xnu-2422.110.17.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..2f662c53c1a3e3ee0c595c697c6b598d33da7620 100644 (file)
--- a/bsd/vfs/vfs_cluster.c
+++ b/bsd/vfs/vfs_cluster.c
@@ -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>


@@ -82,6 +83,7 @@
 #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 <vm/vm_kern.h>
 #include <vm/vm_map.h>
 
 #include <vm/vm_kern.h>
 #include <vm/vm_map.h>


@@ -90,6 +92,8 @@
 #include <sys/kdebug.h>
 #include <libkern/OSAtomic.h>  
 
 #include <sys/kdebug.h>
 #include <libkern/OSAtomic.h>  
 

+#include <sys/sdt.h>
+

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


@@ -111,6 +115,10 @@
 #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_SIZE    (2 * MAX_UPL_SIZE) * PAGE_SIZE
 
 #define MAX_VECTOR_UPL_ELEMENTS        8
 #define MAX_VECTOR_UPL_SIZE    (2 * MAX_UPL_SIZE) * PAGE_SIZE


@@ -122,6 +130,7 @@ 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 */
         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 */


@@ -131,7 +140,7 @@ struct clios {
 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_UNKNOWN      0


@@ -153,8 +162,10 @@ 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);
 


@@ -182,10 +193,10 @@ 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);

 
 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 void    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);


@@ -193,6 +204,30 @@ 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
 /*
  * limit the internal I/O size so that we
  * can represent it in a 32 bit int


@@ -202,23 +237,36 @@ static kern_return_t vfs_drt_control(void **cmapp, int op_type);
 #define MAX_VECTS              16
 #define MIN_DIRECT_WRITE_SIZE  (4 * PAGE_SIZE)
 
 #define MAX_VECTS              16
 #define MIN_DIRECT_WRITE_SIZE  (4 * PAGE_SIZE)
 

-#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           1
+uint32_t speculative_prefetch_max = (MAX_UPL_SIZE * 3);
+uint32_t speculative_prefetch_max_iosize = (512 * 1024);       /* maximum I/O size to use for 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
+
+uint32_t throttle_max_iosize = (128 * 1024);
+
+#define THROTTLE_MAX_IOSIZE (throttle_max_iosize)

 
 

-int hard_throttle_on_root = 0;
-struct timeval priority_IO_timestamp_for_root;
+SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_max_iosize, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_max_iosize, 0, "");

 
 
 void
 
 
 void


@@ -234,14 +282,10 @@ 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");

 }
 
 
 }
 
 


@@ -406,7 +450,7 @@ 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 | PUSH_SYNC, 0, callback, callback_arg);

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


@@ -414,44 +458,63 @@ 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)

 {
 {

-       struct uthread  *ut;

 
 

-       if (hard_throttle) {
-               static struct timeval hard_throttle_maxelapsed = { 0, 200000 };
+       lck_mtx_lock(&iostate->io_mtxp);

 
 

-               if (vp->v_mount->mnt_kern_flag & MNTK_ROOTDEV) {
-                       struct timeval elapsed;
+       while ((iostate->io_issued - iostate->io_completed) > target) {

 
 

-                       if (hard_throttle_on_root)
-                               return(1);
+               KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 95)) | DBG_FUNC_START,
+                            iostate->io_issued, iostate->io_completed, target, 0, 0);

 
 

-                       microuptime(&elapsed);
-                       timevalsub(&elapsed, &priority_IO_timestamp_for_root);
+               iostate->io_wanted = 1;
+               msleep((caddr_t)&iostate->io_wanted, &iostate->io_mtxp, PRIBIO + 1, wait_name, NULL);

 
 

-                       if (timevalcmp(&elapsed, &hard_throttle_maxelapsed, <))
-                               return(1);
-               }
-       }
-       if (throttle_get_io_policy(&ut) == IOPOL_THROTTLE) {
-               if (throttle_io_will_be_throttled(-1, vp->v_mount)) {
-                       return(1);
-               }
-       }
-       return(0);
+               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 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 +531,7 @@ 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
                         */


@@ -502,6 +565,7 @@ 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;
 


@@ -510,32 +574,54 @@ cluster_iodone(buf_t bp, void *callback_arg)
        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)) {
+               boolean_t       need_wakeup = FALSE;

 
 

-                       KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 20)) | DBG_FUNC_END,
-                                    cbp_head, cbp, cbp->b_bcount, cbp->b_flags, 0);
+               lck_mtx_lock_spin(cl_transaction_mtxp);

 
 

-                       return 0;
+               bp->b_flags |= B_TDONE;
+               
+               if (bp->b_flags & B_TWANTED) {
+                       CLR(bp->b_flags, B_TWANTED);
+                       need_wakeup = 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);
+               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 (need_wakeup == TRUE)
+                                       wakeup(bp);
+
+                               return 0;
+                       }
+                       if (cbp->b_flags & B_EOT)
+                               transaction_complete = TRUE;
+               }
+               lck_mtx_unlock(cl_transaction_mtxp);
+
+               if (need_wakeup == TRUE)
+                       wakeup(bp);
+
+               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;


@@ -593,7 +679,7 @@ 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;


@@ -608,7 +694,7 @@ 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 (need_wakeup)
                        wakeup((caddr_t)&iostate->io_wanted);


@@ -620,7 +706,7 @@ cluster_iodone(buf_t bp, void *callback_arg)
                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 {
                        upl_flags = UPL_COMMIT_FREE_ON_EMPTY;
 
                else {
                        upl_flags = UPL_COMMIT_FREE_ON_EMPTY;
 


@@ -633,7 +719,7 @@ 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 +736,10 @@ 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;   


@@ -719,27 +805,36 @@ cluster_wait_IO(buf_t cbp_head, int async)
                /*
                 * async callback completion will not normally
                 * generate a wakeup upon I/O completion...
                /*
                 * async callback completion will not normally
                 * generate a wakeup upon I/O completion...

-                * by setting BL_WANTED, we will force a wakeup
+                * by setting B_TWANTED, we will force a wakeup

                 * to occur as any outstanding I/Os complete... 
                 * 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..
+                * I/Os already completed will have B_TDONE already
+                * set and we won't cause us to block

                 * 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
                 * 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

                 */
                 */

-               buf_list_lock();
+               lck_mtx_lock_spin(cl_transaction_mtxp);

 
                for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next)
 
                for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next)

-                     cbp->b_lflags |= BL_WANTED;
+                     cbp->b_flags |= B_TWANTED;

 
 

-               buf_list_unlock();
+               lck_mtx_unlock(cl_transaction_mtxp);

        }
        for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) {
        }
        for (cbp = cbp_head; cbp; cbp = cbp->b_trans_next) {

-               if (async)
-                       buf_biowait_callback(cbp);
-               else
+
+               if (async) {
+                       while (!ISSET(cbp->b_flags, B_TDONE)) {
+
+                               lck_mtx_lock_spin(cl_transaction_mtxp);
+
+                               if (!ISSET(cbp->b_flags, B_TDONE)) {
+                                       DTRACE_IO1(wait__start, buf_t, cbp);
+                                       (void) msleep(cbp, cl_transaction_mtxp, PDROP | (PRIBIO+1), "cluster_wait_IO", NULL);
+                                       DTRACE_IO1(wait__done, buf_t, cbp);
+                               } else
+                                       lck_mtx_unlock(cl_transaction_mtxp);
+                       }
+               } else

                        buf_biowait(cbp);
        }
 }
                        buf_biowait(cbp);
        }
 }


@@ -749,6 +844,7 @@ 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 +855,25 @@ 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;
 }


@@ -876,20 +986,19 @@ 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;
-                               
+                               u_int scale;
+

                                max_cluster_size = MAX_CLUSTER_SIZE(vp);
                                max_cluster_size = MAX_CLUSTER_SIZE(vp);

-                               max_prefetch = MAX_PREFETCH(vp, cluster_max_io_size(vp->v_mount, CL_READ));

 
                                if (max_iosize > max_cluster_size)
                                        max_cluster = max_cluster_size;
 
                                if (max_iosize > max_cluster_size)
                                        max_cluster = max_cluster_size;


@@ -898,8 +1007,16 @@ 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;

+                               
+                               if ((vp->v_mount->mnt_kern_flag & MNTK_SSD) && !ignore_is_ssd)
+                                       scale = WRITE_THROTTLE_SSD;
+                               else
+                                       scale = WRITE_THROTTLE;
+
+                               if (flags & CL_CLOSE)
+                                       scale += MAX_CLUSTERS;

 
 

-                               async_throttle = min(IO_SCALE(vp, VNODE_ASYNC_THROTTLE), (max_prefetch / max_cluster) - 1);
+                               async_throttle = min(IO_SCALE(vp, VNODE_ASYNC_THROTTLE), ((scale * max_cluster_size) / max_cluster) - 1);

                        }
                }
        }
                        }
                }
        }


@@ -911,12 +1028,14 @@ 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;
 


@@ -973,7 +1092,7 @@ 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'


@@ -1080,7 +1199,6 @@ 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;


@@ -1109,6 +1227,11 @@ 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;
                }
                lblkno = (daddr64_t)(f_offset / PAGE_SIZE_64);
                        continue;
                }
                lblkno = (daddr64_t)(f_offset / PAGE_SIZE_64);


@@ -1313,6 +1436,8 @@ 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;


@@ -1346,10 +1471,8 @@ 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 +1530,14 @@ 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) {


@@ -1455,7 +1586,7 @@ 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;


@@ -1469,7 +1600,7 @@ 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);


@@ -1480,7 +1611,7 @@ cluster_io(vnode_t vp, upl_t upl, vm_offset_t upl_offset, off_t f_offset, int no
                        pg_offset  = upl_offset & PAGE_MASK;
                        abort_size = (upl_end_offset - upl_offset + PAGE_MASK) & ~PAGE_MASK;
                        
                        pg_offset  = upl_offset & 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);


@@ -1580,8 +1711,16 @@ 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 ((max_prefetch / PAGE_SIZE) > speculative_prefetch_max)
+               max_prefetch = (speculative_prefetch_max * PAGE_SIZE);
+
+       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) {
                if ((rap->cl_maxra - extent->e_addr) > ((max_prefetch / PAGE_SIZE) / 4)) {
 
        if (extent->e_addr < rap->cl_maxra) {
                if ((rap->cl_maxra - extent->e_addr) > ((max_prefetch / PAGE_SIZE) / 4)) {
 


@@ -1643,18 +1782,7 @@ 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 +1790,8 @@ 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 +1868,8 @@ 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 +1963,10 @@ 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 +1978,12 @@ 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)


@@ -1975,7 +2108,10 @@ 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;
+       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 +2133,17 @@ 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;
+       

        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;
 


@@ -2044,6 +2185,33 @@ next_dwrite:
         }
 
        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) {
                        cluster_syncup(vp, newEOF, callback, callback_arg);
 
                if (first_IO) {
                        cluster_syncup(vp, newEOF, callback, callback_arg);


@@ -2052,8 +2220,8 @@ next_dwrite:
                io_size  = io_req_size & ~PAGE_MASK;
                iov_base = uio_curriovbase(uio);
 
                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)) {
                        /*


@@ -2183,23 +2351,9 @@ next_dwrite:
                 * if there are already too many outstanding writes
                 * wait until some complete before issuing the next
                 */
                 * if there are already too many outstanding writes
                 * wait until some complete before issuing the next
                 */

-               if (iostate.io_issued > iostate.io_completed) {
+               if (iostate.io_issued > iostate.io_completed)
+                       cluster_iostate_wait(&iostate, max_upl_size * IO_SCALE(vp, 2), "cluster_write_direct");

 
 

-                       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);
-
-                               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, 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 +2387,7 @@ 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();
                        }


@@ -2279,7 +2433,7 @@ 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();       
        }


@@ -2289,23 +2443,16 @@ wait_for_dwrites:
                 * make sure all async writes issued as part of this stream
                 * have completed before we return
                 */
                 * 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);
+               cluster_iostate_wait(&iostate, 0, "cluster_write_direct");

        }
        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


@@ -2368,6 +2515,8 @@ 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;
 


@@ -2456,22 +2605,9 @@ 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))) {
+               if (iostate.io_issued > iostate.io_completed)
+                       cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_write_contig");

 
 

-                               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);
-
-                               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 +2651,14 @@ 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);
-
-                       iostate.io_wanted = 1;
-                       msleep((caddr_t)&iostate.io_wanted, cl_mtxp, PRIBIO + 1, "cluster_write_contig", NULL);
+       if (iostate.io_issued > iostate.io_completed)
+               cluster_iostate_wait(&iostate, 0, "cluster_write_contig");

 
 

-                       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 +2733,9 @@ 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;


@@ -2629,7 +2757,9 @@ 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;
+       

        zero_cnt  = 0;
        zero_cnt1 = 0;
        zero_off  = 0;
        zero_cnt  = 0;
        zero_cnt1 = 0;
        zero_off  = 0;


@@ -2689,7 +2819,16 @@ 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 +2857,7 @@ 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 +3123,7 @@ 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 +3132,17 @@ 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 +3257,27 @@ 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_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);
+                       }
+                       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 +3295,7 @@ 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);

                        }
 
                        /*
                        }
 
                        /*


@@ -3152,18 +3314,6 @@ 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;


@@ -3224,17 +3374,40 @@ cluster_read_ext(vnode_t vp, struct uio *uio, off_t filesize, int xflags, int (*
                flags |= IO_NOCACHE;
        if ((vp->v_flag & VRAOFF) || speculative_reads_disabled)
                flags |= IO_RAOFF;
                flags |= IO_NOCACHE;
        if ((vp->v_flag & VRAOFF) || speculative_reads_disabled)
                flags |= IO_RAOFF;

+       
+       /* 
+        * If we're doing an encrypted IO, then first check to see
+        * if the IO requested was page aligned.  If not, then bail 
+        * out immediately.
+        */
+       if (flags & IO_ENCRYPTED) {             
+               if (read_length & PAGE_MASK) {
+                       retval = EINVAL;
+                       return retval;
+               }
+       }

 
 

-        /*
+       /*

         * 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) || (flags & IO_ENCRYPTED)) && UIO_SEG_IS_USER_SPACE(uio->uio_segflg)) {
+
+               boolean_t       check_io_type = TRUE;
+

 
 

+               if (check_io_type) {
+                       retval = cluster_io_type(uio, &read_type, &read_length, 0);
+               }
+       }
+       

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


@@ -3318,21 +3491,19 @@ 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;
-
-       ut = get_bsdthread_info(current_thread());
-
-       if (ut->uu_iopol_disk != IOPOL_DEFAULT)
-               policy = ut->uu_iopol_disk;
+       policy = throttle_get_io_policy(NULL);

 
 

-       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 +3511,11 @@ 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;
+

        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 +3527,16 @@ 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 +3616,7 @@ 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 +3655,30 @@ 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


@@ -3540,10 +3740,6 @@ 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 +3748,11 @@ 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 +3765,18 @@ 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 +3855,9 @@ 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);
-
-                               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);
+                       if (iostate.io_issued > iostate.io_completed)
+                               cluster_iostate_wait(&iostate, 0, "cluster_read_copy");

 
 

-                                       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 +3869,9 @@ 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 (iostate.io_issued > iostate.io_completed)
+                               cluster_iostate_wait(&iostate, 0, "cluster_read_copy");

                }
                if (start_pg < last_pg) {
                        /*
                }
                if (start_pg < last_pg) {
                        /*


@@ -3728,17 +3931,21 @@ 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 +3956,20 @@ cluster_read_copy(vnode_t vp, struct uio *uio, u_int32_t io_req_size, off_t file
                        }
                }
        }
                        }
                }
        }

+       if (iolock_inited == TRUE) {
+               if (iostate.io_issued > iostate.io_completed) {
+                       /*
+                        * 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);


@@ -3780,7 +4001,6 @@ cluster_read_direct(vnode_t vp, struct uio *uio, off_t filesize, int *read_type,
        int              force_data_sync;
        int              retval = 0;
        int              no_zero_fill = 0;
        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;


@@ -3796,6 +4016,9 @@ 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);
 
        u_int32_t        vector_upl_iosize = 0;
        int              issueVectorUPL = 0,useVectorUPL = (uio->uio_iovcnt > 1);


@@ -3812,14 +4035,25 @@ 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;
+       }
+

        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;
 


@@ -3839,6 +4073,9 @@ 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);
+

 next_dread:
        io_req_size = *read_length;
        iov_base = uio_curriovbase(uio);
 next_dread:
        io_req_size = *read_length;
        iov_base = uio_curriovbase(uio);


@@ -3865,7 +4102,17 @@ 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;

-        }
+    }
+
+       /* 
+        * 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) && (misaligned)) {
+               retval = EINVAL;
+       }
+

        /*
         * 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 +4121,35 @@ 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 +4180,14 @@ 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


@@ -3950,13 +4213,31 @@ next_dread:
                 * (which overlaps the end of the direct read) in order to 
                 * get at the overhang bytes
                 */
                 * (which overlaps the end of the direct read) in order to 
                 * 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
-                        */
-                       io_size &= ~PAGE_MASK;
-                       io_min = PAGE_SIZE;
+               if (io_size & (devblocksize - 1)) {                     
+                       if (flags & IO_ENCRYPTED) {
+                               /* 
+                                * Normally, we'd round down to the previous page boundary to 
+                                * let the UBC manage the zero-filling of the file past the EOF.
+                                * But if we're doing encrypted IO, we can't let any of
+                                * the data hit the UBC.  This means we have to do the full
+                                * IO to the upper block boundary of the device block that
+                                * contains the EOF. The user will be responsible for not
+                                * interpreting data PAST the EOF in its buffer.
+                                *
+                                * So just bump the IO back up to a multiple of devblocksize
+                                */
+                               io_size = ((io_size + devblocksize) & ~(devblocksize - 1));
+                               io_min = io_size;                                       
+                       }
+                       else {
+                               /* 
+                                * Clip the request to the previous page size boundary
+                                * since request does NOT end on a device block boundary
+                                */
+                               io_size &= ~PAGE_MASK;
+                               io_min = PAGE_SIZE;
+                       }
+                       

                }
                if (retval || io_size < io_min) {
                        /*
                }
                if (retval || io_size < io_min) {
                        /*


@@ -3968,21 +4249,49 @@ 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;
+
+                       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);


@@ -3991,13 +4300,11 @@ 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;
-               }
+

                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;


@@ -4028,13 +4335,13 @@ 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 +4359,7 @@ 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 +4383,9 @@ 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) {
+               if (iostate.io_issued > iostate.io_completed)
+                       cluster_iostate_wait(&iostate, max_rd_ahead, "cluster_read_direct");

 
 

-                       lck_mtx_lock(cl_mtxp);
-
-                       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


@@ -4100,7 +4394,7 @@ 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;
                }
 
                        goto wait_for_dreads;
                }


@@ -4130,7 +4424,7 @@ 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();       
                        }


@@ -4138,9 +4432,24 @@ next_dread:
                /*
                 * update the uio structure
                 */
                /*
                 * update the uio structure
                 */

-               uio_update(uio, (user_size_t)io_size);
-
-               io_req_size -= 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);
+               }
+               /*
+                * Under normal circumstances, the io_size should not be
+                * bigger than the io_req_size, but we may have had to round up
+                * to the end of the page in the encrypted IO case.  In that case only,
+                * ensure that we only decrement io_req_size to 0. 
+                */
+               if ((flags & IO_ENCRYPTED) && (io_size > io_req_size)) {
+                       io_req_size = 0;
+               }
+               else {
+                       io_req_size -= io_size;
+               }

 
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_END,
                             upl, (int)uio->uio_offset, io_req_size, retval, 0);
 
                KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 73)) | DBG_FUNC_END,
                             upl, (int)uio->uio_offset, io_req_size, retval, 0);


@@ -4170,25 +4479,17 @@ wait_for_dreads:
         * 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) {
-
-               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_read_direct", NULL);
+       if (iostate.io_issued > iostate.io_completed)
+               cluster_iostate_wait(&iostate, 0, "cluster_read_direct");

 
 

-                       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


@@ -4236,7 +4537,10 @@ 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
        /*
         * When we enter this routine, we know
         *  -- the read_length will not exceed the current iov_len


@@ -4252,6 +4556,8 @@ 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;
 


@@ -4349,21 +4655,9 @@ 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);
-
-                       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);
+               if (iostate.io_issued > iostate.io_completed)
+                       cluster_iostate_wait(&iostate, MAX_IO_CONTIG_SIZE * IO_SCALE(vp, 2), "cluster_read_contig");

 
 

-                               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 +4698,14 @@ 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);
-
-               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);
+       if (iostate.io_issued > iostate.io_completed)
+               cluster_iostate_wait(&iostate, 0, "cluster_read_contig");

 
 

-                       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);
 


@@ -4541,6 +4824,11 @@ 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);
 


@@ -4766,7 +5054,7 @@ 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);
+                       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 +5063,11 @@ 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);
+                       sparse_cluster_push(&(wbp->cl_scmap), vp, ubc_getsize(vp), PUSH_ALL, flags, callback, callback_arg);

                }
                retval = 1;
        } else  {
                }
                retval = 1;
        } else  {

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

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


@@ -4840,7 +5128,7 @@ 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 cl_index;
        int cl_index1;
 {
         int cl_index;
        int cl_index1;


@@ -4923,15 +5211,15 @@ cluster_try_push(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int push_fla
                int     flags;
                struct  cl_extent cl;
 
                int     flags;
                struct  cl_extent cl;
 

+               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)


@@ -5036,9 +5324,9 @@ 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;

 
        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 +5453,12 @@ 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);
 


@@ -5216,7 +5510,7 @@ sparse_cluster_switch(struct cl_writebehind *wbp, vnode_t vp, off_t EOF, int (*c
  * from the write-behind context (the cluster_push case), the wb lock is not held
  */
 static void
  * 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)
+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;
 {
         struct cl_extent cl;
         off_t          offset;


@@ -5234,7 +5528,7 @@ sparse_cluster_push(void **scmap, vnode_t vp, off_t EOF, int push_flag, int (*ca
                cl.b_addr = (daddr64_t)(offset / PAGE_SIZE_64);
                cl.e_addr = (daddr64_t)((offset + length) / PAGE_SIZE_64);
 
                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);
+               cluster_push_now(vp, &cl, EOF, io_flags & (IO_PASSIVE|IO_CLOSE), callback, callback_arg);

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


@@ -5264,7 +5558,7 @@ 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);


@@ -5287,9 +5581,12 @@ 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;
 


@@ -5458,7 +5755,7 @@ 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);
 


@@ -5614,6 +5911,14 @@ 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 */
 


@@ -5756,8 +6061,12 @@ 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;