I/O size so that we + * can represent it in a 32 bit int + */ +#define MAX_IO_REQUEST_SIZE (1024 * 1024 * 512) +#define MAX_IO_CONTIG_SIZE MAX_UPL_SIZE_BYTES +#define MAX_VECTS 16 +/* + * The MIN_DIRECT_WRITE_SIZE governs how much I/O should be issued before we consider + * allowing the caller to bypass the buffer cache. For small I/Os (less than 16k), + * we have not historically allowed the write to bypass the UBC. + */ +#define MIN_DIRECT_WRITE_SIZE (16384) + +#define WRITE_THROTTLE 6 +#define WRITE_THROTTLE_SSD 2 +#define WRITE_BEHIND 1 +#define WRITE_BEHIND_SSD 1 + +#define PREFETCH 3 +#define PREFETCH_SSD 2 +uint32_t speculative_prefetch_max = (MAX_UPL_SIZE_BYTES * 3); /* maximum bytes in a specluative read-ahead */ +uint32_t speculative_prefetch_max_iosize = (512 * 1024); /* maximum I/O size to use in a specluative read-ahead on SSDs*/ + + +#define IO_SCALE(vp, base) (vp->v_mount->mnt_ioscale * (base)) +#define MAX_CLUSTER_SIZE(vp) (cluster_max_io_size(vp->v_mount, CL_WRITE)) +#define MAX_PREFETCH(vp, size, is_ssd) (size * IO_SCALE(vp, ((is_ssd && !ignore_is_ssd) ? PREFETCH_SSD : PREFETCH))) + +int ignore_is_ssd = 0; +int speculative_reads_disabled = 0; /* * throttle the number of async writes that * can be outstanding on a single vnode * before we issue a synchronous write */ -#define HARD_THROTTLE_MAXCNT 0 -#define HARD_THROTTLE_MAXSIZE (64 * 1024) +#define THROTTLE_MAXCNT 0 + +uint32_t throttle_max_iosize = (128 * 1024); -int hard_throttle_on_root = 0; -struct timeval priority_IO_timestamp_for_root; +#define THROTTLE_MAX_IOSIZE (throttle_max_iosize) + +SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_max_iosize, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_max_iosize, 0, ""); void cluster_init(void) { - /* + /* * allocate lock group attribute and group */ - cl_mtx_grp_attr = lck_grp_attr_alloc_init(); + cl_mtx_grp_attr = lck_grp_attr_alloc_init(); cl_mtx_grp = lck_grp_alloc_init("cluster I/O", cl_mtx_grp_attr); /* @@ -175,20 +306,70 @@ cluster_init(void) { */ cl_mtx_attr = lck_attr_alloc_init(); - /* - * allocate and initialize mutex's used to protect updates and waits - * on the cluster_io context - */ - cl_mtxp = lck_mtx_alloc_init(cl_mtx_grp, cl_mtx_attr); + cl_transaction_mtxp = lck_mtx_alloc_init(cl_mtx_grp, cl_mtx_attr); + + if (cl_transaction_mtxp == NULL) + panic("cluster_init: failed to allocate cl_transaction_mtxp"); + + lck_spin_init(&cl_direct_read_spin_lock, cl_mtx_grp, cl_mtx_attr); + + for (int i = 0; i < CL_DIRECT_READ_LOCK_BUCKETS; ++i) + LIST_INIT(&cl_direct_read_locks[i]); +} + + +uint32_t +cluster_max_io_size(mount_t mp, int type) +{ + uint32_t max_io_size; + uint32_t segcnt; + uint32_t maxcnt; - if (cl_mtxp == NULL) - panic("cluster_init: failed to allocate cl_mtxp"); + switch(type) { + + case CL_READ: + segcnt = mp->mnt_segreadcnt; + maxcnt = mp->mnt_maxreadcnt; + break; + case CL_WRITE: + segcnt = mp->mnt_segwritecnt; + maxcnt = mp->mnt_maxwritecnt; + break; + default: + segcnt = min(mp->mnt_segreadcnt, mp->mnt_segwritecnt); + maxcnt = min(mp->mnt_maxreadcnt, mp->mnt_maxwritecnt); + break; + } + if (segcnt > (MAX_UPL_SIZE_BYTES >> PAGE_SHIFT)) { + /* + * don't allow a size beyond the max UPL size we can create + */ + segcnt = MAX_UPL_SIZE_BYTES >> PAGE_SHIFT; + } + max_io_size = min((segcnt * PAGE_SIZE), maxcnt); + + if (max_io_size < MAX_UPL_TRANSFER_BYTES) { + /* + * don't allow a size smaller than the old fixed limit + */ + max_io_size = MAX_UPL_TRANSFER_BYTES; + } else { + /* + * make sure the size specified is a multiple of PAGE_SIZE + */ + max_io_size &= ~PAGE_MASK; + } + return (max_io_size); } + #define CLW_ALLOCATE 0x01 #define CLW_RETURNLOCKED 0x02 +#define CLW_IONOCACHE 0x04 +#define CLW_IOPASSIVE 0x08 + /* * if the read ahead context doesn't yet exist, * allocate and initialize it... @@ -228,7 +409,7 @@ cluster_get_rap(vnode_t vp) else { lck_mtx_destroy(&rap->cl_lockr, cl_mtx_grp); FREE_ZONE((void *)rap, sizeof *rap, M_CLRDAHEAD); - rap = ubc->cl_rahead; + rap = ubc->cl_rahead; } vnode_unlock(vp); } @@ -277,7 +458,7 @@ cluster_get_wbp(vnode_t vp, int flags) else { lck_mtx_destroy(&wbp->cl_lockw, cl_mtx_grp); FREE_ZONE((void *)wbp, sizeof *wbp, M_CLWRBEHIND); - wbp = ubc->cl_wbehind; + wbp = ubc->cl_wbehind; } vnode_unlock(vp); } @@ -288,29 +469,251 @@ cluster_get_wbp(vnode_t vp, int flags) } -static int -cluster_hard_throttle_on(vnode_t vp) +static void +cluster_syncup(vnode_t vp, off_t newEOF, int (*callback)(buf_t, void *), void *callback_arg, int flags) { - static struct timeval hard_throttle_maxelapsed = { 0, 200000 }; + struct cl_writebehind *wbp; - if (vp->v_mount->mnt_kern_flag & MNTK_ROOTDEV) { - struct timeval elapsed; + if ((wbp = cluster_get_wbp(vp, 0)) != NULL) { + + if (wbp->cl_number) { + lck_mtx_lock(&wbp->cl_lockw); - if (hard_throttle_on_root) - return(1); + cluster_try_push(wbp, vp, newEOF, PUSH_ALL | flags, 0, callback, callback_arg, NULL); - microuptime(&elapsed); - timevalsub(&elapsed, &priority_IO_timestamp_for_root); + lck_mtx_unlock(&wbp->cl_lockw); + } + } +} + + +static int +cluster_io_present_in_BC(vnode_t vp, off_t f_offset) +{ + daddr64_t blkno; + size_t io_size; + int (*bootcache_check_fn)(dev_t device, u_int64_t blkno) = bootcache_contains_block; + + if (bootcache_check_fn) { + if (VNOP_BLOCKMAP(vp, f_offset, PAGE_SIZE, &blkno, &io_size, NULL, VNODE_READ, NULL)) + return(0); - if (timevalcmp(&elapsed, &hard_throttle_maxelapsed, <)) - return(1); + if (io_size == 0) + return (0); + + if (bootcache_check_fn(vp->v_mount->mnt_devvp->v_rdev, blkno)) + return(1); } return(0); } +static int +cluster_is_throttled(vnode_t vp) +{ + return (throttle_io_will_be_throttled(-1, vp->v_mount)); +} + + +static void +cluster_iostate_wait(struct clios *iostate, u_int target, const char *wait_name) +{ + + lck_mtx_lock(&iostate->io_mtxp); + + while ((iostate->io_issued - iostate->io_completed) > target) { + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START, + iostate->io_issued, iostate->io_completed, target, 0, 0); + + iostate->io_wanted = 1; + msleep((caddr_t)&iostate->io_wanted, &iostate->io_mtxp, PRIBIO + 1, wait_name, NULL); + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END, + iostate->io_issued, iostate->io_completed, target, 0, 0); + } + lck_mtx_unlock(&iostate->io_mtxp); +} + +static void cluster_handle_associated_upl(struct clios *iostate, upl_t upl, + upl_offset_t upl_offset, upl_size_t size) +{ + if (!size) + return; + + upl_t associated_upl = upl_associated_upl(upl); + + if (!associated_upl) + return; + +#if 0 + printf("1: %d %d\n", upl_offset, upl_offset + size); +#endif + + /* + * The associated UPL is page aligned to file offsets whereas the + * UPL it's attached to has different alignment requirements. The + * upl_offset that we have refers to @upl. The code that follows + * has to deal with the first and last pages in this transaction + * which might straddle pages in the associated UPL. To keep + * track of these pages, we use the mark bits: if the mark bit is + * set, we know another transaction has completed its part of that + * page and so we can unlock that page here. + * + * The following illustrates what we have to deal with: + * + * MEM u <------------ 1 PAGE ------------> e + * +-------------+----------------------+----------------- + * | |######################|################# + * +-------------+----------------------+----------------- + * FILE | <--- a ---> o <------------ 1 PAGE ------------> + * + * So here we show a write to offset @o. The data that is to be + * written is in a buffer that is not page aligned; it has offset + * @a in the page. The upl that carries the data starts in memory + * at @u. The associated upl starts in the file at offset @o. A + * transaction will always end on a page boundary (like @e above) + * except for the very last transaction in the group. We cannot + * unlock the page at @o in the associated upl until both the + * transaction ending at @e and the following transaction (that + * starts at @e) has completed. + */ + + /* + * We record whether or not the two UPLs are aligned as the mark + * bit in the first page of @upl. + */ + upl_page_info_t *pl = UPL_GET_INTERNAL_PAGE_LIST(upl); + bool is_unaligned = upl_page_get_mark(pl, 0); + + if (is_unaligned) { + upl_page_info_t *assoc_pl = UPL_GET_INTERNAL_PAGE_LIST(associated_upl); + + upl_offset_t upl_end = upl_offset + size; + assert(upl_end >= PAGE_SIZE); + + upl_size_t assoc_upl_size = upl_get_size(associated_upl); + + /* + * In the very first transaction in the group, upl_offset will + * not be page aligned, but after that it will be and in that + * case we want the preceding page in the associated UPL hence + * the minus one. + */ + assert(upl_offset); + if (upl_offset) + upl_offset = trunc_page_32(upl_offset - 1); + + lck_mtx_lock_spin(&iostate->io_mtxp); + + // Look at the first page... + if (upl_offset + && !upl_page_get_mark(assoc_pl, upl_offset >> PAGE_SHIFT)) { + /* + * The first page isn't marked so let another transaction + * completion handle it. + */ + upl_page_set_mark(assoc_pl, upl_offset >> PAGE_SHIFT, true); + upl_offset += PAGE_SIZE; + } + + // And now the last page... + + /* + * This needs to be > rather than >= because if it's equal, it + * means there's another transaction that is sharing the last + * page. + */ + if (upl_end > assoc_upl_size) + upl_end = assoc_upl_size; + else { + upl_end = trunc_page_32(upl_end); + const int last_pg = (upl_end >> PAGE_SHIFT) - 1; + + if (!upl_page_get_mark(assoc_pl, last_pg)) { + /* + * The last page isn't marked so mark the page and let another + * transaction completion handle it. + */ + upl_page_set_mark(assoc_pl, last_pg, true); + upl_end -= PAGE_SIZE; + } + } + + lck_mtx_unlock(&iostate->io_mtxp); + +#if 0 + printf("2: %d %d\n", upl_offset, upl_end); +#endif + + if (upl_end <= upl_offset) + return; + + size = upl_end - upl_offset; + } else { + assert(!(upl_offset & PAGE_MASK)); + assert(!(size & PAGE_MASK)); + } + + boolean_t empty; + + /* + * We can unlock these pages now and as this is for a + * direct/uncached write, we want to dump the pages too. + */ + kern_return_t kr = upl_abort_range(associated_upl, upl_offset, size, + UPL_ABORT_DUMP_PAGES, &empty); + + assert(!kr); + + if (!kr && empty) { + upl_set_associated_upl(upl, NULL); + upl_deallocate(associated_upl); + } +} + static int -cluster_iodone(buf_t bp, __unused void *dummy) +cluster_ioerror(upl_t upl, int upl_offset, int abort_size, int error, int io_flags, vnode_t vp) +{ + int upl_abort_code = 0; + int page_in = 0; + int page_out = 0; + + if ((io_flags & (B_PHYS | B_CACHE)) == (B_PHYS | B_CACHE)) + /* + * direct write of any flavor, or a direct read that wasn't aligned + */ + ubc_upl_commit_range(upl, upl_offset, abort_size, UPL_COMMIT_FREE_ON_EMPTY); + else { + if (io_flags & B_PAGEIO) { + if (io_flags & B_READ) + page_in = 1; + else + page_out = 1; + } + if (io_flags & B_CACHE) + /* + * leave pages in the cache unchanged on error + */ + upl_abort_code = UPL_ABORT_FREE_ON_EMPTY; + else if (page_out && ((error != ENXIO) || vnode_isswap(vp))) + /* + * transient error... leave pages unchanged + */ + upl_abort_code = UPL_ABORT_FREE_ON_EMPTY; + else if (page_in) + upl_abort_code = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR; + else + upl_abort_code = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_DUMP_PAGES; + + ubc_upl_abort_range(upl, upl_offset, abort_size, upl_abort_code); + } + return (upl_abort_code); +} + + +static int +cluster_iodone(buf_t bp, void *callback_arg) { int b_flags; int error; @@ -318,31 +721,63 @@ cluster_iodone(buf_t bp, __unused void *dummy) int total_resid; int upl_offset; int zero_offset; + int pg_offset = 0; + int commit_size = 0; + int upl_flags = 0; + int transaction_size = 0; upl_t upl; buf_t cbp; buf_t cbp_head; buf_t cbp_next; buf_t real_bp; + vnode_t vp; struct clios *iostate; - int commit_size; - int pg_offset; + boolean_t transaction_complete = FALSE; - cbp_head = (buf_t)(bp->b_trans_head); + __IGNORE_WCASTALIGN(cbp_head = (buf_t)(bp->b_trans_head)); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_START, - (int)cbp_head, bp->b_lblkno, bp->b_bcount, bp->b_flags, 0); + cbp_head, bp->b_lblkno, bp->b_bcount, bp->b_flags, 0); - for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) { - /* - * all I/O requests that are part of this transaction - * have to complete before we can process it - */ - if ( !(cbp->b_flags & B_DONE)) { + if (cbp_head->b_trans_next || !(cbp_head->b_flags & B_EOT)) { + lck_mtx_lock_spin(cl_transaction_mtxp); + + bp->b_flags |= B_TDONE; + + for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) { + /* + * all I/O requests that are part of this transaction + * have to complete before we can process it + */ + if ( !(cbp->b_flags & B_TDONE)) { + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, + cbp_head, cbp, cbp->b_bcount, cbp->b_flags, 0); + + lck_mtx_unlock(cl_transaction_mtxp); + + return 0; + } + + if (cbp->b_trans_next == CLUSTER_IO_WAITING) { + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, + cbp_head, cbp, cbp->b_bcount, cbp->b_flags, 0); + + lck_mtx_unlock(cl_transaction_mtxp); + wakeup(cbp); + + return 0; + } - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, - (int)cbp_head, (int)cbp, cbp->b_bcount, cbp->b_flags, 0); + if (cbp->b_flags & B_EOT) + transaction_complete = TRUE; + } + lck_mtx_unlock(cl_transaction_mtxp); - return 0; + if (transaction_complete == FALSE) { + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, + cbp_head, 0, 0, 0, 0); + return 0; } } error = 0; @@ -350,6 +785,7 @@ cluster_iodone(buf_t bp, __unused void *dummy) total_resid = 0; cbp = cbp_head; + vp = cbp->b_vp; upl_offset = cbp->b_uploffset; upl = cbp->b_upl; b_flags = cbp->b_flags; @@ -369,13 +805,45 @@ cluster_iodone(buf_t bp, __unused void *dummy) cbp_next = cbp->b_trans_next; - free_io_buf(cbp); + if (cbp_next == NULL) + /* + * compute the overall size of the transaction + * in case we created one that has 'holes' in it + * 'total_size' represents the amount of I/O we + * did, not the span of the transaction w/r to the UPL + */ + transaction_size = cbp->b_uploffset + cbp->b_bcount - upl_offset; + + if (cbp != cbp_head) + free_io_buf(cbp); cbp = cbp_next; } + + if (ISSET(b_flags, B_COMMIT_UPL)) { + cluster_handle_associated_upl(iostate, + cbp_head->b_upl, + upl_offset, + transaction_size); + } + + if (error == 0 && total_resid) + error = EIO; + + if (error == 0) { + int (*cliodone_func)(buf_t, void *) = (int (*)(buf_t, void *))(cbp_head->b_cliodone); + + if (cliodone_func != NULL) { + cbp_head->b_bcount = transaction_size; + + error = (*cliodone_func)(cbp_head, callback_arg); + } + } if (zero_offset) cluster_zero(upl, zero_offset, PAGE_SIZE - (zero_offset & PAGE_MASK), real_bp); + free_io_buf(cbp_head); + if (iostate) { int need_wakeup = 0; @@ -383,7 +851,7 @@ cluster_iodone(buf_t bp, __unused void *dummy) * someone has issued multiple I/Os asynchrounsly * and is waiting for them to complete (streaming) */ - lck_mtx_lock(cl_mtxp); + lck_mtx_lock_spin(&iostate->io_mtxp); if (error && iostate->io_error == 0) iostate->io_error = error; @@ -398,107 +866,91 @@ cluster_iodone(buf_t bp, __unused void *dummy) iostate->io_wanted = 0; need_wakeup = 1; } - lck_mtx_unlock(cl_mtxp); + lck_mtx_unlock(&iostate->io_mtxp); if (need_wakeup) wakeup((caddr_t)&iostate->io_wanted); } - if ((b_flags & B_NEED_IODONE) && real_bp) { - if (error) { - real_bp->b_flags |= B_ERROR; - real_bp->b_error = error; - } - real_bp->b_resid = total_resid; - - buf_biodone(real_bp); - } - if (error == 0 && total_resid) - error = EIO; if (b_flags & B_COMMIT_UPL) { - pg_offset = upl_offset & PAGE_MASK; - commit_size = (pg_offset + total_size + (PAGE_SIZE - 1)) & ~PAGE_MASK; + pg_offset = upl_offset & PAGE_MASK; + commit_size = (pg_offset + transaction_size + (PAGE_SIZE - 1)) & ~PAGE_MASK; - if (error || (b_flags & B_NOCACHE)) { - int upl_abort_code; - int page_in = 0; - int page_out = 0; - - if (b_flags & B_PAGEIO) { - if (b_flags & B_READ) - page_in = 1; - else - page_out = 1; - } - if (b_flags & B_CACHE) /* leave pages in the cache unchanged on error */ - upl_abort_code = UPL_ABORT_FREE_ON_EMPTY; - else if (page_out && (error != ENXIO)) /* transient error */ - upl_abort_code = UPL_ABORT_FREE_ON_EMPTY; - else if (page_in) - upl_abort_code = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR; - else - upl_abort_code = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_DUMP_PAGES; - - ubc_upl_abort_range(upl, upl_offset - pg_offset, commit_size, - upl_abort_code); - - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, - (int)upl, upl_offset - pg_offset, commit_size, - 0x80000000|upl_abort_code, 0); - - } else { - int upl_commit_flags = UPL_COMMIT_FREE_ON_EMPTY; + if (error) + upl_flags = cluster_ioerror(upl, upl_offset - pg_offset, commit_size, error, b_flags, vp); + else { + upl_flags = UPL_COMMIT_FREE_ON_EMPTY; if ((b_flags & B_PHYS) && (b_flags & B_READ)) - upl_commit_flags |= UPL_COMMIT_SET_DIRTY; + upl_flags |= UPL_COMMIT_SET_DIRTY; if (b_flags & B_AGE) - upl_commit_flags |= UPL_COMMIT_INACTIVATE; + upl_flags |= UPL_COMMIT_INACTIVATE; - ubc_upl_commit_range(upl, upl_offset - pg_offset, commit_size, - upl_commit_flags); - - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, - (int)upl, upl_offset - pg_offset, commit_size, - upl_commit_flags, 0); + ubc_upl_commit_range(upl, upl_offset - pg_offset, commit_size, upl_flags); } - } else { - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, - (int)upl, upl_offset, 0, error, 0); } + if (real_bp) { + if (error) { + real_bp->b_flags |= B_ERROR; + real_bp->b_error = error; + } + real_bp->b_resid = total_resid; + + buf_biodone(real_bp); + } + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END, + upl, upl_offset - pg_offset, commit_size, (error << 24) | upl_flags, 0); return (error); } +uint32_t +cluster_throttle_io_limit(vnode_t vp, uint32_t *limit) +{ + if (cluster_is_throttled(vp)) { + *limit = THROTTLE_MAX_IOSIZE; + return 1; + } + return 0; +} + + void -cluster_zero(upl_t upl, vm_offset_t upl_offset, int size, buf_t bp) +cluster_zero(upl_t upl, upl_offset_t upl_offset, int size, buf_t bp) { - upl_page_info_t *pl; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 23)) | DBG_FUNC_START, - upl_offset, size, (int)bp, 0, 0); + upl_offset, size, bp, 0, 0); if (bp == NULL || bp->b_datap == 0) { + upl_page_info_t *pl; + addr64_t zero_addr; pl = ubc_upl_pageinfo(upl); - while (size) { - int page_offset; - int page_index; - addr64_t zero_addr; - int zero_cnt; + if (upl_device_page(pl) == TRUE) { + zero_addr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + upl_offset; + + bzero_phys_nc(zero_addr, size); + } else { + while (size) { + int page_offset; + int page_index; + int zero_cnt; - page_index = upl_offset / PAGE_SIZE; - page_offset = upl_offset & PAGE_MASK; + page_index = upl_offset / PAGE_SIZE; + page_offset = upl_offset & PAGE_MASK; - zero_addr = ((addr64_t)upl_phys_page(pl, page_index) << 12) + page_offset; - zero_cnt = min(PAGE_SIZE - page_offset, size); + zero_addr = ((addr64_t)upl_phys_page(pl, page_index) << PAGE_SHIFT) + page_offset; + zero_cnt = min(PAGE_SIZE - page_offset, size); - bzero_phys(zero_addr, zero_cnt); + bzero_phys(zero_addr, zero_cnt); - size -= zero_cnt; - upl_offset += zero_cnt; + size -= zero_cnt; + upl_offset += zero_cnt; + } } } else bzero((caddr_t)((vm_offset_t)bp->b_datap + upl_offset), size); @@ -508,9 +960,112 @@ cluster_zero(upl_t upl, vm_offset_t upl_offset, int size, buf_t bp) } +static void +cluster_EOT(buf_t cbp_head, buf_t cbp_tail, int zero_offset) +{ + cbp_head->b_validend = zero_offset; + cbp_tail->b_flags |= B_EOT; +} + +static void +cluster_wait_IO(buf_t cbp_head, int async) +{ + buf_t cbp; + + if (async) { + /* + * Async callback completion will not normally generate a + * wakeup upon I/O completion. To get woken up, we set + * b_trans_next (which is safe for us to modify) on the last + * buffer to CLUSTER_IO_WAITING so that cluster_iodone knows + * to wake us up when all buffers as part of this transaction + * are completed. This is done under the umbrella of + * cl_transaction_mtxp which is also taken in cluster_iodone. + */ + bool done = true; + buf_t last = NULL; + + lck_mtx_lock_spin(cl_transaction_mtxp); + + for (cbp = cbp_head; cbp; last = cbp, cbp = cbp->b_trans_next) { + if (!ISSET(cbp->b_flags, B_TDONE)) + done = false; + } + + if (!done) { + last->b_trans_next = CLUSTER_IO_WAITING; + + DTRACE_IO1(wait__start, buf_t, last); + do { + msleep(last, cl_transaction_mtxp, PSPIN | (PRIBIO+1), "cluster_wait_IO", NULL); + + /* + * We should only have been woken up if all the + * buffers are completed, but just in case... + */ + done = true; + for (cbp = cbp_head; cbp != CLUSTER_IO_WAITING; cbp = cbp->b_trans_next) { + if (!ISSET(cbp->b_flags, B_TDONE)) { + done = false; + break; + } + } + } while (!done); + DTRACE_IO1(wait__done, buf_t, last); + + last->b_trans_next = NULL; + } + + lck_mtx_unlock(cl_transaction_mtxp); + } else { // !async + for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) + buf_biowait(cbp); + } +} + +static void +cluster_complete_transaction(buf_t *cbp_head, void *callback_arg, int *retval, int flags, int needwait) +{ + buf_t cbp; + int error; + boolean_t isswapout = FALSE; + + /* + * cluster_complete_transaction will + * only be called if we've issued a complete chain in synchronous mode + * or, we've already done a cluster_wait_IO on an incomplete chain + */ + if (needwait) { + for (cbp = *cbp_head; cbp; cbp = cbp->b_trans_next) + buf_biowait(cbp); + } + /* + * we've already waited on all of the I/Os in this transaction, + * so mark all of the buf_t's in this transaction as B_TDONE + * so that cluster_iodone sees the transaction as completed + */ + for (cbp = *cbp_head; cbp; cbp = cbp->b_trans_next) + cbp->b_flags |= B_TDONE; + cbp = *cbp_head; + + if ((flags & (CL_ASYNC | CL_PAGEOUT)) == CL_PAGEOUT && vnode_isswap(cbp->b_vp)) + isswapout = TRUE; + + error = cluster_iodone(cbp, callback_arg); + + if ( !(flags & CL_ASYNC) && error && *retval == 0) { + if (((flags & (CL_PAGEOUT | CL_KEEPCACHED)) != CL_PAGEOUT) || (error != ENXIO)) + *retval = error; + else if (isswapout == TRUE) + *retval = error; + } + *cbp_head = (buf_t)NULL; +} + + static int cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int non_rounded_size, - int flags, buf_t real_bp, struct clios *iostate) + int flags, buf_t real_bp, struct clios *iostate, int (*callback)(buf_t, void *), void *callback_arg) { buf_t cbp; u_int size; @@ -522,6 +1077,7 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no buf_t cbp_head = NULL; buf_t cbp_tail = NULL; int trans_count = 0; + int max_trans_count; u_int pg_count; int pg_offset; u_int max_iosize; @@ -530,10 +1086,27 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no int zero_offset = 0; int async_throttle = 0; mount_t mp; + vm_offset_t upl_end_offset; + boolean_t need_EOT = FALSE; + + /* + * we currently don't support buffers larger than a page + */ + if (real_bp && non_rounded_size > PAGE_SIZE) + panic("%s(): Called with real buffer of size %d bytes which " + "is greater than the maximum allowed size of " + "%d bytes (the system PAGE_SIZE).\n", + __FUNCTION__, non_rounded_size, PAGE_SIZE); mp = vp->v_mount; - if (mp->mnt_devblocksize > 1) { + /* + * we don't want to do any funny rounding of the size for IO requests + * coming through the DIRECT or CONTIGUOUS paths... those pages don't + * belong to us... we can't extend (nor do we need to) the I/O to fill + * out a page + */ + if (mp->mnt_devblocksize > 1 && !(flags & (CL_DEV_MEMORY | CL_DIRECT_IO))) { /* * round the requested size up so that this I/O ends on a * page boundary in case this is a 'write'... if the filesystem @@ -557,17 +1130,26 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no */ size = non_rounded_size; } - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 22)) | DBG_FUNC_START, - (int)f_offset, size, upl_offset, flags, 0); + upl_end_offset = upl_offset + size; + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 22)) | DBG_FUNC_START, (int)f_offset, size, upl_offset, flags, 0); + + /* + * Set the maximum transaction size to the maximum desired number of + * buffers. + */ + max_trans_count = 8; + if (flags & CL_DEV_MEMORY) + max_trans_count = 16; if (flags & CL_READ) { - io_flags = (B_READ); + io_flags = B_READ; bmap_flags = VNODE_READ; max_iosize = mp->mnt_maxreadcnt; max_vectors = mp->mnt_segreadcnt; } else { - io_flags = 0; + io_flags = B_WRITE; bmap_flags = VNODE_WRITE; max_iosize = mp->mnt_maxwritecnt; @@ -581,26 +1163,71 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no */ max_iosize &= ~PAGE_MASK; + /* + * Ensure the maximum iosize is sensible. + */ + if (!max_iosize) + max_iosize = PAGE_SIZE; + if (flags & CL_THROTTLE) { - if ( !(flags & CL_PAGEOUT) && cluster_hard_throttle_on(vp)) { - if (max_iosize > HARD_THROTTLE_MAXSIZE) - max_iosize = HARD_THROTTLE_MAXSIZE; - async_throttle = HARD_THROTTLE_MAXCNT; - } else - async_throttle = VNODE_ASYNC_THROTTLE; + if ( !(flags & CL_PAGEOUT) && cluster_is_throttled(vp)) { + if (max_iosize > THROTTLE_MAX_IOSIZE) + max_iosize = THROTTLE_MAX_IOSIZE; + async_throttle = THROTTLE_MAXCNT; + } else { + if ( (flags & CL_DEV_MEMORY) ) + async_throttle = IO_SCALE(vp, VNODE_ASYNC_THROTTLE); + else { + u_int max_cluster; + u_int max_cluster_size; + u_int scale; + + if (vp->v_mount->mnt_minsaturationbytecount) { + max_cluster_size = vp->v_mount->mnt_minsaturationbytecount; + + scale = 1; + } else { + max_cluster_size = MAX_CLUSTER_SIZE(vp); + + if ((vp->v_mount->mnt_kern_flag & MNTK_SSD) && !ignore_is_ssd) + scale = WRITE_THROTTLE_SSD; + else + scale = WRITE_THROTTLE; + } + if (max_iosize > max_cluster_size) + max_cluster = max_cluster_size; + else + max_cluster = max_iosize; + + if (size < max_cluster) + max_cluster = size; + + if (flags & CL_CLOSE) + scale += MAX_CLUSTERS; + + async_throttle = min(IO_SCALE(vp, VNODE_ASYNC_THROTTLE), ((scale * max_cluster_size) / max_cluster) - 1); + } + } } if (flags & CL_AGE) io_flags |= B_AGE; - if (flags & CL_DUMP) - io_flags |= B_NOCACHE; if (flags & (CL_PAGEIN | CL_PAGEOUT)) io_flags |= B_PAGEIO; + if (flags & (CL_IOSTREAMING)) + io_flags |= B_IOSTREAMING; if (flags & CL_COMMIT) io_flags |= B_COMMIT_UPL; - if (flags & CL_PRESERVE) + if (flags & CL_DIRECT_IO) io_flags |= B_PHYS; - if (flags & CL_KEEPCACHED) - io_flags |= B_CACHE; + if (flags & (CL_PRESERVE | CL_KEEPCACHED)) + io_flags |= B_CACHE; + if (flags & CL_PASSIVE) + io_flags |= B_PASSIVE; + if (flags & CL_ENCRYPTED) + io_flags |= B_ENCRYPTED_IO; + + if (vp->v_flag & VSYSTEM) + io_flags |= B_META; if ((flags & CL_READ) && ((upl_offset + non_rounded_size) & PAGE_MASK) && (!(flags & CL_NOZERO))) { /* @@ -611,25 +1238,64 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no * read in from the file */ zero_offset = upl_offset + non_rounded_size; + } else if (!ISSET(flags, CL_READ) && ISSET(flags, CL_DIRECT_IO)) { + assert(ISSET(flags, CL_COMMIT)); + + // For a direct/uncached write, we need to lock pages... + + upl_t cached_upl; + + /* + * Create a UPL to lock the pages in the cache whilst the + * write is in progress. + */ + ubc_create_upl(vp, f_offset, non_rounded_size, &cached_upl, + NULL, UPL_SET_LITE); + + /* + * Attach this UPL to the other UPL so that we can find it + * later. + */ + upl_set_associated_upl(upl, cached_upl); + + if (upl_offset & PAGE_MASK) { + /* + * The two UPLs are not aligned, so mark the first page in + * @upl so that cluster_handle_associated_upl can handle + * it accordingly. + */ + upl_page_info_t *pl = UPL_GET_INTERNAL_PAGE_LIST(upl); + upl_page_set_mark(pl, 0, true); + } } + while (size) { - int pg_resid; daddr64_t blkno; daddr64_t lblkno; + u_int io_size_wanted; + size_t io_size_tmp; if (size > max_iosize) io_size = max_iosize; else io_size = size; + + io_size_wanted = io_size; + io_size_tmp = (size_t)io_size; - if ((error = VNOP_BLOCKMAP(vp, f_offset, io_size, &blkno, (size_t *)&io_size, NULL, bmap_flags, NULL))) { + if ((error = VNOP_BLOCKMAP(vp, f_offset, io_size, &blkno, &io_size_tmp, NULL, bmap_flags, NULL))) break; - } + + if (io_size_tmp > io_size_wanted) + io_size = io_size_wanted; + else + io_size = (u_int)io_size_tmp; + if (real_bp && (real_bp->b_blkno == real_bp->b_lblkno)) real_bp->b_blkno = blkno; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 24)) | DBG_FUNC_NONE, - (int)f_offset, (int)blkno, io_size, zero_offset, 0); + (int)f_offset, (int)(blkno>>32), (int)blkno, io_size, 0); if (io_size == 0) { /* @@ -643,8 +1309,11 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no break; } if ( !(flags & CL_READ) && blkno == -1) { - off_t e_offset; + off_t e_offset; + int pageout_flags; + if (upl_get_internal_vectorupl(upl)) + panic("Vector UPLs should not take this code-path\n"); /* * we're writing into a 'hole' */ @@ -657,17 +1326,6 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no error = EINVAL; break; } - if ( !(flags & CL_COMMIT)) { - /* - * currently writes always request the commit to happen - * as part of the io completion... however, if the CL_COMMIT - * flag isn't specified, than we can't issue the abort_range - * since the call site is going to abort or commit the same upl.. - * in this case we can only return an error - */ - error = EINVAL; - break; - } /* * we can get here if the cluster code happens to * pick up a page that was dirtied via mmap vs @@ -686,15 +1344,105 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no * giving up entirely and dumping it (the pageout * path will insure that the zero extent accounting * has been taken care of before we get back into cluster_io) + * + * go direct to vnode_pageout so that we don't have to + * unbusy the page from the UPL... we used to do this + * so that we could call ubc_msync, but that results + * in a potential deadlock if someone else races us to acquire + * that page and wins and in addition needs one of the pages + * we're continuing to hold in the UPL */ - ubc_upl_abort_range(upl, trunc_page(upl_offset), PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY); + pageout_flags = UPL_MSYNC | UPL_VNODE_PAGER | UPL_NESTED_PAGEOUT; - e_offset = round_page_64(f_offset + 1); + if ( !(flags & CL_ASYNC)) + pageout_flags |= UPL_IOSYNC; + if ( !(flags & CL_COMMIT)) + pageout_flags |= UPL_NOCOMMIT; + + if (cbp_head) { + buf_t prev_cbp; + int bytes_in_last_page; + + /* + * first we have to wait for the the current outstanding I/Os + * to complete... EOT hasn't been set yet on this transaction + * so the pages won't be released + */ + cluster_wait_IO(cbp_head, (flags & CL_ASYNC)); + + bytes_in_last_page = cbp_head->b_uploffset & PAGE_MASK; + for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) + bytes_in_last_page += cbp->b_bcount; + bytes_in_last_page &= PAGE_MASK; + + while (bytes_in_last_page) { + /* + * we've got a transcation that + * includes the page we're about to push out through vnode_pageout... + * find the bp's in the list which intersect this page and either + * remove them entirely from the transaction (there could be multiple bp's), or + * round it's iosize down to the page boundary (there can only be one)... + * + * find the last bp in the list and act on it + */ + for (prev_cbp = cbp = cbp_head; cbp->b_trans_next; cbp = cbp->b_trans_next) + prev_cbp = cbp; + + if (bytes_in_last_page >= cbp->b_bcount) { + /* + * this buf no longer has any I/O associated with it + */ + bytes_in_last_page -= cbp->b_bcount; + cbp->b_bcount = 0; + + free_io_buf(cbp); + + if (cbp == cbp_head) { + assert(bytes_in_last_page == 0); + /* + * the buf we just freed was the only buf in + * this transaction... so there's no I/O to do + */ + cbp_head = NULL; + cbp_tail = NULL; + } else { + /* + * remove the buf we just freed from + * the transaction list + */ + prev_cbp->b_trans_next = NULL; + cbp_tail = prev_cbp; + } + } else { + /* + * this is the last bp that has I/O + * intersecting the page of interest + * only some of the I/O is in the intersection + * so clip the size but keep it in the transaction list + */ + cbp->b_bcount -= bytes_in_last_page; + cbp_tail = cbp; + bytes_in_last_page = 0; + } + } + if (cbp_head) { + /* + * there was more to the current transaction + * than just the page we are pushing out via vnode_pageout... + * mark it as finished and complete it... we've already + * waited for the I/Os to complete above in the call to cluster_wait_IO + */ + cluster_EOT(cbp_head, cbp_tail, 0); + + cluster_complete_transaction(&cbp_head, callback_arg, &retval, flags, 0); - if (ubc_sync_range(vp, f_offset, e_offset, UBC_PUSHDIRTY) == 0) { + trans_count = 0; + } + } + if (vnode_pageout(vp, upl, trunc_page(upl_offset), trunc_page_64(f_offset), PAGE_SIZE, pageout_flags, NULL) != PAGER_SUCCESS) { error = EINVAL; - break; } + e_offset = round_page_64(f_offset + 1); io_size = e_offset - f_offset; f_offset += io_size; @@ -721,9 +1469,14 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no */ size = 0; } + if (error) { + if (size == 0) + flags &= ~CL_COMMIT; + break; + } continue; } - lblkno = (daddr64_t)(f_offset / PAGE_SIZE_64); + lblkno = (daddr64_t)(f_offset / 0x1000); /* * we have now figured out how much I/O we can do - this is in 'io_size' * pg_offset is the starting point in the first page for the I/O @@ -732,13 +1485,6 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no pg_offset = upl_offset & PAGE_MASK; if (flags & CL_DEV_MEMORY) { - /* - * currently, can't deal with reading 'holes' in file - */ - if (blkno == -1) { - error = EINVAL; - break; - } /* * treat physical requests as one 'giant' page */ @@ -747,14 +1493,16 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no pg_count = (io_size + pg_offset + (PAGE_SIZE - 1)) / PAGE_SIZE; if ((flags & CL_READ) && blkno == -1) { + vm_offset_t commit_offset; int bytes_to_zero; + int complete_transaction_now = 0; /* * if we're reading and blkno == -1, then we've got a * 'hole' in the file that we need to deal with by zeroing * out the affected area in the upl */ - if (zero_offset && io_size == size) { + if (io_size >= (u_int)non_rounded_size) { /* * if this upl contains the EOF and it is not a multiple of PAGE_SIZE * than 'zero_offset' will be non-zero @@ -764,56 +1512,77 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no * last page in this upl... we need to zero both the hole and the tail * of the page beyond the EOF, since the delayed zero-fill won't kick in */ - bytes_to_zero = (((upl_offset + io_size) + (PAGE_SIZE - 1)) & ~PAGE_MASK) - upl_offset; + bytes_to_zero = non_rounded_size; + if (!(flags & CL_NOZERO)) + bytes_to_zero = (((upl_offset + io_size) + (PAGE_SIZE - 1)) & ~PAGE_MASK) - upl_offset; zero_offset = 0; } else bytes_to_zero = io_size; - cluster_zero(upl, upl_offset, bytes_to_zero, real_bp); + pg_count = 0; + + cluster_zero(upl, upl_offset, bytes_to_zero, real_bp); - if (cbp_head) + if (cbp_head) { + int pg_resid; + /* * if there is a current I/O chain pending * then the first page of the group we just zero'd * will be handled by the I/O completion if the zero * fill started in the middle of the page */ - pg_count = (io_size - pg_offset) / PAGE_SIZE; - else { + commit_offset = (upl_offset + (PAGE_SIZE - 1)) & ~PAGE_MASK; + + pg_resid = commit_offset - upl_offset; + + if (bytes_to_zero >= pg_resid) { + /* + * the last page of the current I/O + * has been completed... + * compute the number of fully zero'd + * pages that are beyond it + * plus the last page if its partial + * and we have no more I/O to issue... + * otherwise a partial page is left + * to begin the next I/O + */ + if ((int)io_size >= non_rounded_size) + pg_count = (bytes_to_zero - pg_resid + (PAGE_SIZE - 1)) / PAGE_SIZE; + else + pg_count = (bytes_to_zero - pg_resid) / PAGE_SIZE; + + complete_transaction_now = 1; + } + } else { /* - * no pending I/O to pick up that first page - * so, we have to make sure it gets committed - * here. - * set the pg_offset to 0 so that the upl_commit_range - * starts with this page + * no pending I/O to deal with + * so, commit all of the fully zero'd pages + * plus the last page if its partial + * and we have no more I/O to issue... + * otherwise a partial page is left + * to begin the next I/O */ - pg_count = (io_size + pg_offset) / PAGE_SIZE; - pg_offset = 0; + if ((int)io_size >= non_rounded_size) + pg_count = (pg_offset + bytes_to_zero + (PAGE_SIZE - 1)) / PAGE_SIZE; + else + pg_count = (pg_offset + bytes_to_zero) / PAGE_SIZE; + + commit_offset = upl_offset & ~PAGE_MASK; } - if (io_size == size && ((upl_offset + io_size) & PAGE_MASK)) - /* - * if we're done with the request for this UPL - * then we have to make sure to commit the last page - * even if we only partially zero-filled it - */ - pg_count++; - if (pg_count) { - if (pg_offset) - pg_resid = PAGE_SIZE - pg_offset; - else - pg_resid = 0; + // Associated UPL is currently only used in the direct write path + assert(!upl_associated_upl(upl)); - if (flags & CL_COMMIT) - ubc_upl_commit_range(upl, - (upl_offset + pg_resid) & ~PAGE_MASK, - pg_count * PAGE_SIZE, - UPL_COMMIT_CLEAR_DIRTY | UPL_COMMIT_FREE_ON_EMPTY); + if ( (flags & CL_COMMIT) && pg_count) { + ubc_upl_commit_range(upl, commit_offset, pg_count * PAGE_SIZE, + UPL_COMMIT_CLEAR_DIRTY | UPL_COMMIT_FREE_ON_EMPTY); } upl_offset += io_size; f_offset += io_size; size -= io_size; + /* * keep track of how much of the original request * that we've actually completed... non_rounded_size @@ -831,10 +1600,16 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no */ size = 0; } - if (cbp_head && pg_count) - goto start_io; - continue; + if (cbp_head && (complete_transaction_now || size == 0)) { + cluster_wait_IO(cbp_head, (flags & CL_ASYNC)); + cluster_EOT(cbp_head, cbp_tail, size == 0 ? zero_offset : 0); + + cluster_complete_transaction(&cbp_head, callback_arg, &retval, flags, 0); + + trans_count = 0; + } + continue; } if (pg_count > max_vectors) { if (((pg_count - max_vectors) * PAGE_SIZE) > io_size) { @@ -845,6 +1620,35 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no pg_count = max_vectors; } } + /* + * If the transaction is going to reach the maximum number of + * desired elements, truncate the i/o to the nearest page so + * that the actual i/o is initiated after this buffer is + * created and added to the i/o chain. + * + * I/O directed to physically contiguous memory + * doesn't have a requirement to make sure we 'fill' a page + */ + if ( !(flags & CL_DEV_MEMORY) && trans_count >= max_trans_count && + ((upl_offset + io_size) & PAGE_MASK)) { + vm_offset_t aligned_ofs; + + aligned_ofs = (upl_offset + io_size) & ~PAGE_MASK; + /* + * If the io_size does not actually finish off even a + * single page we have to keep adding buffers to the + * transaction despite having reached the desired limit. + * + * Eventually we get here with the page being finished + * off (and exceeded) and then we truncate the size of + * this i/o request so that it is page aligned so that + * we can finally issue the i/o on the transaction. + */ + if (aligned_ofs > upl_offset) { + io_size = aligned_ofs - upl_offset; + pg_count--; + } + } if ( !(mp->mnt_kern_flag & MNTK_VIRTUALDEV)) /* @@ -867,16 +1671,23 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no if (flags & CL_PAGEOUT) { u_int i; - for (i = 0; i < pg_count; i++) { - if (buf_invalblkno(vp, lblkno + i, 0) == EBUSY) - panic("BUSY bp found in cluster_io"); + /* + * since blocks are in offsets of 0x1000, scale + * iteration to (PAGE_SIZE * pg_count) of blks. + */ + for (i = 0; i < (PAGE_SIZE * pg_count)/0x1000; i++) { + if (buf_invalblkno(vp, lblkno + i, 0) == EBUSY) + panic("BUSY bp found in cluster_io"); } } if (flags & CL_ASYNC) { - if (buf_setcallback(cbp, (void *)cluster_iodone, NULL)) + if (buf_setcallback(cbp, (void *)cluster_iodone, callback_arg)) panic("buf_setcallback failed\n"); } + cbp->b_cliodone = (void *)callback; cbp->b_flags |= io_flags; + if (flags & CL_NOCACHE) + cbp->b_attr.ba_flags |= BA_NOCACHE; cbp->b_lblkno = lblkno; cbp->b_blkno = blkno; @@ -884,7 +1695,9 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no if (buf_setupl(cbp, upl, upl_offset)) panic("buf_setupl failed\n"); - +#if CONFIG_IOSCHED + upl_set_blkno(upl, upl_offset, io_size, blkno); +#endif cbp->b_trans_next = (buf_t)NULL; if ((cbp->b_iostate = (void *)iostate)) @@ -909,8 +1722,12 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no } else { cbp_head = cbp; cbp_tail = cbp; + + if ( (cbp_head->b_real_bp = real_bp) ) + real_bp = (buf_t)NULL; } - (buf_t)(cbp->b_trans_head) = cbp_head; + *(buf_t *)(&cbp->b_trans_head) = cbp_head; + trans_count++; upl_offset += io_size; @@ -933,80 +1750,85 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no */ size = 0; } - if ( (!(upl_offset & PAGE_MASK) && !(flags & CL_DEV_MEMORY) && ((flags & CL_ASYNC) || trans_count > 8)) || size == 0) { + if (size == 0) { + /* + * we have no more I/O to issue, so go + * finish the final transaction + */ + need_EOT = TRUE; + } else if ( ((flags & CL_DEV_MEMORY) || (upl_offset & PAGE_MASK) == 0) && + ((flags & CL_ASYNC) || trans_count > max_trans_count) ) { /* - * if we have no more I/O to issue or + * I/O directed to physically contiguous memory... + * which doesn't have a requirement to make sure we 'fill' a page + * or... * the current I/O we've prepared fully * completes the last page in this request - * and it's either an ASYNC request or + * and ... + * it's either an ASYNC request or * we've already accumulated more than 8 I/O's into - * this transaction and it's not an I/O directed to - * special DEVICE memory - * then go ahead and issue the I/O + * this transaction so mark it as complete so that + * it can finish asynchronously or via the cluster_complete_transaction + * below if the request is synchronous */ -start_io: - if (real_bp) { - cbp_head->b_flags |= B_NEED_IODONE; - cbp_head->b_real_bp = real_bp; - } else - cbp_head->b_real_bp = (buf_t)NULL; - - if (size == 0) { - /* - * we're about to issue the last I/O for this upl - * if this was a read to the eof and the eof doesn't - * finish on a page boundary, than we need to zero-fill - * the rest of the page.... - */ - cbp_head->b_validend = zero_offset; - } else - cbp_head->b_validend = 0; - - if (flags & CL_THROTTLE) - (void)vnode_waitforwrites(vp, async_throttle, 0, 0, (char *)"cluster_io"); - - for (cbp = cbp_head; cbp;) { - buf_t cbp_next; + need_EOT = TRUE; + } + if (need_EOT == TRUE) + cluster_EOT(cbp_head, cbp_tail, size == 0 ? zero_offset : 0); - if ( !(io_flags & B_READ)) - vnode_startwrite(vp); + if (flags & CL_THROTTLE) + (void)vnode_waitforwrites(vp, async_throttle, 0, 0, "cluster_io"); - cbp_next = cbp->b_trans_next; + if ( !(io_flags & B_READ)) + vnode_startwrite(vp); - (void) VNOP_STRATEGY(cbp); - cbp = cbp_next; - } - if ( !(flags & CL_ASYNC)) { - int dummy; - - for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) - buf_biowait(cbp); - - if ((error = cluster_iodone(cbp_head, (void *)&dummy))) { - if (((flags & (CL_PAGEOUT | CL_KEEPCACHED)) == CL_PAGEOUT) && (error == ENXIO)) - error = 0; /* drop the error */ - else { - if (retval == 0) - retval = error; - error = 0; - } - } - } - cbp_head = (buf_t)NULL; - cbp_tail = (buf_t)NULL; + if (flags & CL_RAW_ENCRYPTED) { + /* + * User requested raw encrypted bytes. + * Twiddle the bit in the ba_flags for the buffer + */ + cbp->b_attr.ba_flags |= BA_RAW_ENCRYPTED_IO; + } + + (void) VNOP_STRATEGY(cbp); + if (need_EOT == TRUE) { + if ( !(flags & CL_ASYNC)) + cluster_complete_transaction(&cbp_head, callback_arg, &retval, flags, 1); + + need_EOT = FALSE; trans_count = 0; + cbp_head = NULL; } - } + } if (error) { - int abort_size; + int abort_size; io_size = 0; - - for (cbp = cbp_head; cbp;) { + + if (cbp_head) { + /* + * Wait until all of the outstanding I/O + * for this partial transaction has completed + */ + cluster_wait_IO(cbp_head, (flags & CL_ASYNC)); + + /* + * Rewind the upl offset to the beginning of the + * transaction. + */ + upl_offset = cbp_head->b_uploffset; + } + + if (ISSET(flags, CL_COMMIT)) { + cluster_handle_associated_upl(iostate, upl, upl_offset, + upl_end_offset - upl_offset); + } + + // Free all the IO buffers in this transaction + for (cbp = cbp_head; cbp;) { buf_t cbp_next; - upl_offset -= cbp->b_bcount; size += cbp->b_bcount; io_size += cbp->b_bcount; @@ -1014,6 +1836,7 @@ start_io: free_io_buf(cbp); cbp = cbp_next; } + if (iostate) { int need_wakeup = 0; @@ -1022,7 +1845,7 @@ start_io: * since we never really issued the io * just go ahead and adjust it back */ - lck_mtx_lock(cl_mtxp); + lck_mtx_lock_spin(&iostate->io_mtxp); if (iostate->io_error == 0) iostate->io_error = error; @@ -1034,54 +1857,68 @@ start_io: * this io stream to change */ iostate->io_wanted = 0; - need_wakeup = 0; + need_wakeup = 1; } - lck_mtx_unlock(cl_mtxp); + lck_mtx_unlock(&iostate->io_mtxp); if (need_wakeup) wakeup((caddr_t)&iostate->io_wanted); } - pg_offset = upl_offset & PAGE_MASK; - abort_size = (size + pg_offset + (PAGE_SIZE - 1)) & ~PAGE_MASK; if (flags & CL_COMMIT) { - int upl_abort_code; + int upl_flags; - if (flags & CL_PRESERVE) { - ubc_upl_commit_range(upl, upl_offset - pg_offset, abort_size, - UPL_COMMIT_FREE_ON_EMPTY); - } else { - if ((flags & CL_PAGEOUT) && (error != ENXIO)) /* transient error */ - upl_abort_code = UPL_ABORT_FREE_ON_EMPTY; - else if (flags & CL_PAGEIN) - upl_abort_code = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR; - else - upl_abort_code = UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_DUMP_PAGES; + pg_offset = upl_offset & PAGE_MASK; + abort_size = (upl_end_offset - upl_offset + PAGE_MASK) & ~PAGE_MASK; - ubc_upl_abort_range(upl, upl_offset - pg_offset, abort_size, - upl_abort_code); - } + upl_flags = cluster_ioerror(upl, upl_offset - pg_offset, abort_size, error, io_flags, vp); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 28)) | DBG_FUNC_NONE, - (int)upl, upl_offset - pg_offset, abort_size, error, 0); - } - if (real_bp) { - real_bp->b_flags |= B_ERROR; - real_bp->b_error = error; - - buf_biodone(real_bp); + upl, upl_offset - pg_offset, abort_size, (error << 24) | upl_flags, 0); } if (retval == 0) retval = error; + } else if (cbp_head) + panic("%s(): cbp_head is not NULL.\n", __FUNCTION__); + + if (real_bp) { + /* + * can get here if we either encountered an error + * or we completely zero-filled the request and + * no I/O was issued + */ + if (error) { + real_bp->b_flags |= B_ERROR; + real_bp->b_error = error; + } + buf_biodone(real_bp); } - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 22)) | DBG_FUNC_END, - (int)f_offset, size, upl_offset, retval, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 22)) | DBG_FUNC_END, (int)f_offset, size, upl_offset, retval, 0); return (retval); } +#define reset_vector_run_state() \ + issueVectorUPL = vector_upl_offset = vector_upl_index = vector_upl_iosize = vector_upl_size = 0; + +static int +vector_cluster_io(vnode_t vp, upl_t vector_upl, vm_offset_t vector_upl_offset, off_t v_upl_uio_offset, int vector_upl_iosize, + int io_flag, buf_t real_bp, struct clios *iostate, int (*callback)(buf_t, void *), void *callback_arg) +{ + vector_upl_set_pagelist(vector_upl); + + if(io_flag & CL_READ) { + if(vector_upl_offset == 0 && ((vector_upl_iosize & PAGE_MASK)==0)) + io_flag &= ~CL_PRESERVE; /*don't zero fill*/ + else + io_flag |= CL_PRESERVE; /*zero fill*/ + } + return (cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, real_bp, iostate, callback, callback_arg)); + +} static int -cluster_rd_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize) +cluster_read_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize, int (*callback)(buf_t, void *), void *callback_arg, int bflag) { int pages_in_prefetch; @@ -1093,16 +1930,11 @@ cluster_rd_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize) (int)f_offset, 0, 0, 0, 0); return(0); } - if (size > (MAX_UPL_TRANSFER * PAGE_SIZE)) - size = (MAX_UPL_TRANSFER * PAGE_SIZE); - else - size = (size + (PAGE_SIZE - 1)) & ~PAGE_MASK; - if ((off_t)size > (filesize - f_offset)) size = filesize - f_offset; pages_in_prefetch = (size + (PAGE_SIZE - 1)) / PAGE_SIZE; - advisory_read(vp, filesize, f_offset, size); + advisory_read_ext(vp, filesize, f_offset, size, callback, callback_arg, bflag); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 49)) | DBG_FUNC_END, (int)f_offset + size, pages_in_prefetch, 0, 1, 0); @@ -1113,11 +1945,13 @@ cluster_rd_prefetch(vnode_t vp, off_t f_offset, u_int size, off_t filesize) static void -cluster_rd_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl_readahead *rap) +cluster_read_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl_readahead *rap, int (*callback)(buf_t, void *), void *callback_arg, + int bflag) { daddr64_t r_addr; off_t f_offset; int size_of_prefetch; + u_int max_prefetch; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_START, @@ -1128,8 +1962,7 @@ cluster_rd_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 0, 0); return; } - if (rap->cl_lastr == -1 || (extent->b_addr != rap->cl_lastr && extent->b_addr != (rap->cl_lastr + 1) && - (extent->b_addr != (rap->cl_maxra + 1) || rap->cl_ralen == 0))) { + if (rap->cl_lastr == -1 || (extent->b_addr != rap->cl_lastr && extent->b_addr != (rap->cl_lastr + 1))) { rap->cl_ralen = 0; rap->cl_maxra = 0; @@ -1138,8 +1971,18 @@ cluster_rd_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl return; } - if (extent->e_addr < rap->cl_maxra) { - if ((rap->cl_maxra - extent->e_addr) > (MAX_UPL_TRANSFER / 4)) { + max_prefetch = MAX_PREFETCH(vp, cluster_max_io_size(vp->v_mount, CL_READ), (vp->v_mount->mnt_kern_flag & MNTK_SSD)); + + if (max_prefetch > speculative_prefetch_max) + max_prefetch = speculative_prefetch_max; + + if (max_prefetch <= PAGE_SIZE) { + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END, + rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 6, 0); + return; + } + if (extent->e_addr < rap->cl_maxra && rap->cl_ralen >= 4) { + if ((rap->cl_maxra - extent->e_addr) > (rap->cl_ralen / 4)) { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END, rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 2, 0); @@ -1161,17 +2004,17 @@ cluster_rd_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl if (f_offset < filesize) { daddr64_t read_size; - rap->cl_ralen = rap->cl_ralen ? min(MAX_UPL_TRANSFER, rap->cl_ralen << 1) : 1; + rap->cl_ralen = rap->cl_ralen ? min(max_prefetch / PAGE_SIZE, rap->cl_ralen << 1) : 1; read_size = (extent->e_addr + 1) - extent->b_addr; if (read_size > rap->cl_ralen) { - if (read_size > MAX_UPL_TRANSFER) - rap->cl_ralen = MAX_UPL_TRANSFER; + if (read_size > max_prefetch / PAGE_SIZE) + rap->cl_ralen = max_prefetch / PAGE_SIZE; else rap->cl_ralen = read_size; } - size_of_prefetch = cluster_rd_prefetch(vp, f_offset, rap->cl_ralen * PAGE_SIZE, filesize); + size_of_prefetch = cluster_read_prefetch(vp, f_offset, rap->cl_ralen * PAGE_SIZE, filesize, callback, callback_arg, bflag); if (size_of_prefetch) rap->cl_maxra = (r_addr + size_of_prefetch) - 1; @@ -1180,28 +2023,26 @@ cluster_rd_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct cl rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 4, 0); } + int -cluster_pageout(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, +cluster_pageout(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset, int size, off_t filesize, int flags) +{ + return cluster_pageout_ext(vp, upl, upl_offset, f_offset, size, filesize, flags, NULL, NULL); + +} + + +int +cluster_pageout_ext(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset, + int size, off_t filesize, int flags, int (*callback)(buf_t, void *), void *callback_arg) { int io_size; int rounded_size; off_t max_size; int local_flags; - struct cl_writebehind *wbp; - if (vp->v_mount->mnt_kern_flag & MNTK_VIRTUALDEV) - /* - * if we know we're issuing this I/O to a virtual device (i.e. disk image) - * then we don't want to enforce this throttle... if we do, we can - * potentially deadlock since we're stalling the pageout thread at a time - * when the disk image might need additional memory (which won't be available - * if the pageout thread can't run)... instead we'll just depend on the throttle - * that the pageout thread now has in place to deal with external files - */ - local_flags = CL_PAGEOUT; - else - local_flags = CL_PAGEOUT | CL_THROTTLE; + local_flags = CL_PAGEOUT | CL_THROTTLE; if ((flags & UPL_IOSYNC) == 0) local_flags |= CL_ASYNC; @@ -1209,6 +2050,8 @@ cluster_pageout(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, local_flags |= CL_COMMIT; if ((flags & UPL_KEEPCACHED)) local_flags |= CL_KEEPCACHED; + if (flags & UPL_PAGING_ENCRYPTED) + local_flags |= CL_ENCRYPTED; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 52)) | DBG_FUNC_NONE, @@ -1253,16 +2096,22 @@ cluster_pageout(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, ubc_upl_abort_range(upl, upl_offset + rounded_size, size - rounded_size, UPL_ABORT_FREE_ON_EMPTY); } - if ((wbp = cluster_get_wbp(vp, 0)) != NULL) - wbp->cl_hasbeenpaged = 1; - return (cluster_io(vp, upl, upl_offset, f_offset, io_size, - local_flags, (buf_t)NULL, (struct clios *)NULL)); + local_flags, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg)); } + int -cluster_pagein(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, +cluster_pagein(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset, int size, off_t filesize, int flags) +{ + return cluster_pagein_ext(vp, upl, upl_offset, f_offset, size, filesize, flags, NULL, NULL); +} + + +int +cluster_pagein_ext(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offset, + int size, off_t filesize, int flags, int (*callback)(buf_t, void *), void *callback_arg) { u_int io_size; int rounded_size; @@ -1277,6 +2126,10 @@ cluster_pagein(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, local_flags |= CL_ASYNC; if ((flags & UPL_NOCOMMIT) == 0) local_flags |= CL_COMMIT; + if (flags & UPL_IOSTREAMING) + local_flags |= CL_IOSTREAMING; + if (flags & UPL_PAGING_ENCRYPTED) + local_flags |= CL_ENCRYPTED; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 56)) | DBG_FUNC_NONE, @@ -1308,299 +2161,365 @@ cluster_pagein(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, size - rounded_size, UPL_ABORT_FREE_ON_EMPTY | UPL_ABORT_ERROR); retval = cluster_io(vp, upl, upl_offset, f_offset, io_size, - local_flags | CL_READ | CL_PAGEIN, (buf_t)NULL, (struct clios *)NULL); - - if (retval == 0 && !(flags & UPL_NORDAHEAD) && !(vp->v_flag & VRAOFF)) { - struct cl_readahead *rap; - - rap = cluster_get_rap(vp); + local_flags | CL_READ | CL_PAGEIN, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg); - if (rap != NULL) { - struct cl_extent extent; - - extent.b_addr = (daddr64_t)(f_offset / PAGE_SIZE_64); - extent.e_addr = (daddr64_t)((f_offset + ((off_t)io_size - 1)) / PAGE_SIZE_64); - - if (rounded_size == PAGE_SIZE) { - /* - * we haven't read the last page in of the file yet - * so let's try to read ahead if we're in - * a sequential access pattern - */ - cluster_rd_ahead(vp, &extent, filesize, rap); - } - rap->cl_lastr = extent.e_addr; - - lck_mtx_unlock(&rap->cl_lockr); - } - } return (retval); } + int cluster_bp(buf_t bp) +{ + return cluster_bp_ext(bp, NULL, NULL); +} + + +int +cluster_bp_ext(buf_t bp, int (*callback)(buf_t, void *), void *callback_arg) { off_t f_offset; int flags; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 19)) | DBG_FUNC_START, - (int)bp, (int)bp->b_lblkno, bp->b_bcount, bp->b_flags, 0); + bp, (int)bp->b_lblkno, bp->b_bcount, bp->b_flags, 0); if (bp->b_flags & B_READ) flags = CL_ASYNC | CL_READ; else flags = CL_ASYNC; + if (bp->b_flags & B_PASSIVE) + flags |= CL_PASSIVE; f_offset = ubc_blktooff(bp->b_vp, bp->b_lblkno); - return (cluster_io(bp->b_vp, bp->b_upl, 0, f_offset, bp->b_bcount, flags, bp, (struct clios *)NULL)); + return (cluster_io(bp->b_vp, bp->b_upl, 0, f_offset, bp->b_bcount, flags, bp, (struct clios *)NULL, callback, callback_arg)); } + + int cluster_write(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t headOff, off_t tailOff, int xflags) { - int prev_resid; - u_int clip_size; - off_t max_io_size; - int upl_size; - int upl_flags; - upl_t upl; - int retval = 0; - int flags; + return cluster_write_ext(vp, uio, oldEOF, newEOF, headOff, tailOff, xflags, NULL, NULL); +} + + +int +cluster_write_ext(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t headOff, off_t tailOff, + int xflags, int (*callback)(buf_t, void *), void *callback_arg) +{ + user_ssize_t cur_resid; + int retval = 0; + int flags; + int zflags; + int bflag; + int write_type = IO_COPY; + u_int32_t write_length; flags = xflags; - if (vp->v_flag & VNOCACHE_DATA) - flags |= IO_NOCACHE; + if (flags & IO_PASSIVE) + bflag = CL_PASSIVE; + else + bflag = 0; - if ( (!(flags & IO_NOCACHE)) || (!uio) || (!UIO_SEG_IS_USER_SPACE(uio->uio_segflg))) { + if (vp->v_flag & VNOCACHE_DATA){ + flags |= IO_NOCACHE; + bflag |= CL_NOCACHE; + } + if (uio == NULL) { /* - * go do a write through the cache if one of the following is true.... - * NOCACHE is not true - * there is no uio structure or it doesn't target USERSPACE + * no user data... + * this call is being made to zero-fill some range in the file */ - return (cluster_write_x(vp, uio, oldEOF, newEOF, headOff, tailOff, flags)); - } + retval = cluster_write_copy(vp, NULL, (u_int32_t)0, oldEOF, newEOF, headOff, tailOff, flags, callback, callback_arg); -#if LP64_DEBUG - if (IS_VALID_UIO_SEGFLG(uio->uio_segflg) == 0) { - panic("%s :%d - invalid uio_segflg\n", __FILE__, __LINE__); + return(retval); } -#endif /* LP64_DEBUG */ - - while (uio_resid(uio) && uio->uio_offset < newEOF && retval == 0) { - user_size_t iov_len; - user_addr_t iov_base; - - /* - * we know we have a resid, so this is safe - * skip over any emtpy vectors - */ - uio_update(uio, (user_size_t)0); - - iov_len = uio_curriovlen(uio); - iov_base = uio_curriovbase(uio); - - upl_size = PAGE_SIZE; - upl_flags = UPL_QUERY_OBJECT_TYPE; - - // LP64todo - fix this! - if ((vm_map_get_upl(current_map(), - (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)), - &upl_size, &upl, NULL, NULL, &upl_flags, 0)) != KERN_SUCCESS) { - /* - * the user app must have passed in an invalid address - */ - return (EFAULT); - } - - /* - * We check every vector target but if it is physically - * contiguous space, we skip the sanity checks. + /* + * do a write through the cache if one of the following is true.... + * NOCACHE is not true or NODIRECT is true + * the uio request doesn't target USERSPACE + * otherwise, find out if we want the direct or contig variant for + * the first vector in the uio request + */ + if ( ((flags & (IO_NOCACHE | IO_NODIRECT)) == IO_NOCACHE) && UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ) + retval = cluster_io_type(uio, &write_type, &write_length, MIN_DIRECT_WRITE_SIZE); + + if ( (flags & (IO_TAILZEROFILL | IO_HEADZEROFILL)) && write_type == IO_DIRECT) + /* + * must go through the cached variant in this case */ - if (upl_flags & UPL_PHYS_CONTIG) { - int zflags; - - zflags = flags & ~IO_TAILZEROFILL; - zflags |= IO_HEADZEROFILL; + write_type = IO_COPY; - if (flags & IO_HEADZEROFILL) { - /* - * in case we have additional vectors, we don't want to do this again - */ - flags &= ~IO_HEADZEROFILL; - - if ((retval = cluster_write_x(vp, (struct uio *)0, 0, uio->uio_offset, headOff, 0, zflags))) - return(retval); - } - retval = cluster_phys_write(vp, uio, newEOF); + while ((cur_resid = uio_resid(uio)) && uio->uio_offset < newEOF && retval == 0) { + + switch (write_type) { - if (uio_resid(uio) == 0 && (flags & IO_TAILZEROFILL)) { - return (cluster_write_x(vp, (struct uio *)0, 0, tailOff, uio->uio_offset, 0, zflags)); - } - } - else if ((uio_resid(uio) < PAGE_SIZE) || (flags & (IO_TAILZEROFILL | IO_HEADZEROFILL))) { + case IO_COPY: /* - * we're here because we're don't have a physically contiguous target buffer - * go do a write through the cache if one of the following is true.... - * the total xfer size is less than a page... - * we're being asked to ZEROFILL either the head or the tail of the I/O... + * make sure the uio_resid isn't too big... + * internally, we want to handle all of the I/O in + * chunk sizes that fit in a 32 bit int */ - return (cluster_write_x(vp, uio, oldEOF, newEOF, headOff, tailOff, flags)); - } - // LP64todo - fix this! - else if (((int)uio->uio_offset & PAGE_MASK) || (CAST_DOWN(int, iov_base) & PAGE_MASK)) { - if (((int)uio->uio_offset & PAGE_MASK) == (CAST_DOWN(int, iov_base) & PAGE_MASK)) { + if (cur_resid > (user_ssize_t)(MAX_IO_REQUEST_SIZE)) { /* - * Bring the file offset write up to a pagesize boundary - * this will also bring the base address to a page boundary - * since they both are currently on the same offset within a page - * note: if we get here, uio->uio_resid is greater than PAGE_SIZE - * so the computed clip_size must always be less than the current uio_resid + * we're going to have to call cluster_write_copy + * more than once... + * + * only want the last call to cluster_write_copy to + * have the IO_TAILZEROFILL flag set and only the + * first call should have IO_HEADZEROFILL */ - clip_size = (PAGE_SIZE - (uio->uio_offset & PAGE_MASK_64)); + zflags = flags & ~IO_TAILZEROFILL; + flags &= ~IO_HEADZEROFILL; - /* - * Fake the resid going into the cluster_write_x call - * and restore it on the way out. + write_length = MAX_IO_REQUEST_SIZE; + } else { + /* + * last call to cluster_write_copy */ - // LP64todo - fix this - prev_resid = uio_resid(uio); - uio_setresid(uio, clip_size); + zflags = flags; + + write_length = (u_int32_t)cur_resid; + } + retval = cluster_write_copy(vp, uio, write_length, oldEOF, newEOF, headOff, tailOff, zflags, callback, callback_arg); + break; - retval = cluster_write_x(vp, uio, oldEOF, newEOF, headOff, tailOff, flags); + case IO_CONTIG: + zflags = flags & ~(IO_TAILZEROFILL | IO_HEADZEROFILL); - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); - } else { - /* - * can't get both the file offset and the buffer offset aligned to a page boundary - * so fire an I/O through the cache for this entire vector + if (flags & IO_HEADZEROFILL) { + /* + * only do this once per request */ - // LP64todo - fix this - clip_size = iov_len; - // LP64todo - fix this - prev_resid = uio_resid(uio); - uio_setresid(uio, clip_size); - - retval = cluster_write_x(vp, uio, oldEOF, newEOF, headOff, tailOff, flags); + flags &= ~IO_HEADZEROFILL; - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); + retval = cluster_write_copy(vp, (struct uio *)0, (u_int32_t)0, (off_t)0, uio->uio_offset, + headOff, (off_t)0, zflags | IO_HEADZEROFILL | IO_SYNC, callback, callback_arg); + if (retval) + break; } - } else { - /* - * If we come in here, we know the offset into - * the file is on a pagesize boundary and the - * target buffer address is also on a page boundary - */ - max_io_size = newEOF - uio->uio_offset; - // LP64todo - fix this - clip_size = uio_resid(uio); - if (iov_len < clip_size) - // LP64todo - fix this! - clip_size = iov_len; - if (max_io_size < clip_size) - clip_size = max_io_size; - - if (clip_size < PAGE_SIZE) { - /* - * Take care of tail end of write in this vector + retval = cluster_write_contig(vp, uio, newEOF, &write_type, &write_length, callback, callback_arg, bflag); + + if (retval == 0 && (flags & IO_TAILZEROFILL) && uio_resid(uio) == 0) { + /* + * we're done with the data from the user specified buffer(s) + * and we've been requested to zero fill at the tail + * treat this as an IO_HEADZEROFILL which doesn't require a uio + * by rearranging the args and passing in IO_HEADZEROFILL */ - // LP64todo - fix this - prev_resid = uio_resid(uio); - uio_setresid(uio, clip_size); - - retval = cluster_write_x(vp, uio, oldEOF, newEOF, headOff, tailOff, flags); - - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); - } else { - /* round clip_size down to a multiple of pagesize */ - clip_size = clip_size & ~(PAGE_MASK); - // LP64todo - fix this - prev_resid = uio_resid(uio); - uio_setresid(uio, clip_size); - - retval = cluster_nocopy_write(vp, uio, newEOF); - - if ((retval == 0) && uio_resid(uio)) - retval = cluster_write_x(vp, uio, oldEOF, newEOF, headOff, tailOff, flags); - - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); + retval = cluster_write_copy(vp, (struct uio *)0, (u_int32_t)0, (off_t)0, tailOff, uio->uio_offset, + (off_t)0, zflags | IO_HEADZEROFILL | IO_SYNC, callback, callback_arg); } - } /* end else */ - } /* end while */ + break; - return(retval); + case IO_DIRECT: + /* + * cluster_write_direct is never called with IO_TAILZEROFILL || IO_HEADZEROFILL + */ + retval = cluster_write_direct(vp, uio, oldEOF, newEOF, &write_type, &write_length, flags, callback, callback_arg); + break; + + case IO_UNKNOWN: + retval = cluster_io_type(uio, &write_type, &write_length, MIN_DIRECT_WRITE_SIZE); + break; + } + /* + * in case we end up calling cluster_write_copy (from cluster_write_direct) + * multiple times to service a multi-vector request that is not aligned properly + * we need to update the oldEOF so that we + * don't zero-fill the head of a page if we've successfully written + * data to that area... 'cluster_write_copy' will zero-fill the head of a + * page that is beyond the oldEOF if the write is unaligned... we only + * want that to happen for the very first page of the cluster_write, + * NOT the first page of each vector making up a multi-vector write. + */ + if (uio->uio_offset > oldEOF) + oldEOF = uio->uio_offset; + } + return (retval); } static int -cluster_nocopy_write(vnode_t vp, struct uio *uio, off_t newEOF) +cluster_write_direct(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, int *write_type, u_int32_t *write_length, + int flags, int (*callback)(buf_t, void *), void *callback_arg) { upl_t upl; upl_page_info_t *pl; vm_offset_t upl_offset; - int io_size; - int io_flag; - int upl_size; - int upl_needed_size; - int pages_in_pl; - int upl_flags; + vm_offset_t vector_upl_offset = 0; + u_int32_t io_req_size; + u_int32_t offset_in_file; + u_int32_t offset_in_iovbase; + u_int32_t io_size; + int io_flag = 0; + upl_size_t upl_size, vector_upl_size = 0; + vm_size_t upl_needed_size; + mach_msg_type_number_t pages_in_pl; + upl_control_flags_t upl_flags; kern_return_t kret; - int i; + mach_msg_type_number_t i; int force_data_sync; - int error = 0; + int retval = 0; + int first_IO = 1; struct clios iostate; - struct cl_writebehind *wbp; + user_addr_t iov_base; + u_int32_t mem_alignment_mask; + u_int32_t devblocksize; + u_int32_t max_io_size; + u_int32_t max_upl_size; + u_int32_t max_vector_size; + u_int32_t bytes_outstanding_limit; + boolean_t io_throttled = FALSE; + u_int32_t vector_upl_iosize = 0; + int issueVectorUPL = 0,useVectorUPL = (uio->uio_iovcnt > 1); + off_t v_upl_uio_offset = 0; + int vector_upl_index=0; + upl_t vector_upl = NULL; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 75)) | DBG_FUNC_START, - (int)uio->uio_offset, (int)uio_resid(uio), - (int)newEOF, 0, 0); /* * When we enter this routine, we know - * -- the offset into the file is on a pagesize boundary - * -- the resid is a page multiple * -- the resid will not exceed iov_len */ - - if ((wbp = cluster_get_wbp(vp, CLW_RETURNLOCKED)) != NULL) { + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 75)) | DBG_FUNC_START, + (int)uio->uio_offset, *write_length, (int)newEOF, 0, 0); - cluster_try_push(wbp, vp, newEOF, 0, 1); + max_upl_size = cluster_max_io_size(vp->v_mount, CL_WRITE); + + io_flag = CL_ASYNC | CL_PRESERVE | CL_COMMIT | CL_THROTTLE | CL_DIRECT_IO; + + if (flags & IO_PASSIVE) + io_flag |= CL_PASSIVE; + + if (flags & IO_NOCACHE) + io_flag |= CL_NOCACHE; + + if (flags & IO_SKIP_ENCRYPTION) + io_flag |= CL_ENCRYPTED; - lck_mtx_unlock(&wbp->cl_lockw); - } iostate.io_completed = 0; iostate.io_issued = 0; iostate.io_error = 0; iostate.io_wanted = 0; - while (uio_resid(uio) && uio->uio_offset < newEOF && error == 0) { - user_addr_t iov_base; + lck_mtx_init(&iostate.io_mtxp, cl_mtx_grp, cl_mtx_attr); + + mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask; + devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize; + + if (devblocksize == 1) { + /* + * the AFP client advertises a devblocksize of 1 + * however, its BLOCKMAP routine maps to physical + * blocks that are PAGE_SIZE in size... + * therefore we can't ask for I/Os that aren't page aligned + * or aren't multiples of PAGE_SIZE in size + * by setting devblocksize to PAGE_SIZE, we re-instate + * the old behavior we had before the mem_alignment_mask + * changes went in... + */ + devblocksize = PAGE_SIZE; + } + +next_dwrite: + io_req_size = *write_length; + iov_base = uio_curriovbase(uio); + + offset_in_file = (u_int32_t)uio->uio_offset & PAGE_MASK; + offset_in_iovbase = (u_int32_t)iov_base & mem_alignment_mask; + + if (offset_in_file || offset_in_iovbase) { + /* + * one of the 2 important offsets is misaligned + * so fire an I/O through the cache for this entire vector + */ + goto wait_for_dwrites; + } + if (iov_base & (devblocksize - 1)) { + /* + * the offset in memory must be on a device block boundary + * so that we can guarantee that we can generate an + * I/O that ends on a page boundary in cluster_io + */ + goto wait_for_dwrites; + } - io_size = uio_resid(uio); + task_update_logical_writes(current_task(), (io_req_size & ~PAGE_MASK), TASK_WRITE_IMMEDIATE, vp); + while (io_req_size >= PAGE_SIZE && uio->uio_offset < newEOF && retval == 0) { + int throttle_type; + + if ( (throttle_type = cluster_is_throttled(vp)) ) { + /* + * we're in the throttle window, at the very least + * we want to limit the size of the I/O we're about + * to issue + */ + if ( (flags & IO_RETURN_ON_THROTTLE) && throttle_type == THROTTLE_NOW) { + /* + * we're in the throttle window and at least 1 I/O + * has already been issued by a throttleable thread + * in this window, so return with EAGAIN to indicate + * to the FS issuing the cluster_write call that it + * should now throttle after dropping any locks + */ + throttle_info_update_by_mount(vp->v_mount); - if (io_size > (MAX_UPL_TRANSFER * PAGE_SIZE)) - io_size = MAX_UPL_TRANSFER * PAGE_SIZE; + io_throttled = TRUE; + goto wait_for_dwrites; + } + max_vector_size = THROTTLE_MAX_IOSIZE; + max_io_size = THROTTLE_MAX_IOSIZE; + } else { + max_vector_size = MAX_VECTOR_UPL_SIZE; + max_io_size = max_upl_size; + } + if (first_IO) { + cluster_syncup(vp, newEOF, callback, callback_arg, callback ? PUSH_SYNC : 0); + first_IO = 0; + } + io_size = io_req_size & ~PAGE_MASK; iov_base = uio_curriovbase(uio); - // LP64todo - fix this! - upl_offset = CAST_DOWN(vm_offset_t, iov_base) & PAGE_MASK; - + if (io_size > max_io_size) + io_size = max_io_size; + + if(useVectorUPL && (iov_base & PAGE_MASK)) { + /* + * We have an iov_base that's not page-aligned. + * Issue all I/O's that have been collected within + * this Vectored UPL. + */ + if(vector_upl_index) { + retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + reset_vector_run_state(); + } + + /* + * After this point, if we are using the Vector UPL path and the base is + * not page-aligned then the UPL with that base will be the first in the vector UPL. + */ + } + + upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK); upl_needed_size = (upl_offset + io_size + (PAGE_SIZE -1)) & ~PAGE_MASK; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_START, (int)upl_offset, upl_needed_size, (int)iov_base, io_size, 0); + vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map; for (force_data_sync = 0; force_data_sync < 3; force_data_sync++) { pages_in_pl = 0; upl_size = upl_needed_size; upl_flags = UPL_FILE_IO | UPL_COPYOUT_FROM | UPL_NO_SYNC | - UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE; + UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE + | UPL_MEMORY_TAG_MAKE(VM_KERN_MEMORY_FILE); - // LP64todo - fix this! - kret = vm_map_get_upl(current_map(), + kret = vm_map_get_upl(map, (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)), &upl_size, &upl, @@ -1613,13 +2532,13 @@ cluster_nocopy_write(vnode_t vp, struct uio *uio, off_t newEOF) KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_END, 0, 0, 0, kret, 0); /* - * cluster_nocopy_write: failed to get pagelist + * failed to get pagelist * * we may have already spun some portion of this request * off as async requests... we need to wait for the I/O * to complete before returning */ - goto wait_for_writes; + goto wait_for_dwrites; } pl = UPL_GET_INTERNAL_PAGE_LIST(upl); pages_in_pl = upl_size / PAGE_SIZE; @@ -1635,8 +2554,7 @@ cluster_nocopy_write(vnode_t vp, struct uio *uio, off_t newEOF) * didn't get all the pages back that we * needed... release this upl and try again */ - ubc_upl_abort_range(upl, (upl_offset & ~PAGE_MASK), upl_size, - UPL_ABORT_FREE_ON_EMPTY); + ubc_upl_abort(upl, 0); } if (force_data_sync >= 3) { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_END, @@ -1649,35 +2567,42 @@ cluster_nocopy_write(vnode_t vp, struct uio *uio, off_t newEOF) * off as async requests... we need to wait for the I/O * to complete before returning */ - goto wait_for_writes; + goto wait_for_dwrites; } /* * Consider the possibility that upl_size wasn't satisfied. */ - if (upl_size != upl_needed_size) - io_size = (upl_size - (int)upl_offset) & ~PAGE_MASK; - + if (upl_size < upl_needed_size) { + if (upl_size && upl_offset == 0) + io_size = upl_size; + else + io_size = 0; + } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_END, (int)upl_offset, upl_size, (int)iov_base, io_size, 0); if (io_size == 0) { - ubc_upl_abort_range(upl, (upl_offset & ~PAGE_MASK), upl_size, - UPL_ABORT_FREE_ON_EMPTY); + ubc_upl_abort(upl, 0); /* * we may have already spun some portion of this request * off as async requests... we need to wait for the I/O * to complete before returning */ - goto wait_for_writes; + goto wait_for_dwrites; + } + + if(useVectorUPL) { + vm_offset_t end_off = ((iov_base + io_size) & PAGE_MASK); + if(end_off) + issueVectorUPL = 1; + /* + * After this point, if we are using a vector UPL, then + * either all the UPL elements end on a page boundary OR + * this UPL is the last element because it does not end + * on a page boundary. + */ } - /* - * Now look for pages already in the cache - * and throw them away. - * uio->uio_offset is page aligned within the file - * io_size is a multiple of PAGE_SIZE - */ - ubc_range_op(vp, uio->uio_offset, uio->uio_offset + io_size, UPL_ROP_DUMP, NULL); /* * we want push out these writes asynchronously so that we can overlap @@ -1685,13 +2610,12 @@ cluster_nocopy_write(vnode_t vp, struct uio *uio, off_t newEOF) * if there are already too many outstanding writes * wait until some complete before issuing the next */ - lck_mtx_lock(cl_mtxp); + if (vp->v_mount->mnt_minsaturationbytecount) + bytes_outstanding_limit = vp->v_mount->mnt_minsaturationbytecount; + else + bytes_outstanding_limit = max_upl_size * IO_SCALE(vp, 2); - while ((iostate.io_issued - iostate.io_completed) > (2 * MAX_UPL_TRANSFER * PAGE_SIZE)) { - iostate.io_wanted = 1; - msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_nocopy_write", 0); - } - lck_mtx_unlock(cl_mtxp); + cluster_iostate_wait(&iostate, bytes_outstanding_limit, "cluster_write_direct"); if (iostate.io_error) { /* @@ -1701,180 +2625,359 @@ cluster_nocopy_write(vnode_t vp, struct uio *uio, off_t newEOF) * go wait for all writes that are part of this stream * to complete before returning the error to the caller */ - ubc_upl_abort_range(upl, (upl_offset & ~PAGE_MASK), upl_size, - UPL_ABORT_FREE_ON_EMPTY); + ubc_upl_abort(upl, 0); - goto wait_for_writes; + goto wait_for_dwrites; } - io_flag = CL_ASYNC | CL_PRESERVE | CL_COMMIT | CL_THROTTLE; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 77)) | DBG_FUNC_START, (int)upl_offset, (int)uio->uio_offset, io_size, io_flag, 0); - error = cluster_io(vp, upl, upl_offset, uio->uio_offset, - io_size, io_flag, (buf_t)NULL, &iostate); + if(!useVectorUPL) + retval = cluster_io(vp, upl, upl_offset, uio->uio_offset, + io_size, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + + else { + if(!vector_upl_index) { + vector_upl = vector_upl_create(upl_offset); + v_upl_uio_offset = uio->uio_offset; + vector_upl_offset = upl_offset; + } + + vector_upl_set_subupl(vector_upl,upl,upl_size); + vector_upl_set_iostate(vector_upl, upl, vector_upl_size, upl_size); + vector_upl_index++; + vector_upl_iosize += io_size; + vector_upl_size += upl_size; + + if(issueVectorUPL || vector_upl_index == MAX_VECTOR_UPL_ELEMENTS || vector_upl_size >= max_vector_size) { + retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + reset_vector_run_state(); + } + } + /* + * update the uio structure to + * reflect the I/O that we just issued + */ uio_update(uio, (user_size_t)io_size); + /* + * in case we end up calling through to cluster_write_copy to finish + * the tail of this request, we need to update the oldEOF so that we + * don't zero-fill the head of a page if we've successfully written + * data to that area... 'cluster_write_copy' will zero-fill the head of a + * page that is beyond the oldEOF if the write is unaligned... we only + * want that to happen for the very first page of the cluster_write, + * NOT the first page of each vector making up a multi-vector write. + */ + if (uio->uio_offset > oldEOF) + oldEOF = uio->uio_offset; + + io_req_size -= io_size; + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 77)) | DBG_FUNC_END, - (int)upl_offset, (int)uio->uio_offset, (int)uio_resid(uio), error, 0); + (int)upl_offset, (int)uio->uio_offset, io_req_size, retval, 0); } /* end while */ -wait_for_writes: + if (retval == 0 && iostate.io_error == 0 && io_req_size == 0) { + + retval = cluster_io_type(uio, write_type, write_length, MIN_DIRECT_WRITE_SIZE); + + if (retval == 0 && *write_type == IO_DIRECT) { + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 75)) | DBG_FUNC_NONE, + (int)uio->uio_offset, *write_length, (int)newEOF, 0, 0); + + goto next_dwrite; + } + } + +wait_for_dwrites: + + if (retval == 0 && iostate.io_error == 0 && useVectorUPL && vector_upl_index) { + retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + reset_vector_run_state(); + } /* * make sure all async writes issued as part of this stream * have completed before we return */ - lck_mtx_lock(cl_mtxp); + cluster_iostate_wait(&iostate, 0, "cluster_write_direct"); - while (iostate.io_issued != iostate.io_completed) { - iostate.io_wanted = 1; - msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_nocopy_write", 0); - } - lck_mtx_unlock(cl_mtxp); + if (iostate.io_error) + retval = iostate.io_error; + + lck_mtx_destroy(&iostate.io_mtxp, cl_mtx_grp); + + if (io_throttled == TRUE && retval == 0) + retval = EAGAIN; + + if (io_req_size && retval == 0) { + /* + * we couldn't handle the tail of this request in DIRECT mode + * so fire it through the copy path + * + * note that flags will never have IO_HEADZEROFILL or IO_TAILZEROFILL set + * so we can just pass 0 in for the headOff and tailOff + */ + if (uio->uio_offset > oldEOF) + oldEOF = uio->uio_offset; - if (iostate.io_error) - error = iostate.io_error; + retval = cluster_write_copy(vp, uio, io_req_size, oldEOF, newEOF, (off_t)0, (off_t)0, flags, callback, callback_arg); + *write_type = IO_UNKNOWN; + } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 75)) | DBG_FUNC_END, - (int)uio->uio_offset, (int)uio->uio_resid, error, 4, 0); + (int)uio->uio_offset, io_req_size, retval, 4, 0); - return (error); + return (retval); } static int -cluster_phys_write(vnode_t vp, struct uio *uio, off_t newEOF) +cluster_write_contig(vnode_t vp, struct uio *uio, off_t newEOF, int *write_type, u_int32_t *write_length, + int (*callback)(buf_t, void *), void *callback_arg, int bflag) { upl_page_info_t *pl; - addr64_t src_paddr; - upl_t upl; + addr64_t src_paddr = 0; + upl_t upl[MAX_VECTS]; vm_offset_t upl_offset; - int tail_size; - int io_size; - int upl_size; - int upl_needed_size; - int pages_in_pl; - int upl_flags; + u_int32_t tail_size = 0; + u_int32_t io_size; + u_int32_t xsize; + upl_size_t upl_size; + vm_size_t upl_needed_size; + mach_msg_type_number_t pages_in_pl; + upl_control_flags_t upl_flags; kern_return_t kret; + struct clios iostate; int error = 0; + int cur_upl = 0; + int num_upl = 0; + int n; user_addr_t iov_base; - int devblocksize; - struct cl_writebehind *wbp; + u_int32_t devblocksize; + u_int32_t mem_alignment_mask; - devblocksize = vp->v_mount->mnt_devblocksize; /* * When we enter this routine, we know - * -- the resid will not exceed iov_len - * -- the vector target address is physcially contiguous + * -- the io_req_size will not exceed iov_len + * -- the target address is physically contiguous */ - if ((wbp = cluster_get_wbp(vp, CLW_RETURNLOCKED)) != NULL) { + cluster_syncup(vp, newEOF, callback, callback_arg, callback ? PUSH_SYNC : 0); - cluster_try_push(wbp, vp, newEOF, 0, 1); + devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize; + mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask; - lck_mtx_unlock(&wbp->cl_lockw); - } -#if LP64_DEBUG - if (IS_VALID_UIO_SEGFLG(uio->uio_segflg) == 0) { - panic("%s :%d - invalid uio_segflg\n", __FILE__, __LINE__); - } -#endif /* LP64_DEBUG */ + iostate.io_completed = 0; + iostate.io_issued = 0; + iostate.io_error = 0; + iostate.io_wanted = 0; + + lck_mtx_init(&iostate.io_mtxp, cl_mtx_grp, cl_mtx_attr); + +next_cwrite: + io_size = *write_length; - // LP64todo - fix this! - io_size = (int)uio_curriovlen(uio); iov_base = uio_curriovbase(uio); - upl_offset = CAST_DOWN(upl_offset_t, iov_base) & PAGE_MASK; + upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK); upl_needed_size = upl_offset + io_size; pages_in_pl = 0; upl_size = upl_needed_size; upl_flags = UPL_FILE_IO | UPL_COPYOUT_FROM | UPL_NO_SYNC | - UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE; + UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE + | UPL_MEMORY_TAG_MAKE(VM_KERN_MEMORY_FILE); - // LP64todo - fix this! - kret = vm_map_get_upl(current_map(), + vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map; + kret = vm_map_get_upl(map, (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)), - &upl_size, &upl, NULL, &pages_in_pl, &upl_flags, 0); + &upl_size, &upl[cur_upl], NULL, &pages_in_pl, &upl_flags, 0); if (kret != KERN_SUCCESS) { /* - * cluster_phys_write: failed to get pagelist - * note: return kret here + * failed to get pagelist */ - return(EINVAL); + error = EINVAL; + goto wait_for_cwrites; } + num_upl++; + /* * Consider the possibility that upl_size wasn't satisfied. - * This is a failure in the physical memory case. */ if (upl_size < upl_needed_size) { - ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); - return(EINVAL); + /* + * This is a failure in the physical memory case. + */ + error = EINVAL; + goto wait_for_cwrites; } - pl = ubc_upl_pageinfo(upl); + pl = ubc_upl_pageinfo(upl[cur_upl]); - src_paddr = ((addr64_t)upl_phys_page(pl, 0) << 12) + (addr64_t)upl_offset; + src_paddr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + (addr64_t)upl_offset; while (((uio->uio_offset & (devblocksize - 1)) || io_size < devblocksize) && io_size) { - int head_size; + u_int32_t head_size; - head_size = devblocksize - (int)(uio->uio_offset & (devblocksize - 1)); + head_size = devblocksize - (u_int32_t)(uio->uio_offset & (devblocksize - 1)); if (head_size > io_size) head_size = io_size; - error = cluster_align_phys_io(vp, uio, src_paddr, head_size, 0); + error = cluster_align_phys_io(vp, uio, src_paddr, head_size, 0, callback, callback_arg); - if (error) { - ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); + if (error) + goto wait_for_cwrites; - return(EINVAL); - } upl_offset += head_size; src_paddr += head_size; io_size -= head_size; + + iov_base += head_size; + } + if ((u_int32_t)iov_base & mem_alignment_mask) { + /* + * request doesn't set up on a memory boundary + * the underlying DMA engine can handle... + * return an error instead of going through + * the slow copy path since the intent of this + * path is direct I/O from device memory + */ + error = EINVAL; + goto wait_for_cwrites; } + tail_size = io_size & (devblocksize - 1); io_size -= tail_size; - if (io_size) { - /* - * issue a synchronous write to cluster_io + while (io_size && error == 0) { + + if (io_size > MAX_IO_CONTIG_SIZE) + xsize = MAX_IO_CONTIG_SIZE; + else + xsize = io_size; + /* + * request asynchronously so that we can overlap + * the preparation of the next I/O... we'll do + * the commit after all the I/O has completed + * since its all issued against the same UPL + * if there are already too many outstanding writes + * wait until some have completed before issuing the next */ - error = cluster_io(vp, upl, upl_offset, uio->uio_offset, - io_size, CL_DEV_MEMORY, (buf_t)NULL, (struct clios *)NULL); - } - if (error == 0) { + cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_write_contig"); + + if (iostate.io_error) { + /* + * one of the earlier writes we issued ran into a hard error + * don't issue any more writes... + * go wait for all writes that are part of this stream + * to complete before returning the error to the caller + */ + goto wait_for_cwrites; + } /* - * The cluster_io write completed successfully, - * update the uio structure + * issue an asynchronous write to cluster_io */ - uio_update(uio, (user_size_t)io_size); + error = cluster_io(vp, upl[cur_upl], upl_offset, uio->uio_offset, + xsize, CL_DEV_MEMORY | CL_ASYNC | bflag, (buf_t)NULL, (struct clios *)&iostate, callback, callback_arg); - src_paddr += io_size; + if (error == 0) { + /* + * The cluster_io write completed successfully, + * update the uio structure + */ + uio_update(uio, (user_size_t)xsize); - if (tail_size) - error = cluster_align_phys_io(vp, uio, src_paddr, tail_size, 0); + upl_offset += xsize; + src_paddr += xsize; + io_size -= xsize; + } } + if (error == 0 && iostate.io_error == 0 && tail_size == 0 && num_upl < MAX_VECTS) { + + error = cluster_io_type(uio, write_type, write_length, 0); + + if (error == 0 && *write_type == IO_CONTIG) { + cur_upl++; + goto next_cwrite; + } + } else + *write_type = IO_UNKNOWN; + +wait_for_cwrites: /* - * just release our hold on the physically contiguous - * region without changing any state - */ - ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); + * make sure all async writes that are part of this stream + * have completed before we proceed + */ + cluster_iostate_wait(&iostate, 0, "cluster_write_contig"); + + if (iostate.io_error) + error = iostate.io_error; + + lck_mtx_destroy(&iostate.io_mtxp, cl_mtx_grp); + + if (error == 0 && tail_size) + error = cluster_align_phys_io(vp, uio, src_paddr, tail_size, 0, callback, callback_arg); + + for (n = 0; n < num_upl; n++) + /* + * just release our hold on each physically contiguous + * region without changing any state + */ + ubc_upl_abort(upl[n], 0); return (error); } +/* + * need to avoid a race between an msync of a range of pages dirtied via mmap + * vs a filesystem such as HFS deciding to write a 'hole' to disk via cluster_write's + * zerofill mechanism before it has seen the VNOP_PAGEOUTs for the pages being msync'd + * + * we should never force-zero-fill pages that are already valid in the cache... + * the entire page contains valid data (either from disk, zero-filled or dirtied + * via an mmap) so we can only do damage by trying to zero-fill + * + */ +static int +cluster_zero_range(upl_t upl, upl_page_info_t *pl, int flags, int io_offset, off_t zero_off, off_t upl_f_offset, int bytes_to_zero) +{ + int zero_pg_index; + boolean_t need_cluster_zero = TRUE; + + if ((flags & (IO_NOZEROVALID | IO_NOZERODIRTY))) { + + bytes_to_zero = min(bytes_to_zero, PAGE_SIZE - (int)(zero_off & PAGE_MASK_64)); + zero_pg_index = (int)((zero_off - upl_f_offset) / PAGE_SIZE_64); + + if (upl_valid_page(pl, zero_pg_index)) { + /* + * never force zero valid pages - dirty or clean + * we'll leave these in the UPL for cluster_write_copy to deal with + */ + need_cluster_zero = FALSE; + } + } + if (need_cluster_zero == TRUE) + cluster_zero(upl, io_offset, bytes_to_zero, NULL); + + return (bytes_to_zero); +} + + static int -cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t headOff, off_t tailOff, int flags) +cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t oldEOF, off_t newEOF, off_t headOff, + off_t tailOff, int flags, int (*callback)(buf_t, void *), void *callback_arg) { upl_page_info_t *pl; upl_t upl; vm_offset_t upl_offset = 0; - int upl_size; + vm_size_t upl_size; off_t upl_f_offset; int pages_in_upl; int start_offset; @@ -1891,50 +2994,44 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h off_t zero_off; long long zero_cnt1; off_t zero_off1; + off_t write_off = 0; + int write_cnt = 0; + boolean_t first_pass = FALSE; struct cl_extent cl; - int intersection; struct cl_writebehind *wbp; + int bflag; + u_int max_cluster_pgcount; + u_int max_io_size; - if ((wbp = cluster_get_wbp(vp, 0)) != NULL) - { - if (wbp->cl_hasbeenpaged) { - /* - * this vnode had pages cleaned to it by - * the pager which indicates that either - * it's not very 'hot', or the system is - * being overwhelmed by a lot of dirty - * data being delayed in the VM cache... - * in either event, we'll push our remaining - * delayed data at this point... this will - * be more efficient than paging out 1 page at - * a time, and will also act as a throttle - * by delaying this client from writing any - * more data until all his delayed data has - * at least been queued to the uderlying driver. - */ - if (wbp->cl_number || wbp->cl_scmap) - cluster_push_EOF(vp, newEOF); - - wbp->cl_hasbeenpaged = 0; - } - } if (uio) { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_START, - (int)uio->uio_offset, uio_resid(uio), (int)oldEOF, (int)newEOF, 0); + (int)uio->uio_offset, io_req_size, (int)oldEOF, (int)newEOF, 0); - // LP64todo - fix this - io_resid = uio_resid(uio); + io_resid = io_req_size; } else { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_START, 0, 0, (int)oldEOF, (int)newEOF, 0); io_resid = 0; } + if (flags & IO_PASSIVE) + bflag = CL_PASSIVE; + else + bflag = 0; + if (flags & IO_NOCACHE) + bflag |= CL_NOCACHE; + + if (flags & IO_SKIP_ENCRYPTION) + bflag |= CL_ENCRYPTED; + zero_cnt = 0; zero_cnt1 = 0; zero_off = 0; zero_off1 = 0; + max_cluster_pgcount = MAX_CLUSTER_SIZE(vp) / PAGE_SIZE; + max_io_size = cluster_max_io_size(vp->v_mount, CL_WRITE); + if (flags & IO_HEADZEROFILL) { /* * some filesystems (HFS is one) don't support unallocated holes within a file... @@ -1952,22 +3049,50 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h zero_cnt = newEOF - headOff; zero_off = headOff; } + } else { + if (uio && uio->uio_offset > oldEOF) { + zero_off = uio->uio_offset & ~PAGE_MASK_64; + + if (zero_off >= oldEOF) { + zero_cnt = uio->uio_offset - zero_off; + + flags |= IO_HEADZEROFILL; + } + } } if (flags & IO_TAILZEROFILL) { if (uio) { - // LP64todo - fix this - zero_off1 = uio->uio_offset + uio_resid(uio); + zero_off1 = uio->uio_offset + io_req_size; if (zero_off1 < tailOff) zero_cnt1 = tailOff - zero_off1; } + } else { + if (uio && newEOF > oldEOF) { + zero_off1 = uio->uio_offset + io_req_size; + + if (zero_off1 == newEOF && (zero_off1 & PAGE_MASK_64)) { + zero_cnt1 = PAGE_SIZE_64 - (zero_off1 & PAGE_MASK_64); + + flags |= IO_TAILZEROFILL; + } + } } if (zero_cnt == 0 && uio == (struct uio *) 0) { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_END, retval, 0, 0, 0, 0); return (0); } - + if (uio) { + write_off = uio->uio_offset; + write_cnt = uio_resid(uio); + /* + * delay updating the sequential write info + * in the control block until we've obtained + * the lock for it + */ + first_pass = TRUE; + } while ((total_size = (io_resid + zero_cnt + zero_cnt1)) && retval == 0) { /* * for this iteration of the loop, figure out where our starting point is @@ -1985,26 +3110,26 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 46)) | DBG_FUNC_NONE, (int)zero_off, (int)zero_cnt, (int)zero_off1, (int)zero_cnt1, 0); - if (total_size > (MAX_UPL_TRANSFER * PAGE_SIZE)) - total_size = MAX_UPL_TRANSFER * PAGE_SIZE; + if (total_size > max_io_size) + total_size = max_io_size; cl.b_addr = (daddr64_t)(upl_f_offset / PAGE_SIZE_64); - if (uio && ((flags & (IO_NOCACHE | IO_SYNC | IO_HEADZEROFILL | IO_TAILZEROFILL)) == 0)) { + if (uio && ((flags & (IO_SYNC | IO_HEADZEROFILL | IO_TAILZEROFILL)) == 0)) { /* * assumption... total_size <= io_resid * because IO_HEADZEROFILL and IO_TAILZEROFILL not set */ - if ((start_offset + total_size) > (MAX_UPL_TRANSFER * PAGE_SIZE)) - total_size -= start_offset; + if ((start_offset + total_size) > max_io_size) + total_size = max_io_size - start_offset; xfer_resid = total_size; - retval = cluster_copy_ubc_data(vp, uio, &xfer_resid, 1); - + retval = cluster_copy_ubc_data_internal(vp, uio, &xfer_resid, 1, 1); + if (retval) break; - io_resid -= (total_size - xfer_resid); + io_resid -= (total_size - xfer_resid); total_size = xfer_resid; start_offset = (int)(uio->uio_offset & PAGE_MASK_64); upl_f_offset = uio->uio_offset - start_offset; @@ -2036,8 +3161,8 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h */ upl_size = (start_offset + total_size + (PAGE_SIZE - 1)) & ~PAGE_MASK; - if (upl_size > (MAX_UPL_TRANSFER * PAGE_SIZE)) - upl_size = MAX_UPL_TRANSFER * PAGE_SIZE; + if (upl_size > max_io_size) + upl_size = max_io_size; pages_in_upl = upl_size / PAGE_SIZE; io_size = upl_size - start_offset; @@ -2058,14 +3183,14 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h upl_size, &upl, &pl, - UPL_SET_LITE | UPL_WILL_MODIFY); + UPL_SET_LITE | (( uio!=NULL && (uio->uio_flags & UIO_FLAGS_IS_COMPRESSED_FILE)) ? 0 : UPL_WILL_MODIFY)); if (kret != KERN_SUCCESS) - panic("cluster_write: failed to get pagelist"); + panic("cluster_write_copy: failed to get pagelist"); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 41)) | DBG_FUNC_END, - (int)upl, (int)upl_f_offset, start_offset, 0, 0); + upl, (int)upl_f_offset, start_offset, 0, 0); - if (start_offset && !upl_valid_page(pl, 0)) { + if (start_offset && upl_f_offset < oldEOF && !upl_valid_page(pl, 0)) { int read_size; /* @@ -2075,11 +3200,11 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h */ read_size = PAGE_SIZE; - if ((upl_f_offset + read_size) > newEOF) - read_size = newEOF - upl_f_offset; + if ((upl_f_offset + read_size) > oldEOF) + read_size = oldEOF - upl_f_offset; retval = cluster_io(vp, upl, 0, upl_f_offset, read_size, - CL_READ, (buf_t)NULL, (struct clios *)NULL); + CL_READ | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg); if (retval) { /* * we had an error during the read which causes us to abort @@ -2087,13 +3212,13 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h * to release the rest of the pages in the upl without modifying * there state and mark the failed page in error */ - ubc_upl_abort_range(upl, 0, PAGE_SIZE, UPL_ABORT_DUMP_PAGES); + ubc_upl_abort_range(upl, 0, PAGE_SIZE, UPL_ABORT_DUMP_PAGES|UPL_ABORT_FREE_ON_EMPTY); if (upl_size > PAGE_SIZE) ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 45)) | DBG_FUNC_NONE, - (int)upl, 0, 0, retval, 0); + upl, 0, 0, retval, 0); break; } } @@ -2111,11 +3236,11 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h read_size = PAGE_SIZE; - if ((upl_f_offset + upl_offset + read_size) > newEOF) - read_size = newEOF - (upl_f_offset + upl_offset); + if ((off_t)(upl_f_offset + upl_offset + read_size) > oldEOF) + read_size = oldEOF - (upl_f_offset + upl_offset); retval = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, read_size, - CL_READ, (buf_t)NULL, (struct clios *)NULL); + CL_READ | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg); if (retval) { /* * we had an error during the read which causes us to abort @@ -2123,13 +3248,13 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h * need to release the rest of the pages in the upl without * modifying there state and mark the failed page in error */ - ubc_upl_abort_range(upl, upl_offset, PAGE_SIZE, UPL_ABORT_DUMP_PAGES); + ubc_upl_abort_range(upl, upl_offset, PAGE_SIZE, UPL_ABORT_DUMP_PAGES|UPL_ABORT_FREE_ON_EMPTY); if (upl_size > PAGE_SIZE) ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 45)) | DBG_FUNC_NONE, - (int)upl, 0, 0, retval, 0); + upl, 0, 0, retval, 0); break; } } @@ -2144,40 +3269,28 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h else bytes_to_zero = xfer_resid; - if ( !(flags & (IO_NOZEROVALID | IO_NOZERODIRTY))) { - cluster_zero(upl, io_offset, bytes_to_zero, NULL); - } else { - int zero_pg_index; - - bytes_to_zero = min(bytes_to_zero, PAGE_SIZE - (int)(zero_off & PAGE_MASK_64)); - zero_pg_index = (int)((zero_off - upl_f_offset) / PAGE_SIZE_64); - - if ( !upl_valid_page(pl, zero_pg_index)) { - cluster_zero(upl, io_offset, bytes_to_zero, NULL); + bytes_to_zero = cluster_zero_range(upl, pl, flags, io_offset, zero_off, upl_f_offset, bytes_to_zero); - } else if ((flags & (IO_NOZERODIRTY | IO_NOZEROVALID)) == IO_NOZERODIRTY && - !upl_dirty_page(pl, zero_pg_index)) { - cluster_zero(upl, io_offset, bytes_to_zero, NULL); - } - } xfer_resid -= bytes_to_zero; zero_cnt -= bytes_to_zero; zero_off += bytes_to_zero; io_offset += bytes_to_zero; } if (xfer_resid && io_resid) { + u_int32_t io_requested; + bytes_to_move = min(io_resid, xfer_resid); + io_requested = bytes_to_move; - retval = cluster_copy_upl_data(uio, upl, io_offset, bytes_to_move); + retval = cluster_copy_upl_data(uio, upl, io_offset, (int *)&io_requested); if (retval) { - ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 45)) | DBG_FUNC_NONE, - (int)upl, 0, 0, retval, 0); + upl, 0, 0, retval, 0); } else { - io_resid -= bytes_to_move; + io_resid -= bytes_to_move; xfer_resid -= bytes_to_move; io_offset += bytes_to_move; } @@ -2189,34 +3302,20 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h else bytes_to_zero = xfer_resid; - if ( !(flags & (IO_NOZEROVALID | IO_NOZERODIRTY))) { - cluster_zero(upl, io_offset, bytes_to_zero, NULL); - } else { - int zero_pg_index; - - bytes_to_zero = min(bytes_to_zero, PAGE_SIZE - (int)(zero_off1 & PAGE_MASK_64)); - zero_pg_index = (int)((zero_off1 - upl_f_offset) / PAGE_SIZE_64); - - if ( !upl_valid_page(pl, zero_pg_index)) { - cluster_zero(upl, io_offset, bytes_to_zero, NULL); - } else if ((flags & (IO_NOZERODIRTY | IO_NOZEROVALID)) == IO_NOZERODIRTY && - !upl_dirty_page(pl, zero_pg_index)) { - cluster_zero(upl, io_offset, bytes_to_zero, NULL); - } - } + bytes_to_zero = cluster_zero_range(upl, pl, flags, io_offset, zero_off1, upl_f_offset, bytes_to_zero); + xfer_resid -= bytes_to_zero; zero_cnt1 -= bytes_to_zero; zero_off1 += bytes_to_zero; io_offset += bytes_to_zero; } - if (retval == 0) { int cl_index; - int can_delay; + int ret_cluster_try_push; io_size += start_offset; - if ((upl_f_offset + io_size) >= newEOF && io_size < upl_size) { + if ((upl_f_offset + io_size) >= newEOF && (u_int)io_size < upl_size) { /* * if we're extending the file with this write * we'll zero fill the rest of the page so that @@ -2225,43 +3324,55 @@ cluster_write_x(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t h */ cluster_zero(upl, io_size, upl_size - io_size, NULL); } - if (flags & IO_SYNC) + /* + * release the upl now if we hold one since... + * 1) pages in it may be present in the sparse cluster map + * and may span 2 separate buckets there... if they do and + * we happen to have to flush a bucket to make room and it intersects + * this upl, a deadlock may result on page BUSY + * 2) we're delaying the I/O... from this point forward we're just updating + * the cluster state... no need to hold the pages, so commit them + * 3) IO_SYNC is set... + * because we had to ask for a UPL that provides currenty non-present pages, the + * UPL has been automatically set to clear the dirty flags (both software and hardware) + * upon committing it... this is not the behavior we want since it's possible for + * pages currently present as part of a mapped file to be dirtied while the I/O is in flight. + * we'll pick these pages back up later with the correct behavior specified. + * 4) we don't want to hold pages busy in a UPL and then block on the cluster lock... if a flush + * of this vnode is in progress, we will deadlock if the pages being flushed intersect the pages + * we hold since the flushing context is holding the cluster lock. + */ + ubc_upl_commit_range(upl, 0, upl_size, + UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); +check_cluster: + /* + * calculate the last logical block number + * that this delayed I/O encompassed + */ + cl.e_addr = (daddr64_t)((upl_f_offset + (off_t)upl_size) / PAGE_SIZE_64); + + if (flags & IO_SYNC) { /* * if the IO_SYNC flag is set than we need to * bypass any clusters and immediately issue * the I/O */ goto issue_io; -check_cluster: + } /* * take the lock to protect our accesses * of the writebehind and sparse cluster state */ wbp = cluster_get_wbp(vp, CLW_ALLOCATE | CLW_RETURNLOCKED); - /* - * calculate the last logical block number - * that this delayed I/O encompassed - */ - cl.e_addr = (daddr64_t)((upl_f_offset + (off_t)upl_size) / PAGE_SIZE_64); - if (wbp->cl_scmap) { if ( !(flags & IO_NOCACHE)) { /* * we've fallen into the sparse * cluster method of delaying dirty pages - * first, we need to release the upl if we hold one - * since pages in it may be present in the sparse cluster map - * and may span 2 separate buckets there... if they do and - * we happen to have to flush a bucket to make room and it intersects - * this upl, a deadlock may result on page BUSY */ - if (upl_size) - ubc_upl_commit_range(upl, 0, upl_size, - UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); - - sparse_cluster_add(wbp, vp, &cl, newEOF); + sparse_cluster_add(&(wbp->cl_scmap), vp, &cl, newEOF, callback, callback_arg); lck_mtx_unlock(&wbp->cl_lockw); @@ -2273,21 +3384,10 @@ check_cluster: * to uncached writes on the file, so go ahead * and push whatever's in the sparse map * and switch back to normal clustering - * - * see the comment above concerning a possible deadlock... */ - if (upl_size) { - ubc_upl_commit_range(upl, 0, upl_size, - UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); - /* - * setting upl_size to 0 keeps us from committing a - * second time in the start_new_cluster path - */ - upl_size = 0; - } - sparse_cluster_push(wbp, vp, newEOF, 1); - wbp->cl_number = 0; + + sparse_cluster_push(&(wbp->cl_scmap), vp, newEOF, PUSH_ALL, 0, callback, callback_arg); /* * no clusters of either type present at this point * so just go directly to start_new_cluster since @@ -2296,9 +3396,17 @@ check_cluster: * to avoid the deadlock with sparse_cluster_push */ goto start_new_cluster; - } - upl_offset = 0; + } + if (first_pass) { + if (write_off == wbp->cl_last_write) + wbp->cl_seq_written += write_cnt; + else + wbp->cl_seq_written = write_cnt; + wbp->cl_last_write = write_off + write_cnt; + + first_pass = FALSE; + } if (wbp->cl_number == 0) /* * no clusters currently present @@ -2317,7 +3425,7 @@ check_cluster: /* * the current write starts at or after the current cluster */ - if (cl.e_addr <= (wbp->cl_clusters[cl_index].b_addr + MAX_UPL_TRANSFER)) { + if (cl.e_addr <= (wbp->cl_clusters[cl_index].b_addr + max_cluster_pgcount)) { /* * we have a write that fits entirely * within the existing cluster limits @@ -2329,7 +3437,7 @@ check_cluster: wbp->cl_clusters[cl_index].e_addr = cl.e_addr; break; } - if (cl.b_addr < (wbp->cl_clusters[cl_index].b_addr + MAX_UPL_TRANSFER)) { + if (cl.b_addr < (wbp->cl_clusters[cl_index].b_addr + max_cluster_pgcount)) { /* * we have a write that starts in the middle of the current cluster * but extends beyond the cluster's limit... we know this because @@ -2340,23 +3448,8 @@ check_cluster: * note that we'll always have a leftover tail in this case since * full absorbtion would have occurred in the clause above */ - wbp->cl_clusters[cl_index].e_addr = wbp->cl_clusters[cl_index].b_addr + MAX_UPL_TRANSFER; - - if (upl_size) { - daddr64_t start_pg_in_upl; - - start_pg_in_upl = (daddr64_t)(upl_f_offset / PAGE_SIZE_64); - - if (start_pg_in_upl < wbp->cl_clusters[cl_index].e_addr) { - intersection = (int)((wbp->cl_clusters[cl_index].e_addr - start_pg_in_upl) * PAGE_SIZE); - - ubc_upl_commit_range(upl, upl_offset, intersection, - UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); - upl_f_offset += intersection; - upl_offset += intersection; - upl_size -= intersection; - } - } + wbp->cl_clusters[cl_index].e_addr = wbp->cl_clusters[cl_index].b_addr + max_cluster_pgcount; + cl.b_addr = wbp->cl_clusters[cl_index].e_addr; } /* @@ -2371,7 +3464,7 @@ check_cluster: /* * the current write starts in front of the cluster we're currently considering */ - if ((wbp->cl_clusters[cl_index].e_addr - cl.b_addr) <= MAX_UPL_TRANSFER) { + if ((wbp->cl_clusters[cl_index].e_addr - cl.b_addr) <= max_cluster_pgcount) { /* * we can just merge the new request into * this cluster and leave it in the cache @@ -2384,7 +3477,7 @@ check_cluster: /* * the current write completely * envelops the existing cluster and since - * each write is limited to at most MAX_UPL_TRANSFER bytes + * each write is limited to at most max_cluster_pgcount pages * we can just use the start and last blocknos of the write * to generate the cluster limits */ @@ -2402,29 +3495,14 @@ check_cluster: * get an intersection with the current write * */ - if (cl.e_addr > wbp->cl_clusters[cl_index].e_addr - MAX_UPL_TRANSFER) { + if (cl.e_addr > wbp->cl_clusters[cl_index].e_addr - max_cluster_pgcount) { /* * the current write extends into the proposed cluster * clip the length of the current write after first combining it's * tail with the newly shaped cluster */ - wbp->cl_clusters[cl_index].b_addr = wbp->cl_clusters[cl_index].e_addr - MAX_UPL_TRANSFER; - - if (upl_size) { - intersection = (int)((cl.e_addr - wbp->cl_clusters[cl_index].b_addr) * PAGE_SIZE); - - if (intersection > upl_size) - /* - * because the current write may consist of a number of pages found in the cache - * which are not part of the UPL, we may have an intersection that exceeds - * the size of the UPL that is also part of this write - */ - intersection = upl_size; - - ubc_upl_commit_range(upl, upl_offset + (upl_size - intersection), intersection, - UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); - upl_size -= intersection; - } + wbp->cl_clusters[cl_index].b_addr = wbp->cl_clusters[cl_index].e_addr - max_cluster_pgcount; + cl.e_addr = wbp->cl_clusters[cl_index].b_addr; } /* @@ -2443,14 +3521,36 @@ check_cluster: */ goto delay_io; - if (wbp->cl_number < MAX_CLUSTERS && !(flags & IO_NOCACHE)) + if (!((unsigned int)vfs_flags(vp->v_mount) & MNT_DEFWRITE) && + wbp->cl_number == MAX_CLUSTERS && + wbp->cl_seq_written >= (MAX_CLUSTERS * (max_cluster_pgcount * PAGE_SIZE))) { + uint32_t n; + + if (vp->v_mount->mnt_minsaturationbytecount) { + n = vp->v_mount->mnt_minsaturationbytecount / MAX_CLUSTER_SIZE(vp); + + if (n > MAX_CLUSTERS) + n = MAX_CLUSTERS; + } else + n = 0; + + if (n == 0) { + if (vp->v_mount->mnt_kern_flag & MNTK_SSD) + n = WRITE_BEHIND_SSD; + else + n = WRITE_BEHIND; + } + while (n--) + cluster_try_push(wbp, vp, newEOF, 0, 0, callback, callback_arg, NULL); + } + if (wbp->cl_number < MAX_CLUSTERS) { /* * we didn't find an existing cluster to * merge into, but there's room to start * a new one */ goto start_new_cluster; - + } /* * no exisitng cluster to merge with and no * room to start a new one... we'll try @@ -2461,263 +3561,170 @@ check_cluster: * number of remaining clusters... and * returns the number of currently unused clusters */ - int ret_cluster_try_push = 0; - /* if writes are not deferred, call cluster push immediately */ + ret_cluster_try_push = 0; + + /* + * if writes are not deferred, call cluster push immediately + */ if (!((unsigned int)vfs_flags(vp->v_mount) & MNT_DEFWRITE)) { - if (flags & IO_NOCACHE) - can_delay = 0; - else - can_delay = 1; - ret_cluster_try_push = cluster_try_push(wbp, vp, newEOF, can_delay, 0); + ret_cluster_try_push = cluster_try_push(wbp, vp, newEOF, (flags & IO_NOCACHE) ? 0 : PUSH_DELAY, 0, callback, callback_arg, NULL); } - /* execute following regardless writes are deferred or not */ + /* + * execute following regardless of writes being deferred or not + */ if (ret_cluster_try_push == 0) { /* * no more room in the normal cluster mechanism * so let's switch to the more expansive but expensive * sparse mechanism.... - * first, we need to release the upl if we hold one - * since pages in it may be present in the sparse cluster map (after the cluster_switch) - * and may span 2 separate buckets there... if they do and - * we happen to have to flush a bucket to make room and it intersects - * this upl, a deadlock may result on page BUSY */ - if (upl_size) - ubc_upl_commit_range(upl, upl_offset, upl_size, - UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); - - sparse_cluster_switch(wbp, vp, newEOF); - sparse_cluster_add(wbp, vp, &cl, newEOF); + sparse_cluster_switch(wbp, vp, newEOF, callback, callback_arg); + sparse_cluster_add(&(wbp->cl_scmap), vp, &cl, newEOF, callback, callback_arg); lck_mtx_unlock(&wbp->cl_lockw); continue; } - /* - * we pushed one cluster successfully, so we must be sequentially writing this file - * otherwise, we would have failed and fallen into the sparse cluster support - * so let's take the opportunity to push out additional clusters as long as we - * remain below the throttle... this will give us better I/O locality if we're - * in a copy loop (i.e. we won't jump back and forth between the read and write points - * however, we don't want to push so much out that the write throttle kicks in and - * hangs this thread up until some of the I/O completes... - */ - if (!((unsigned int)vfs_flags(vp->v_mount) & MNT_DEFWRITE)) { - while (wbp->cl_number && (vp->v_numoutput <= (VNODE_ASYNC_THROTTLE / 2))) - cluster_try_push(wbp, vp, newEOF, 0, 0); - } - start_new_cluster: wbp->cl_clusters[wbp->cl_number].b_addr = cl.b_addr; wbp->cl_clusters[wbp->cl_number].e_addr = cl.e_addr; + wbp->cl_clusters[wbp->cl_number].io_flags = 0; + if (flags & IO_NOCACHE) - wbp->cl_clusters[wbp->cl_number].io_nocache = 1; - else - wbp->cl_clusters[wbp->cl_number].io_nocache = 0; + wbp->cl_clusters[wbp->cl_number].io_flags |= CLW_IONOCACHE; + + if (bflag & CL_PASSIVE) + wbp->cl_clusters[wbp->cl_number].io_flags |= CLW_IOPASSIVE; + wbp->cl_number++; delay_io: - if (upl_size) - ubc_upl_commit_range(upl, upl_offset, upl_size, - UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); - lck_mtx_unlock(&wbp->cl_lockw); continue; issue_io: /* - * we don't hold the vnode lock at this point + * we don't hold the lock at this point * - * because we had to ask for a UPL that provides currenty non-present pages, the - * UPL has been automatically set to clear the dirty flags (both software and hardware) - * upon committing it... this is not the behavior we want since it's possible for - * pages currently present as part of a mapped file to be dirtied while the I/O is in flight. - * in order to maintain some semblance of coherency with mapped writes - * we need to drop the current upl and pick it back up with COPYOUT_FROM set + * we've already dropped the current upl, so pick it back up with COPYOUT_FROM set * so that we correctly deal with a change in state of the hardware modify bit... - * we do this via cluster_push_x... by passing along the IO_SYNC flag, we force - * cluster_push_x to wait until all the I/Os have completed... cluster_push_x is also + * we do this via cluster_push_now... by passing along the IO_SYNC flag, we force + * cluster_push_now to wait until all the I/Os have completed... cluster_push_now is also * responsible for generating the correct sized I/O(s) */ - ubc_upl_commit_range(upl, 0, upl_size, - UPL_COMMIT_SET_DIRTY | UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); - - cl.e_addr = (upl_f_offset + (off_t)upl_size) / PAGE_SIZE_64; - - retval = cluster_push_x(vp, &cl, newEOF, flags); + retval = cluster_push_now(vp, &cl, newEOF, flags, callback, callback_arg); } } - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_END, - retval, 0, io_resid, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 40)) | DBG_FUNC_END, retval, 0, io_resid, 0, 0); return (retval); } + + int cluster_read(vnode_t vp, struct uio *uio, off_t filesize, int xflags) { - int prev_resid; - u_int clip_size; - off_t max_io_size; - int upl_size; - int upl_flags; - upl_t upl; - int retval = 0; - int flags; + return cluster_read_ext(vp, uio, filesize, xflags, NULL, NULL); +} + + +int +cluster_read_ext(vnode_t vp, struct uio *uio, off_t filesize, int xflags, int (*callback)(buf_t, void *), void *callback_arg) +{ + int retval = 0; + int flags; + user_ssize_t cur_resid; + u_int32_t io_size; + u_int32_t read_length = 0; + int read_type = IO_COPY; flags = xflags; if (vp->v_flag & VNOCACHE_DATA) flags |= IO_NOCACHE; - if (vp->v_flag & VRAOFF) + if ((vp->v_flag & VRAOFF) || speculative_reads_disabled) flags |= IO_RAOFF; - if (!((flags & IO_NOCACHE) && UIO_SEG_IS_USER_SPACE(uio->uio_segflg))) { - /* - * go do a read through the cache if one of the following is true.... - * NOCACHE is not true - * the uio request doesn't target USERSPACE - */ - return (cluster_read_x(vp, uio, filesize, flags)); - } - -#if LP64_DEBUG - if (IS_VALID_UIO_SEGFLG(uio->uio_segflg) == 0) { - panic("%s :%d - invalid uio_segflg\n", __FILE__, __LINE__); - } -#endif /* LP64_DEBUG */ + if (flags & IO_SKIP_ENCRYPTION) + flags |= IO_ENCRYPTED; - while (uio_resid(uio) && uio->uio_offset < filesize && retval == 0) { - user_size_t iov_len; - user_addr_t iov_base; + /* + * do a read through the cache if one of the following is true.... + * NOCACHE is not true + * the uio request doesn't target USERSPACE + * Alternatively, if IO_ENCRYPTED is set, then we want to bypass the cache as well. + * Reading encrypted data from a CP filesystem should never result in the data touching + * the UBC. + * + * otherwise, find out if we want the direct or contig variant for + * the first vector in the uio request + */ + if ( ((flags & IO_NOCACHE) && UIO_SEG_IS_USER_SPACE(uio->uio_segflg)) || (flags & IO_ENCRYPTED) ) { - /* - * we know we have a resid, so this is safe - * skip over any emtpy vectors - */ - uio_update(uio, (user_size_t)0); + retval = cluster_io_type(uio, &read_type, &read_length, 0); + } - iov_len = uio_curriovlen(uio); - iov_base = uio_curriovbase(uio); + while ((cur_resid = uio_resid(uio)) && uio->uio_offset < filesize && retval == 0) { - upl_size = PAGE_SIZE; - upl_flags = UPL_QUERY_OBJECT_TYPE; - - // LP64todo - fix this! - if ((vm_map_get_upl(current_map(), - (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)), - &upl_size, &upl, NULL, NULL, &upl_flags, 0)) != KERN_SUCCESS) { + switch (read_type) { + + case IO_COPY: /* - * the user app must have passed in an invalid address + * make sure the uio_resid isn't too big... + * internally, we want to handle all of the I/O in + * chunk sizes that fit in a 32 bit int */ - return (EFAULT); - } + if (cur_resid > (user_ssize_t)(MAX_IO_REQUEST_SIZE)) + io_size = MAX_IO_REQUEST_SIZE; + else + io_size = (u_int32_t)cur_resid; - /* - * We check every vector target but if it is physically - * contiguous space, we skip the sanity checks. - */ - if (upl_flags & UPL_PHYS_CONTIG) { - retval = cluster_phys_read(vp, uio, filesize); - } - else if (uio_resid(uio) < PAGE_SIZE) { - /* - * we're here because we're don't have a physically contiguous target buffer - * go do a read through the cache if - * the total xfer size is less than a page... - */ - return (cluster_read_x(vp, uio, filesize, flags)); - } - // LP64todo - fix this! - else if (((int)uio->uio_offset & PAGE_MASK) || (CAST_DOWN(int, iov_base) & PAGE_MASK)) { - if (((int)uio->uio_offset & PAGE_MASK) == (CAST_DOWN(int, iov_base) & PAGE_MASK)) { - /* - * Bring the file offset read up to a pagesize boundary - * this will also bring the base address to a page boundary - * since they both are currently on the same offset within a page - * note: if we get here, uio->uio_resid is greater than PAGE_SIZE - * so the computed clip_size must always be less than the current uio_resid - */ - clip_size = (PAGE_SIZE - (int)(uio->uio_offset & PAGE_MASK_64)); - - /* - * Fake the resid going into the cluster_read_x call - * and restore it on the way out. - */ - prev_resid = uio_resid(uio); - // LP64todo - fix this - uio_setresid(uio, clip_size); - - retval = cluster_read_x(vp, uio, filesize, flags); - - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); - } else { - /* - * can't get both the file offset and the buffer offset aligned to a page boundary - * so fire an I/O through the cache for this entire vector - */ - // LP64todo - fix this! - clip_size = iov_len; - prev_resid = uio_resid(uio); - uio_setresid(uio, clip_size); - - retval = cluster_read_x(vp, uio, filesize, flags); - - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); - } - } else { - /* - * If we come in here, we know the offset into - * the file is on a pagesize boundary - */ - max_io_size = filesize - uio->uio_offset; - // LP64todo - fix this - clip_size = uio_resid(uio); - if (iov_len < clip_size) - clip_size = iov_len; - if (max_io_size < clip_size) - clip_size = (int)max_io_size; - - if (clip_size < PAGE_SIZE) { - /* - * Take care of the tail end of the read in this vector. - */ - // LP64todo - fix this - prev_resid = uio_resid(uio); - uio_setresid(uio, clip_size); + retval = cluster_read_copy(vp, uio, io_size, filesize, flags, callback, callback_arg); + break; - retval = cluster_read_x(vp, uio, filesize, flags); + case IO_DIRECT: + retval = cluster_read_direct(vp, uio, filesize, &read_type, &read_length, flags, callback, callback_arg); + break; - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); - } else { - /* round clip_size down to a multiple of pagesize */ - clip_size = clip_size & ~(PAGE_MASK); - // LP64todo - fix this - prev_resid = uio_resid(uio); - uio_setresid(uio, clip_size); + case IO_CONTIG: + retval = cluster_read_contig(vp, uio, filesize, &read_type, &read_length, callback, callback_arg, flags); + break; + + case IO_UNKNOWN: + retval = cluster_io_type(uio, &read_type, &read_length, 0); + break; + } + } + return (retval); +} - retval = cluster_nocopy_read(vp, uio, filesize); - if ((retval==0) && uio_resid(uio)) - retval = cluster_read_x(vp, uio, filesize, flags); - uio_setresid(uio, prev_resid - (clip_size - uio_resid(uio))); - } - } /* end else */ - } /* end while */ +static void +cluster_read_upl_release(upl_t upl, int start_pg, int last_pg, int take_reference) +{ + int range; + int abort_flags = UPL_ABORT_FREE_ON_EMPTY; - return(retval); + if ((range = last_pg - start_pg)) { + if (take_reference) + abort_flags |= UPL_ABORT_REFERENCE; + + ubc_upl_abort_range(upl, start_pg * PAGE_SIZE, range * PAGE_SIZE, abort_flags); + } } + static int -cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) +cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t filesize, int flags, int (*callback)(buf_t, void *), void *callback_arg) { upl_page_info_t *pl; upl_t upl; vm_offset_t upl_offset; - int upl_size; + u_int32_t upl_size; off_t upl_f_offset; int start_offset; int start_pg; @@ -2727,43 +3734,79 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) off_t max_size; off_t last_ioread_offset; off_t last_request_offset; - u_int size_of_prefetch; - u_int io_size; kern_return_t kret; int error = 0; int retval = 0; - u_int max_rd_size = MAX_UPL_TRANSFER * PAGE_SIZE; + u_int32_t size_of_prefetch; + u_int32_t xsize; + u_int32_t io_size; + u_int32_t max_rd_size; + u_int32_t max_io_size; + u_int32_t max_prefetch; u_int rd_ahead_enabled = 1; u_int prefetch_enabled = 1; struct cl_readahead * rap; struct clios iostate; struct cl_extent extent; + int bflag; + int take_reference = 1; + int policy = IOPOL_DEFAULT; + boolean_t iolock_inited = FALSE; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_START, - (int)uio->uio_offset, uio_resid(uio), (int)filesize, 0, 0); + (int)uio->uio_offset, io_req_size, (int)filesize, flags, 0); + + if (flags & IO_ENCRYPTED) { + panic ("encrypted blocks will hit UBC!"); + } - // LP64todo - fix this - last_request_offset = uio->uio_offset + uio_resid(uio); + policy = throttle_get_io_policy(NULL); + + if (policy == THROTTLE_LEVEL_TIER3 || policy == THROTTLE_LEVEL_TIER2 || (flags & IO_NOCACHE)) + take_reference = 0; + + if (flags & IO_PASSIVE) + bflag = CL_PASSIVE; + else + bflag = 0; + + if (flags & IO_NOCACHE) + bflag |= CL_NOCACHE; + + if (flags & IO_SKIP_ENCRYPTION) + bflag |= CL_ENCRYPTED; + + max_io_size = cluster_max_io_size(vp->v_mount, CL_READ); + max_prefetch = MAX_PREFETCH(vp, max_io_size, (vp->v_mount->mnt_kern_flag & MNTK_SSD)); + max_rd_size = max_prefetch; + + last_request_offset = uio->uio_offset + io_req_size; + + if (last_request_offset > filesize) + last_request_offset = filesize; - if ((flags & (IO_RAOFF|IO_NOCACHE)) || - ((last_request_offset & ~PAGE_MASK_64) == (uio->uio_offset & ~PAGE_MASK_64))) { + if ((flags & (IO_RAOFF|IO_NOCACHE)) || ((last_request_offset & ~PAGE_MASK_64) == (uio->uio_offset & ~PAGE_MASK_64))) { rd_ahead_enabled = 0; rap = NULL; } else { - if (cluster_hard_throttle_on(vp)) { + if (cluster_is_throttled(vp)) { + /* + * we're in the throttle window, at the very least + * we want to limit the size of the I/O we're about + * to issue + */ rd_ahead_enabled = 0; prefetch_enabled = 0; - max_rd_size = HARD_THROTTLE_MAXSIZE; + max_rd_size = THROTTLE_MAX_IOSIZE; } if ((rap = cluster_get_rap(vp)) == NULL) rd_ahead_enabled = 0; - } - if (last_request_offset > filesize) - last_request_offset = filesize; - extent.b_addr = uio->uio_offset / PAGE_SIZE_64; - extent.e_addr = (last_request_offset - 1) / PAGE_SIZE_64; - + else { + extent.b_addr = uio->uio_offset / PAGE_SIZE_64; + extent.e_addr = (last_request_offset - 1) / PAGE_SIZE_64; + } + } if (rap != NULL && rap->cl_ralen && (rap->cl_lastr == extent.b_addr || (rap->cl_lastr + 1) == extent.b_addr)) { /* * determine if we already have a read-ahead in the pipe courtesy of the @@ -2781,34 +3824,24 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) } else last_ioread_offset = (off_t)0; - while (uio_resid(uio) && uio->uio_offset < filesize && retval == 0) { - /* - * compute the size of the upl needed to encompass - * the requested read... limit each call to cluster_io - * to the maximum UPL size... cluster_io will clip if - * this exceeds the maximum io_size for the device, - * make sure to account for - * a starting offset that's not page aligned - */ - start_offset = (int)(uio->uio_offset & PAGE_MASK_64); - upl_f_offset = uio->uio_offset - (off_t)start_offset; - max_size = filesize - uio->uio_offset; + while (io_req_size && uio->uio_offset < filesize && retval == 0) { + + max_size = filesize - uio->uio_offset; - // LP64todo - fix this! - if ((off_t)((unsigned int)uio_resid(uio)) < max_size) - io_size = uio_resid(uio); + if ((off_t)(io_req_size) < max_size) + io_size = io_req_size; else io_size = max_size; if (!(flags & IO_NOCACHE)) { while (io_size) { - u_int io_resid; - u_int io_requested; + u_int32_t io_resid; + u_int32_t io_requested; /* * if we keep finding the pages we need already in the cache, then - * don't bother to call cluster_rd_prefetch since it costs CPU cycles + * don't bother to call cluster_read_prefetch since it costs CPU cycles * to determine that we have all the pages we need... once we miss in * the cache and have issued an I/O, than we'll assume that we're likely * to continue to miss in the cache and it's to our advantage to try and prefetch @@ -2825,7 +3858,7 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) if (size_of_prefetch > max_rd_size) size_of_prefetch = max_rd_size; - size_of_prefetch = cluster_rd_prefetch(vp, last_ioread_offset, size_of_prefetch, filesize); + size_of_prefetch = cluster_read_prefetch(vp, last_ioread_offset, size_of_prefetch, filesize, callback, callback_arg, bflag); last_ioread_offset += (off_t)(size_of_prefetch * PAGE_SIZE); @@ -2838,16 +3871,19 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) * we can notice that our I/O pipe is running dry and * get the next I/O issued before it does go dry */ - if (last_ioread_offset && io_size > ((MAX_UPL_TRANSFER * PAGE_SIZE) / 4)) - io_resid = ((MAX_UPL_TRANSFER * PAGE_SIZE) / 4); + if (last_ioread_offset && io_size > (max_io_size / 4)) + io_resid = (max_io_size / 4); else io_resid = io_size; io_requested = io_resid; - retval = cluster_copy_ubc_data(vp, uio, &io_resid, 0); + retval = cluster_copy_ubc_data_internal(vp, uio, (int *)&io_resid, 0, take_reference); - io_size -= (io_requested - io_resid); + xsize = io_requested - io_resid; + + io_size -= xsize; + io_req_size -= xsize; if (retval || io_resid) /* @@ -2857,12 +3893,12 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) */ break; - if ((io_size == 0 || last_ioread_offset == last_request_offset) && rd_ahead_enabled) { + if (rd_ahead_enabled && (io_size == 0 || last_ioread_offset == last_request_offset)) { /* * we're already finished the I/O for this read request * let's see if we should do a read-ahead */ - cluster_rd_ahead(vp, &extent, filesize, rap); + cluster_read_ahead(vp, &extent, filesize, rap, callback, callback_arg, bflag); } } if (retval) @@ -2875,33 +3911,80 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) } break; } - start_offset = (int)(uio->uio_offset & PAGE_MASK_64); - upl_f_offset = uio->uio_offset - (off_t)start_offset; - max_size = filesize - uio->uio_offset; + /* + * recompute max_size since cluster_copy_ubc_data_internal + * may have advanced uio->uio_offset + */ + max_size = filesize - uio->uio_offset; } + + iostate.io_completed = 0; + iostate.io_issued = 0; + iostate.io_error = 0; + iostate.io_wanted = 0; + + if ( (flags & IO_RETURN_ON_THROTTLE) ) { + if (cluster_is_throttled(vp) == THROTTLE_NOW) { + if ( !cluster_io_present_in_BC(vp, uio->uio_offset)) { + /* + * we're in the throttle window and at least 1 I/O + * has already been issued by a throttleable thread + * in this window, so return with EAGAIN to indicate + * to the FS issuing the cluster_read call that it + * should now throttle after dropping any locks + */ + throttle_info_update_by_mount(vp->v_mount); + + retval = EAGAIN; + break; + } + } + } + + /* + * compute the size of the upl needed to encompass + * the requested read... limit each call to cluster_io + * to the maximum UPL size... cluster_io will clip if + * this exceeds the maximum io_size for the device, + * make sure to account for + * a starting offset that's not page aligned + */ + start_offset = (int)(uio->uio_offset & PAGE_MASK_64); + upl_f_offset = uio->uio_offset - (off_t)start_offset; + if (io_size > max_rd_size) io_size = max_rd_size; upl_size = (start_offset + io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK; - if (upl_size > (MAX_UPL_TRANSFER * PAGE_SIZE) / 4) - upl_size = (MAX_UPL_TRANSFER * PAGE_SIZE) / 4; + if (flags & IO_NOCACHE) { + if (upl_size > max_io_size) + upl_size = max_io_size; + } else { + if (upl_size > max_io_size / 4) { + upl_size = max_io_size / 4; + upl_size &= ~PAGE_MASK; + + if (upl_size == 0) + upl_size = PAGE_SIZE; + } + } pages_in_upl = upl_size / PAGE_SIZE; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 33)) | DBG_FUNC_START, - (int)upl, (int)upl_f_offset, upl_size, start_offset, 0); + upl, (int)upl_f_offset, upl_size, start_offset, 0); kret = ubc_create_upl(vp, upl_f_offset, upl_size, &upl, &pl, - UPL_SET_LITE); + UPL_FILE_IO | UPL_SET_LITE); if (kret != KERN_SUCCESS) - panic("cluster_read: failed to get pagelist"); + panic("cluster_read_copy: failed to get pagelist"); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 33)) | DBG_FUNC_END, - (int)upl, (int)upl_f_offset, upl_size, start_offset, 0); + upl, (int)upl_f_offset, upl_size, start_offset, 0); /* * scan from the beginning of the upl looking for the first @@ -2924,10 +4007,6 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) if (upl_valid_page(pl, last_pg)) break; } - iostate.io_completed = 0; - iostate.io_issued = 0; - iostate.io_error = 0; - iostate.io_wanted = 0; if (start_pg < last_pg) { /* @@ -2936,10 +4015,15 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) * we may have to clip the size of it to keep from reading past * the end of the last physical block associated with the file */ + if (iolock_inited == FALSE) { + lck_mtx_init(&iostate.io_mtxp, cl_mtx_grp, cl_mtx_attr); + + iolock_inited = TRUE; + } upl_offset = start_pg * PAGE_SIZE; io_size = (last_pg - start_pg) * PAGE_SIZE; - if ((upl_f_offset + upl_offset + io_size) > filesize) + if ((off_t)(upl_f_offset + upl_offset + io_size) > filesize) io_size = filesize - (upl_f_offset + upl_offset); /* @@ -2947,7 +4031,19 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) */ error = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, - io_size, CL_READ | CL_ASYNC, (buf_t)NULL, &iostate); + io_size, CL_READ | CL_ASYNC | bflag, (buf_t)NULL, &iostate, callback, callback_arg); + + if (rap) { + if (extent.e_addr < rap->cl_maxra) { + /* + * we've just issued a read for a block that should have been + * in the cache courtesy of the read-ahead engine... something + * has gone wrong with the pipeline, so reset the read-ahead + * logic which will cause us to restart from scratch + */ + rap->cl_maxra = 0; + } + } } if (error == 0) { /* @@ -2962,8 +4058,19 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) if (!upl_valid_page(pl, uio_last)) break; } + if (uio_last < pages_in_upl) { + /* + * there were some invalid pages beyond the valid pages + * that we didn't issue an I/O for, just release them + * unchanged now, so that any prefetch/readahed can + * include them + */ + ubc_upl_abort_range(upl, uio_last * PAGE_SIZE, + (pages_in_upl - uio_last) * PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY); + } + /* - * compute size to transfer this round, if uio->uio_resid is + * compute size to transfer this round, if io_req_size is * still non-zero after this attempt, we'll loop around and * set up for another I/O. */ @@ -2972,26 +4079,32 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) if (val_size > max_size) val_size = max_size; - if (val_size > uio_resid(uio)) - // LP64todo - fix this - val_size = uio_resid(uio); + if (val_size > io_req_size) + val_size = io_req_size; - if (last_ioread_offset == 0) + if ((uio->uio_offset + val_size) > last_ioread_offset) last_ioread_offset = uio->uio_offset + val_size; if ((size_of_prefetch = (last_request_offset - last_ioread_offset)) && prefetch_enabled) { - /* - * if there's still I/O left to do for this request, and... - * we're not in hard throttle mode, then issue a - * pre-fetch I/O... the I/O latency will overlap - * with the copying of the data - */ - size_of_prefetch = cluster_rd_prefetch(vp, last_ioread_offset, size_of_prefetch, filesize); - last_ioread_offset += (off_t)(size_of_prefetch * PAGE_SIZE); + if ((last_ioread_offset - (uio->uio_offset + val_size)) <= upl_size) { + /* + * if there's still I/O left to do for this request, and... + * we're not in hard throttle mode, and... + * we're close to using up the previous prefetch, then issue a + * new pre-fetch I/O... the I/O latency will overlap + * with the copying of the data + */ + if (size_of_prefetch > max_rd_size) + size_of_prefetch = max_rd_size; + + size_of_prefetch = cluster_read_prefetch(vp, last_ioread_offset, size_of_prefetch, filesize, callback, callback_arg, bflag); + + last_ioread_offset += (off_t)(size_of_prefetch * PAGE_SIZE); - if (last_ioread_offset > last_request_offset) - last_ioread_offset = last_request_offset; + if (last_ioread_offset > last_request_offset) + last_ioread_offset = last_request_offset; + } } else if ((uio->uio_offset + val_size) == last_request_offset) { /* @@ -3001,7 +4114,7 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) * explicitly disabled it */ if (rd_ahead_enabled) - cluster_rd_ahead(vp, &extent, filesize, rap); + cluster_read_ahead(vp, &extent, filesize, rap, callback, callback_arg, bflag); if (rap != NULL) { if (extent.e_addr < rap->cl_lastr) @@ -3009,46 +4122,51 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) rap->cl_lastr = extent.e_addr; } } - lck_mtx_lock(cl_mtxp); - - while (iostate.io_issued != iostate.io_completed) { - iostate.io_wanted = 1; - msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_read_x", 0); - } - lck_mtx_unlock(cl_mtxp); + if (iolock_inited == TRUE) + cluster_iostate_wait(&iostate, 0, "cluster_read_copy"); if (iostate.io_error) error = iostate.io_error; - else - retval = cluster_copy_upl_data(uio, upl, start_offset, val_size); + else { + u_int32_t io_requested; + + io_requested = val_size; + + retval = cluster_copy_upl_data(uio, upl, start_offset, (int *)&io_requested); + + io_req_size -= (val_size - io_requested); + } + } else { + if (iolock_inited == TRUE) + cluster_iostate_wait(&iostate, 0, "cluster_read_copy"); } if (start_pg < last_pg) { /* * compute the range of pages that we actually issued an I/O for * and either commit them as valid if the I/O succeeded - * or abort them if the I/O failed + * or abort them if the I/O failed or we're not supposed to + * keep them in the cache */ io_size = (last_pg - start_pg) * PAGE_SIZE; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_START, - (int)upl, start_pg * PAGE_SIZE, io_size, error, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_START, upl, start_pg * PAGE_SIZE, io_size, error, 0); if (error || (flags & IO_NOCACHE)) ubc_upl_abort_range(upl, start_pg * PAGE_SIZE, io_size, - UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY); - else - ubc_upl_commit_range(upl, start_pg * PAGE_SIZE, io_size, - UPL_COMMIT_CLEAR_DIRTY | - UPL_COMMIT_FREE_ON_EMPTY | - UPL_COMMIT_INACTIVATE); + UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY); + else { + int commit_flags = UPL_COMMIT_CLEAR_DIRTY | UPL_COMMIT_FREE_ON_EMPTY; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_END, - (int)upl, start_pg * PAGE_SIZE, io_size, error, 0); + if (take_reference) + commit_flags |= UPL_COMMIT_INACTIVATE; + else + commit_flags |= UPL_COMMIT_SPECULATE; + + ubc_upl_commit_range(upl, start_pg * PAGE_SIZE, io_size, commit_flags); + } + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_END, upl, start_pg * PAGE_SIZE, io_size, error, 0); } if ((last_pg - start_pg) < pages_in_upl) { - int cur_pg; - int commit_flags; - /* * the set of pages that we issued an I/O for did not encompass * the entire upl... so just release these without modifying @@ -3057,223 +4175,490 @@ cluster_read_x(vnode_t vp, struct uio *uio, off_t filesize, int flags) if (error) ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); else { + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_START, - (int)upl, -1, pages_in_upl - (last_pg - start_pg), 0, 0); + upl, -1, pages_in_upl - (last_pg - start_pg), 0, 0); - if (start_pg) { - /* - * we found some already valid pages at the beginning of - * the upl commit these back to the inactive list with - * reference cleared - */ - for (cur_pg = 0; cur_pg < start_pg; cur_pg++) { - commit_flags = UPL_COMMIT_FREE_ON_EMPTY - | UPL_COMMIT_INACTIVATE; - - if (upl_dirty_page(pl, cur_pg)) - commit_flags |= UPL_COMMIT_SET_DIRTY; - - if ( !(commit_flags & UPL_COMMIT_SET_DIRTY) && (flags & IO_NOCACHE)) - ubc_upl_abort_range(upl, cur_pg * PAGE_SIZE, PAGE_SIZE, - UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY); - else - ubc_upl_commit_range(upl, cur_pg * PAGE_SIZE, - PAGE_SIZE, commit_flags); - } - } - if (last_pg < uio_last) { - /* - * we found some already valid pages immediately after the - * pages we issued I/O for, commit these back to the - * inactive list with reference cleared - */ - for (cur_pg = last_pg; cur_pg < uio_last; cur_pg++) { - commit_flags = UPL_COMMIT_FREE_ON_EMPTY - | UPL_COMMIT_INACTIVATE; - - if (upl_dirty_page(pl, cur_pg)) - commit_flags |= UPL_COMMIT_SET_DIRTY; - - if ( !(commit_flags & UPL_COMMIT_SET_DIRTY) && (flags & IO_NOCACHE)) - ubc_upl_abort_range(upl, cur_pg * PAGE_SIZE, PAGE_SIZE, - UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY); - else - ubc_upl_commit_range(upl, cur_pg * PAGE_SIZE, - PAGE_SIZE, commit_flags); - } - } - if (uio_last < pages_in_upl) { - /* - * there were some invalid pages beyond the valid pages - * that we didn't issue an I/O for, just release them - * unchanged - */ - ubc_upl_abort_range(upl, uio_last * PAGE_SIZE, - (pages_in_upl - uio_last) * PAGE_SIZE, UPL_ABORT_FREE_ON_EMPTY); - } + /* + * handle any valid pages at the beginning of + * the upl... release these appropriately + */ + cluster_read_upl_release(upl, 0, start_pg, take_reference); + + /* + * handle any valid pages immediately after the + * pages we issued I/O for... ... release these appropriately + */ + cluster_read_upl_release(upl, last_pg, uio_last, take_reference); - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_END, - (int)upl, -1, -1, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 35)) | DBG_FUNC_END, upl, -1, -1, 0, 0); } } if (retval == 0) retval = error; - if ( uio_resid(uio) ) { - if (cluster_hard_throttle_on(vp)) { + if (io_req_size) { + if (cluster_is_throttled(vp)) { + /* + * we're in the throttle window, at the very least + * we want to limit the size of the I/O we're about + * to issue + */ rd_ahead_enabled = 0; prefetch_enabled = 0; - - max_rd_size = HARD_THROTTLE_MAXSIZE; + max_rd_size = THROTTLE_MAX_IOSIZE; } else { - if (rap != NULL) - rd_ahead_enabled = 1; - prefetch_enabled = 1; - - max_rd_size = MAX_UPL_TRANSFER * PAGE_SIZE; + if (max_rd_size == THROTTLE_MAX_IOSIZE) { + /* + * coming out of throttled state + */ + if (policy != THROTTLE_LEVEL_TIER3 && policy != THROTTLE_LEVEL_TIER2) { + if (rap != NULL) + rd_ahead_enabled = 1; + prefetch_enabled = 1; + } + max_rd_size = max_prefetch; + last_ioread_offset = 0; + } } } } + if (iolock_inited == TRUE) { + /* + * cluster_io returned an error after it + * had already issued some I/O. we need + * to wait for that I/O to complete before + * we can destroy the iostate mutex... + * 'retval' already contains the early error + * so no need to pick it up from iostate.io_error + */ + cluster_iostate_wait(&iostate, 0, "cluster_read_copy"); + + lck_mtx_destroy(&iostate.io_mtxp, cl_mtx_grp); + } if (rap != NULL) { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_END, - (int)uio->uio_offset, uio_resid(uio), rap->cl_lastr, retval, 0); + (int)uio->uio_offset, io_req_size, rap->cl_lastr, retval, 0); lck_mtx_unlock(&rap->cl_lockr); } else { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_END, - (int)uio->uio_offset, uio_resid(uio), 0, retval, 0); + (int)uio->uio_offset, io_req_size, 0, retval, 0); } return (retval); } +/* + * We don't want another read/write lock for every vnode in the system + * so we keep a hash of them here. There should never be very many of + * these around at any point in time. + */ +cl_direct_read_lock_t *cluster_lock_direct_read(vnode_t vp, lck_rw_type_t type) +{ + struct cl_direct_read_locks *head + = &cl_direct_read_locks[(uintptr_t)vp / sizeof(*vp) + % CL_DIRECT_READ_LOCK_BUCKETS]; + + struct cl_direct_read_lock *lck, *new_lck = NULL; + + for (;;) { + lck_spin_lock(&cl_direct_read_spin_lock); + + LIST_FOREACH(lck, head, chain) { + if (lck->vp == vp) { + ++lck->ref_count; + lck_spin_unlock(&cl_direct_read_spin_lock); + if (new_lck) { + // Someone beat us to it, ditch the allocation + lck_rw_destroy(&new_lck->rw_lock, cl_mtx_grp); + FREE(new_lck, M_TEMP); + } + lck_rw_lock(&lck->rw_lock, type); + return lck; + } + } + + if (new_lck) { + // Use the lock we allocated + LIST_INSERT_HEAD(head, new_lck, chain); + lck_spin_unlock(&cl_direct_read_spin_lock); + lck_rw_lock(&new_lck->rw_lock, type); + return new_lck; + } + + lck_spin_unlock(&cl_direct_read_spin_lock); + + // Allocate a new lock + MALLOC(new_lck, cl_direct_read_lock_t *, sizeof(*new_lck), + M_TEMP, M_WAITOK); + lck_rw_init(&new_lck->rw_lock, cl_mtx_grp, cl_mtx_attr); + new_lck->vp = vp; + new_lck->ref_count = 1; + + // Got to go round again + } +} + +void cluster_unlock_direct_read(cl_direct_read_lock_t *lck) +{ + lck_rw_done(&lck->rw_lock); + + lck_spin_lock(&cl_direct_read_spin_lock); + if (lck->ref_count == 1) { + LIST_REMOVE(lck, chain); + lck_spin_unlock(&cl_direct_read_spin_lock); + lck_rw_destroy(&lck->rw_lock, cl_mtx_grp); + FREE(lck, M_TEMP); + } else { + --lck->ref_count; + lck_spin_unlock(&cl_direct_read_spin_lock); + } +} static int -cluster_nocopy_read(vnode_t vp, struct uio *uio, off_t filesize) +cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length, + int flags, int (*callback)(buf_t, void *), void *callback_arg) { upl_t upl; upl_page_info_t *pl; - vm_offset_t upl_offset; - off_t max_io_size; - int io_size; - int upl_size; - int upl_needed_size; - int pages_in_pl; - int upl_flags; + off_t max_io_size; + vm_offset_t upl_offset, vector_upl_offset = 0; + upl_size_t upl_size, vector_upl_size = 0; + vm_size_t upl_needed_size; + unsigned int pages_in_pl; + upl_control_flags_t upl_flags; kern_return_t kret; - int i; + unsigned int i; int force_data_sync; int retval = 0; int no_zero_fill = 0; - int abort_flag = 0; + int io_flag = 0; + int misaligned = 0; struct clios iostate; - u_int max_rd_size = MAX_UPL_TRANSFER * PAGE_SIZE; - u_int max_rd_ahead = MAX_UPL_TRANSFER * PAGE_SIZE * 2; - + user_addr_t iov_base; + u_int32_t io_req_size; + u_int32_t offset_in_file; + u_int32_t offset_in_iovbase; + u_int32_t io_size; + u_int32_t io_min; + u_int32_t xsize; + u_int32_t devblocksize; + u_int32_t mem_alignment_mask; + u_int32_t max_upl_size; + u_int32_t max_rd_size; + u_int32_t max_rd_ahead; + u_int32_t max_vector_size; + boolean_t strict_uncached_IO = FALSE; + boolean_t io_throttled = FALSE; + + u_int32_t vector_upl_iosize = 0; + int issueVectorUPL = 0,useVectorUPL = (uio->uio_iovcnt > 1); + off_t v_upl_uio_offset = 0; + int vector_upl_index=0; + upl_t vector_upl = NULL; + cl_direct_read_lock_t *lock = NULL; + + user_addr_t orig_iov_base = 0; + user_addr_t last_iov_base = 0; + user_addr_t next_iov_base = 0; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_START, - (int)uio->uio_offset, uio_resid(uio), (int)filesize, 0, 0); + (int)uio->uio_offset, (int)filesize, *read_type, *read_length, 0); - /* - * When we enter this routine, we know - * -- the offset into the file is on a pagesize boundary - * -- the resid is a page multiple - * -- the resid will not exceed iov_len - */ + max_upl_size = cluster_max_io_size(vp->v_mount, CL_READ); + + max_rd_size = max_upl_size; + max_rd_ahead = max_rd_size * IO_SCALE(vp, 2); + + io_flag = CL_COMMIT | CL_READ | CL_ASYNC | CL_NOZERO | CL_DIRECT_IO; + + if (flags & IO_PASSIVE) + io_flag |= CL_PASSIVE; + + if (flags & IO_ENCRYPTED) { + io_flag |= CL_RAW_ENCRYPTED; + } + + if (flags & IO_NOCACHE) { + io_flag |= CL_NOCACHE; + } + + if (flags & IO_SKIP_ENCRYPTION) + io_flag |= CL_ENCRYPTED; iostate.io_completed = 0; iostate.io_issued = 0; iostate.io_error = 0; iostate.io_wanted = 0; - while (uio_resid(uio) && uio->uio_offset < filesize && retval == 0) { - user_addr_t iov_base; + lck_mtx_init(&iostate.io_mtxp, cl_mtx_grp, cl_mtx_attr); + + devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize; + mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask; + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_NONE, + (int)devblocksize, (int)mem_alignment_mask, 0, 0, 0); + + if (devblocksize == 1) { + /* + * the AFP client advertises a devblocksize of 1 + * however, its BLOCKMAP routine maps to physical + * blocks that are PAGE_SIZE in size... + * therefore we can't ask for I/Os that aren't page aligned + * or aren't multiples of PAGE_SIZE in size + * by setting devblocksize to PAGE_SIZE, we re-instate + * the old behavior we had before the mem_alignment_mask + * changes went in... + */ + devblocksize = PAGE_SIZE; + } + + strict_uncached_IO = ubc_strict_uncached_IO(vp); + + orig_iov_base = uio_curriovbase(uio); + last_iov_base = orig_iov_base; + +next_dread: + io_req_size = *read_length; + iov_base = uio_curriovbase(uio); + + offset_in_file = (u_int32_t)uio->uio_offset & (devblocksize - 1); + offset_in_iovbase = (u_int32_t)iov_base & mem_alignment_mask; + + if (offset_in_file || offset_in_iovbase) { + /* + * one of the 2 important offsets is misaligned + * so fire an I/O through the cache for this entire vector + */ + misaligned = 1; + } + if (iov_base & (devblocksize - 1)) { + /* + * the offset in memory must be on a device block boundary + * so that we can guarantee that we can generate an + * I/O that ends on a page boundary in cluster_io + */ + misaligned = 1; + } + + max_io_size = filesize - uio->uio_offset; + + /* + * The user must request IO in aligned chunks. If the + * offset into the file is bad, or the userland pointer + * is non-aligned, then we cannot service the encrypted IO request. + */ + if (flags & IO_ENCRYPTED) { + if (misaligned || (io_req_size & (devblocksize - 1))) + retval = EINVAL; + + max_io_size = roundup(max_io_size, devblocksize); + } + + if ((off_t)io_req_size > max_io_size) + io_req_size = max_io_size; + + /* + * When we get to this point, we know... + * -- the offset into the file is on a devblocksize boundary + */ + + while (io_req_size && retval == 0) { + u_int32_t io_start; - if (cluster_hard_throttle_on(vp)) { - max_rd_size = HARD_THROTTLE_MAXSIZE; - max_rd_ahead = HARD_THROTTLE_MAXSIZE - 1; + if (cluster_is_throttled(vp)) { + /* + * we're in the throttle window, at the very least + * we want to limit the size of the I/O we're about + * to issue + */ + max_rd_size = THROTTLE_MAX_IOSIZE; + max_rd_ahead = THROTTLE_MAX_IOSIZE - 1; + max_vector_size = THROTTLE_MAX_IOSIZE; } else { - max_rd_size = MAX_UPL_TRANSFER * PAGE_SIZE; - max_rd_ahead = MAX_UPL_TRANSFER * PAGE_SIZE * 8; + max_rd_size = max_upl_size; + max_rd_ahead = max_rd_size * IO_SCALE(vp, 2); + max_vector_size = MAX_VECTOR_UPL_SIZE; } - max_io_size = filesize - uio->uio_offset; - - // LP64todo - fix this - if (max_io_size < (off_t)((unsigned int)uio_resid(uio))) - io_size = max_io_size; - else - io_size = uio_resid(uio); + io_start = io_size = io_req_size; /* * First look for pages already in the cache - * and move them to user space. + * and move them to user space. But only do this + * check if we are not retrieving encrypted data directly + * from the filesystem; those blocks should never + * be in the UBC. + * + * cluster_copy_ubc_data returns the resid + * in io_size */ - retval = cluster_copy_ubc_data(vp, uio, &io_size, 0); - - if (retval) { + if ((strict_uncached_IO == FALSE) && ((flags & IO_ENCRYPTED) == 0)) { + retval = cluster_copy_ubc_data_internal(vp, uio, (int *)&io_size, 0, 0); + } + /* + * calculate the number of bytes actually copied + * starting size - residual + */ + xsize = io_start - io_size; + + io_req_size -= xsize; + + if(useVectorUPL && (xsize || (iov_base & PAGE_MASK))) { /* - * we may have already spun some portion of this request - * off as async requests... we need to wait for the I/O - * to complete before returning + * We found something in the cache or we have an iov_base that's not + * page-aligned. + * + * Issue all I/O's that have been collected within this Vectored UPL. */ - goto wait_for_reads; + if(vector_upl_index) { + retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + reset_vector_run_state(); + } + + if(xsize) + useVectorUPL = 0; + + /* + * After this point, if we are using the Vector UPL path and the base is + * not page-aligned then the UPL with that base will be the first in the vector UPL. + */ } + /* - * If we are already finished with this read, then return + * check to see if we are finished with this request. + * + * If we satisfied this IO already, then io_req_size will be 0. + * Otherwise, see if the IO was mis-aligned and needs to go through + * the UBC to deal with the 'tail'. + * */ - if (io_size == 0) { + if (io_req_size == 0 || (misaligned)) { + /* + * see if there's another uio vector to + * process that's of type IO_DIRECT + * + * break out of while loop to get there + */ + break; + } + /* + * assume the request ends on a device block boundary + */ + io_min = devblocksize; + + /* + * we can handle I/O's in multiples of the device block size + * however, if io_size isn't a multiple of devblocksize we + * want to clip it back to the nearest page boundary since + * we are going to have to go through cluster_read_copy to + * deal with the 'overhang'... by clipping it to a PAGE_SIZE + * multiple, we avoid asking the drive for the same physical + * blocks twice.. once for the partial page at the end of the + * request and a 2nd time for the page we read into the cache + * (which overlaps the end of the direct read) in order to + * get at the overhang bytes + */ + if (io_size & (devblocksize - 1)) { + assert(!(flags & IO_ENCRYPTED)); /* + * Clip the request to the previous page size boundary + * since request does NOT end on a device block boundary + */ + io_size &= ~PAGE_MASK; + io_min = PAGE_SIZE; + } + if (retval || io_size < io_min) { + /* + * either an error or we only have the tail left to + * complete via the copy path... * we may have already spun some portion of this request * off as async requests... we need to wait for the I/O * to complete before returning */ - goto wait_for_reads; + goto wait_for_dreads; } - max_io_size = io_size; - if (max_io_size > max_rd_size) - max_io_size = max_rd_size; + /* + * Don't re-check the UBC data if we are looking for uncached IO + * or asking for encrypted blocks. + */ + if ((strict_uncached_IO == FALSE) && ((flags & IO_ENCRYPTED) == 0)) { - io_size = 0; + if ((xsize = io_size) > max_rd_size) + xsize = max_rd_size; - ubc_range_op(vp, uio->uio_offset, uio->uio_offset + max_io_size, UPL_ROP_ABSENT, &io_size); + io_size = 0; + + if (!lock) { + /* + * We hold a lock here between the time we check the + * cache and the time we issue I/O. This saves us + * from having to lock the pages in the cache. Not + * all clients will care about this lock but some + * clients may want to guarantee stability between + * here and when the I/O is issued in which case they + * will take the lock exclusively. + */ + lock = cluster_lock_direct_read(vp, LCK_RW_TYPE_SHARED); + } + + ubc_range_op(vp, uio->uio_offset, uio->uio_offset + xsize, UPL_ROP_ABSENT, (int *)&io_size); + + if (io_size == 0) { + /* + * a page must have just come into the cache + * since the first page in this range is no + * longer absent, go back and re-evaluate + */ + continue; + } + } + if ( (flags & IO_RETURN_ON_THROTTLE) ) { + if (cluster_is_throttled(vp) == THROTTLE_NOW) { + if ( !cluster_io_present_in_BC(vp, uio->uio_offset)) { + /* + * we're in the throttle window and at least 1 I/O + * has already been issued by a throttleable thread + * in this window, so return with EAGAIN to indicate + * to the FS issuing the cluster_read call that it + * should now throttle after dropping any locks + */ + throttle_info_update_by_mount(vp->v_mount); + + io_throttled = TRUE; + goto wait_for_dreads; + } + } + } + if (io_size > max_rd_size) + io_size = max_rd_size; - if (io_size == 0) - /* - * we may have already spun some portion of this request - * off as async requests... we need to wait for the I/O - * to complete before returning - */ - goto wait_for_reads; - iov_base = uio_curriovbase(uio); - // LP64todo - fix this! - upl_offset = CAST_DOWN(vm_offset_t, iov_base) & PAGE_MASK; + upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK); upl_needed_size = (upl_offset + io_size + (PAGE_SIZE -1)) & ~PAGE_MASK; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_START, (int)upl_offset, upl_needed_size, (int)iov_base, io_size, 0); - if (upl_offset == 0 && ((io_size & PAGE_MASK) == 0)) { + if (upl_offset == 0 && ((io_size & PAGE_MASK) == 0)) no_zero_fill = 1; - abort_flag = UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY; - } else { + else no_zero_fill = 0; - abort_flag = UPL_ABORT_FREE_ON_EMPTY; - } + + vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map; for (force_data_sync = 0; force_data_sync < 3; force_data_sync++) { pages_in_pl = 0; upl_size = upl_needed_size; - upl_flags = UPL_FILE_IO | UPL_NO_SYNC | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE; - + upl_flags = UPL_FILE_IO | UPL_NO_SYNC | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE + | UPL_MEMORY_TAG_MAKE(VM_KERN_MEMORY_FILE); if (no_zero_fill) upl_flags |= UPL_NOZEROFILL; if (force_data_sync) upl_flags |= UPL_FORCE_DATA_SYNC; - // LP64todo - fix this! - kret = vm_map_create_upl(current_map(), + kret = vm_map_create_upl(map, (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)), &upl_size, &upl, NULL, &pages_in_pl, &upl_flags); @@ -3281,59 +4666,68 @@ cluster_nocopy_read(vnode_t vp, struct uio *uio, off_t filesize) KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END, (int)upl_offset, upl_size, io_size, kret, 0); /* - * cluster_nocopy_read: failed to get pagelist + * failed to get pagelist * * we may have already spun some portion of this request * off as async requests... we need to wait for the I/O * to complete before returning */ - goto wait_for_reads; + goto wait_for_dreads; } pages_in_pl = upl_size / PAGE_SIZE; pl = UPL_GET_INTERNAL_PAGE_LIST(upl); for (i = 0; i < pages_in_pl; i++) { - if (!upl_valid_page(pl, i)) + if (!upl_page_present(pl, i)) break; } if (i == pages_in_pl) break; - ubc_upl_abort_range(upl, (upl_offset & ~PAGE_MASK), upl_size, abort_flag); + ubc_upl_abort(upl, 0); } if (force_data_sync >= 3) { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END, (int)upl_offset, upl_size, io_size, kret, 0); - goto wait_for_reads; + goto wait_for_dreads; } /* * Consider the possibility that upl_size wasn't satisfied. */ - if (upl_size != upl_needed_size) - io_size = (upl_size - (int)upl_offset) & ~PAGE_MASK; - + if (upl_size < upl_needed_size) { + if (upl_size && upl_offset == 0) + io_size = upl_size; + else + io_size = 0; + } if (io_size == 0) { - ubc_upl_abort_range(upl, (upl_offset & ~PAGE_MASK), upl_size, abort_flag); - goto wait_for_reads; + ubc_upl_abort(upl, 0); + goto wait_for_dreads; } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END, (int)upl_offset, upl_size, io_size, kret, 0); + if(useVectorUPL) { + vm_offset_t end_off = ((iov_base + io_size) & PAGE_MASK); + if(end_off) + issueVectorUPL = 1; + /* + * After this point, if we are using a vector UPL, then + * either all the UPL elements end on a page boundary OR + * this UPL is the last element because it does not end + * on a page boundary. + */ + } + /* * request asynchronously so that we can overlap * the preparation of the next I/O * if there are already too many outstanding reads * wait until some have completed before issuing the next read */ - lck_mtx_lock(cl_mtxp); + cluster_iostate_wait(&iostate, max_rd_ahead, "cluster_read_direct"); - while ((iostate.io_issued - iostate.io_completed) > max_rd_ahead) { - iostate.io_wanted = 1; - msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_nocopy_read", 0); - } - lck_mtx_unlock(cl_mtxp); - if (iostate.io_error) { /* * one of the earlier reads we issued ran into a hard error @@ -3342,158 +4736,269 @@ cluster_nocopy_read(vnode_t vp, struct uio *uio, off_t filesize) * go wait for any other reads to complete before * returning the error to the caller */ - ubc_upl_abort_range(upl, (upl_offset & ~PAGE_MASK), upl_size, abort_flag); + ubc_upl_abort(upl, 0); - goto wait_for_reads; + goto wait_for_dreads; } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_START, - (int)upl, (int)upl_offset, (int)uio->uio_offset, io_size, 0); + upl, (int)upl_offset, (int)uio->uio_offset, io_size, 0); + + if(!useVectorUPL) { + if (no_zero_fill) + io_flag &= ~CL_PRESERVE; + else + io_flag |= CL_PRESERVE; + + retval = cluster_io(vp, upl, upl_offset, uio->uio_offset, io_size, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + + } else { + + if(!vector_upl_index) { + vector_upl = vector_upl_create(upl_offset); + v_upl_uio_offset = uio->uio_offset; + vector_upl_offset = upl_offset; + } + + vector_upl_set_subupl(vector_upl,upl, upl_size); + vector_upl_set_iostate(vector_upl, upl, vector_upl_size, upl_size); + vector_upl_index++; + vector_upl_size += upl_size; + vector_upl_iosize += io_size; + + if(issueVectorUPL || vector_upl_index == MAX_VECTOR_UPL_ELEMENTS || vector_upl_size >= max_vector_size) { + retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + reset_vector_run_state(); + } + } + last_iov_base = iov_base + io_size; - retval = cluster_io(vp, upl, upl_offset, uio->uio_offset, io_size, - CL_PRESERVE | CL_COMMIT | CL_READ | CL_ASYNC | CL_NOZERO, - (buf_t)NULL, &iostate); + if (lock) { + // We don't need to wait for the I/O to complete + cluster_unlock_direct_read(lock); + lock = NULL; + } /* * update the uio structure */ - uio_update(uio, (user_size_t)io_size); + if ((flags & IO_ENCRYPTED) && (max_io_size < io_size)) { + uio_update(uio, (user_size_t)max_io_size); + } + else { + uio_update(uio, (user_size_t)io_size); + } + + io_req_size -= io_size; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_END, - (int)upl, (int)uio->uio_offset, (int)uio_resid(uio), retval, 0); + upl, (int)uio->uio_offset, io_req_size, retval, 0); } /* end while */ -wait_for_reads: + if (retval == 0 && iostate.io_error == 0 && io_req_size == 0 && uio->uio_offset < filesize) { + + retval = cluster_io_type(uio, read_type, read_length, 0); + + if (retval == 0 && *read_type == IO_DIRECT) { + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_NONE, + (int)uio->uio_offset, (int)filesize, *read_type, *read_length, 0); + + goto next_dread; + } + } + +wait_for_dreads: + + if(retval == 0 && iostate.io_error == 0 && useVectorUPL && vector_upl_index) { + retval = vector_cluster_io(vp, vector_upl, vector_upl_offset, v_upl_uio_offset, vector_upl_iosize, io_flag, (buf_t)NULL, &iostate, callback, callback_arg); + reset_vector_run_state(); + } + + // We don't need to wait for the I/O to complete + if (lock) + cluster_unlock_direct_read(lock); + /* * make sure all async reads that are part of this stream * have completed before we return */ - lck_mtx_lock(cl_mtxp); - - while (iostate.io_issued != iostate.io_completed) { - iostate.io_wanted = 1; - msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_nocopy_read", 0); - } - lck_mtx_unlock(cl_mtxp); + cluster_iostate_wait(&iostate, 0, "cluster_read_direct"); if (iostate.io_error) - retval = iostate.io_error; + retval = iostate.io_error; + + lck_mtx_destroy(&iostate.io_mtxp, cl_mtx_grp); + + if (io_throttled == TRUE && retval == 0) + retval = EAGAIN; + + for (next_iov_base = orig_iov_base; next_iov_base < last_iov_base; next_iov_base += PAGE_SIZE) { + /* + * This is specifically done for pmap accounting purposes. + * vm_pre_fault() will call vm_fault() to enter the page into + * the pmap if there isn't _a_ physical page for that VA already. + */ + vm_pre_fault(vm_map_trunc_page(next_iov_base, PAGE_MASK)); + } + + if (io_req_size && retval == 0) { + /* + * we couldn't handle the tail of this request in DIRECT mode + * so fire it through the copy path + */ + retval = cluster_read_copy(vp, uio, io_req_size, filesize, flags, callback, callback_arg); + *read_type = IO_UNKNOWN; + } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_END, - (int)uio->uio_offset, (int)uio_resid(uio), 6, retval, 0); + (int)uio->uio_offset, (int)uio_resid(uio), io_req_size, retval, 0); return (retval); } static int -cluster_phys_read(vnode_t vp, struct uio *uio, off_t filesize) +cluster_read_contig(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length, + int (*callback)(buf_t, void *), void *callback_arg, int flags) { upl_page_info_t *pl; - upl_t upl; + upl_t upl[MAX_VECTS]; vm_offset_t upl_offset; - addr64_t dst_paddr; - off_t max_size; - int io_size; - user_size_t iov_len; + addr64_t dst_paddr = 0; user_addr_t iov_base; - int tail_size; - int upl_size; - int upl_needed_size; - int pages_in_pl; - int upl_flags; + off_t max_size; + upl_size_t upl_size; + vm_size_t upl_needed_size; + mach_msg_type_number_t pages_in_pl; + upl_control_flags_t upl_flags; kern_return_t kret; struct clios iostate; - int error; - int devblocksize; - - devblocksize = vp->v_mount->mnt_devblocksize; + int error= 0; + int cur_upl = 0; + int num_upl = 0; + int n; + u_int32_t xsize; + u_int32_t io_size; + u_int32_t devblocksize; + u_int32_t mem_alignment_mask; + u_int32_t tail_size = 0; + int bflag; + + if (flags & IO_PASSIVE) + bflag = CL_PASSIVE; + else + bflag = 0; + + if (flags & IO_NOCACHE) + bflag |= CL_NOCACHE; + /* * When we enter this routine, we know - * -- the resid will not exceed iov_len - * -- the target address is physically contiguous + * -- the read_length will not exceed the current iov_len + * -- the target address is physically contiguous for read_length */ + cluster_syncup(vp, filesize, callback, callback_arg, PUSH_SYNC); -#if LP64_DEBUG - if (IS_VALID_UIO_SEGFLG(uio->uio_segflg) == 0) { - panic("%s :%d - invalid uio_segflg\n", __FILE__, __LINE__); - } -#endif /* LP64_DEBUG */ + devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize; + mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask; - iov_len = uio_curriovlen(uio); - iov_base = uio_curriovbase(uio); + iostate.io_completed = 0; + iostate.io_issued = 0; + iostate.io_error = 0; + iostate.io_wanted = 0; + + lck_mtx_init(&iostate.io_mtxp, cl_mtx_grp, cl_mtx_attr); + +next_cread: + io_size = *read_length; max_size = filesize - uio->uio_offset; - // LP64todo - fix this! - if (max_size < 0 || (u_int64_t)max_size > iov_len) - io_size = iov_len; - else + if (io_size > max_size) io_size = max_size; - // LP64todo - fix this! - upl_offset = CAST_DOWN(vm_offset_t, iov_base) & PAGE_MASK; + iov_base = uio_curriovbase(uio); + + upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK); upl_needed_size = upl_offset + io_size; - error = 0; pages_in_pl = 0; upl_size = upl_needed_size; - upl_flags = UPL_FILE_IO | UPL_NO_SYNC | UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE; + upl_flags = UPL_FILE_IO | UPL_NO_SYNC | UPL_CLEAN_IN_PLACE | UPL_SET_INTERNAL | UPL_SET_LITE | UPL_SET_IO_WIRE + | UPL_MEMORY_TAG_MAKE(VM_KERN_MEMORY_FILE); + - kret = vm_map_get_upl(current_map(), + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 92)) | DBG_FUNC_START, + (int)upl_offset, (int)upl_size, (int)iov_base, io_size, 0); + + vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map; + kret = vm_map_get_upl(map, (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)), - &upl_size, &upl, NULL, &pages_in_pl, &upl_flags, 0); + &upl_size, &upl[cur_upl], NULL, &pages_in_pl, &upl_flags, 0); + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 92)) | DBG_FUNC_END, + (int)upl_offset, upl_size, io_size, kret, 0); if (kret != KERN_SUCCESS) { /* - * cluster_phys_read: failed to get pagelist + * failed to get pagelist */ - return(EINVAL); + error = EINVAL; + goto wait_for_creads; } + num_upl++; + if (upl_size < upl_needed_size) { /* * The upl_size wasn't satisfied. */ - ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); - - return(EINVAL); + error = EINVAL; + goto wait_for_creads; } - pl = ubc_upl_pageinfo(upl); + pl = ubc_upl_pageinfo(upl[cur_upl]); - dst_paddr = ((addr64_t)upl_phys_page(pl, 0) << 12) + (addr64_t)upl_offset; + dst_paddr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + (addr64_t)upl_offset; while (((uio->uio_offset & (devblocksize - 1)) || io_size < devblocksize) && io_size) { - int head_size; + u_int32_t head_size; - head_size = devblocksize - (int)(uio->uio_offset & (devblocksize - 1)); + head_size = devblocksize - (u_int32_t)(uio->uio_offset & (devblocksize - 1)); if (head_size > io_size) head_size = io_size; - error = cluster_align_phys_io(vp, uio, dst_paddr, head_size, CL_READ); + error = cluster_align_phys_io(vp, uio, dst_paddr, head_size, CL_READ, callback, callback_arg); - if (error) { - ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); + if (error) + goto wait_for_creads; - return(EINVAL); - } upl_offset += head_size; dst_paddr += head_size; io_size -= head_size; + + iov_base += head_size; } + if ((u_int32_t)iov_base & mem_alignment_mask) { + /* + * request doesn't set up on a memory boundary + * the underlying DMA engine can handle... + * return an error instead of going through + * the slow copy path since the intent of this + * path is direct I/O to device memory + */ + error = EINVAL; + goto wait_for_creads; + } + tail_size = io_size & (devblocksize - 1); - io_size -= tail_size; - iostate.io_completed = 0; - iostate.io_issued = 0; - iostate.io_error = 0; - iostate.io_wanted = 0; + io_size -= tail_size; while (io_size && error == 0) { - int xsize; - if (io_size > (MAX_UPL_TRANSFER * PAGE_SIZE)) - xsize = MAX_UPL_TRANSFER * PAGE_SIZE; + if (io_size > MAX_IO_CONTIG_SIZE) + xsize = MAX_IO_CONTIG_SIZE; else xsize = io_size; /* @@ -3504,17 +5009,20 @@ cluster_phys_read(vnode_t vp, struct uio *uio, off_t filesize) * if there are already too many outstanding reads * wait until some have completed before issuing the next */ - lck_mtx_lock(cl_mtxp); - - while ((iostate.io_issued - iostate.io_completed) > (8 * MAX_UPL_TRANSFER * PAGE_SIZE)) { - iostate.io_wanted = 1; - msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_phys_read", 0); - } - lck_mtx_unlock(cl_mtxp); + cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_read_contig"); - error = cluster_io(vp, upl, upl_offset, uio->uio_offset, xsize, - CL_READ | CL_NOZERO | CL_DEV_MEMORY | CL_ASYNC, - (buf_t)NULL, &iostate); + if (iostate.io_error) { + /* + * one of the earlier reads we issued ran into a hard error + * don't issue any more reads... + * go wait for any other reads to complete before + * returning the error to the caller + */ + goto wait_for_creads; + } + error = cluster_io(vp, upl[cur_upl], upl_offset, uio->uio_offset, xsize, + CL_READ | CL_NOZERO | CL_DEV_MEMORY | CL_ASYNC | bflag, + (buf_t)NULL, &iostate, callback, callback_arg); /* * The cluster_io read was issued successfully, * update the uio structure @@ -3527,45 +5035,126 @@ cluster_phys_read(vnode_t vp, struct uio *uio, off_t filesize) io_size -= xsize; } } + if (error == 0 && iostate.io_error == 0 && tail_size == 0 && num_upl < MAX_VECTS && uio->uio_offset < filesize) { + + error = cluster_io_type(uio, read_type, read_length, 0); + + if (error == 0 && *read_type == IO_CONTIG) { + cur_upl++; + goto next_cread; + } + } else + *read_type = IO_UNKNOWN; + +wait_for_creads: /* * make sure all async reads that are part of this stream * have completed before we proceed */ - lck_mtx_lock(cl_mtxp); - - while (iostate.io_issued != iostate.io_completed) { - iostate.io_wanted = 1; - msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_phys_read", 0); - } - lck_mtx_unlock(cl_mtxp); + cluster_iostate_wait(&iostate, 0, "cluster_read_contig"); if (iostate.io_error) error = iostate.io_error; + lck_mtx_destroy(&iostate.io_mtxp, cl_mtx_grp); + if (error == 0 && tail_size) - error = cluster_align_phys_io(vp, uio, dst_paddr, tail_size, CL_READ); + error = cluster_align_phys_io(vp, uio, dst_paddr, tail_size, CL_READ, callback, callback_arg); - /* - * just release our hold on the physically contiguous - * region without changing any state - */ - ubc_upl_abort_range(upl, 0, upl_size, UPL_ABORT_FREE_ON_EMPTY); + for (n = 0; n < num_upl; n++) + /* + * just release our hold on each physically contiguous + * region without changing any state + */ + ubc_upl_abort(upl[n], 0); return (error); } +static int +cluster_io_type(struct uio *uio, int *io_type, u_int32_t *io_length, u_int32_t min_length) +{ + user_size_t iov_len; + user_addr_t iov_base = 0; + upl_t upl; + upl_size_t upl_size; + upl_control_flags_t upl_flags; + int retval = 0; + + /* + * skip over any emtpy vectors + */ + uio_update(uio, (user_size_t)0); + + iov_len = uio_curriovlen(uio); + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 94)) | DBG_FUNC_START, uio, (int)iov_len, 0, 0, 0); + + if (iov_len) { + iov_base = uio_curriovbase(uio); + /* + * make sure the size of the vector isn't too big... + * internally, we want to handle all of the I/O in + * chunk sizes that fit in a 32 bit int + */ + if (iov_len > (user_size_t)MAX_IO_REQUEST_SIZE) + upl_size = MAX_IO_REQUEST_SIZE; + else + upl_size = (u_int32_t)iov_len; + + upl_flags = UPL_QUERY_OBJECT_TYPE | UPL_MEMORY_TAG_MAKE(VM_KERN_MEMORY_FILE); + + vm_map_t map = UIO_SEG_IS_USER_SPACE(uio->uio_segflg) ? current_map() : kernel_map; + if ((vm_map_get_upl(map, + (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)), + &upl_size, &upl, NULL, NULL, &upl_flags, 0)) != KERN_SUCCESS) { + /* + * the user app must have passed in an invalid address + */ + retval = EFAULT; + } + if (upl_size == 0) + retval = EFAULT; + + *io_length = upl_size; + + if (upl_flags & UPL_PHYS_CONTIG) + *io_type = IO_CONTIG; + else if (iov_len >= min_length) + *io_type = IO_DIRECT; + else + *io_type = IO_COPY; + } else { + /* + * nothing left to do for this uio + */ + *io_length = 0; + *io_type = IO_UNKNOWN; + } + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 94)) | DBG_FUNC_END, iov_base, *io_type, *io_length, retval, 0); + + return (retval); +} + + /* * generate advisory I/O's in the largest chunks possible * the completed pages will be released into the VM cache */ int advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid) +{ + return advisory_read_ext(vp, filesize, f_offset, resid, NULL, NULL, CL_PASSIVE); +} + +int +advisory_read_ext(vnode_t vp, off_t filesize, off_t f_offset, int resid, int (*callback)(buf_t, void *), void *callback_arg, int bflag) { upl_page_info_t *pl; upl_t upl; vm_offset_t upl_offset; - int upl_size; + int upl_size; off_t upl_f_offset; int start_offset; int start_pg; @@ -3577,10 +5166,22 @@ advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid) int retval = 0; int issued_io; int skip_range; + uint32_t max_io_size; + if ( !UBCINFOEXISTS(vp)) return(EINVAL); + if (resid < 0) + return(EINVAL); + + max_io_size = cluster_max_io_size(vp->v_mount, CL_READ); + + if ((vp->v_mount->mnt_kern_flag & MNTK_SSD) && !ignore_is_ssd) { + if (max_io_size > speculative_prefetch_max_iosize) + max_io_size = speculative_prefetch_max_iosize; + } + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 60)) | DBG_FUNC_START, (int)f_offset, resid, (int)filesize, 0, 0); @@ -3603,8 +5204,8 @@ advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid) io_size = max_size; upl_size = (start_offset + io_size + (PAGE_SIZE - 1)) & ~PAGE_MASK; - if (upl_size > (MAX_UPL_TRANSFER * PAGE_SIZE)) - upl_size = MAX_UPL_TRANSFER * PAGE_SIZE; + if ((uint32_t)upl_size > max_io_size) + upl_size = max_io_size; skip_range = 0; /* @@ -3636,7 +5237,7 @@ advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid) pages_in_upl = upl_size / PAGE_SIZE; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 61)) | DBG_FUNC_START, - (int)upl, (int)upl_f_offset, upl_size, start_offset, 0); + upl, (int)upl_f_offset, upl_size, start_offset, 0); kret = ubc_create_upl(vp, upl_f_offset, @@ -3661,7 +5262,7 @@ advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid) KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 61)) | DBG_FUNC_END, - (int)upl, (int)upl_f_offset, upl_size, start_offset, 0); + upl, (int)upl_f_offset, upl_size, start_offset, 0); for (last_pg = 0; last_pg < pages_in_upl; ) { @@ -3697,14 +5298,14 @@ advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid) upl_offset = start_pg * PAGE_SIZE; io_size = (last_pg - start_pg) * PAGE_SIZE; - if ((upl_f_offset + upl_offset + io_size) > filesize) + if ((off_t)(upl_f_offset + upl_offset + io_size) > filesize) io_size = filesize - (upl_f_offset + upl_offset); /* * issue an asynchronous read to cluster_io */ retval = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, io_size, - CL_ASYNC | CL_READ | CL_COMMIT | CL_AGE, (buf_t)NULL, (struct clios *)NULL); + CL_ASYNC | CL_READ | CL_COMMIT | CL_AGE | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg); issued_io = 1; } @@ -3729,12 +5330,30 @@ advisory_read(vnode_t vp, off_t filesize, off_t f_offset, int resid) int cluster_push(vnode_t vp, int flags) +{ + return cluster_push_ext(vp, flags, NULL, NULL); +} + + +int +cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *callback_arg) +{ + return cluster_push_err(vp, flags, callback, callback_arg, NULL); +} + +/* write errors via err, but return the number of clusters written */ +int +cluster_push_err(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *callback_arg, int *err) { int retval; + int my_sparse_wait = 0; struct cl_writebehind *wbp; + if (err) + *err = 0; + if ( !UBCINFOEXISTS(vp)) { - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, (int)vp, flags, 0, -1, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, kdebug_vnode(vp), flags, 0, -1, 0); return (0); } /* return if deferred write is set */ @@ -3742,32 +5361,99 @@ cluster_push(vnode_t vp, int flags) return (0); } if ((wbp = cluster_get_wbp(vp, CLW_RETURNLOCKED)) == NULL) { - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, (int)vp, flags, 0, -2, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, kdebug_vnode(vp), flags, 0, -2, 0); return (0); } - if (wbp->cl_number == 0 && wbp->cl_scmap == NULL) { + if (!ISSET(flags, IO_SYNC) && wbp->cl_number == 0 && wbp->cl_scmap == NULL) { lck_mtx_unlock(&wbp->cl_lockw); - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, (int)vp, flags, 0, -3, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, kdebug_vnode(vp), flags, 0, -3, 0); return(0); } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_START, - (int)wbp->cl_scmap, wbp->cl_number, flags, 0, 0); + wbp->cl_scmap, wbp->cl_number, flags, 0, 0); + + /* + * if we have an fsync in progress, we don't want to allow any additional + * sync/fsync/close(s) to occur until it finishes. + * note that its possible for writes to continue to occur to this file + * while we're waiting and also once the fsync starts to clean if we're + * in the sparse map case + */ + while (wbp->cl_sparse_wait) { + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_START, kdebug_vnode(vp), 0, 0, 0, 0); + + msleep((caddr_t)&wbp->cl_sparse_wait, &wbp->cl_lockw, PRIBIO + 1, "cluster_push_ext", NULL); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_END, kdebug_vnode(vp), 0, 0, 0, 0); + } + if (flags & IO_SYNC) { + my_sparse_wait = 1; + wbp->cl_sparse_wait = 1; + + /* + * this is an fsync (or equivalent)... we must wait for any existing async + * cleaning operations to complete before we evaulate the current state + * and finish cleaning... this insures that all writes issued before this + * fsync actually get cleaned to the disk before this fsync returns + */ + while (wbp->cl_sparse_pushes) { + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 98)) | DBG_FUNC_START, kdebug_vnode(vp), 0, 0, 0, 0); + + msleep((caddr_t)&wbp->cl_sparse_pushes, &wbp->cl_lockw, PRIBIO + 1, "cluster_push_ext", NULL); + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 98)) | DBG_FUNC_END, kdebug_vnode(vp), 0, 0, 0, 0); + } + } if (wbp->cl_scmap) { - sparse_cluster_push(wbp, vp, ubc_getsize(vp), 1); + void *scmap; - retval = 1; - } else - retval = cluster_try_push(wbp, vp, ubc_getsize(vp), 0, 1); + if (wbp->cl_sparse_pushes < SPARSE_PUSH_LIMIT) { + + scmap = wbp->cl_scmap; + wbp->cl_scmap = NULL; + + wbp->cl_sparse_pushes++; + + lck_mtx_unlock(&wbp->cl_lockw); + + retval = sparse_cluster_push(&scmap, vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg); + lck_mtx_lock(&wbp->cl_lockw); + + wbp->cl_sparse_pushes--; + + if (wbp->cl_sparse_wait && wbp->cl_sparse_pushes == 0) + wakeup((caddr_t)&wbp->cl_sparse_pushes); + } else { + retval = sparse_cluster_push(&(wbp->cl_scmap), vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg); + } + if (err) + *err = retval; + retval = 1; + } else { + retval = cluster_try_push(wbp, vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg, err); + } lck_mtx_unlock(&wbp->cl_lockw); if (flags & IO_SYNC) - (void)vnode_waitforwrites(vp, 0, 0, 0, (char *)"cluster_push"); + (void)vnode_waitforwrites(vp, 0, 0, 0, "cluster_push"); + + if (my_sparse_wait) { + /* + * I'm the owner of the serialization token + * clear it and wakeup anyone that is waiting + * for me to finish + */ + lck_mtx_lock(&wbp->cl_lockw); + wbp->cl_sparse_wait = 0; + wakeup((caddr_t)&wbp->cl_sparse_wait); + + lck_mtx_unlock(&wbp->cl_lockw); + } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_END, - (int)wbp->cl_scmap, wbp->cl_number, retval, 0, 0); + wbp->cl_scmap, wbp->cl_number, retval, 0, 0); return (retval); } @@ -3781,12 +5467,12 @@ cluster_release(struct ubc_info *ubc) if ((wbp = ubc->cl_wbehind)) { - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_START, (int)ubc, (int)wbp->cl_scmap, wbp->cl_scdirty, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_START, ubc, wbp->cl_scmap, 0, 0, 0); if (wbp->cl_scmap) vfs_drt_control(&(wbp->cl_scmap), 0); } else { - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_START, (int)ubc, 0, 0, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_START, ubc, 0, 0, 0, 0); } rap = ubc->cl_rahead; @@ -3802,34 +5488,12 @@ cluster_release(struct ubc_info *ubc) ubc->cl_rahead = NULL; ubc->cl_wbehind = NULL; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_END, (int)ubc, (int)rap, (int)wbp, 0, 0); -} - - -static void -cluster_push_EOF(vnode_t vp, off_t EOF) -{ - struct cl_writebehind *wbp; - - wbp = cluster_get_wbp(vp, CLW_ALLOCATE | CLW_RETURNLOCKED); - - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_START, - (int)wbp->cl_scmap, wbp->cl_number, (int)EOF, 0, 0); - - if (wbp->cl_scmap) - sparse_cluster_push(wbp, vp, EOF, 1); - else - cluster_try_push(wbp, vp, EOF, 0, 1); - - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_END, - (int)wbp->cl_scmap, wbp->cl_number, 0, 0, 0); - - lck_mtx_unlock(&wbp->cl_lockw); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 81)) | DBG_FUNC_END, ubc, rap, wbp, 0, 0); } static int -cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int can_delay, int push_all) +cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_flag, int io_flags, int (*callback)(buf_t, void *), void *callback_arg, int *err) { int cl_index; int cl_index1; @@ -3837,11 +5501,22 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int can_dela int cl_len; int cl_pushed = 0; struct cl_wextent l_clusters[MAX_CLUSTERS]; + u_int max_cluster_pgcount; + int error = 0; + max_cluster_pgcount = MAX_CLUSTER_SIZE(vp) / PAGE_SIZE; /* * the write behind context exists and has * already been locked... - * + */ + if (wbp->cl_number == 0) + /* + * no clusters to push + * return number of empty slots + */ + return (MAX_CLUSTERS); + + /* * make a local 'sorted' copy of the clusters * and clear wbp->cl_number so that new clusters can * be developed @@ -3857,9 +5532,10 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int can_dela } if (min_index == -1) break; + l_clusters[cl_index].b_addr = wbp->cl_clusters[min_index].b_addr; l_clusters[cl_index].e_addr = wbp->cl_clusters[min_index].e_addr; - l_clusters[cl_index].io_nocache = wbp->cl_clusters[min_index].io_nocache; + l_clusters[cl_index].io_flags = wbp->cl_clusters[min_index].io_flags; wbp->cl_clusters[min_index].b_addr = wbp->cl_clusters[min_index].e_addr; } @@ -3867,7 +5543,9 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int can_dela cl_len = cl_index; - if (can_delay && cl_len == MAX_CLUSTERS) { + /* skip switching to the sparse cluster mechanism if on diskimage */ + if ( ((push_flag & PUSH_DELAY) && cl_len == MAX_CLUSTERS ) && + !(vp->v_mount->mnt_kern_flag & MNTK_VIRTUALDEV) ) { int i; /* @@ -3877,7 +5555,7 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int can_dela * used for managing more random I/O patterns * * we know that we've got all clusters currently in use and the next write doesn't fit into one of them... - * that's why we're in try_push with can_delay true... + * that's why we're in try_push with PUSH_DELAY... * * check to make sure that all the clusters except the last one are 'full'... and that each cluster * is adjacent to the next (i.e. we're looking for sequential writes) they were sorted above @@ -3890,45 +5568,49 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int can_dela * of order... if this occurs at the tail of the last cluster, we don't want to fall into the sparse cluster world... */ for (i = 0; i < MAX_CLUSTERS - 1; i++) { - if ((l_clusters[i].e_addr - l_clusters[i].b_addr) != MAX_UPL_TRANSFER) + if ((l_clusters[i].e_addr - l_clusters[i].b_addr) != max_cluster_pgcount) goto dont_try; if (l_clusters[i].e_addr != l_clusters[i+1].b_addr) goto dont_try; } } - /* - * drop the lock while we're firing off the I/Os... - * this is safe since I'm working off of a private sorted copy - * of the clusters, and I'm going to re-evaluate the public - * state after I retake the lock - */ - lck_mtx_unlock(&wbp->cl_lockw); - for (cl_index = 0; cl_index < cl_len; cl_index++) { - int flags; - struct cl_extent cl; + int flags; + struct cl_extent cl; + int retval; + + flags = io_flags & (IO_PASSIVE|IO_CLOSE); /* * try to push each cluster in turn... */ - if (l_clusters[cl_index].io_nocache) - flags = IO_NOCACHE; - else - flags = 0; + if (l_clusters[cl_index].io_flags & CLW_IONOCACHE) + flags |= IO_NOCACHE; + + if (l_clusters[cl_index].io_flags & CLW_IOPASSIVE) + flags |= IO_PASSIVE; + + if (push_flag & PUSH_SYNC) + flags |= IO_SYNC; + cl.b_addr = l_clusters[cl_index].b_addr; cl.e_addr = l_clusters[cl_index].e_addr; - cluster_push_x(vp, &cl, EOF, flags); + retval = cluster_push_now(vp, &cl, EOF, flags, callback, callback_arg); + + if (error == 0 && retval) + error = retval; l_clusters[cl_index].b_addr = 0; l_clusters[cl_index].e_addr = 0; cl_pushed++; - if (push_all == 0) + if ( !(push_flag & PUSH_ALL) ) break; } - lck_mtx_lock(&wbp->cl_lockw); + if (err) + *err = error; dont_try: if (cl_len > cl_pushed) { @@ -3946,14 +5628,14 @@ dont_try: * * collect the active public clusters... */ - sparse_cluster_switch(wbp, vp, EOF); + sparse_cluster_switch(wbp, vp, EOF, callback, callback_arg); for (cl_index = 0, cl_index1 = 0; cl_index < cl_len; cl_index++) { if (l_clusters[cl_index].b_addr == l_clusters[cl_index].e_addr) continue; wbp->cl_clusters[cl_index1].b_addr = l_clusters[cl_index].b_addr; wbp->cl_clusters[cl_index1].e_addr = l_clusters[cl_index].e_addr; - wbp->cl_clusters[cl_index1].io_nocache = l_clusters[cl_index].io_nocache; + wbp->cl_clusters[cl_index1].io_flags = l_clusters[cl_index].io_flags; cl_index1++; } @@ -3966,7 +5648,7 @@ dont_try: * and collect the original clusters that were moved into the * local storage for sorting purposes */ - sparse_cluster_switch(wbp, vp, EOF); + sparse_cluster_switch(wbp, vp, EOF, callback, callback_arg); } else { /* @@ -3980,7 +5662,7 @@ dont_try: wbp->cl_clusters[cl_index1].b_addr = l_clusters[cl_index].b_addr; wbp->cl_clusters[cl_index1].e_addr = l_clusters[cl_index].e_addr; - wbp->cl_clusters[cl_index1].io_nocache = l_clusters[cl_index].io_nocache; + wbp->cl_clusters[cl_index1].io_flags = l_clusters[cl_index].io_flags; cl_index1++; } @@ -3990,13 +5672,13 @@ dont_try: wbp->cl_number = cl_index1; } } - return(MAX_CLUSTERS - wbp->cl_number); + return (MAX_CLUSTERS - wbp->cl_number); } static int -cluster_push_x(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags) +cluster_push_now(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags, int (*callback)(buf_t, void *), void *callback_arg) { upl_page_info_t *pl; upl_t upl; @@ -4009,11 +5691,19 @@ cluster_push_x(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags) int io_size; int io_flags; int upl_flags; + int bflag; int size; int error = 0; int retval; kern_return_t kret; + if (flags & IO_PASSIVE) + bflag = CL_PASSIVE; + else + bflag = 0; + + if (flags & IO_SKIP_ENCRYPTION) + bflag |= CL_ENCRYPTED; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 51)) | DBG_FUNC_START, (int)cl->b_addr, (int)cl->e_addr, (int)EOF, flags, 0); @@ -4073,7 +5763,7 @@ cluster_push_x(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags) if (kret != KERN_SUCCESS) panic("cluster_push: failed to get pagelist"); - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 41)) | DBG_FUNC_END, (int)upl, upl_f_offset, 0, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 41)) | DBG_FUNC_END, upl, upl_f_offset, 0, 0, 0); /* * since we only asked for the dirty pages back @@ -4135,13 +5825,19 @@ cluster_push_x(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags) io_size = min(size, (last_pg - start_pg) * PAGE_SIZE); - io_flags = CL_THROTTLE | CL_COMMIT; + io_flags = CL_THROTTLE | CL_COMMIT | CL_AGE | bflag; if ( !(flags & IO_SYNC)) io_flags |= CL_ASYNC; + if (flags & IO_CLOSE) + io_flags |= CL_CLOSE; + + if (flags & IO_NOCACHE) + io_flags |= CL_NOCACHE; + retval = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, io_size, - io_flags, (buf_t)NULL, (struct clios *)NULL); + io_flags, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg); if (error == 0 && retval) error = retval; @@ -4158,14 +5854,11 @@ cluster_push_x(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags) * sparse_cluster_switch is called with the write behind lock held */ static void -sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF) +sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int (*callback)(buf_t, void *), void *callback_arg) { int cl_index; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_START, (int)vp, (int)wbp->cl_scmap, wbp->cl_scdirty, 0, 0); - - if (wbp->cl_scmap == NULL) - wbp->cl_scdirty = 0; + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_START, kdebug_vnode(vp), wbp->cl_scmap, 0, 0, 0); for (cl_index = 0; cl_index < wbp->cl_number; cl_index++) { int flags; @@ -4173,51 +5866,57 @@ sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF) for (cl.b_addr = wbp->cl_clusters[cl_index].b_addr; cl.b_addr < wbp->cl_clusters[cl_index].e_addr; cl.b_addr++) { - if (ubc_page_op(vp, (off_t)(cl.b_addr * PAGE_SIZE_64), 0, 0, &flags) == KERN_SUCCESS) { + if (ubc_page_op(vp, (off_t)(cl.b_addr * PAGE_SIZE_64), 0, NULL, &flags) == KERN_SUCCESS) { if (flags & UPL_POP_DIRTY) { cl.e_addr = cl.b_addr + 1; - sparse_cluster_add(wbp, vp, &cl, EOF); + sparse_cluster_add(&(wbp->cl_scmap), vp, &cl, EOF, callback, callback_arg); } } } } wbp->cl_number = 0; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_END, (int)vp, (int)wbp->cl_scmap, wbp->cl_scdirty, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_END, kdebug_vnode(vp), wbp->cl_scmap, 0, 0, 0); } /* - * sparse_cluster_push is called with the write behind lock held + * sparse_cluster_push must be called with the write-behind lock held if the scmap is + * still associated with the write-behind context... however, if the scmap has been disassociated + * from the write-behind context (the cluster_push case), the wb lock is not held */ -static void -sparse_cluster_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_all) +static int +sparse_cluster_push(void **scmap, vnode_t vp, off_t EOF, int push_flag, int io_flags, int (*callback)(buf_t, void *), void *callback_arg) { struct cl_extent cl; off_t offset; u_int length; + int error = 0; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_START, (int)vp, (int)wbp->cl_scmap, wbp->cl_scdirty, push_all, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_START, kdebug_vnode(vp), (*scmap), 0, push_flag, 0); - if (push_all) - vfs_drt_control(&(wbp->cl_scmap), 1); + if (push_flag & PUSH_ALL) + vfs_drt_control(scmap, 1); for (;;) { - if (vfs_drt_get_cluster(&(wbp->cl_scmap), &offset, &length) != KERN_SUCCESS) + int retval; + if (vfs_drt_get_cluster(scmap, &offset, &length) != KERN_SUCCESS) break; cl.b_addr = (daddr64_t)(offset / PAGE_SIZE_64); cl.e_addr = (daddr64_t)((offset + length) / PAGE_SIZE_64); - wbp->cl_scdirty -= (int)(cl.e_addr - cl.b_addr); - - cluster_push_x(vp, &cl, EOF, 0); + retval = cluster_push_now(vp, &cl, EOF, io_flags & (IO_PASSIVE|IO_CLOSE), callback, callback_arg); + if (error == 0 && retval) + error = retval; - if (push_all == 0) + if ( !(push_flag & PUSH_ALL) ) break; } - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_END, (int)vp, (int)wbp->cl_scmap, wbp->cl_scdirty, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_END, kdebug_vnode(vp), (*scmap), 0, 0, 0); + + return error; } @@ -4225,38 +5924,34 @@ sparse_cluster_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_ * sparse_cluster_add is called with the write behind lock held */ static void -sparse_cluster_add(struct cl_writebehind *wbp, vnode_t vp, struct cl_extent *cl, off_t EOF) +sparse_cluster_add(void **scmap, vnode_t vp, struct cl_extent *cl, off_t EOF, int (*callback)(buf_t, void *), void *callback_arg) { u_int new_dirty; u_int length; off_t offset; - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_START, (int)wbp->cl_scmap, wbp->cl_scdirty, (int)cl->b_addr, (int)cl->e_addr, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_START, (*scmap), 0, cl->b_addr, (int)cl->e_addr, 0); offset = (off_t)(cl->b_addr * PAGE_SIZE_64); length = ((u_int)(cl->e_addr - cl->b_addr)) * PAGE_SIZE; - while (vfs_drt_mark_pages(&(wbp->cl_scmap), offset, length, &new_dirty) != KERN_SUCCESS) { + while (vfs_drt_mark_pages(scmap, offset, length, &new_dirty) != KERN_SUCCESS) { /* * no room left in the map * only a partial update was done * push out some pages and try again */ - wbp->cl_scdirty += new_dirty; - - sparse_cluster_push(wbp, vp, EOF, 0); + sparse_cluster_push(scmap, vp, EOF, 0, 0, callback, callback_arg); offset += (new_dirty * PAGE_SIZE_64); length -= (new_dirty * PAGE_SIZE); } - wbp->cl_scdirty += new_dirty; - - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_END, (int)vp, (int)wbp->cl_scmap, wbp->cl_scdirty, 0, 0); + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_END, kdebug_vnode(vp), (*scmap), 0, 0, 0); } static int -cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, int xsize, int flags) +cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, u_int32_t xsize, int flags, int (*callback)(buf_t, void *), void *callback_arg) { upl_page_info_t *pl; upl_t upl; @@ -4266,17 +5961,33 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, int xsize int did_read = 0; int abort_flags; int upl_flags; + int bflag; + + if (flags & IO_PASSIVE) + bflag = CL_PASSIVE; + else + bflag = 0; + + if (flags & IO_NOCACHE) + bflag |= CL_NOCACHE; upl_flags = UPL_SET_LITE; - if (! (flags & CL_READ)) { + + if ( !(flags & CL_READ) ) { /* * "write" operation: let the UPL subsystem know * that we intend to modify the buffer cache pages * we're gathering. */ upl_flags |= UPL_WILL_MODIFY; + } else { + /* + * indicate that there is no need to pull the + * mapping for this page... we're only going + * to read from it, not modify it. + */ + upl_flags |= UPL_FILE_IO; } - kret = ubc_create_upl(vp, uio->uio_offset & ~PAGE_MASK_64, PAGE_SIZE, @@ -4292,7 +6003,7 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, int xsize * issue a synchronous read to cluster_io */ error = cluster_io(vp, upl, 0, uio->uio_offset & ~PAGE_MASK_64, PAGE_SIZE, - CL_READ, (buf_t)NULL, (struct clios *)NULL); + CL_READ | bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg); if (error) { ubc_upl_abort_range(upl, 0, PAGE_SIZE, UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY); @@ -4300,7 +6011,7 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, int xsize } did_read = 1; } - ubc_paddr = ((addr64_t)upl_phys_page(pl, 0) << 12) + (addr64_t)(uio->uio_offset & PAGE_MASK_64); + ubc_paddr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + (addr64_t)(uio->uio_offset & PAGE_MASK_64); /* * NOTE: There is no prototype for the following in BSD. It, and the definitions @@ -4320,9 +6031,9 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, int xsize * issue a synchronous write to cluster_io */ error = cluster_io(vp, upl, 0, uio->uio_offset & ~PAGE_MASK_64, PAGE_SIZE, - 0, (buf_t)NULL, (struct clios *)NULL); + bflag, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg); } - if (error == 0) + if (error == 0) uio_update(uio, (user_size_t)xsize); if (did_read) @@ -4335,20 +6046,22 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, int xsize return (error); } - - int -cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int xsize) +cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int *io_resid) { int pg_offset; int pg_index; int csize; int segflg; int retval = 0; + int xsize; upl_page_info_t *pl; + int dirty_count; + + xsize = *io_resid; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_START, - (int)uio->uio_offset, uio_resid(uio), upl_offset, xsize, 0); + (int)uio->uio_offset, upl_offset, xsize, 0, 0); segflg = uio->uio_segflg; @@ -4369,17 +6082,10 @@ cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int xsize) uio->uio_segflg = UIO_PHYS_USERSPACE64; break; - case UIO_SYSSPACE32: - uio->uio_segflg = UIO_PHYS_SYSSPACE32; - break; - case UIO_SYSSPACE: uio->uio_segflg = UIO_PHYS_SYSSPACE; break; - case UIO_SYSSPACE64: - uio->uio_segflg = UIO_PHYS_SYSSPACE64; - break; } pl = ubc_upl_pageinfo(upl); @@ -4387,10 +6093,13 @@ cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int xsize) pg_offset = upl_offset & PAGE_MASK; csize = min(PAGE_SIZE - pg_offset, xsize); + dirty_count = 0; while (xsize && retval == 0) { addr64_t paddr; - paddr = ((addr64_t)upl_phys_page(pl, pg_index) << 12) + pg_offset; + paddr = ((addr64_t)upl_phys_page(pl, pg_index) << PAGE_SHIFT) + pg_offset; + if ((uio->uio_rw == UIO_WRITE) && (upl_dirty_page(pl, pg_index) == FALSE)) + dirty_count++; retval = uiomove64(paddr, csize, uio); @@ -4399,17 +6108,28 @@ cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int xsize) xsize -= csize; csize = min(PAGE_SIZE, xsize); } + *io_resid = xsize; + uio->uio_segflg = segflg; + task_update_logical_writes(current_task(), (dirty_count * PAGE_SIZE), TASK_WRITE_DEFERRED, upl_lookup_vnode(upl)); KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_END, - (int)uio->uio_offset, uio_resid(uio), retval, segflg, 0); - + (int)uio->uio_offset, xsize, retval, segflg, 0); + return (retval); } int cluster_copy_ubc_data(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty) +{ + + return (cluster_copy_ubc_data_internal(vp, uio, io_resid, mark_dirty, 1)); +} + + +static int +cluster_copy_ubc_data_internal(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty, int take_reference) { int segflg; int io_size; @@ -4418,14 +6138,16 @@ cluster_copy_ubc_data(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty int retval = 0; memory_object_control_t control; + io_size = *io_resid; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_START, - (int)uio->uio_offset, uio_resid(uio), 0, *io_resid, 0); + (int)uio->uio_offset, io_size, mark_dirty, take_reference, 0); control = ubc_getobject(vp, UBC_FLAGS_NONE); + if (control == MEMORY_OBJECT_CONTROL_NULL) { KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_END, - (int)uio->uio_offset, uio_resid(uio), retval, 3, 0); + (int)uio->uio_offset, io_size, retval, 3, 0); return(0); } @@ -4443,14 +6165,6 @@ cluster_copy_ubc_data(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty uio->uio_segflg = UIO_PHYS_USERSPACE64; break; - case UIO_SYSSPACE32: - uio->uio_segflg = UIO_PHYS_SYSSPACE32; - break; - - case UIO_SYSSPACE64: - uio->uio_segflg = UIO_PHYS_SYSSPACE64; - break; - case UIO_USERSPACE: case UIO_USERISPACE: uio->uio_segflg = UIO_PHYS_USERSPACE; @@ -4465,8 +6179,8 @@ cluster_copy_ubc_data(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty start_offset = (int)(uio->uio_offset & PAGE_MASK_64); xsize = uio_resid(uio); - retval = memory_object_control_uiomove(control, uio->uio_offset - start_offset, - uio, start_offset, io_size, mark_dirty); + retval = memory_object_control_uiomove(control, uio->uio_offset - start_offset, uio, + start_offset, io_size, mark_dirty, take_reference); xsize -= uio_resid(uio); io_size -= xsize; } @@ -4474,7 +6188,7 @@ cluster_copy_ubc_data(vnode_t vp, struct uio *uio, int *io_resid, int mark_dirty *io_resid = io_size; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_END, - (int)uio->uio_offset, uio_resid(uio), retval, 0x80000000 | segflg, 0); + (int)uio->uio_offset, io_size, retval, 0x80000000 | segflg, 0); return(retval); } @@ -4488,7 +6202,7 @@ is_file_clean(vnode_t vp, off_t filesize) int total_dirty = 0; for (f_offset = 0; f_offset < filesize; f_offset += PAGE_SIZE_64) { - if (ubc_page_op(vp, f_offset, 0, 0, &flags) == KERN_SUCCESS) { + if (ubc_page_op(vp, f_offset, 0, NULL, &flags) == KERN_SUCCESS) { if (flags & UPL_POP_DIRTY) { total_dirty++; } @@ -4520,15 +6234,15 @@ is_file_clean(vnode_t vp, off_t filesize) * single hashtable entry. Each hashtable entry is aligned to this * size within the file. */ -#define DRT_BITVECTOR_PAGES 256 +#define DRT_BITVECTOR_PAGES ((1024 * 1024) / PAGE_SIZE) /* * File offset handling. * - * DRT_ADDRESS_MASK is dependent on DRT_BITVECTOR_PAGES; - * the correct formula is (~(DRT_BITVECTOR_PAGES * PAGE_SIZE) - 1) + * DRT_ADDRESS_MASK is dependent on DRT_BITVECTOR_PAGES; + * the correct formula is (~((DRT_BITVECTOR_PAGES * PAGE_SIZE) - 1)) */ -#define DRT_ADDRESS_MASK (~((1 << 20) - 1)) +#define DRT_ADDRESS_MASK (~((DRT_BITVECTOR_PAGES * PAGE_SIZE) - 1)) #define DRT_ALIGN_ADDRESS(addr) ((addr) & DRT_ADDRESS_MASK) /* @@ -4583,6 +6297,14 @@ is_file_clean(vnode_t vp, off_t filesize) #define DRT_HASH_SMALL_MODULUS 23 #define DRT_HASH_LARGE_MODULUS 401 +/* + * Physical memory required before the large hash modulus is permitted. + * + * On small memory systems, the large hash modulus can lead to phsyical + * memory starvation, so we avoid using it there. + */ +#define DRT_HASH_LARGE_MEMORY_REQUIRED (1024LL * 1024LL * 1024LL) /* 1GiB */ + #define DRT_SMALL_ALLOCATION 1024 /* 104 bytes spare */ #define DRT_LARGE_ALLOCATION 16384 /* 344 bytes spare */ @@ -4618,7 +6340,15 @@ is_file_clean(vnode_t vp, off_t filesize) */ struct vfs_drt_hashentry { u_int64_t dhe_control; - u_int32_t dhe_bitvector[DRT_BITVECTOR_PAGES / 32]; +/* +* dhe_bitvector was declared as dhe_bitvector[DRT_BITVECTOR_PAGES / 32]; +* DRT_BITVECTOR_PAGES is defined as ((1024 * 1024) / PAGE_SIZE) +* Since PAGE_SIZE is only known at boot time, +* -define MAX_DRT_BITVECTOR_PAGES for smallest supported page size (4k) +* -declare dhe_bitvector array for largest possible length +*/ +#define MAX_DRT_BITVECTOR_PAGES (1024 * 1024)/( 4 * 1024) + u_int32_t dhe_bitvector[MAX_DRT_BITVECTOR_PAGES/32]; }; /* @@ -4675,7 +6405,7 @@ static kern_return_t vfs_drt_do_mark_pages( void **cmapp, u_int64_t offset, u_int length, - int *setcountp, + u_int *setcountp, int dirty); static void vfs_drt_trace( struct vfs_drt_clustermap *cmap, @@ -4700,7 +6430,8 @@ vfs_drt_alloc_map(struct vfs_drt_clustermap **cmapp) struct vfs_drt_clustermap *cmap, *ocmap; kern_return_t kret; u_int64_t offset; - int nsize, i, active_buckets, index, copycount; + u_int32_t i; + int nsize, active_buckets, index, copycount; ocmap = NULL; if (cmapp != NULL) @@ -4724,8 +6455,12 @@ vfs_drt_alloc_map(struct vfs_drt_clustermap **cmapp) * see whether we should grow to the large one. */ if (ocmap->scm_modulus == DRT_HASH_SMALL_MODULUS) { - /* if the ring is nearly full */ - if (active_buckets > (DRT_HASH_SMALL_MODULUS - 5)) { + /* + * If the ring is nearly full and we are allowed to + * use the large modulus, upgrade. + */ + if ((active_buckets > (DRT_HASH_SMALL_MODULUS - 5)) && + (max_mem >= DRT_HASH_LARGE_MEMORY_REQUIRED)) { nsize = DRT_HASH_LARGE_MODULUS; } else { nsize = DRT_HASH_SMALL_MODULUS; @@ -4749,7 +6484,7 @@ vfs_drt_alloc_map(struct vfs_drt_clustermap **cmapp) */ kret = kmem_alloc(kernel_map, (vm_offset_t *)&cmap, - (nsize == DRT_HASH_SMALL_MODULUS) ? DRT_SMALL_ALLOCATION : DRT_LARGE_ALLOCATION); + (nsize == DRT_HASH_SMALL_MODULUS) ? DRT_SMALL_ALLOCATION : DRT_LARGE_ALLOCATION, VM_KERN_MEMORY_FILE); if (kret != KERN_SUCCESS) return(kret); cmap->scm_magic = DRT_SCM_MAGIC; @@ -4779,6 +6514,7 @@ vfs_drt_alloc_map(struct vfs_drt_clustermap **cmapp) if (kret != KERN_SUCCESS) { /* XXX need to bail out gracefully here */ panic("vfs_drt: new cluster map mysteriously too small"); + index = 0; } /* copy */ DRT_HASH_COPY(ocmap, i, cmap, index); @@ -4827,7 +6563,8 @@ vfs_drt_free_map(struct vfs_drt_clustermap *cmap) static kern_return_t vfs_drt_search_index(struct vfs_drt_clustermap *cmap, u_int64_t offset, int *indexp) { - int index, i; + int index; + u_int32_t i; offset = DRT_ALIGN_ADDRESS(offset); index = DRT_HASH(cmap, offset); @@ -4871,7 +6608,8 @@ vfs_drt_get_index(struct vfs_drt_clustermap **cmapp, u_int64_t offset, int *inde { struct vfs_drt_clustermap *cmap; kern_return_t kret; - int index, i; + u_int32_t index; + u_int32_t i; cmap = *cmapp; @@ -4926,7 +6664,7 @@ vfs_drt_do_mark_pages( void **private, u_int64_t offset, u_int length, - int *setcountp, + u_int *setcountp, int dirty) { struct vfs_drt_clustermap *cmap, **cmapp; @@ -5041,7 +6779,7 @@ vfs_drt_do_mark_pages( * Returns KERN_SUCCESS if all the pages were successfully marked. */ static kern_return_t -vfs_drt_mark_pages(void **cmapp, off_t offset, u_int length, int *setcountp) +vfs_drt_mark_pages(void **cmapp, off_t offset, u_int length, u_int *setcountp) { /* XXX size unused, drop from interface */ return(vfs_drt_do_mark_pages(cmapp, offset, length, setcountp, 1)); @@ -5081,7 +6819,8 @@ vfs_drt_get_cluster(void **cmapp, off_t *offsetp, u_int *lengthp) struct vfs_drt_clustermap *cmap; u_int64_t offset; u_int length; - int index, i, j, fs, ls; + u_int32_t j; + int index, i, fs, ls; /* sanity */ if ((cmapp == NULL) || (*cmapp == NULL))