xnu-3789.70.16.tar.gz

[apple/xnu.git] / bsd / vfs / vfs_cluster.c

diff --git a/bsd/vfs/vfs_cluster.c b/bsd/vfs/vfs_cluster.c
index 5aec1498a48a26d7480d7a8ec7c7176177659475..70eecc5ff44759d87818dc61b6ea475697195ea0 100644 (file)
--- a/bsd/vfs/vfs_cluster.c
+++ b/bsd/vfs/vfs_cluster.c
@@ -1,5 +1,5 @@
 /*
 /*

- * Copyright (c) 2000-2008 Apple Inc. All rights reserved.
+ * Copyright (c) 2000-2014 Apple Inc. All rights reserved.

  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 
  *
  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  * 


@@ -71,6 +71,7 @@
 #include <sys/time.h>
 #include <sys/kernel.h>
 #include <sys/resourcevar.h>
 #include <sys/time.h>
 #include <sys/kernel.h>
 #include <sys/resourcevar.h>

+#include <miscfs/specfs/specdev.h>

 #include <sys/uio_internal.h>
 #include <libkern/libkern.h>
 #include <machine/machine_routines.h>
 #include <sys/uio_internal.h>
 #include <libkern/libkern.h>
 #include <machine/machine_routines.h>


@@ -82,14 +83,21 @@
 #include <mach/memory_object_types.h>
 #include <mach/vm_map.h>
 #include <mach/upl.h>
 #include <mach/memory_object_types.h>
 #include <mach/vm_map.h>
 #include <mach/upl.h>

+#include <kern/task.h>
+#include <kern/policy_internal.h>

 
 #include <vm/vm_kern.h>
 #include <vm/vm_map.h>
 #include <vm/vm_pageout.h>
 
 #include <vm/vm_kern.h>
 #include <vm/vm_map.h>
 #include <vm/vm_pageout.h>

+#include <vm/vm_fault.h>

 
 #include <sys/kdebug.h>
 #include <libkern/OSAtomic.h>  
 
 
 #include <sys/kdebug.h>
 #include <libkern/OSAtomic.h>  
 

+#include <sys/sdt.h>
+
+#include <stdbool.h>
+

 #if 0
 #undef KERNEL_DEBUG
 #define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT
 #if 0
 #undef KERNEL_DEBUG
 #define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT


@@ -111,9 +119,15 @@
 #define CL_DIRECT_IO   0x1000
 #define CL_PASSIVE     0x2000
 #define CL_IOSTREAMING 0x4000
 #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_ELEMENTS        8

-#define MAX_VECTOR_UPL_SIZE    (2 * MAX_UPL_SIZE) * PAGE_SIZE
+#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 upl_t vector_upl_create(vm_offset_t);
 extern boolean_t vector_upl_is_valid(upl_t);


@@ -122,17 +136,31 @@ 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 {
 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 */
 };
 
         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_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_UNKNOWN      0
 #define        IO_DIRECT       1


@@ -153,10 +181,12 @@ static int cluster_io_type(struct uio *uio, int *io_type, u_int32_t *io_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, int (*)(buf_t, void *), void *callback_arg);
 static int cluster_iodone(buf_t bp, void *callback_arg);
 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 (*)(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);
-static int cluster_hard_throttle_on(vnode_t vp, uint32_t);
+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 void cluster_iostate_wait(struct clios *iostate, u_int target, const char *wait_name);

 
 

-static void cluster_syncup(vnode_t vp, off_t newEOF, int (*)(buf_t, void *), void *callback_arg);
+static void cluster_syncup(vnode_t vp, off_t newEOF, int (*)(buf_t, void *), void *callback_arg, int flags);

 
 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 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);


@@ -182,10 +212,10 @@ static void       cluster_read_ahead(vnode_t vp, struct cl_extent *extent, off_t files
 
 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_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 (*)(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 void    sparse_cluster_switch(struct cl_writebehind *, vnode_t vp, off_t EOF, int (*)(buf_t, void *), void *callback_arg);

-static void    sparse_cluster_push(void **cmapp, vnode_t vp, off_t EOF, int push_flag, 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 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);


@@ -193,32 +223,74 @@ static kern_return_t vfs_drt_get_cluster(void **cmapp, off_t *offsetp, u_int *le
 static kern_return_t vfs_drt_control(void **cmapp, int op_type);
 
 
 static kern_return_t vfs_drt_control(void **cmapp, int op_type);
 
 

+/*
+ * 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)
 /*
  * 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 * PAGE_SIZE)
+#define MAX_IO_CONTIG_SIZE     MAX_UPL_SIZE_BYTES

 #define MAX_VECTS              16
 #define MAX_VECTS              16

-#define MIN_DIRECT_WRITE_SIZE  (4 * PAGE_SIZE)
+/*
+ * 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 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, io_size)      (io_size * IO_SCALE(vp, 3))
+#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*/

 
 
 
 

-int speculative_reads_disabled = 0;
+#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 
  */
 
 /*
  * 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  (32 * 1024)
+#define THROTTLE_MAXCNT        0

 
 

-int hard_throttle_on_root = 0;
-struct timeval priority_IO_timestamp_for_root;
+uint32_t throttle_max_iosize = (128 * 1024);
+
+#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
 
 
 void


@@ -234,14 +306,15 @@ cluster_init(void) {
         */
        cl_mtx_attr = lck_attr_alloc_init();
 
         */
        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_mtxp == NULL)
-               panic("cluster_init: failed to allocate cl_mtxp");
+       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]);

 }
 
 
 }
 
 


@@ -267,19 +340,19 @@ cluster_max_io_size(mount_t mp, int type)
                maxcnt = min(mp->mnt_maxreadcnt, mp->mnt_maxwritecnt);
                break;
        }
                maxcnt = min(mp->mnt_maxreadcnt, mp->mnt_maxwritecnt);
                break;
        }

-       if (segcnt > MAX_UPL_SIZE) {
+       if (segcnt > (MAX_UPL_SIZE_BYTES >> PAGE_SHIFT)) {

               /*
                * don't allow a size beyond the max UPL size we can create
                */
               /*
                * don't allow a size beyond the max UPL size we can create
                */

-               segcnt = MAX_UPL_SIZE;
+               segcnt = MAX_UPL_SIZE_BYTES >> PAGE_SHIFT;

        }
        max_io_size = min((segcnt * PAGE_SIZE), maxcnt);
 
        }
        max_io_size = min((segcnt * PAGE_SIZE), maxcnt);
 

-       if (max_io_size < (MAX_UPL_TRANSFER * PAGE_SIZE)) {
+       if (max_io_size < MAX_UPL_TRANSFER_BYTES) {

               /*
                * don't allow a size smaller than the old fixed limit
                */
               /*
                * don't allow a size smaller than the old fixed limit
                */

-              max_io_size = (MAX_UPL_TRANSFER * PAGE_SIZE);
+              max_io_size = MAX_UPL_TRANSFER_BYTES;

        } else {
               /*
                * make sure the size specified is a multiple of PAGE_SIZE
        } else {
               /*
                * make sure the size specified is a multiple of PAGE_SIZE


@@ -397,7 +470,7 @@ cluster_get_wbp(vnode_t vp, int flags)
 
 
 static void
 
 
 static void

-cluster_syncup(vnode_t vp, off_t newEOF, int (*callback)(buf_t, void *), void *callback_arg)
+cluster_syncup(vnode_t vp, off_t newEOF, int (*callback)(buf_t, void *), void *callback_arg, int flags)

 {
        struct cl_writebehind *wbp;
 
 {
        struct cl_writebehind *wbp;
 


@@ -406,7 +479,7 @@ cluster_syncup(vnode_t vp, off_t newEOF, int (*callback)(buf_t, void *), void *c
                if (wbp->cl_number) {
                        lck_mtx_lock(&wbp->cl_lockw);
 
                if (wbp->cl_number) {
                        lck_mtx_lock(&wbp->cl_lockw);
 

-                       cluster_try_push(wbp, vp, newEOF, PUSH_ALL | PUSH_SYNC, callback, callback_arg);
+                       cluster_try_push(wbp, vp, newEOF, PUSH_ALL | flags, 0, callback, callback_arg, NULL);

 
                        lck_mtx_unlock(&wbp->cl_lockw);
                }
 
                        lck_mtx_unlock(&wbp->cl_lockw);
                }


@@ -414,44 +487,199 @@ cluster_syncup(vnode_t vp, off_t newEOF, int (*callback)(buf_t, void *), void *c
 }
 
 
 }
 
 

+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 (io_size == 0)
+                       return (0);
+
+               if (bootcache_check_fn(vp->v_mount->mnt_devvp->v_rdev, blkno))
+                       return(1);
+       }
+       return(0);
+}
+
+

 static int 
 static int 

-cluster_hard_throttle_on(vnode_t vp, uint32_t hard_throttle)
+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)

 {
 {

-       struct uthread  *ut;
+       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);

 
 

-       if (hard_throttle) {
-               static struct timeval hard_throttle_maxelapsed = { 0, 200000 };
+               upl_offset_t upl_end = upl_offset + size;
+               assert(upl_end >= PAGE_SIZE);

 
 

-               if (vp->v_mount->mnt_kern_flag & MNTK_ROOTDEV) {
-                       struct timeval elapsed;
+               upl_size_t assoc_upl_size = upl_get_size(associated_upl);

 
 

-                       if (hard_throttle_on_root)
-                               return(1);
+               /*
+                * 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);

 
 

-                       microuptime(&elapsed);
-                       timevalsub(&elapsed, &priority_IO_timestamp_for_root);
+               lck_mtx_lock_spin(&iostate->io_mtxp);

 
 

-                       if (timevalcmp(&elapsed, &hard_throttle_maxelapsed, <))
-                               return(1);
+               // 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;

                }
                }

-       }
-       if (throttle_get_io_policy(&ut) == IOPOL_THROTTLE) {
-               if (throttle_io_will_be_throttled(-1, vp->v_mount)) {
-                       return(1);
+
+               // 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));

        }
        }

-       return(0);
-}

 
 

+       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
 
 static int

-cluster_ioerror(upl_t upl, int upl_offset, int abort_size, int error, int io_flags)
+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;
 
 {
         int upl_abort_code = 0;
        int page_in  = 0;
        int page_out = 0;
 

-       if (io_flags & B_PHYS)
+       if ((io_flags & (B_PHYS | B_CACHE)) == (B_PHYS | B_CACHE))

                /*
                 * direct write of any flavor, or a direct read that wasn't aligned
                 */
                /*
                 * direct write of any flavor, or a direct read that wasn't aligned
                 */


@@ -468,7 +696,7 @@ cluster_ioerror(upl_t upl, int upl_offset, int abort_size, int error, int io_fla
                         * leave pages in the cache unchanged on error
                         */
                        upl_abort_code = UPL_ABORT_FREE_ON_EMPTY;
                         * leave pages in the cache unchanged on error
                         */
                        upl_abort_code = UPL_ABORT_FREE_ON_EMPTY;

-               else if (page_out && (error != ENXIO))
+               else if (page_out && ((error != ENXIO) || vnode_isswap(vp)))

                        /*
                         * transient error... leave pages unchanged
                         */
                        /*
                         * transient error... leave pages unchanged
                         */


@@ -502,40 +730,62 @@ cluster_iodone(buf_t bp, void *callback_arg)
        buf_t   cbp_head;
        buf_t   cbp_next;
        buf_t   real_bp;
        buf_t   cbp_head;
        buf_t   cbp_next;
        buf_t   real_bp;

+       vnode_t vp;

        struct  clios *iostate;
        boolean_t       transaction_complete = FALSE;
 
        struct  clios *iostate;
        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,
                     cbp_head, bp->b_lblkno, bp->b_bcount, bp->b_flags, 0);
 
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_START,
                     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);

 
 

-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END,
-                                    cbp_head, cbp, cbp->b_bcount, cbp->b_flags, 0);
+               bp->b_flags |= B_TDONE;

 
 

-                       return 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_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;
+                       }
+
+                       if (cbp->b_flags & B_EOT)
+                               transaction_complete = TRUE;

                }
                }

-               if (cbp->b_flags & B_EOT)
-                       transaction_complete = TRUE;
-       }
-       if (transaction_complete == FALSE) {
-               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END,
-                            cbp_head, 0, 0, 0, 0);
+               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;
        total_size  = 0;
        total_resid = 0;
 
        cbp        = cbp_head;
        }
        error       = 0;
        total_size  = 0;
        total_resid = 0;
 
        cbp        = cbp_head;

+       vp         = cbp->b_vp;

        upl_offset = cbp->b_uploffset;
        upl        = cbp->b_upl;
        b_flags    = cbp->b_flags;
        upl_offset = cbp->b_uploffset;
        upl        = cbp->b_upl;
        b_flags    = cbp->b_flags;


@@ -569,6 +819,14 @@ cluster_iodone(buf_t bp, void *callback_arg)
 
                cbp = cbp_next;
        }
 
                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 && total_resid)
                error = EIO;
 


@@ -593,7 +851,7 @@ cluster_iodone(buf_t bp, void *callback_arg)
                 * someone has issued multiple I/Os asynchrounsly
                 * and is waiting for them to complete (streaming)
                 */
                 * someone has issued multiple I/Os asynchrounsly
                 * and is waiting for them to complete (streaming)
                 */

-               lck_mtx_lock_spin(cl_mtxp);
+               lck_mtx_lock_spin(&iostate->io_mtxp);

 
                if (error && iostate->io_error == 0)
                        iostate->io_error = error;
 
                if (error && iostate->io_error == 0)
                        iostate->io_error = error;


@@ -608,21 +866,20 @@ cluster_iodone(buf_t bp, void *callback_arg)
                        iostate->io_wanted = 0;
                        need_wakeup = 1;
                }
                        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_COMMIT_UPL) {
 
                if (need_wakeup)
                        wakeup((caddr_t)&iostate->io_wanted);
        }
 
        if (b_flags & B_COMMIT_UPL) {

-
-               pg_offset   = upl_offset & PAGE_MASK;
+               pg_offset   = upl_offset & PAGE_MASK;

                commit_size = (pg_offset + transaction_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;
 
                if (error)
                commit_size = (pg_offset + transaction_size + (PAGE_SIZE - 1)) & ~PAGE_MASK;
 
                if (error)

-                       upl_flags = cluster_ioerror(upl, upl_offset - pg_offset, commit_size, error, b_flags);
+                       upl_flags = cluster_ioerror(upl, upl_offset - pg_offset, commit_size, error, b_flags, vp);

                else {
                else {

-                       upl_flags = UPL_COMMIT_FREE_ON_EMPTY;
+                       upl_flags = UPL_COMMIT_FREE_ON_EMPTY;

 
                        if ((b_flags & B_PHYS) && (b_flags & B_READ)) 
                                upl_flags |= UPL_COMMIT_SET_DIRTY;
 
                        if ((b_flags & B_PHYS) && (b_flags & B_READ)) 
                                upl_flags |= UPL_COMMIT_SET_DIRTY;


@@ -633,7 +890,7 @@ cluster_iodone(buf_t bp, void *callback_arg)
                        ubc_upl_commit_range(upl, upl_offset - pg_offset, commit_size, upl_flags);
                }
        }
                        ubc_upl_commit_range(upl, upl_offset - pg_offset, commit_size, upl_flags);
                }
        }

-       if ((b_flags & B_NEED_IODONE) && real_bp) {
+       if (real_bp) {

                if (error) {
                        real_bp->b_flags |= B_ERROR;
                        real_bp->b_error = error;
                if (error) {
                        real_bp->b_flags |= B_ERROR;
                        real_bp->b_error = error;


@@ -650,10 +907,10 @@ cluster_iodone(buf_t bp, void *callback_arg)
 
 
 uint32_t
 
 
 uint32_t

-cluster_hard_throttle_limit(vnode_t vp, uint32_t *limit, uint32_t hard_throttle)
+cluster_throttle_io_limit(vnode_t vp, uint32_t *limit)

 {
 {

-       if (cluster_hard_throttle_on(vp, hard_throttle)) {
-               *limit = HARD_THROTTLE_MAXSIZE;
+       if (cluster_is_throttled(vp)) {
+               *limit = THROTTLE_MAX_IOSIZE;

                return 1;
        }
        return 0;   
                return 1;
        }
        return 0;   


@@ -674,7 +931,7 @@ cluster_zero(upl_t upl, upl_offset_t upl_offset, int size, buf_t bp)
                pl = ubc_upl_pageinfo(upl);
 
                if (upl_device_page(pl) == TRUE) {
                pl = ubc_upl_pageinfo(upl);
 
                if (upl_device_page(pl) == TRUE) {

-                       zero_addr = ((addr64_t)upl_phys_page(pl, 0) << 12) + upl_offset;
+                       zero_addr = ((addr64_t)upl_phys_page(pl, 0) << PAGE_SHIFT) + upl_offset;

 
                        bzero_phys_nc(zero_addr, size);
                } else {
 
                        bzero_phys_nc(zero_addr, size);
                } else {


@@ -686,7 +943,7 @@ cluster_zero(upl_t upl, upl_offset_t upl_offset, int size, buf_t bp)
                                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_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);
                                zero_cnt  = min(PAGE_SIZE - page_offset, size);
 
                                bzero_phys(zero_addr, zero_cnt);


@@ -716,31 +973,53 @@ cluster_wait_IO(buf_t cbp_head, int async)
         buf_t  cbp;
 
        if (async) {
         buf_t  cbp;
 
        if (async) {

-               /*
-                * async callback completion will not normally
-                * generate a wakeup upon I/O completion...
-                * by setting BL_WANTED, we will force a wakeup
-                * to occur as any outstanding I/Os complete... 
-                * I/Os already completed will have BL_CALLDONE already
-                * set and we won't block in buf_biowait_callback..
-                * note that we're actually waiting for the bp to have
-                * completed the callback function... only then
-                * can we safely take back ownership of the bp
-                * need the main buf mutex in order to safely
-                * update b_lflags
+               /*
+                * 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.

                 */
                 */

-               buf_list_lock();
+               bool done = true;
+               buf_t last = NULL;

 
 

-               for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next)
-                     cbp->b_lflags |= BL_WANTED;
+               lck_mtx_lock_spin(cl_transaction_mtxp);

 
 

-               buf_list_unlock();
-       }
-       for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) {
-               if (async)
-                       buf_biowait_callback(cbp);
-               else
-                       buf_biowait(cbp);
+               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);

        }
 }
 
        }
 }
 


@@ -749,6 +1028,7 @@ cluster_complete_transaction(buf_t *cbp_head, void *callback_arg, int *retval, i
 {
         buf_t  cbp;
        int     error;
 {
         buf_t  cbp;
        int     error;

+       boolean_t isswapout = FALSE;

 
        /*
         * cluster_complete_transaction will
 
        /*
         * cluster_complete_transaction will


@@ -759,11 +1039,25 @@ cluster_complete_transaction(buf_t *cbp_head, void *callback_arg, int *retval, i
                for (cbp = *cbp_head; cbp; cbp = cbp->b_trans_next)
                        buf_biowait(cbp);
        }
                for (cbp = *cbp_head; cbp; cbp = cbp->b_trans_next)
                        buf_biowait(cbp);
        }

-       error = cluster_iodone(*cbp_head, callback_arg);
+       /*
+        * 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_ASYNC) && error && *retval == 0) {

-               if (((flags & (CL_PAGEOUT | CL_KEEPCACHED)) != CL_PAGEOUT) || (error != ENXIO))
-                       *retval = error;
+               if (((flags & (CL_PAGEOUT | CL_KEEPCACHED)) != CL_PAGEOUT) || (error != ENXIO))
+                       *retval = error;
+               else if (isswapout == TRUE)
+                       *retval = error;

        }
        *cbp_head = (buf_t)NULL;
 }
        }
        *cbp_head = (buf_t)NULL;
 }


@@ -876,21 +1170,30 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                max_iosize = PAGE_SIZE;
 
        if (flags & CL_THROTTLE) {
                max_iosize = PAGE_SIZE;
 
        if (flags & CL_THROTTLE) {

-               if ( !(flags & CL_PAGEOUT) && cluster_hard_throttle_on(vp, 1)) {
-                       if (max_iosize > HARD_THROTTLE_MAXSIZE)
-                               max_iosize = HARD_THROTTLE_MAXSIZE;
-                       async_throttle = HARD_THROTTLE_MAXCNT;
+               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;
                } 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 max_prefetch;
-                               
-                               max_cluster_size = MAX_CLUSTER_SIZE(vp);
-                               max_prefetch = MAX_PREFETCH(vp, cluster_max_io_size(vp->v_mount, CL_READ));
+                               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
                                if (max_iosize > max_cluster_size)
                                        max_cluster = max_cluster_size;
                                else


@@ -898,8 +1201,11 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
 
                                if (size < max_cluster)
                                        max_cluster = size;
 
                                if (size < max_cluster)
                                        max_cluster = size;

-
-                               async_throttle = min(IO_SCALE(vp, VNODE_ASYNC_THROTTLE), (max_prefetch / max_cluster) - 1);
+                               
+                               if (flags & CL_CLOSE)
+                                       scale += MAX_CLUSTERS;
+                               
+                               async_throttle = min(IO_SCALE(vp, VNODE_ASYNC_THROTTLE), ((scale * max_cluster_size) / max_cluster) - 1);

                        }
                }
        }
                        }
                }
        }


@@ -911,12 +1217,15 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                io_flags |= B_IOSTREAMING;
        if (flags & CL_COMMIT)
                io_flags |= B_COMMIT_UPL;
                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;
                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_PASSIVE)
                io_flags |= B_PASSIVE;

+       if (flags & CL_ENCRYPTED)
+               io_flags |= B_ENCRYPTED_IO;     
+

        if (vp->v_flag & VSYSTEM)
                io_flags |= B_META;
 
        if (vp->v_flag & VSYSTEM)
                io_flags |= B_META;
 


@@ -929,7 +1238,37 @@ 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;
                 * 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) {
                daddr64_t blkno;
                daddr64_t lblkno;
        while (size) {
                daddr64_t blkno;
                daddr64_t lblkno;


@@ -973,7 +1312,7 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                        off_t   e_offset;
                        int     pageout_flags;
 
                        off_t   e_offset;
                        int     pageout_flags;
 

-                       if(upl_get_internal_vectorupl(upl))
+                       if (upl_get_internal_vectorupl(upl))

                                panic("Vector UPLs should not take this code-path\n");
                        /*
                         * we're writing into a 'hole'
                                panic("Vector UPLs should not take this code-path\n");
                        /*
                         * we're writing into a 'hole'


@@ -1008,7 +1347,7 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                         *
                         * go direct to vnode_pageout so that we don't have to
                         * unbusy the page from the UPL... we used to do this
                         *
                         * 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_sync_range, but that results
+                        * 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
                         * 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


@@ -1021,47 +1360,69 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                                pageout_flags |= UPL_NOCOMMIT;
 
                        if (cbp_head) {
                                pageout_flags |= UPL_NOCOMMIT;
 
                        if (cbp_head) {

-                               buf_t last_cbp;
+                               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
 
                                /*
                                 * 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 just because all of the current
-                                * I/O linked to this transaction has completed...
+                                * so the pages won't be released

                                 */
                                cluster_wait_IO(cbp_head, (flags & CL_ASYNC));
 
                                 */
                                cluster_wait_IO(cbp_head, (flags & CL_ASYNC));
 

-                               /*
-                                * we've got a transcation that
-                                * includes the page we're about to push out through vnode_pageout...
-                                * find the last bp in the list which will be the one that
-                                * includes the head of this page and round it's iosize down
-                                * to a page boundary...
-                                */
-                                for (last_cbp = cbp = cbp_head; cbp->b_trans_next; cbp = cbp->b_trans_next)
-                                       last_cbp = cbp;
-
-                               cbp->b_bcount &= ~PAGE_MASK;
-
-                               if (cbp->b_bcount == 0) {
-                                       /*
-                                        * this buf no longer has any I/O associated with it
+                               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

                                         */
                                         */

-                                       free_io_buf(cbp);
+                                       for (prev_cbp = cbp = cbp_head; cbp->b_trans_next; cbp = cbp->b_trans_next)
+                                               prev_cbp = cbp;

 
 

-                                       if (cbp == cbp_head) {
-                                               /*
-                                                * the buf we just freed was the only buf in
-                                                * this transaction... so there's no I/O to do
+                                       if (bytes_in_last_page >= cbp->b_bcount) {
+                                               /*
+                                                * this buf no longer has any I/O associated with it

                                                 */
                                                 */

-                                               cbp_head = NULL;
+                                               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 {
                                        } else {

-                                               /*
-                                                * remove the buf we just freed from
-                                                * the transaction list
+                                               /*
+                                                * 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

                                                 */
                                                 */

-                                               last_cbp->b_trans_next = NULL;
-                                               cbp_tail = last_cbp;
+                                               cbp->b_bcount -= bytes_in_last_page;
+                                               cbp_tail = cbp;
+                                               bytes_in_last_page = 0;

                                        }
                                }
                                if (cbp_head) {
                                        }
                                }
                                if (cbp_head) {


@@ -1080,7 +1441,6 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                        }
                        if (vnode_pageout(vp, upl, trunc_page(upl_offset), trunc_page_64(f_offset), PAGE_SIZE, pageout_flags, NULL) != PAGER_SUCCESS) {
                                error = EINVAL;
                        }
                        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;
                        }
                        e_offset = round_page_64(f_offset + 1);
                        io_size = e_offset - f_offset;


@@ -1109,9 +1469,14 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                                 */
                                size = 0;
                        }
                                 */
                                size = 0;
                        }

+                       if (error) {
+                               if (size == 0)
+                                       flags &= ~CL_COMMIT;
+                               break;
+                       }

                        continue;
                }
                        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
                /*
                 * 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


@@ -1206,6 +1571,10 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
 
                                commit_offset = upl_offset & ~PAGE_MASK;
                        }
 
                                commit_offset = upl_offset & ~PAGE_MASK;
                        }

+
+                       // Associated UPL is currently only used in the direct write path
+                       assert(!upl_associated_upl(upl));
+

                        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);
                        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);


@@ -1302,9 +1671,13 @@ 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;
 
                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 (flags & CL_ASYNC) {


@@ -1313,6 +1686,8 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                }
                cbp->b_cliodone = (void *)callback;
                cbp->b_flags |= io_flags;
                }
                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;
 
                cbp->b_lblkno = lblkno;
                cbp->b_blkno  = blkno;


@@ -1320,7 +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 (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))
                cbp->b_trans_next = (buf_t)NULL;
 
                if ((cbp->b_iostate = (void *)iostate))


@@ -1346,10 +1723,8 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                        cbp_head = cbp;
                        cbp_tail = cbp;
 
                        cbp_head = cbp;
                        cbp_tail = cbp;
 

-                       if ( (cbp_head->b_real_bp = real_bp) ) {
-                               cbp_head->b_flags |= B_NEED_IODONE;
+                       if ( (cbp_head->b_real_bp = real_bp) )

                                real_bp = (buf_t)NULL;
                                real_bp = (buf_t)NULL;

-                       }

                }
                *(buf_t *)(&cbp->b_trans_head) = cbp_head;
 
                }
                *(buf_t *)(&cbp->b_trans_head) = cbp_head;
 


@@ -1407,6 +1782,14 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                if ( !(io_flags & B_READ))
                        vnode_startwrite(vp);
                                
                if ( !(io_flags & B_READ))
                        vnode_startwrite(vp);
                                

+               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) {
                (void) VNOP_STRATEGY(cbp);
 
                if (need_EOT == TRUE) {


@@ -1419,34 +1802,41 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                }
         }
        if (error) {
                }
         }
        if (error) {

-               int abort_size;
+               int abort_size;

 
                io_size = 0;
 
                io_size = 0;

-               
+

                if (cbp_head) {
                if (cbp_head) {

-                        /*
-                         * first wait until all of the outstanding I/O
-                         * for this partial transaction has completed
-                         */
-                       cluster_wait_IO(cbp_head, (flags & CL_ASYNC));
+                       /*
+                        * 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;
 
                        /*
                         * Rewind the upl offset to the beginning of the
                         * transaction.
                         */
                        upl_offset = cbp_head->b_uploffset;

+               }

 
 

-                       for (cbp = cbp_head; cbp;) {
-                               buf_t   cbp_next;
-        
-                               size       += cbp->b_bcount;
-                               io_size    += cbp->b_bcount;
+               if (ISSET(flags, CL_COMMIT)) {
+                       cluster_handle_associated_upl(iostate, upl, upl_offset,
+                                                                                 upl_end_offset - upl_offset);
+               }

 
 

-                               cbp_next = cbp->b_trans_next;
-                               free_io_buf(cbp);
-                               cbp = cbp_next;
-                       }
+               // Free all the IO buffers in this transaction
+               for (cbp = cbp_head; cbp;) {
+                       buf_t   cbp_next;
+ 
+                       size       += cbp->b_bcount;
+                       io_size    += cbp->b_bcount;
+
+                       cbp_next = cbp->b_trans_next;
+                       free_io_buf(cbp);
+                       cbp = cbp_next;

                }
                }

+

                if (iostate) {
                        int need_wakeup = 0;
 
                if (iostate) {
                        int need_wakeup = 0;
 


@@ -1455,7 +1845,7 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                         * since we never really issued the io
                         * just go ahead and adjust it back
                         */
                         * since we never really issued the io
                         * just go ahead and adjust it back
                         */

-                       lck_mtx_lock_spin(cl_mtxp);
+                       lck_mtx_lock_spin(&iostate->io_mtxp);

 
                        if (iostate->io_error == 0)
                                iostate->io_error = error;
 
                        if (iostate->io_error == 0)
                                iostate->io_error = error;


@@ -1469,18 +1859,19 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                                iostate->io_wanted = 0;
                                need_wakeup = 1;
                        }
                                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 (need_wakeup)
                                wakeup((caddr_t)&iostate->io_wanted);
                }

+

                if (flags & CL_COMMIT) {
                        int     upl_flags;
 
                if (flags & CL_COMMIT) {
                        int     upl_flags;
 

-                       pg_offset  = upl_offset & PAGE_MASK;
+                       pg_offset  = upl_offset & PAGE_MASK;

                        abort_size = (upl_end_offset - upl_offset + PAGE_MASK) & ~PAGE_MASK;
                        abort_size = (upl_end_offset - upl_offset + PAGE_MASK) & ~PAGE_MASK;

-                       
-                       upl_flags = cluster_ioerror(upl, upl_offset - pg_offset, abort_size, error, io_flags);
+
+                       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,
                                     upl, upl_offset - pg_offset, abort_size, (error << 24) | upl_flags, 0);
                        
                        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 28)) | DBG_FUNC_NONE,
                                     upl, upl_offset - pg_offset, abort_size, (error << 24) | upl_flags, 0);


@@ -1580,10 +1971,18 @@ cluster_read_ahead(vnode_t vp, struct cl_extent *extent, off_t filesize, struct
 
                return;
        }
 
                return;
        }

-       max_prefetch = MAX_PREFETCH(vp, cluster_max_io_size(vp->v_mount, CL_READ));
+       max_prefetch = MAX_PREFETCH(vp, cluster_max_io_size(vp->v_mount, CL_READ), (vp->v_mount->mnt_kern_flag & MNTK_SSD));

 
 

-       if (extent->e_addr < rap->cl_maxra) {
-               if ((rap->cl_maxra - extent->e_addr) > ((max_prefetch / PAGE_SIZE) / 4)) {
+       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);
 
                        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 48)) | DBG_FUNC_END,
                                     rap->cl_ralen, (int)rap->cl_maxra, (int)rap->cl_lastr, 2, 0);


@@ -1643,18 +2042,7 @@ cluster_pageout_ext(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offs
         off_t         max_size;
        int           local_flags;
 
         off_t         max_size;
        int           local_flags;
 

-       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;
 
        if ((flags & UPL_IOSYNC) == 0) 
                local_flags |= CL_ASYNC;


@@ -1662,6 +2050,8 @@ cluster_pageout_ext(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offs
                local_flags |= CL_COMMIT;
        if ((flags & UPL_KEEPCACHED))
                local_flags |= CL_KEEPCACHED;
                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,
 
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 52)) | DBG_FUNC_NONE,


@@ -1738,6 +2128,8 @@ cluster_pagein_ext(vnode_t vp, upl_t upl, upl_offset_t upl_offset, off_t f_offse
                local_flags |= CL_COMMIT;
        if (flags & UPL_IOSTREAMING)
                local_flags |= CL_IOSTREAMING;
                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,
 
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 56)) | DBG_FUNC_NONE,


@@ -1831,9 +2223,10 @@ cluster_write_ext(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t
        else
                bflag = 0;
 
        else
                bflag = 0;
 

-       if (vp->v_flag & VNOCACHE_DATA)
+       if (vp->v_flag & VNOCACHE_DATA){

                flags |= IO_NOCACHE;
                flags |= IO_NOCACHE;

-
+               bflag |= CL_NOCACHE;
+       }

         if (uio == NULL) {
                /*
                 * no user data...
         if (uio == NULL) {
                /*
                 * no user data...


@@ -1845,12 +2238,12 @@ cluster_write_ext(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, off_t
        }
         /*
          * do a write through the cache if one of the following is true....
        }
         /*
          * do a write through the cache if one of the following is true....

-         *   NOCACHE is not true and
+         *   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
          */
          *   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) && UIO_SEG_IS_USER_SPACE(uio->uio_segflg) )
+       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)
                retval = cluster_io_type(uio, &write_type, &write_length, MIN_DIRECT_WRITE_SIZE);
 
         if ( (flags & (IO_TAILZEROFILL | IO_HEADZEROFILL)) && write_type == IO_DIRECT)


@@ -1965,7 +2358,7 @@ cluster_write_direct(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, in
        upl_size_t       upl_size, vector_upl_size = 0;
        vm_size_t        upl_needed_size;
        mach_msg_type_number_t  pages_in_pl;
        upl_size_t       upl_size, vector_upl_size = 0;
        vm_size_t        upl_needed_size;
        mach_msg_type_number_t  pages_in_pl;

-       int              upl_flags;
+       upl_control_flags_t upl_flags;

        kern_return_t    kret;
        mach_msg_type_number_t  i;
        int              force_data_sync;
        kern_return_t    kret;
        mach_msg_type_number_t  i;
        int              force_data_sync;


@@ -1975,7 +2368,11 @@ cluster_write_direct(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, in
        user_addr_t      iov_base;
        u_int32_t        mem_alignment_mask;
        u_int32_t        devblocksize;
        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_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);
 
        u_int32_t        vector_upl_iosize = 0;
        int              issueVectorUPL = 0,useVectorUPL = (uio->uio_iovcnt > 1);


@@ -1997,12 +2394,20 @@ cluster_write_direct(vnode_t vp, struct uio *uio, off_t oldEOF, off_t newEOF, in
 
        if (flags & IO_PASSIVE)
                io_flag |= CL_PASSIVE;
 
        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;

 
        iostate.io_completed = 0;
        iostate.io_issued = 0;
        iostate.io_error = 0;
        iostate.io_wanted = 0;
 
 
        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);
+

        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;
        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
 
        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;
        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
 


@@ -2043,17 +2448,45 @@ next_dwrite:
                goto wait_for_dwrites;
         }
 
                goto wait_for_dwrites;
         }
 

+       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) {
        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);
+
+                               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) {
 
                if (first_IO) {

-                       cluster_syncup(vp, newEOF, callback, callback_arg);
+                       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);
 
                        first_IO = 0;
                }
                io_size  = io_req_size & ~PAGE_MASK;
                iov_base = uio_curriovbase(uio);
 

-               if (io_size > max_upl_size)
-                       io_size = max_upl_size;
+               if (io_size > max_io_size)
+                       io_size = max_io_size;

 
                if(useVectorUPL && (iov_base & PAGE_MASK)) {
                        /*
 
                if(useVectorUPL && (iov_base & PAGE_MASK)) {
                        /*


@@ -2078,13 +2511,15 @@ next_dwrite:
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 76)) | DBG_FUNC_START,
                             (int)upl_offset, upl_needed_size, (int)iov_base, io_size, 0);
 
                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 |
                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);

 
 

-                       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, 
                                              (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)),
                                              &upl_size,
                                              &upl, 


@@ -2169,37 +2604,19 @@ next_dwrite:
                         */
                }
 
                         */
                }
 

-               /*
-                * 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
                 * the preparation of the next I/O
                 * if there are already too many outstanding writes
                 * wait until some complete before issuing the next
                 */
                /*
                 * we want push out these writes asynchronously so that we can overlap
                 * the preparation of the next I/O
                 * if there are already too many outstanding writes
                 * wait until some complete before issuing the next
                 */

-               if (iostate.io_issued > iostate.io_completed) {
-
-                       lck_mtx_lock(cl_mtxp);
-
-                       while ((iostate.io_issued - iostate.io_completed) > (max_upl_size * IO_SCALE(vp, 2))) {
-
-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                            iostate.io_issued, iostate.io_completed, max_upl_size * IO_SCALE(vp, 2), 0, 0);
+               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);

 
 

-                               iostate.io_wanted = 1;
-                               msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_write_direct", NULL);
+               cluster_iostate_wait(&iostate, bytes_outstanding_limit, "cluster_write_direct");

 
 

-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                            iostate.io_issued, iostate.io_completed, max_upl_size * IO_SCALE(vp, 2), 0, 0);
-                       }       
-                       lck_mtx_unlock(cl_mtxp);
-               }

                if (iostate.io_error) {
                        /*
                         * one of the earlier writes we issued ran into a hard error
                if (iostate.io_error) {
                        /*
                         * one of the earlier writes we issued ran into a hard error


@@ -2233,7 +2650,7 @@ next_dwrite:
                        vector_upl_iosize += io_size;
                        vector_upl_size += upl_size;
 
                        vector_upl_iosize += io_size;
                        vector_upl_size += upl_size;
 

-                       if(issueVectorUPL || vector_upl_index ==  MAX_VECTOR_UPL_ELEMENTS || vector_upl_size >= MAX_VECTOR_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();
                        }
                                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();
                        }


@@ -2279,33 +2696,24 @@ next_dwrite:
 
 wait_for_dwrites:
 
 
 wait_for_dwrites:
 

-       if(retval == 0 && iostate.io_error == 0 && useVectorUPL && vector_upl_index) {
+       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();       
        }
                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
+        */
+       cluster_iostate_wait(&iostate, 0, "cluster_write_direct");

 
 

-       if (iostate.io_issued > iostate.io_completed) {
-               /*
-                * make sure all async writes issued as part of this stream
-                * have completed before we return
-                */
-               lck_mtx_lock(cl_mtxp);
-
-               while (iostate.io_issued != iostate.io_completed) {
-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-
-                       iostate.io_wanted = 1;
-                       msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_write_direct", NULL);
-
-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-               }       
-               lck_mtx_unlock(cl_mtxp);
-       }

        if (iostate.io_error)
                retval = iostate.io_error;
 
        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
        if (io_req_size && retval == 0) {
                /*
                 * we couldn't handle the tail of this request in DIRECT mode


@@ -2342,7 +2750,7 @@ cluster_write_contig(vnode_t vp, struct uio *uio, off_t newEOF, int *write_type,
        upl_size_t       upl_size;
        vm_size_t        upl_needed_size;
        mach_msg_type_number_t  pages_in_pl;
        upl_size_t       upl_size;
        vm_size_t        upl_needed_size;
        mach_msg_type_number_t  pages_in_pl;

-       int              upl_flags;
+       upl_control_flags_t upl_flags;

        kern_return_t    kret;
         struct clios     iostate;
        int              error  = 0;
        kern_return_t    kret;
         struct clios     iostate;
        int              error  = 0;


@@ -2358,7 +2766,7 @@ cluster_write_contig(vnode_t vp, struct uio *uio, off_t newEOF, int *write_type,
         *  -- the io_req_size will not exceed iov_len
         *  -- the target address is physically contiguous
         */
         *  -- the io_req_size will not exceed iov_len
         *  -- the target address is physically contiguous
         */

-       cluster_syncup(vp, newEOF, callback, callback_arg);
+       cluster_syncup(vp, newEOF, callback, callback_arg, callback ? PUSH_SYNC : 0);

 
        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;
 
        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;


@@ -2368,6 +2776,8 @@ cluster_write_contig(vnode_t vp, struct uio *uio, off_t newEOF, int *write_type,
         iostate.io_error = 0;
         iostate.io_wanted = 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;
 
 next_cwrite:
        io_size = *write_length;
 


@@ -2379,9 +2789,11 @@ next_cwrite:
        pages_in_pl = 0;
        upl_size = upl_needed_size;
        upl_flags = UPL_FILE_IO | UPL_COPYOUT_FROM | UPL_NO_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);

 
 

-       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[cur_upl], NULL, &pages_in_pl, &upl_flags, 0);
 
                              (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)),
                              &upl_size, &upl[cur_upl], NULL, &pages_in_pl, &upl_flags, 0);
 


@@ -2406,7 +2818,7 @@ next_cwrite:
        }
        pl = ubc_upl_pageinfo(upl[cur_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) {
                u_int32_t   head_size;
 
        while (((uio->uio_offset & (devblocksize - 1)) || io_size < devblocksize) && io_size) {
                u_int32_t   head_size;


@@ -2456,22 +2868,8 @@ next_cwrite:
                 * if there are already too many outstanding writes
                 * wait until some have completed before issuing the next
                 */
                 * if there are already too many outstanding writes
                 * wait until some have completed before issuing the next
                 */

-               if (iostate.io_issued > iostate.io_completed) {
-                       lck_mtx_lock(cl_mtxp);
-
-                       while ((iostate.io_issued - iostate.io_completed) > (MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2))) {
-
-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                            iostate.io_issued, iostate.io_completed, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), 0, 0);
-
-                               iostate.io_wanted = 1;
-                               msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_write_contig", NULL);
+               cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_write_contig");

 
 

-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                            iostate.io_issued, iostate.io_completed, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), 0, 0);
-                       }
-                       lck_mtx_unlock(cl_mtxp);
-               }

                 if (iostate.io_error) {
                         /*
                          * one of the earlier writes we issued ran into a hard error
                 if (iostate.io_error) {
                         /*
                          * one of the earlier writes we issued ran into a hard error


@@ -2515,25 +2913,13 @@ wait_for_cwrites:
          * make sure all async writes that are part of this stream
          * have completed before we proceed
          */
          * make sure all async writes that are part of this stream
          * have completed before we proceed
          */

-       if (iostate.io_issued > iostate.io_completed) {
-               
-               lck_mtx_lock(cl_mtxp);
-
-               while (iostate.io_issued != iostate.io_completed) {
-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
+       cluster_iostate_wait(&iostate, 0, "cluster_write_contig");

 
 

-                       iostate.io_wanted = 1;
-                       msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_write_contig", NULL);
-
-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-               }
-               lck_mtx_unlock(cl_mtxp);
-       }

         if (iostate.io_error)
                error = iostate.io_error;
 
         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);
 
        if (error == 0 && tail_size)
                error = cluster_align_phys_io(vp, uio, src_paddr, tail_size, 0, callback, callback_arg);
 


@@ -2608,6 +2994,9 @@ cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t old
        off_t            zero_off;
        long long        zero_cnt1;
        off_t            zero_off1;
        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;
        struct cl_writebehind *wbp;
        int              bflag;
        struct cl_extent cl;
        struct cl_writebehind *wbp;
        int              bflag;


@@ -2629,6 +3018,11 @@ cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t old
                bflag = CL_PASSIVE;
        else
                bflag = 0;
                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_cnt  = 0;
        zero_cnt1 = 0;


@@ -2689,7 +3083,16 @@ cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t old
                             retval, 0, 0, 0, 0);
                return (0);
        }
                             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
        while ((total_size = (io_resid + zero_cnt + zero_cnt1)) && retval == 0) {
                /*
                 * for this iteration of the loop, figure out where our starting point is


@@ -2718,7 +3121,7 @@ cluster_write_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t old
                         * because IO_HEADZEROFILL and IO_TAILZEROFILL not set
                         */
                        if ((start_offset + total_size) > max_io_size)
                         * because IO_HEADZEROFILL and IO_TAILZEROFILL not set
                         */
                        if ((start_offset + total_size) > max_io_size)

-                               total_size -= start_offset;
+                               total_size = max_io_size - start_offset;

                        xfer_resid = total_size;
 
                        retval = cluster_copy_ubc_data_internal(vp, uio, &xfer_resid, 1, 1);
                        xfer_resid = total_size;
 
                        retval = cluster_copy_ubc_data_internal(vp, uio, &xfer_resid, 1, 1);


@@ -2984,7 +3387,7 @@ check_cluster:
                                 */
                                wbp->cl_number = 0;
 
                                 */
                                wbp->cl_number = 0;
 

-                               sparse_cluster_push(&(wbp->cl_scmap), vp, newEOF, PUSH_ALL, callback, callback_arg);
+                               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
                                /*
                                 * no clusters of either type present at this point
                                 * so just go directly to start_new_cluster since


@@ -2993,7 +3396,17 @@ check_cluster:
                                 * to avoid the deadlock with sparse_cluster_push
                                 */
                                goto start_new_cluster;
                                 * to avoid the deadlock with sparse_cluster_push
                                 */
                                goto start_new_cluster;

-                       }                   
+                       }
+                       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
                        if (wbp->cl_number == 0)
                                /*
                                 * no clusters currently present


@@ -3108,14 +3521,36 @@ check_cluster:
                                 */
                                goto delay_io;
 
                                 */
                                goto delay_io;
 

-                       if (wbp->cl_number < MAX_CLUSTERS)
+                       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;
                                /*
                                 * 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 
                        /*
                         * no exisitng cluster to merge with and no
                         * room to start a new one... we'll try 


@@ -3133,7 +3568,7 @@ check_cluster:
                         */
                        if (!((unsigned int)vfs_flags(vp->v_mount) & MNT_DEFWRITE)) {
                                
                         */
                        if (!((unsigned int)vfs_flags(vp->v_mount) & MNT_DEFWRITE)) {
                                

-                               ret_cluster_try_push = cluster_try_push(wbp, vp, newEOF, (flags & IO_NOCACHE) ? 0 : PUSH_DELAY, callback, callback_arg);
+                               ret_cluster_try_push = cluster_try_push(wbp, vp, newEOF, (flags & IO_NOCACHE) ? 0 : PUSH_DELAY, 0, callback, callback_arg, NULL);

                        }
 
                        /*
                        }
 
                        /*


@@ -3152,18 +3587,6 @@ check_cluster:
 
                                continue;
                        }
 
                                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...
-                        * 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
-                        */
-                       if (!((unsigned int)vfs_flags(vp->v_mount) & MNT_DEFWRITE)) {
-                               while (wbp->cl_number)
-                                       cluster_try_push(wbp, vp, newEOF, 0, callback, callback_arg);
-                       }
-

 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;
 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;


@@ -3225,15 +3648,24 @@ cluster_read_ext(vnode_t vp, struct uio *uio, off_t filesize, int xflags, int (*
        if ((vp->v_flag & VRAOFF) || speculative_reads_disabled)
                flags |= IO_RAOFF;
 
        if ((vp->v_flag & VRAOFF) || speculative_reads_disabled)
                flags |= IO_RAOFF;
 

-        /*
+       if (flags & IO_SKIP_ENCRYPTION)
+               flags |= IO_ENCRYPTED;
+
+       /*

         * do a read through the cache if one of the following is true....
         *   NOCACHE is not true
         *   the uio request doesn't target USERSPACE
         * 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
         */
         * 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) )
-               retval = cluster_io_type(uio, &read_type, &read_length, 0);
+       if ( ((flags & IO_NOCACHE) && UIO_SEG_IS_USER_SPACE(uio->uio_segflg)) || (flags & IO_ENCRYPTED) ) {
+
+               retval = cluster_io_type(uio, &read_type, &read_length, 0);
+       }

 
        while ((cur_resid = uio_resid(uio)) && uio->uio_offset < filesize && retval == 0) {
 
 
        while ((cur_resid = uio_resid(uio)) && uio->uio_offset < filesize && retval == 0) {
 


@@ -3318,21 +3750,19 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
        struct cl_extent        extent;
        int              bflag;
        int              take_reference = 1;
        struct cl_extent        extent;
        int              bflag;
        int              take_reference = 1;

-       struct uthread  *ut;

        int              policy = IOPOL_DEFAULT;
        int              policy = IOPOL_DEFAULT;

-
+       boolean_t        iolock_inited = FALSE;

 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_START,
                     (int)uio->uio_offset, io_req_size, (int)filesize, flags, 0);
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_START,
                     (int)uio->uio_offset, io_req_size, (int)filesize, flags, 0);

+       
+       if (flags & IO_ENCRYPTED) {
+               panic ("encrypted blocks will hit UBC!");
+       }

                         
                         

-       policy = current_proc()->p_iopol_disk;
+       policy = throttle_get_io_policy(NULL);

 
 

-       ut = get_bsdthread_info(current_thread());
-
-       if (ut->uu_iopol_disk != IOPOL_DEFAULT)
-               policy = ut->uu_iopol_disk;
-
-       if (policy == IOPOL_THROTTLE || (flags & IO_NOCACHE))
+       if (policy == THROTTLE_LEVEL_TIER3 || policy == THROTTLE_LEVEL_TIER2 || (flags & IO_NOCACHE))

                take_reference = 0;
 
        if (flags & IO_PASSIVE)
                take_reference = 0;
 
        if (flags & IO_PASSIVE)


@@ -3340,8 +3770,14 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
        else
                bflag = 0;
 
        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_io_size = cluster_max_io_size(vp->v_mount, CL_READ);

-       max_prefetch = MAX_PREFETCH(vp, max_io_size);
+       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;
        max_rd_size = max_prefetch;
 
        last_request_offset = uio->uio_offset + io_req_size;


@@ -3353,14 +3789,16 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                rd_ahead_enabled = 0;
                rap = NULL;
        } else {
                rd_ahead_enabled = 0;
                rap = NULL;
        } else {

-               if (cluster_hard_throttle_on(vp, 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
+                        */

                        rd_ahead_enabled = 0;
                        prefetch_enabled = 0;
 
                        rd_ahead_enabled = 0;
                        prefetch_enabled = 0;
 

-                       max_rd_size = HARD_THROTTLE_MAXSIZE;
-               } else if (policy == IOPOL_THROTTLE) {
-                       rd_ahead_enabled = 0;
-                       prefetch_enabled = 0;
+                       max_rd_size = THROTTLE_MAX_IOSIZE;

                }
                if ((rap = cluster_get_rap(vp)) == NULL)
                        rd_ahead_enabled = 0;
                }
                if ((rap = cluster_get_rap(vp)) == NULL)
                        rd_ahead_enabled = 0;


@@ -3440,7 +3878,7 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
 
                                io_requested = io_resid;
 
 
                                io_requested = io_resid;
 

-                               retval = cluster_copy_ubc_data_internal(vp, uio, (int *)&io_resid, 0, last_ioread_offset == 0 ? take_reference : 0);
+                               retval = cluster_copy_ubc_data_internal(vp, uio, (int *)&io_resid, 0, take_reference);

 
                                xsize = io_requested - io_resid;
 
 
                                xsize = io_requested - io_resid;
 


@@ -3479,6 +3917,30 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                         */
                        max_size = filesize - 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
                /*
                 * compute the size of the upl needed to encompass
                 * the requested read... limit each call to cluster_io


@@ -3499,8 +3961,13 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                        if (upl_size > max_io_size)
                                upl_size = max_io_size;
                } else {
                        if (upl_size > max_io_size)
                                upl_size = max_io_size;
                } else {

-                       if (upl_size > max_io_size / 4)
+                       if (upl_size > max_io_size / 4) {

                                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;
 
                }
                pages_in_upl = upl_size / PAGE_SIZE;
 


@@ -3540,10 +4007,6 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                        if (upl_valid_page(pl, last_pg))
                                break;
                }
                        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) {               
                        /*
 
                if (start_pg < last_pg) {               
                        /*


@@ -3552,6 +4015,11 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                         * 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
                         */
                         * 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;
 
                        upl_offset = start_pg * PAGE_SIZE;
                        io_size    = (last_pg - start_pg) * PAGE_SIZE;
 


@@ -3564,6 +4032,18 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
 
                        error = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset,
                                           io_size, CL_READ | CL_ASYNC | bflag, (buf_t)NULL, &iostate, callback, callback_arg);
 
                        error = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset,
                                           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) {
                        /*
                }
                if (error == 0) {
                        /*


@@ -3642,22 +4122,9 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                                        rap->cl_lastr = extent.e_addr;
                                }
                        }
                                        rap->cl_lastr = extent.e_addr;
                                }
                        }

-                       if (iostate.io_issued > iostate.io_completed) {
-
-                               lck_mtx_lock(cl_mtxp);
+                       if (iolock_inited == TRUE)
+                               cluster_iostate_wait(&iostate, 0, "cluster_read_copy");

 
 

-                               while (iostate.io_issued != iostate.io_completed) {
-                                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-
-                                       iostate.io_wanted = 1;
-                                       msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_read_copy", NULL);
-
-                                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-                               }       
-                               lck_mtx_unlock(cl_mtxp);
-                       }

                        if (iostate.io_error)
                                error = iostate.io_error;
                        else {
                        if (iostate.io_error)
                                error = iostate.io_error;
                        else {


@@ -3669,6 +4136,9 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                                
                                io_req_size -= (val_size - 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) {
                        /*
                }
                if (start_pg < last_pg) {
                        /*


@@ -3728,17 +4198,21 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                        retval = error;
 
                if (io_req_size) {
                        retval = error;
 
                if (io_req_size) {

-                       if (cluster_hard_throttle_on(vp, 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
+                                */

                                rd_ahead_enabled = 0;
                                prefetch_enabled = 0;
                                rd_ahead_enabled = 0;
                                prefetch_enabled = 0;

-
-                               max_rd_size = HARD_THROTTLE_MAXSIZE;
+                               max_rd_size = THROTTLE_MAX_IOSIZE;

                        } else {
                        } else {

-                               if (max_rd_size == HARD_THROTTLE_MAXSIZE) {
+                               if (max_rd_size == THROTTLE_MAX_IOSIZE) {

                                        /*
                                         * coming out of throttled state
                                         */
                                        /*
                                         * coming out of throttled state
                                         */

-                                       if (policy != IOPOL_THROTTLE) {
+                                       if (policy != THROTTLE_LEVEL_TIER3 && policy != THROTTLE_LEVEL_TIER2) {

                                                if (rap != NULL)
                                                        rd_ahead_enabled = 1;
                                                prefetch_enabled = 1;
                                                if (rap != NULL)
                                                        rd_ahead_enabled = 1;
                                                prefetch_enabled = 1;


@@ -3749,6 +4223,19 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                        }
                }
        }
                        }
                }
        }

+       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, io_req_size, rap->cl_lastr, retval, 0);
        if (rap != NULL) {
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 32)) | DBG_FUNC_END,
                             (int)uio->uio_offset, io_req_size, rap->cl_lastr, retval, 0);


@@ -3762,6 +4249,72 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
        return (retval);
 }
 
        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_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length,
 
 static int
 cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type, u_int32_t *read_length,


@@ -3774,13 +4327,12 @@ cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
        upl_size_t       upl_size, vector_upl_size = 0;
        vm_size_t        upl_needed_size;
        unsigned int     pages_in_pl;
        upl_size_t       upl_size, vector_upl_size = 0;
        vm_size_t        upl_needed_size;
        unsigned int     pages_in_pl;

-       int              upl_flags;
+       upl_control_flags_t upl_flags;

        kern_return_t    kret;
        unsigned int     i;
        int              force_data_sync;
        int              retval = 0;
        int              no_zero_fill = 0;
        kern_return_t    kret;
        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;
        int              io_flag = 0;
        int              misaligned = 0;
        struct clios     iostate;


@@ -3796,12 +4348,20 @@ cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
        u_int32_t        max_upl_size;
        u_int32_t        max_rd_size;
        u_int32_t        max_rd_ahead;
        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;
 
        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, (int)filesize, *read_type, *read_length, 0);
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 70)) | DBG_FUNC_START,
                     (int)uio->uio_offset, (int)filesize, *read_type, *read_length, 0);


@@ -3812,14 +4372,28 @@ cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
        max_rd_ahead = max_rd_size * IO_SCALE(vp, 2);
 
        io_flag = CL_COMMIT | CL_READ | CL_ASYNC | CL_NOZERO | CL_DIRECT_IO;
        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_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;
 
        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);
+

        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;
 
        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;
 


@@ -3839,15 +4413,16 @@ cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
                 */
                devblocksize = PAGE_SIZE;
        }
                 */
                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);
 
 next_dread:
        io_req_size = *read_length;
        iov_base = uio_curriovbase(uio);
 

-        max_io_size = filesize - uio->uio_offset;
-
-       if ((off_t)io_req_size > max_io_size)
-               io_req_size = max_io_size;
-

        offset_in_file = (u_int32_t)uio->uio_offset & (devblocksize - 1);
        offset_in_iovbase = (u_int32_t)iov_base & mem_alignment_mask;
 
        offset_in_file = (u_int32_t)uio->uio_offset & (devblocksize - 1);
        offset_in_iovbase = (u_int32_t)iov_base & mem_alignment_mask;
 


@@ -3865,7 +4440,25 @@ next_dread:
                 * I/O that ends on a page boundary in cluster_io
                 */
                misaligned = 1;
                 * 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
        /*
         * When we get to this point, we know...
         *  -- the offset into the file is on a devblocksize boundary


@@ -3874,24 +4467,35 @@ next_dread:
        while (io_req_size && retval == 0) {
                u_int32_t io_start;
 
        while (io_req_size && retval == 0) {
                u_int32_t io_start;
 

-               if (cluster_hard_throttle_on(vp, 1)) {
-                       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_size;
                        max_rd_ahead = max_rd_size * IO_SCALE(vp, 2);
                } else {
                        max_rd_size  = max_upl_size;
                        max_rd_ahead = max_rd_size * IO_SCALE(vp, 2);

+                       max_vector_size = MAX_VECTOR_UPL_SIZE;

                }
                io_start = io_size = io_req_size;
 
                /*
                 * First look for pages already in the cache
                }
                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
                 */
                 *
                 * cluster_copy_ubc_data returns the resid
                 * in io_size
                 */

-               retval = cluster_copy_ubc_data_internal(vp, uio, (int *)&io_size, 0, 0);
-                       
+               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
                /*
                 * calculate the number of bytes actually copied
                 * starting size - residual


@@ -3922,9 +4526,14 @@ next_dread:
                }
 
                /*
                }
 
                /*

-                * check to see if we are finished with this request...
+                * 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_req_size == 0 || misaligned) {
+               if (io_req_size == 0 || (misaligned)) {

                        /*
                         * see if there's another uio vector to
                         * process that's of type IO_DIRECT
                        /*
                         * see if there's another uio vector to
                         * process that's of type IO_DIRECT


@@ -3951,11 +4560,12 @@ next_dread:
                 * get at the overhang bytes
                 */
                if (io_size & (devblocksize - 1)) {
                 * get at the overhang bytes
                 */
                if (io_size & (devblocksize - 1)) {

-                       /*
-                        * request does NOT end on a device block boundary
-                        * so clip it back to a PAGE_SIZE boundary
+                       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_size &= ~PAGE_MASK;

                        io_min = PAGE_SIZE;
                }
                if (retval || io_size < io_min) {
                        io_min = PAGE_SIZE;
                }
                if (retval || io_size < io_min) {


@@ -3968,21 +4578,62 @@ next_dread:
                         */
                        goto wait_for_dreads;
                }
                         */
                        goto wait_for_dreads;
                }

-               if ((xsize = io_size) > max_rd_size)
-                       xsize = max_rd_size;

 
 

-               io_size = 0;
+               /*
+                * 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)) {

 
 

-               ubc_range_op(vp, uio->uio_offset, uio->uio_offset + xsize, UPL_ROP_ABSENT, (int *)&io_size);
+                       if ((xsize = io_size) > max_rd_size)
+                               xsize = max_rd_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;
+                       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;
+

                iov_base = uio_curriovbase(uio);
 
                upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK);
                iov_base = uio_curriovbase(uio);
 
                upl_offset = (vm_offset_t)((u_int32_t)iov_base & PAGE_MASK);


@@ -3991,24 +4642,23 @@ next_dread:
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_START,
                             (int)upl_offset, upl_needed_size, (int)iov_base, io_size, 0);
 
                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;
                        no_zero_fill = 1;

-                       abort_flag = UPL_ABORT_DUMP_PAGES | UPL_ABORT_FREE_ON_EMPTY;
-               } else {
+               else

                        no_zero_fill = 0;
                        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;
                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;
 
                        if (no_zero_fill)
                                upl_flags |= UPL_NOZEROFILL;
                        if (force_data_sync)
                                upl_flags |= UPL_FORCE_DATA_SYNC;
 

-                       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);
 
                                                 (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)),
                                                 &upl_size, &upl, NULL, &pages_in_pl, &upl_flags);
 


@@ -4028,13 +4678,13 @@ next_dread:
                        pl = UPL_GET_INTERNAL_PAGE_LIST(upl);
 
                        for (i = 0; i < pages_in_pl; i++) {
                        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;
 
                                        break;            
                        }
                        if (i == pages_in_pl)
                                break;
 

-                       ubc_upl_abort(upl, abort_flag);
+                       ubc_upl_abort(upl, 0);

                }
                if (force_data_sync >= 3) {
                        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END,
                }
                if (force_data_sync >= 3) {
                        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END,


@@ -4052,7 +4702,7 @@ next_dread:
                                io_size = 0;
                }
                if (io_size == 0) {
                                io_size = 0;
                }
                if (io_size == 0) {

-                       ubc_upl_abort(upl, abort_flag);
+                       ubc_upl_abort(upl, 0);

                        goto wait_for_dreads;
                }
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END,
                        goto wait_for_dreads;
                }
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 72)) | DBG_FUNC_END,


@@ -4076,22 +4726,8 @@ next_dread:
                 * if there are already too many outstanding reads
                 * wait until some have completed before issuing the next read
                 */
                 * if there are already too many outstanding reads
                 * wait until some have completed before issuing the next read
                 */

-               if (iostate.io_issued > iostate.io_completed) {
-
-                       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) {
-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                            iostate.io_issued, iostate.io_completed, max_rd_ahead, 0, 0);
-
-                               iostate.io_wanted = 1;
-                               msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_read_direct", NULL);
-
-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                            iostate.io_issued, iostate.io_completed, max_rd_ahead, 0, 0);
-                       }       
-                       lck_mtx_unlock(cl_mtxp);
-               }

                if (iostate.io_error) {
                        /*
                         * one of the earlier reads we issued ran into a hard error
                if (iostate.io_error) {
                        /*
                         * one of the earlier reads we issued ran into a hard error


@@ -4100,14 +4736,13 @@ next_dread:
                         * go wait for any other reads to complete before
                         * returning the error to the caller
                         */
                         * go wait for any other reads to complete before
                         * returning the error to the caller
                         */

-                       ubc_upl_abort(upl, abort_flag);
+                       ubc_upl_abort(upl, 0);

 
                        goto wait_for_dreads;
                }
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_START,
                             upl, (int)upl_offset, (int)uio->uio_offset, io_size, 0);
 
 
                        goto wait_for_dreads;
                }
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_START,
                             upl, (int)upl_offset, (int)uio->uio_offset, io_size, 0);
 

-

                if(!useVectorUPL) {
                        if (no_zero_fill)
                                io_flag &= ~CL_PRESERVE;
                if(!useVectorUPL) {
                        if (no_zero_fill)
                                io_flag &= ~CL_PRESERVE;


@@ -4130,15 +4765,28 @@ next_dread:
                        vector_upl_size += upl_size;
                        vector_upl_iosize += io_size;
                        
                        vector_upl_size += upl_size;
                        vector_upl_iosize += io_size;
                        

-                       if(issueVectorUPL || vector_upl_index ==  MAX_VECTOR_UPL_ELEMENTS || vector_upl_size >= MAX_VECTOR_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();       
                        }
                                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;
+
+               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
                 */
                /*
                 * 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;
 
 
                io_req_size -= io_size;
 


@@ -4166,28 +4814,33 @@ wait_for_dreads:
                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();
        }
                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
         */
        /*
         * make sure all async reads that are part of this stream
         * have completed before we return
         */

-       if (iostate.io_issued > iostate.io_completed) {
+       cluster_iostate_wait(&iostate, 0, "cluster_read_direct");

 
 

-               lck_mtx_lock(cl_mtxp);
+       if (iostate.io_error)
+               retval = iostate.io_error;

 
 

-               while (iostate.io_issued != iostate.io_completed) {
-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
+       lck_mtx_destroy(&iostate.io_mtxp, cl_mtx_grp);

 
 

-                       iostate.io_wanted = 1;
-                       msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_read_direct", NULL);
+       if (io_throttled == TRUE && retval == 0)
+               retval = EAGAIN;

 
 

-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-               }       
-               lck_mtx_unlock(cl_mtxp);
+       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 (iostate.io_error)
-               retval = iostate.io_error;

 
        if (io_req_size && retval == 0) {
                /*
 
        if (io_req_size && retval == 0) {
                /*


@@ -4218,7 +4871,7 @@ cluster_read_contig(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
        upl_size_t       upl_size;
        vm_size_t        upl_needed_size;
        mach_msg_type_number_t  pages_in_pl;
        upl_size_t       upl_size;
        vm_size_t        upl_needed_size;
        mach_msg_type_number_t  pages_in_pl;

-       int              upl_flags;
+       upl_control_flags_t upl_flags;

        kern_return_t    kret;
        struct clios     iostate;
        int              error= 0;
        kern_return_t    kret;
        struct clios     iostate;
        int              error= 0;


@@ -4236,13 +4889,16 @@ cluster_read_contig(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
                bflag = CL_PASSIVE;
        else
                bflag = 0;
                bflag = CL_PASSIVE;
        else
                bflag = 0;

-
+       
+       if (flags & IO_NOCACHE)
+               bflag |= CL_NOCACHE;
+       

        /*
         * When we enter this routine, we know
         *  -- the read_length will not exceed the current iov_len
         *  -- the target address is physically contiguous for read_length
         */
        /*
         * When we enter this routine, we know
         *  -- 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);
+       cluster_syncup(vp, filesize, callback, callback_arg, PUSH_SYNC);

 
        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;
 
        devblocksize = (u_int32_t)vp->v_mount->mnt_devblocksize;
        mem_alignment_mask = (u_int32_t)vp->v_mount->mnt_alignmentmask;


@@ -4252,6 +4908,8 @@ cluster_read_contig(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
        iostate.io_error = 0;
        iostate.io_wanted = 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;
 
 next_cread:
        io_size = *read_length;
 


@@ -4267,13 +4925,15 @@ next_cread:
 
        pages_in_pl = 0;
        upl_size = upl_needed_size;
 
        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);

 
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 92)) | DBG_FUNC_START,
                     (int)upl_offset, (int)upl_size, (int)iov_base, io_size, 0);
 
 
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 92)) | DBG_FUNC_START,
                     (int)upl_offset, (int)upl_size, (int)iov_base, io_size, 0);
 

-       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[cur_upl], NULL, &pages_in_pl, &upl_flags, 0);
 
                              (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)),
                              &upl_size, &upl[cur_upl], NULL, &pages_in_pl, &upl_flags, 0);
 


@@ -4298,7 +4958,7 @@ next_cread:
        }
        pl = ubc_upl_pageinfo(upl[cur_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) {
                u_int32_t   head_size;
 
        while (((uio->uio_offset & (devblocksize - 1)) || io_size < devblocksize) && io_size) {
                u_int32_t   head_size;


@@ -4349,21 +5009,8 @@ next_cread:
                 * if there are already too many outstanding reads
                 * wait until some have completed before issuing the next
                 */
                 * if there are already too many outstanding reads
                 * wait until some have completed before issuing the next
                 */

-               if (iostate.io_issued > iostate.io_completed) {
-                       lck_mtx_lock(cl_mtxp);
+               cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_read_contig");

 
 

-                       while ((iostate.io_issued - iostate.io_completed) > (MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2))) {
-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                            iostate.io_issued, iostate.io_completed, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), 0, 0);
-
-                               iostate.io_wanted = 1;
-                               msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_read_contig", NULL);
-
-                               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                            iostate.io_issued, iostate.io_completed, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), 0, 0);
-                       }       
-                       lck_mtx_unlock(cl_mtxp);
-               }

                if (iostate.io_error) {
                        /*
                         * one of the earlier reads we issued ran into a hard error
                if (iostate.io_error) {
                        /*
                         * one of the earlier reads we issued ran into a hard error


@@ -4404,25 +5051,13 @@ wait_for_creads:
         * make sure all async reads that are part of this stream
         * have completed before we proceed
         */
         * make sure all async reads that are part of this stream
         * have completed before we proceed
         */

-       if (iostate.io_issued > iostate.io_completed) {
-
-               lck_mtx_lock(cl_mtxp);
+       cluster_iostate_wait(&iostate, 0, "cluster_read_contig");

 
 

-               while (iostate.io_issued != iostate.io_completed) {
-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-
-                       iostate.io_wanted = 1;
-                       msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_read_contig", NULL);
-
-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_END,
-                                    iostate.io_issued, iostate.io_completed, 0, 0, 0);
-               }       
-               lck_mtx_unlock(cl_mtxp);
-       }

        if (iostate.io_error)
                error = iostate.io_error;
 
        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, callback, callback_arg);
 
        if (error == 0 && tail_size)
                error = cluster_align_phys_io(vp, uio, dst_paddr, tail_size, CL_READ, callback, callback_arg);
 


@@ -4444,7 +5079,7 @@ cluster_io_type(struct uio *uio, int *io_type, u_int32_t *io_length, u_int32_t m
        user_addr_t      iov_base = 0;
        upl_t            upl;
        upl_size_t       upl_size;
        user_addr_t      iov_base = 0;
        upl_t            upl;
        upl_size_t       upl_size;

-       int              upl_flags;
+       upl_control_flags_t upl_flags;

        int              retval = 0;
 
         /*
        int              retval = 0;
 
         /*


@@ -4468,9 +5103,10 @@ cluster_io_type(struct uio *uio, int *io_type, u_int32_t *io_length, u_int32_t m
                else
                        upl_size = (u_int32_t)iov_len;
 
                else
                        upl_size = (u_int32_t)iov_len;
 

-               upl_flags = UPL_QUERY_OBJECT_TYPE;
-  
-               if ((vm_map_get_upl(current_map(),
+               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) {
                        /*
                                    (vm_map_offset_t)(iov_base & ~((user_addr_t)PAGE_MASK)),
                                    &upl_size, &upl, NULL, NULL, &upl_flags, 0)) != KERN_SUCCESS) {
                        /*


@@ -4541,6 +5177,11 @@ advisory_read_ext(vnode_t vp, off_t filesize, off_t f_offset, int resid, int (*c
 
        max_io_size = cluster_max_io_size(vp->v_mount, CL_READ);
 
 
        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);
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 60)) | DBG_FUNC_START,
                     (int)f_offset, resid, (int)filesize, 0, 0);
 


@@ -4696,13 +5337,23 @@ cluster_push(vnode_t vp, int flags)
 
 int
 cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *callback_arg)
 
 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;
 
 {
         int    retval;
        int     my_sparse_wait = 0;
        struct  cl_writebehind *wbp;
 

+       if (err)
+               *err = 0;
+

        if ( !UBCINFOEXISTS(vp)) {
        if ( !UBCINFOEXISTS(vp)) {

-               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, 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 */
                return (0);
        }
        /* return if deferred write is set */


@@ -4710,13 +5361,13 @@ cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *ca
                return (0);
        }
        if ((wbp = cluster_get_wbp(vp, CLW_RETURNLOCKED)) == NULL) {
                return (0);
        }
        if ((wbp = cluster_get_wbp(vp, CLW_RETURNLOCKED)) == NULL) {

-               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, vp, flags, 0, -2, 0);
+               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, kdebug_vnode(vp), flags, 0, -2, 0);

                return (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);
 
                lck_mtx_unlock(&wbp->cl_lockw);
 

-               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_NONE, 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,
                return(0);
        }
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 53)) | DBG_FUNC_START,


@@ -4730,11 +5381,11 @@ cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *ca
         * in the sparse map case
         */
        while (wbp->cl_sparse_wait) {
         * in the sparse map case
         */
        while (wbp->cl_sparse_wait) {

-               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_START, vp, 0, 0, 0, 0);
+               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);
 
 
                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, vp, 0, 0, 0, 0);
+               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;
        }
        if (flags & IO_SYNC) {
                my_sparse_wait = 1;


@@ -4747,11 +5398,11 @@ cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *ca
                 * fsync actually get cleaned to the disk before this fsync returns
                 */
                while (wbp->cl_sparse_pushes) {
                 * 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, vp, 0, 0, 0, 0);
+                       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);
 
 
                        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, vp, 0, 0, 0, 0);
+                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 98)) | DBG_FUNC_END, kdebug_vnode(vp), 0, 0, 0, 0);

                }
        }
        if (wbp->cl_scmap) {
                }
        }
        if (wbp->cl_scmap) {


@@ -4766,7 +5417,7 @@ cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *ca
 
                        lck_mtx_unlock(&wbp->cl_lockw);
 
 
                        lck_mtx_unlock(&wbp->cl_lockw);
 

-                       sparse_cluster_push(&scmap, vp, ubc_getsize(vp), PUSH_ALL | IO_PASSIVE, callback, callback_arg);
+                       retval = sparse_cluster_push(&scmap, vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg);

 
                        lck_mtx_lock(&wbp->cl_lockw);
 
 
                        lck_mtx_lock(&wbp->cl_lockw);
 


@@ -4775,11 +5426,13 @@ cluster_push_ext(vnode_t vp, int flags, int (*callback)(buf_t, void *), void *ca
                        if (wbp->cl_sparse_wait && wbp->cl_sparse_pushes == 0)
                                wakeup((caddr_t)&wbp->cl_sparse_pushes);
                } else {
                        if (wbp->cl_sparse_wait && wbp->cl_sparse_pushes == 0)
                                wakeup((caddr_t)&wbp->cl_sparse_pushes);
                } else {

-                       sparse_cluster_push(&(wbp->cl_scmap), vp, ubc_getsize(vp), PUSH_ALL | IO_PASSIVE, callback, callback_arg);
+                       retval = sparse_cluster_push(&(wbp->cl_scmap), vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg);

                }
                }

+               if (err)
+                       *err = retval;

                retval = 1;
                retval = 1;

-       } else  {
-               retval = cluster_try_push(wbp, vp, ubc_getsize(vp), PUSH_ALL | IO_PASSIVE, callback, callback_arg);
+       } else {
+               retval = cluster_try_push(wbp, vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg, err);

        }
        lck_mtx_unlock(&wbp->cl_lockw);
 
        }
        lck_mtx_unlock(&wbp->cl_lockw);
 


@@ -4840,7 +5493,7 @@ cluster_release(struct ubc_info *ubc)
 
 
 static int
 
 
 static int

-cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_flag, int (*callback)(buf_t, void *), void *callback_arg)
+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;
 {
         int cl_index;
        int cl_index1;


@@ -4849,7 +5502,7 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_fla
        int cl_pushed = 0;
        struct cl_wextent l_clusters[MAX_CLUSTERS];
        u_int  max_cluster_pgcount;
        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;
        /*
 
        max_cluster_pgcount = MAX_CLUSTER_SIZE(vp) / PAGE_SIZE;
        /*


@@ -4890,7 +5543,9 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_fla
 
        cl_len = cl_index;
 
 
        cl_len = cl_index;
 

-       if ( (push_flag & PUSH_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;
                
                /*
                int   i;
                
                /*


@@ -4922,16 +5577,17 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_fla
        for (cl_index = 0; cl_index < cl_len; cl_index++) {
                int     flags;
                struct  cl_extent cl;
        for (cl_index = 0; cl_index < cl_len; cl_index++) {
                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_flags & CLW_IONOCACHE)
 
                /*
                 * try to push each cluster in turn...
                 */
                if (l_clusters[cl_index].io_flags & CLW_IONOCACHE)

-                       flags = IO_NOCACHE;
-               else
-                       flags = 0;
+                       flags |= IO_NOCACHE;

 
 

-               if ((l_clusters[cl_index].io_flags & CLW_IOPASSIVE) || (push_flag & IO_PASSIVE))
+               if (l_clusters[cl_index].io_flags & CLW_IOPASSIVE)

                        flags |= IO_PASSIVE;
 
                if (push_flag & PUSH_SYNC)
                        flags |= IO_PASSIVE;
 
                if (push_flag & PUSH_SYNC)


@@ -4940,7 +5596,10 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_fla
                cl.b_addr = l_clusters[cl_index].b_addr;
                cl.e_addr = l_clusters[cl_index].e_addr;
 
                cl.b_addr = l_clusters[cl_index].b_addr;
                cl.e_addr = l_clusters[cl_index].e_addr;
 

-               cluster_push_now(vp, &cl, EOF, flags, callback, callback_arg);
+               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;
 
                l_clusters[cl_index].b_addr = 0;
                l_clusters[cl_index].e_addr = 0;


@@ -4950,6 +5609,9 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_fla
                if ( !(push_flag & PUSH_ALL) )
                        break;
        }
                if ( !(push_flag & PUSH_ALL) )
                        break;
        }

+       if (err)
+               *err = error;
+

 dont_try:
        if (cl_len > cl_pushed) {
               /*
 dont_try:
        if (cl_len > cl_pushed) {
               /*


@@ -5036,9 +5698,12 @@ cluster_push_now(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags, int (*c
        kern_return_t    kret;
 
        if (flags & IO_PASSIVE)
        kern_return_t    kret;
 
        if (flags & IO_PASSIVE)

-           bflag = CL_PASSIVE;
+               bflag = CL_PASSIVE;

        else
        else

-           bflag = 0;
+               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);
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 51)) | DBG_FUNC_START,
                     (int)cl->b_addr, (int)cl->e_addr, (int)EOF, flags, 0);


@@ -5165,6 +5830,12 @@ cluster_push_now(vnode_t vp, struct cl_extent *cl, off_t EOF, int flags, int (*c
                if ( !(flags & IO_SYNC))
                        io_flags |= CL_ASYNC;
 
                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, callback, callback_arg);
 
                retval = cluster_io(vp, upl, upl_offset, upl_f_offset + upl_offset, io_size,
                                    io_flags, (buf_t)NULL, (struct clios *)NULL, callback, callback_arg);
 


@@ -5187,7 +5858,7 @@ sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int (*c
 {
         int    cl_index;
 
 {
         int    cl_index;
 

-       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_START, vp, wbp->cl_scmap, 0, 0, 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;
 
        for (cl_index = 0; cl_index < wbp->cl_number; cl_index++) {
                int       flags;


@@ -5206,7 +5877,7 @@ sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int (*c
        }
        wbp->cl_number = 0;
 
        }
        wbp->cl_number = 0;
 

-       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_END, vp, wbp->cl_scmap, 0, 0, 0);
+       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 78)) | DBG_FUNC_END, kdebug_vnode(vp), wbp->cl_scmap, 0, 0, 0);

 }
 
 
 }
 
 


@@ -5215,31 +5886,37 @@ sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int (*c
  * 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
  */
  * 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(void **scmap, vnode_t vp, off_t EOF, int push_flag, int (*callback)(buf_t, void *), void *callback_arg)
+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;
 {
         struct cl_extent cl;
         off_t          offset;
        u_int           length;

+       int error = 0;

 
 

-       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_START, vp, (*scmap), 0, push_flag, 0);
+       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_START, kdebug_vnode(vp), (*scmap), 0, push_flag, 0);

 
        if (push_flag & PUSH_ALL)
                vfs_drt_control(scmap, 1);
 
        for (;;) {
 
        if (push_flag & PUSH_ALL)
                vfs_drt_control(scmap, 1);
 
        for (;;) {

+               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);
 
                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);
 

-               cluster_push_now(vp, &cl, EOF, push_flag & IO_PASSIVE, callback, callback_arg);
+               retval = cluster_push_now(vp, &cl, EOF, io_flags & (IO_PASSIVE|IO_CLOSE), callback, callback_arg);
+               if (error == 0 && retval)
+                       error = retval;

 
                if ( !(push_flag & PUSH_ALL) )
                        break;
        }
 
                if ( !(push_flag & PUSH_ALL) )
                        break;
        }

-       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_END, vp, (*scmap), 0, 0, 0);
+       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 79)) | DBG_FUNC_END, kdebug_vnode(vp), (*scmap), 0, 0, 0);
+
+       return error;

 }
 
 
 }
 
 


@@ -5264,12 +5941,12 @@ sparse_cluster_add(void **scmap, vnode_t vp, struct cl_extent *cl, off_t EOF, in
                 * only a partial update was done
                 * push out some pages and try again
                 */
                 * only a partial update was done
                 * push out some pages and try again
                 */

-               sparse_cluster_push(scmap, vp, EOF, 0, callback, callback_arg);
+               sparse_cluster_push(scmap, vp, EOF, 0, 0, callback, callback_arg);

 
                offset += (new_dirty * PAGE_SIZE_64);
                length -= (new_dirty * PAGE_SIZE);
        }
 
                offset += (new_dirty * PAGE_SIZE_64);
                length -= (new_dirty * PAGE_SIZE);
        }

-       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_END, vp, (*scmap), 0, 0, 0);
+       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 80)) | DBG_FUNC_END, kdebug_vnode(vp), (*scmap), 0, 0, 0);

 }
 
 
 }
 
 


@@ -5287,9 +5964,12 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, u_int32_t
        int              bflag;
 
        if (flags & IO_PASSIVE)
        int              bflag;
 
        if (flags & IO_PASSIVE)

-           bflag = CL_PASSIVE;
+               bflag = CL_PASSIVE;

        else
        else

-           bflag = 0;
+               bflag = 0;
+
+       if (flags & IO_NOCACHE)
+               bflag |= CL_NOCACHE;

 
        upl_flags = UPL_SET_LITE;
 
 
        upl_flags = UPL_SET_LITE;
 


@@ -5331,7 +6011,7 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, u_int32_t
                 }
                did_read = 1;
         }
                 }
                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
 
 /*
  *     NOTE:  There is no prototype for the following in BSD. It, and the definitions


@@ -5366,8 +6046,6 @@ cluster_align_phys_io(vnode_t vp, struct uio *uio, addr64_t usr_paddr, u_int32_t
        return (error);
 }
 
        return (error);
 }
 

-
-

 int
 cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int *io_resid)
 {
 int
 cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int *io_resid)
 {


@@ -5378,6 +6056,7 @@ cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int *io_resid)
        int       retval = 0;
        int       xsize;
        upl_page_info_t *pl;
        int       retval = 0;
        int       xsize;
        upl_page_info_t *pl;

+       int       dirty_count;

 
        xsize = *io_resid;
 
 
        xsize = *io_resid;
 


@@ -5414,10 +6093,13 @@ cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int *io_resid)
        pg_offset = upl_offset & PAGE_MASK;
        csize     = min(PAGE_SIZE - pg_offset, xsize);
 
        pg_offset = upl_offset & PAGE_MASK;
        csize     = min(PAGE_SIZE - pg_offset, xsize);
 

+       dirty_count = 0;

        while (xsize && retval == 0) {
                addr64_t  paddr;
 
        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);
 
 
                retval = uiomove64(paddr, csize, uio);
 


@@ -5430,9 +6112,10 @@ cluster_copy_upl_data(struct uio *uio, upl_t upl, int upl_offset, int *io_resid)
 
        uio->uio_segflg = segflg;
 
 
        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, xsize, retval, segflg, 0);
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_END,
                     (int)uio->uio_offset, xsize, retval, segflg, 0);

-
+       

        return (retval);
 }
 
        return (retval);
 }
 


@@ -5458,7 +6141,7 @@ cluster_copy_ubc_data_internal(vnode_t vp, struct uio *uio, int *io_resid, int m
        io_size = *io_resid;
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_START,
        io_size = *io_resid;
 
        KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 34)) | DBG_FUNC_START,

-                    (int)uio->uio_offset, 0, io_size, 0, 0);
+                    (int)uio->uio_offset, io_size, mark_dirty, take_reference, 0);

 
        control = ubc_getobject(vp, UBC_FLAGS_NONE);
 
 
        control = ubc_getobject(vp, UBC_FLAGS_NONE);
 


@@ -5551,15 +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.
  */
  * 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.
  *
 
 /*
  * 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)
 
 /*
 #define DRT_ALIGN_ADDRESS(addr)                ((addr) & DRT_ADDRESS_MASK)
 
 /*


@@ -5614,6 +6297,14 @@ is_file_clean(vnode_t vp, off_t filesize)
 #define DRT_HASH_SMALL_MODULUS 23
 #define DRT_HASH_LARGE_MODULUS 401
 
 #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 */
 
 #define DRT_SMALL_ALLOCATION   1024    /* 104 bytes spare */
 #define DRT_LARGE_ALLOCATION   16384   /* 344 bytes spare */
 


@@ -5649,7 +6340,15 @@ is_file_clean(vnode_t vp, off_t filesize)
  */
 struct vfs_drt_hashentry {
        u_int64_t       dhe_control;
  */
 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];

 };
 
 /*
 };
 
 /*


@@ -5756,8 +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) {
                 * 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;
                                nsize = DRT_HASH_LARGE_MODULUS;
                        } else {
                                nsize = DRT_HASH_SMALL_MODULUS;


@@ -5781,7 +6484,7 @@ vfs_drt_alloc_map(struct vfs_drt_clustermap **cmapp)
         */
 
        kret = kmem_alloc(kernel_map, (vm_offset_t *)&cmap,
         */
 
        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;
        if (kret != KERN_SUCCESS)
                return(kret);
        cmap->scm_magic = DRT_SCM_MAGIC;