X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosc%2Fosc_cache.c;h=288853062728e7fb4d20c46fda8cff7f7202597e;hp=5394a755c18576ac673849a74210cced069e688f;hb=05f326a7988a7a0d6954d1b0d318315526209ae6;hpb=b4a959eb61bc7e6a64261c704f3f3f5e220c2f02

diff --git a/lustre/osc/osc_cache.c b/lustre/osc/osc_cache.c
index 5394a75..2888530 100644
--- a/lustre/osc/osc_cache.c
+++ b/lustre/osc/osc_cache.c
@@ -38,6 +38,7 @@
 #define DEBUG_SUBSYSTEM S_OSC
 
 #include <lustre_osc.h>
+#include <lustre_dlm.h>
 
 #include "osc_internal.h"
 
@@ -57,10 +58,10 @@ static int osc_io_unplug_async(const struct lu_env *env,
 static void osc_free_grant(struct client_obd *cli, unsigned int nr_pages,
 			   unsigned int lost_grant, unsigned int dirty_grant);
 
-static void osc_extent_tree_dump0(int level, struct osc_object *obj,
+static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
 				  const char *func, int line);
-#define osc_extent_tree_dump(lvl, obj) \
-	osc_extent_tree_dump0(lvl, obj, __func__, __LINE__)
+#define osc_extent_tree_dump(mask, obj) \
+	osc_extent_tree_dump0(mask, obj, __func__, __LINE__)
 
 static void osc_unreserve_grant(struct client_obd *cli, unsigned int reserved,
 				unsigned int unused);
@@ -74,7 +75,7 @@ static inline char *ext_flags(struct osc_extent *ext, char *flags)
 {
 	char *buf = flags;
 	*buf++ = ext->oe_rw ? 'r' : 'w';
-	if (ext->oe_intree)
+	if (!RB_EMPTY_NODE(&ext->oe_node))
 		*buf++ = 'i';
 	if (ext->oe_sync)
 		*buf++ = 'S';
@@ -94,28 +95,24 @@ static inline char *ext_flags(struct osc_extent *ext, char *flags)
 	return flags;
 }
 
-static inline char list_empty_marker(struct list_head *list)
-{
-	return list_empty(list) ? '-' : '+';
-}
-
 #define EXTSTR       "[%lu -> %lu/%lu]"
 #define EXTPARA(ext) (ext)->oe_start, (ext)->oe_end, (ext)->oe_max_end
 static const char *oes_strings[] = {
 	"inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 
-#define OSC_EXTENT_DUMP(lvl, extent, fmt, ...) do {			      \
+#define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
+	static struct cfs_debug_limit_state cdls;			      \
 	struct osc_extent *__ext = (extent);				      \
 	char __buf[16];							      \
 									      \
-	CDEBUG(lvl,							      \
+	__CDEBUG_WITH_LOC(file, func, line, mask, &cdls,		      \
 		"extent %p@{" EXTSTR ", "				      \
 		"[%d|%d|%c|%s|%s|%p], [%d|%d|%c|%c|%p|%u|%p]} " fmt,	      \
 		/* ----- extent part 0 ----- */				      \
 		__ext, EXTPARA(__ext),					      \
 		/* ----- part 1 ----- */				      \
-		atomic_read(&__ext->oe_refc),			      \
-		atomic_read(&__ext->oe_users),			      \
+		kref_read(&__ext->oe_refc),				      \
+		atomic_read(&__ext->oe_users),				      \
 		list_empty_marker(&__ext->oe_link),			      \
 		oes_strings[__ext->oe_state], ext_flags(__ext, __buf),	      \
 		__ext->oe_obj,						      \
@@ -126,12 +123,16 @@ static const char *oes_strings[] = {
 		__ext->oe_dlmlock, __ext->oe_mppr, __ext->oe_owner,	      \
 		/* ----- part 4 ----- */				      \
 		## __VA_ARGS__);					      \
-	if (lvl == D_ERROR && __ext->oe_dlmlock != NULL)		      \
+	if (mask == D_ERROR && __ext->oe_dlmlock != NULL)		      \
 		LDLM_ERROR(__ext->oe_dlmlock, "extent: %p", __ext);	      \
 	else								      \
 		LDLM_DEBUG(__ext->oe_dlmlock, "extent: %p", __ext);	      \
 } while (0)
 
+#define OSC_EXTENT_DUMP(mask, ext, fmt, ...)				\
+	OSC_EXTENT_DUMP_WITH_LOC(__FILE__, __func__, __LINE__,		\
+				 mask, ext, fmt, ## __VA_ARGS__)
+
 #undef EASSERTF
 #define EASSERTF(expr, ext, fmt, args...) do {				\
 	if (!(expr)) {							\
@@ -146,10 +147,7 @@ static const char *oes_strings[] = {
 
 static inline struct osc_extent *rb_extent(struct rb_node *n)
 {
-	if (n == NULL)
-		return NULL;
-
-	return container_of(n, struct osc_extent, oe_node);
+	return rb_entry_safe(n, struct osc_extent, oe_node);
 }
 
 static inline struct osc_extent *next_extent(struct osc_extent *ext)
@@ -157,7 +155,7 @@ static inline struct osc_extent *next_extent(struct osc_extent *ext)
 	if (ext == NULL)
 		return NULL;
 
-	LASSERT(ext->oe_intree);
+	LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
 	return rb_extent(rb_next(&ext->oe_node));
 }
 
@@ -166,7 +164,7 @@ static inline struct osc_extent *prev_extent(struct osc_extent *ext)
 	if (ext == NULL)
 		return NULL;
 
-	LASSERT(ext->oe_intree);
+	LASSERT(!RB_EMPTY_NODE(&ext->oe_node));
 	return rb_extent(rb_prev(&ext->oe_node));
 }
 
@@ -184,16 +182,15 @@ static int osc_extent_sanity_check0(struct osc_extent *ext,
 	size_t page_count;
 	int rc = 0;
 
-	if (!osc_object_is_locked(obj))
-		GOTO(out, rc = 9);
+	assert_osc_object_is_locked(obj);
 
 	if (ext->oe_state >= OES_STATE_MAX)
 		GOTO(out, rc = 10);
 
-	if (atomic_read(&ext->oe_refc) <= 0)
+	if (kref_read(&ext->oe_refc) <= 0)
 		GOTO(out, rc = 20);
 
-	if (atomic_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
+	if (kref_read(&ext->oe_refc) < atomic_read(&ext->oe_users))
 		GOTO(out, rc = 30);
 
 	switch (ext->oe_state) {
@@ -215,6 +212,7 @@ static int osc_extent_sanity_check0(struct osc_extent *ext,
 			GOTO(out, rc = 60);
 		if (ext->oe_fsync_wait && !ext->oe_urgent && !ext->oe_hp)
 			GOTO(out, rc = 65);
+		/* fallthrough */
 	default:
 		if (atomic_read(&ext->oe_users) > 0)
 			GOTO(out, rc = 70);
@@ -263,9 +261,9 @@ static int osc_extent_sanity_check0(struct osc_extent *ext,
 
 out:
 	if (rc != 0)
-		OSC_EXTENT_DUMP(D_ERROR, ext,
-				"%s:%d sanity check %p failed with rc = %d\n",
-				func, line, ext, rc);
+		OSC_EXTENT_DUMP_WITH_LOC(__FILE__, func, line, D_ERROR, ext,
+					 "sanity check %p failed: rc = %d\n",
+					 ext, rc);
 	return rc;
 }
 
@@ -280,6 +278,11 @@ out:
 	__res;                                                                 \
 })
 
+static inline bool
+overlapped(const struct osc_extent *ex1, const struct osc_extent *ex2)
+{
+	return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
+}
 
 /**
  * sanity check - to make sure there is no overlapped extent in the tree.
@@ -289,7 +292,7 @@ static int osc_extent_is_overlapped(struct osc_object *obj,
 {
 	struct osc_extent *tmp;
 
-	LASSERT(osc_object_is_locked(obj));
+	assert_osc_object_is_locked(obj);
 
 	if (!extent_debug)
 		return 0;
@@ -297,8 +300,7 @@ static int osc_extent_is_overlapped(struct osc_object *obj,
 	for (tmp = first_extent(obj); tmp != NULL; tmp = next_extent(tmp)) {
 		if (tmp == ext)
 			continue;
-		if (tmp->oe_end >= ext->oe_start &&
-		    tmp->oe_start <= ext->oe_end)
+		if (overlapped(tmp, ext))
 			return 1;
 	}
 	return 0;
@@ -306,15 +308,15 @@ static int osc_extent_is_overlapped(struct osc_object *obj,
 
 static void osc_extent_state_set(struct osc_extent *ext, int state)
 {
-	LASSERT(osc_object_is_locked(ext->oe_obj));
+	assert_osc_object_is_locked(ext->oe_obj);
 	LASSERT(state >= OES_INV && state < OES_STATE_MAX);
 
 	/* Never try to sanity check a state changing extent :-) */
 	/* LASSERT(sanity_check_nolock(ext) == 0); */
 
 	/* TODO: validate the state machine */
-	ext->oe_state = state;
-	wake_up_all(&ext->oe_waitq);
+	smp_store_release(&ext->oe_state, state);
+	wake_up(&ext->oe_waitq);
 }
 
 static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
@@ -328,7 +330,7 @@ static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 	RB_CLEAR_NODE(&ext->oe_node);
 	ext->oe_obj = obj;
 	cl_object_get(osc2cl(obj));
-	atomic_set(&ext->oe_refc, 1);
+	kref_init(&ext->oe_refc);
 	atomic_set(&ext->oe_users, 0);
 	INIT_LIST_HEAD(&ext->oe_link);
 	ext->oe_state = OES_INV;
@@ -339,35 +341,50 @@ static struct osc_extent *osc_extent_alloc(struct osc_object *obj)
 	return ext;
 }
 
-static void osc_extent_free(struct osc_extent *ext)
+static void osc_extent_free(struct kref *kref)
 {
+	struct osc_extent *ext = container_of(kref, struct osc_extent,
+					      oe_refc);
+
+	LASSERT(list_empty(&ext->oe_link));
+	LASSERT(atomic_read(&ext->oe_users) == 0);
+	LASSERT(ext->oe_state == OES_INV);
+	LASSERT(RB_EMPTY_NODE(&ext->oe_node));
+
+	if (ext->oe_dlmlock) {
+		lu_ref_del(&ext->oe_dlmlock->l_reference,
+			   "osc_extent", ext);
+		LDLM_LOCK_PUT(ext->oe_dlmlock);
+		ext->oe_dlmlock = NULL;
+	}
+#if 0
+	/* If/When cl_object_put drops the need for 'env',
+	 * this code can be enabled, and matching code in
+	 * osc_extent_put removed.
+	 */
+	cl_object_put(osc2cl(ext->oe_obj));
+
 	OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
+#endif
 }
 
 static struct osc_extent *osc_extent_get(struct osc_extent *ext)
 {
-	LASSERT(atomic_read(&ext->oe_refc) >= 0);
-	atomic_inc(&ext->oe_refc);
+	LASSERT(kref_read(&ext->oe_refc) >= 0);
+	kref_get(&ext->oe_refc);
 	return ext;
 }
 
 static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
 {
-	LASSERT(atomic_read(&ext->oe_refc) > 0);
-	if (atomic_dec_and_test(&ext->oe_refc)) {
-		LASSERT(list_empty(&ext->oe_link));
-		LASSERT(atomic_read(&ext->oe_users) == 0);
-		LASSERT(ext->oe_state == OES_INV);
-		LASSERT(!ext->oe_intree);
-
-		if (ext->oe_dlmlock != NULL) {
-			lu_ref_add(&ext->oe_dlmlock->l_reference,
-				   "osc_extent", ext);
-			LDLM_LOCK_PUT(ext->oe_dlmlock);
-			ext->oe_dlmlock = NULL;
-		}
+	LASSERT(kref_read(&ext->oe_refc) > 0);
+	if (kref_put(&ext->oe_refc, osc_extent_free)) {
+		/* This should be in osc_extent_free(), but
+		 * while we need to pass 'env' it cannot be.
+		 */
 		cl_object_put(env, osc2cl(ext->oe_obj));
-		osc_extent_free(ext);
+
+		OBD_SLAB_FREE_PTR(ext, osc_extent_kmem);
 	}
 }
 
@@ -378,9 +395,9 @@ static void osc_extent_put(const struct lu_env *env, struct osc_extent *ext)
  */
 static void osc_extent_put_trust(struct osc_extent *ext)
 {
-	LASSERT(atomic_read(&ext->oe_refc) > 1);
-	LASSERT(osc_object_is_locked(ext->oe_obj));
-	atomic_dec(&ext->oe_refc);
+	LASSERT(kref_read(&ext->oe_refc) > 1);
+	assert_osc_object_is_locked(ext->oe_obj);
+	osc_extent_put(NULL, ext);
 }
 
 /**
@@ -393,7 +410,7 @@ static struct osc_extent *osc_extent_search(struct osc_object *obj,
 	struct rb_node    *n = obj->oo_root.rb_node;
 	struct osc_extent *tmp, *p = NULL;
 
-	LASSERT(osc_object_is_locked(obj));
+	assert_osc_object_is_locked(obj);
 	while (n != NULL) {
 		tmp = rb_extent(n);
 		if (index < tmp->oe_start) {
@@ -430,9 +447,9 @@ static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 	struct rb_node    *parent = NULL;
 	struct osc_extent *tmp;
 
-	LASSERT(ext->oe_intree == 0);
+	LASSERT(RB_EMPTY_NODE(&ext->oe_node));
 	LASSERT(ext->oe_obj == obj);
-	LASSERT(osc_object_is_locked(obj));
+	assert_osc_object_is_locked(obj);
 	while (*n != NULL) {
 		tmp = rb_extent(*n);
 		parent = *n;
@@ -447,17 +464,16 @@ static void osc_extent_insert(struct osc_object *obj, struct osc_extent *ext)
 	rb_link_node(&ext->oe_node, parent, n);
 	rb_insert_color(&ext->oe_node, &obj->oo_root);
 	osc_extent_get(ext);
-	ext->oe_intree = 1;
 }
 
 /* caller must have held object lock. */
 static void osc_extent_erase(struct osc_extent *ext)
 {
 	struct osc_object *obj = ext->oe_obj;
-	LASSERT(osc_object_is_locked(obj));
-	if (ext->oe_intree) {
+	assert_osc_object_is_locked(obj);
+	if (!RB_EMPTY_NODE(&ext->oe_node)) {
 		rb_erase(&ext->oe_node, &obj->oo_root);
-		ext->oe_intree = 0;
+		RB_CLEAR_NODE(&ext->oe_node);
 		/* rbtree held a refcount */
 		osc_extent_put_trust(ext);
 	}
@@ -467,7 +483,7 @@ static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 {
 	struct osc_object *obj = ext->oe_obj;
 
-	LASSERT(osc_object_is_locked(obj));
+	assert_osc_object_is_locked(obj);
 	LASSERT(ext->oe_state == OES_ACTIVE || ext->oe_state == OES_CACHE);
 	if (ext->oe_state == OES_CACHE) {
 		osc_extent_state_set(ext, OES_ACTIVE);
@@ -480,7 +496,7 @@ static struct osc_extent *osc_extent_hold(struct osc_extent *ext)
 
 static void __osc_extent_remove(struct osc_extent *ext)
 {
-	LASSERT(osc_object_is_locked(ext->oe_obj));
+	assert_osc_object_is_locked(ext->oe_obj);
 	LASSERT(list_empty(&ext->oe_pages));
 	osc_extent_erase(ext);
 	list_del_init(&ext->oe_link);
@@ -511,7 +527,7 @@ static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 	int			 ppc_bits;
 
 	LASSERT(cur->oe_state == OES_CACHE);
-	LASSERT(osc_object_is_locked(obj));
+	assert_osc_object_is_locked(obj);
 	if (victim == NULL)
 		return -EINVAL;
 
@@ -521,6 +537,14 @@ static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 	if (cur->oe_max_end != victim->oe_max_end)
 		return -ERANGE;
 
+	/*
+	 * In the rare case max_pages_per_rpc (mppr) is changed, don't
+	 * merge extents until after old ones have been sent, or the
+	 * "extents are aligned to RPCs" checks are unhappy.
+	 */
+	if (cur->oe_mppr != victim->oe_mppr)
+		return -ERANGE;
+
 	LASSERT(cur->oe_dlmlock == victim->oe_dlmlock);
 	ppc_bits = osc_cli(obj)->cl_chunkbits - PAGE_SHIFT;
 	chunk_start = cur->oe_start >> ppc_bits;
@@ -546,7 +570,6 @@ static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 	cur->oe_urgent   |= victim->oe_urgent;
 	cur->oe_memalloc |= victim->oe_memalloc;
 	list_splice_init(&victim->oe_pages, &cur->oe_pages);
-	list_del_init(&victim->oe_link);
 	victim->oe_nr_pages = 0;
 
 	osc_extent_get(victim);
@@ -579,6 +602,7 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 			 * osc_cache_truncate_start(). */
 			osc_extent_state_set(ext, OES_TRUNC);
 			ext->oe_trunc_pending = 0;
+			osc_object_unlock(obj);
 		} else {
 			int grant = 0;
 
@@ -591,18 +615,21 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 				grant += cli->cl_grant_extent_tax;
 			if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 				grant += cli->cl_grant_extent_tax;
-			if (grant > 0)
-				osc_unreserve_grant(cli, 0, grant);
 
-			if (ext->oe_urgent)
+			if (ext->oe_hp)
+				list_move_tail(&ext->oe_link,
+					       &obj->oo_hp_exts);
+			else if (ext->oe_urgent)
 				list_move_tail(&ext->oe_link,
 					       &obj->oo_urgent_exts);
 			else if (ext->oe_nr_pages == ext->oe_mppr) {
 				list_move_tail(&ext->oe_link,
 					       &obj->oo_full_exts);
 			}
+			osc_object_unlock(obj);
+			if (grant > 0)
+				osc_unreserve_grant(cli, 0, grant);
 		}
-		osc_object_unlock(obj);
 
 		osc_io_unplug_async(env, cli, obj);
 	}
@@ -610,11 +637,6 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 	RETURN(rc);
 }
 
-static inline int overlapped(struct osc_extent *ex1, struct osc_extent *ex2)
-{
-	return !(ex1->oe_end < ex2->oe_start || ex2->oe_end < ex1->oe_start);
-}
-
 /**
  * Find or create an extent which includes @index, core function to manage
  * extent tree.
@@ -675,7 +697,7 @@ static struct osc_extent *osc_extent_find(const struct lu_env *env,
 		cur->oe_start = descr->cld_start;
 	if (cur->oe_end > max_end)
 		cur->oe_end = max_end;
-	cur->oe_grants  = 0;
+	cur->oe_grants  = chunksize + cli->cl_grant_extent_tax;
 	cur->oe_mppr    = max_pages;
 	if (olck->ols_dlmlock != NULL) {
 		LASSERT(olck->ols_hold);
@@ -692,11 +714,11 @@ static struct osc_extent *osc_extent_find(const struct lu_env *env,
 restart:
 	osc_object_lock(obj);
 	ext = osc_extent_search(obj, cur->oe_start);
-	if (ext == NULL)
+	if (!ext)
 		ext = first_extent(obj);
-	while (ext != NULL) {
+	for (; ext; ext = next_extent(ext)) {
 		pgoff_t ext_chk_start = ext->oe_start >> ppc_bits;
-		pgoff_t ext_chk_end   = ext->oe_end   >> ppc_bits;
+		pgoff_t ext_chk_end = ext->oe_end >> ppc_bits;
 
 		LASSERT(sanity_check_nolock(ext) == 0);
 		if (chunk > ext_chk_end + 1 || chunk < ext_chk_start)
@@ -707,15 +729,12 @@ restart:
 			EASSERTF(!overlapped(ext, cur), ext,
 				 EXTSTR"\n", EXTPARA(cur));
 
-			ext = next_extent(ext);
 			continue;
 		}
 
 		/* discontiguous chunks? */
-		if (chunk + 1 < ext_chk_start) {
-			ext = next_extent(ext);
+		if (chunk + 1 < ext_chk_start)
 			continue;
-		}
 
 		/* ok, from now on, ext and cur have these attrs:
 		 * 1. covered by the same lock
@@ -740,68 +759,27 @@ restart:
 		}
 
 		/* non-overlapped extent */
-		if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait) {
+		if (ext->oe_state != OES_CACHE || ext->oe_fsync_wait)
 			/* we can't do anything for a non OES_CACHE extent, or
 			 * if there is someone waiting for this extent to be
 			 * flushed, try next one. */
-			ext = next_extent(ext);
 			continue;
-		}
 
-		/* check if they belong to the same rpc slot before trying to
-		 * merge. the extents are not overlapped and contiguous at
-		 * chunk level to get here. */
-		if (ext->oe_max_end != max_end) {
-			/* if they don't belong to the same RPC slot or
-			 * max_pages_per_rpc has ever changed, do not merge. */
-			ext = next_extent(ext);
-			continue;
-		}
-
-		/* check whether maximum extent size will be hit */
-		if ((ext_chk_end - ext_chk_start + 1 + 1) << ppc_bits >
-		    cli->cl_max_extent_pages) {
-			ext = next_extent(ext);
-			continue;
-		}
-
-		/* it's required that an extent must be contiguous at chunk
-		 * level so that we know the whole extent is covered by grant
-		 * (the pages in the extent are NOT required to be contiguous).
-		 * Otherwise, it will be too much difficult to know which
-		 * chunks have grants allocated. */
-
-		/* try to do front merge - extend ext's start */
-		if (chunk + 1 == ext_chk_start) {
-			/* ext must be chunk size aligned */
-			EASSERT((ext->oe_start & ~chunk_mask) == 0, ext);
-
-			/* pull ext's start back to cover cur */
-			ext->oe_start   = cur->oe_start;
-			ext->oe_grants += chunksize;
+		if (osc_extent_merge(env, ext, cur) == 0) {
 			LASSERT(*grants >= chunksize);
 			*grants -= chunksize;
-
 			found = osc_extent_hold(ext);
-		} else if (chunk == ext_chk_end + 1) {
-			/* rear merge */
-			ext->oe_end     = cur->oe_end;
-			ext->oe_grants += chunksize;
-			LASSERT(*grants >= chunksize);
-			*grants -= chunksize;
 
-			/* try to merge with the next one because we just fill
-			 * in a gap */
+			/*
+			 * Try to merge with the next one too because we
+			 * might have just filled in a gap.
+			 */
 			if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 				/* we can save extent tax from next extent */
 				*grants += cli->cl_grant_extent_tax;
 
-			found = osc_extent_hold(ext);
-		}
-		if (found != NULL)
 			break;
-
-		ext = next_extent(ext);
+		}
 	}
 
 	osc_extent_tree_dump(D_CACHE, obj);
@@ -815,7 +793,6 @@ restart:
 	} else if (conflict == NULL) {
 		/* create a new extent */
 		EASSERT(osc_extent_is_overlapped(obj, cur) == 0, cur);
-		cur->oe_grants = chunksize + cli->cl_grant_extent_tax;
 		LASSERT(*grants >= cur->oe_grants);
 		*grants -= cur->oe_grants;
 
@@ -907,17 +884,6 @@ int osc_extent_finish(const struct lu_env *env, struct osc_extent *ext,
 	RETURN(0);
 }
 
-static int extent_wait_cb(struct osc_extent *ext, enum osc_extent_state state)
-{
-	int ret;
-
-	osc_object_lock(ext->oe_obj);
-	ret = ext->oe_state == state;
-	osc_object_unlock(ext->oe_obj);
-
-	return ret;
-}
-
 /**
  * Wait for the extent's state to become @state.
  */
@@ -925,8 +891,6 @@ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 			   enum osc_extent_state state)
 {
 	struct osc_object *obj = ext->oe_obj;
-	struct l_wait_info lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(600), NULL,
-						  LWI_ON_SIGNAL_NOOP, NULL);
 	int rc = 0;
 	ENTRY;
 
@@ -948,18 +912,21 @@ static int osc_extent_wait(const struct lu_env *env, struct osc_extent *ext,
 		osc_extent_release(env, ext);
 
 	/* wait for the extent until its state becomes @state */
-	rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state), &lwi);
-	if (rc == -ETIMEDOUT) {
+	rc = wait_event_idle_timeout(ext->oe_waitq,
+				     smp_load_acquire(&ext->oe_state) == state,
+				     cfs_time_seconds(600));
+	if (rc == 0) {
 		OSC_EXTENT_DUMP(D_ERROR, ext,
 			"%s: wait ext to %u timedout, recovery in progress?\n",
 			cli_name(osc_cli(obj)), state);
 
-		lwi = LWI_INTR(NULL, NULL);
-		rc = l_wait_event(ext->oe_waitq, extent_wait_cb(ext, state),
-				  &lwi);
+		wait_event_idle(ext->oe_waitq,
+				smp_load_acquire(&ext->oe_state) == state);
 	}
-	if (rc == 0 && ext->oe_rc < 0)
+	if (ext->oe_rc < 0)
 		rc = ext->oe_rc;
+	else
+		rc = 0;
 	RETURN(rc);
 }
 
@@ -1228,34 +1195,34 @@ out:
 	RETURN(rc);
 }
 
-static void osc_extent_tree_dump0(int level, struct osc_object *obj,
+static void osc_extent_tree_dump0(int mask, struct osc_object *obj,
 				  const char *func, int line)
 {
 	struct osc_extent *ext;
 	int cnt;
 
-	if (!cfs_cdebug_show(level, DEBUG_SUBSYSTEM))
+	if (!cfs_cdebug_show(mask, DEBUG_SUBSYSTEM))
 		return;
 
-	CDEBUG(level, "Dump object %p extents at %s:%d, mppr: %u.\n",
+	CDEBUG(mask, "Dump object %p extents at %s:%d, mppr: %u.\n",
 	       obj, func, line, osc_cli(obj)->cl_max_pages_per_rpc);
 
 	/* osc_object_lock(obj); */
 	cnt = 1;
 	for (ext = first_extent(obj); ext != NULL; ext = next_extent(ext))
-		OSC_EXTENT_DUMP(level, ext, "in tree %d.\n", cnt++);
+		OSC_EXTENT_DUMP(mask, ext, "in tree %d.\n", cnt++);
 
 	cnt = 1;
 	list_for_each_entry(ext, &obj->oo_hp_exts, oe_link)
-		OSC_EXTENT_DUMP(level, ext, "hp %d.\n", cnt++);
+		OSC_EXTENT_DUMP(mask, ext, "hp %d.\n", cnt++);
 
 	cnt = 1;
 	list_for_each_entry(ext, &obj->oo_urgent_exts, oe_link)
-		OSC_EXTENT_DUMP(level, ext, "urgent %d.\n", cnt++);
+		OSC_EXTENT_DUMP(mask, ext, "urgent %d.\n", cnt++);
 
 	cnt = 1;
 	list_for_each_entry(ext, &obj->oo_reading_exts, oe_link)
-		OSC_EXTENT_DUMP(level, ext, "reading %d.\n", cnt++);
+		OSC_EXTENT_DUMP(mask, ext, "reading %d.\n", cnt++);
 	/* osc_object_unlock(obj); */
 }
 
@@ -1321,7 +1288,7 @@ static int osc_refresh_count(const struct lu_env *env,
 		return 0;
 	else if (cl_offset(obj, index + 1) > kms)
 		/* catch sub-page write at end of file */
-		return kms % PAGE_SIZE;
+		return kms & ~PAGE_MASK;
 	else
 		return PAGE_SIZE;
 }
@@ -1375,9 +1342,9 @@ static int osc_completion(const struct lu_env *env, struct osc_async_page *oap,
 	RETURN(0);
 }
 
-#define OSC_DUMP_GRANT(lvl, cli, fmt, args...) do {			\
+#define OSC_DUMP_GRANT(mask, cli, fmt, args...) do {			\
 	struct client_obd *__tmp = (cli);				\
-	CDEBUG(lvl, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "	\
+	CDEBUG(mask, "%s: grant { dirty: %ld/%ld dirty_pages: %ld/%lu "	\
 	       "dropped: %ld avail: %ld, dirty_grant: %ld, "		\
 	       "reserved: %ld, flight: %d } lru {in list: %ld, "	\
 	       "left: %ld, waiters: %d }" fmt "\n",			\
@@ -1398,7 +1365,6 @@ static void osc_consume_write_grant(struct client_obd *cli,
 {
 	assert_spin_locked(&cli->cl_loi_list_lock);
 	LASSERT(!(pga->flag & OBD_BRW_FROM_GRANT));
-	atomic_long_inc(&obd_dirty_pages);
 	cli->cl_dirty_pages++;
 	pga->flag |= OBD_BRW_FROM_GRANT;
 	CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
@@ -1422,11 +1388,6 @@ static void osc_release_write_grant(struct client_obd *cli,
 	pga->flag &= ~OBD_BRW_FROM_GRANT;
 	atomic_long_dec(&obd_dirty_pages);
 	cli->cl_dirty_pages--;
-	if (pga->flag & OBD_BRW_NOCACHE) {
-		pga->flag &= ~OBD_BRW_NOCACHE;
-		atomic_long_dec(&obd_dirty_transit_pages);
-		cli->cl_dirty_transit--;
-	}
 	EXIT;
 }
 
@@ -1466,13 +1427,20 @@ static void __osc_unreserve_grant(struct client_obd *cli,
 	}
 }
 
-static void osc_unreserve_grant(struct client_obd *cli,
-				unsigned int reserved, unsigned int unused)
+static void osc_unreserve_grant_nolock(struct client_obd *cli,
+				       unsigned int reserved,
+				       unsigned int unused)
 {
-	spin_lock(&cli->cl_loi_list_lock);
 	__osc_unreserve_grant(cli, reserved, unused);
 	if (unused > 0)
 		osc_wake_cache_waiters(cli);
+}
+
+static void osc_unreserve_grant(struct client_obd *cli,
+				unsigned int reserved, unsigned int unused)
+{
+	spin_lock(&cli->cl_loi_list_lock);
+	osc_unreserve_grant_nolock(cli, reserved, unused);
 	spin_unlock(&cli->cl_loi_list_lock);
 }
 
@@ -1532,7 +1500,7 @@ static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap)
  */
 static int osc_enter_cache_try(struct client_obd *cli,
 			       struct osc_async_page *oap,
-			       int bytes, int transient)
+			       int bytes)
 {
 	int rc;
 
@@ -1542,31 +1510,41 @@ static int osc_enter_cache_try(struct client_obd *cli,
 	if (rc < 0)
 		return 0;
 
-	if (cli->cl_dirty_pages < cli->cl_dirty_max_pages &&
-	    1 + atomic_long_read(&obd_dirty_pages) <= obd_max_dirty_pages) {
-		osc_consume_write_grant(cli, &oap->oap_brw_page);
-		if (transient) {
-			cli->cl_dirty_transit++;
-			atomic_long_inc(&obd_dirty_transit_pages);
-			oap->oap_brw_flags |= OBD_BRW_NOCACHE;
-		}
-		rc = 1;
-	} else {
-		__osc_unreserve_grant(cli, bytes, bytes);
-		rc = 0;
+	if (cli->cl_dirty_pages < cli->cl_dirty_max_pages) {
+		if (atomic_long_add_return(1, &obd_dirty_pages) <=
+		    obd_max_dirty_pages) {
+			osc_consume_write_grant(cli, &oap->oap_brw_page);
+			rc = 1;
+			goto out;
+		} else
+			atomic_long_dec(&obd_dirty_pages);
 	}
+	__osc_unreserve_grant(cli, bytes, bytes);
+
+out:
 	return rc;
 }
 
-static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
+/* Following two inlines exist to pass code fragments
+ * to wait_event_idle_exclusive_timeout_cmd().  Passing
+ * code fragments as macro args can look confusing, so
+ * we provide inlines to encapsulate them.
+ */
+static inline void cli_unlock_and_unplug(const struct lu_env *env,
+					 struct client_obd *cli,
+					 struct osc_async_page *oap)
 {
-	int rc;
-	spin_lock(&cli->cl_loi_list_lock);
-	rc = list_empty(&ocw->ocw_entry);
 	spin_unlock(&cli->cl_loi_list_lock);
-	return rc;
+	osc_io_unplug_async(env, cli, NULL);
+	CDEBUG(D_CACHE,
+	       "%s: sleeping for cache space for %p\n",
+	       cli_name(cli), oap);
 }
 
+static inline void cli_lock_after_unplug(struct client_obd *cli)
+{
+	spin_lock(&cli->cl_loi_list_lock);
+}
 /**
  * The main entry to reserve dirty page accounting. Usually the grant reserved
  * in this function will be freed in bulk in osc_free_grant() unless it fails
@@ -1577,15 +1555,23 @@ static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
 static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
 			   struct osc_async_page *oap, int bytes)
 {
-	struct osc_object	*osc = oap->oap_obj;
-	struct lov_oinfo	*loi = osc->oo_oinfo;
-	struct osc_cache_waiter	 ocw;
-	struct l_wait_info	 lwi;
-	int			 rc = -EDQUOT;
-	ENTRY;
+	struct osc_object *osc = oap->oap_obj;
+	struct lov_oinfo *loi = osc->oo_oinfo;
+	int rc = -EDQUOT;
+	int remain;
+	bool entered = false;
+	/* We cannot wait for a long time here since we are holding ldlm lock
+	 * across the actual IO. If no requests complete fast (e.g. due to
+	 * overloaded OST that takes a long time to process everything, we'd
+	 * get evicted if we wait for a normal obd_timeout or some such.
+	 * So we try to wait half the time it would take the client to be
+	 * evicted by server which is half obd_timeout when AT is off
+	 * or at least ldlm_enqueue_min with AT on.
+	 * See LU-13131 */
+	unsigned long timeout = cfs_time_seconds(AT_OFF ? obd_timeout / 2 :
+							  ldlm_enqueue_min / 2);
 
-	lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(AT_OFF ? obd_timeout : at_max),
-			       NULL, LWI_ON_SIGNAL_NOOP, NULL);
+	ENTRY;
 
 	OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
 
@@ -1600,76 +1586,40 @@ static int osc_enter_cache(const struct lu_env *env, struct client_obd *cli,
 		GOTO(out, rc = -EDQUOT);
 	}
 
-	/* Hopefully normal case - cache space and write credits available */
-	if (osc_enter_cache_try(cli, oap, bytes, 0)) {
-		OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
-		GOTO(out, rc = 0);
-	}
-
-	/* We can get here for two reasons: too many dirty pages in cache, or
+	/*
+	 * We can wait here for two reasons: too many dirty pages in cache, or
 	 * run out of grants. In both cases we should write dirty pages out.
 	 * Adding a cache waiter will trigger urgent write-out no matter what
 	 * RPC size will be.
-	 * The exiting condition is no avail grants and no dirty pages caching,
-	 * that really means there is no space on the OST. */
-	init_waitqueue_head(&ocw.ocw_waitq);
-	ocw.ocw_oap   = oap;
-	ocw.ocw_grant = bytes;
-	while (cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0) {
-		list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
-		ocw.ocw_rc = 0;
-		spin_unlock(&cli->cl_loi_list_lock);
-
-		osc_io_unplug_async(env, cli, NULL);
-
-		CDEBUG(D_CACHE, "%s: sleeping for cache space @ %p for %p\n",
-		       cli_name(cli), &ocw, oap);
-
-		rc = l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
-
-		spin_lock(&cli->cl_loi_list_lock);
-
-		if (rc < 0) {
-			/* l_wait_event is interrupted by signal or timed out */
-			list_del_init(&ocw.ocw_entry);
-			break;
-		}
-		LASSERT(list_empty(&ocw.ocw_entry));
-		rc = ocw.ocw_rc;
-
-		if (rc != -EDQUOT)
-			break;
-		if (osc_enter_cache_try(cli, oap, bytes, 0)) {
-			rc = 0;
-			break;
-		}
-	}
-
-	switch (rc) {
-	case 0:
-		OSC_DUMP_GRANT(D_CACHE, cli, "finally got grant space\n");
-		break;
-	case -ETIMEDOUT:
+	 * The exiting condition (other than success) is no avail grants
+	 * and no dirty pages caching, that really means there is no space
+	 * on the OST.
+	 */
+	remain = wait_event_idle_exclusive_timeout_cmd(
+		cli->cl_cache_waiters,
+		(entered = osc_enter_cache_try(cli, oap, bytes)) ||
+		(cli->cl_dirty_pages == 0 && cli->cl_w_in_flight == 0),
+		timeout,
+		cli_unlock_and_unplug(env, cli, oap),
+		cli_lock_after_unplug(cli));
+
+	if (entered) {
+		if (remain == timeout)
+			OSC_DUMP_GRANT(D_CACHE, cli, "granted from cache\n");
+		else
+			OSC_DUMP_GRANT(D_CACHE, cli,
+				       "finally got grant space\n");
+		wake_up(&cli->cl_cache_waiters);
+		rc = 0;
+	} else if (remain == 0) {
 		OSC_DUMP_GRANT(D_CACHE, cli,
 			       "timeout, fall back to sync i/o\n");
 		osc_extent_tree_dump(D_CACHE, osc);
 		/* fall back to synchronous I/O */
-		rc = -EDQUOT;
-		break;
-	case -EINTR:
-		/* Ensures restartability - LU-3581 */
-		OSC_DUMP_GRANT(D_CACHE, cli, "interrupted\n");
-		rc = -ERESTARTSYS;
-		break;
-	case -EDQUOT:
+	} else {
 		OSC_DUMP_GRANT(D_CACHE, cli,
 			       "no grant space, fall back to sync i/o\n");
-		break;
-	default:
-		CDEBUG(D_CACHE, "%s: event for cache space @ %p never arrived "
-		       "due to %d, fall back to sync i/o\n",
-		       cli_name(cli), &ocw, rc);
-		break;
+		wake_up_all(&cli->cl_cache_waiters);
 	}
 	EXIT;
 out:
@@ -1677,42 +1627,6 @@ out:
 	RETURN(rc);
 }
 
-/* caller must hold loi_list_lock */
-void osc_wake_cache_waiters(struct client_obd *cli)
-{
-	struct list_head *l, *tmp;
-	struct osc_cache_waiter *ocw;
-
-	ENTRY;
-	list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
-		ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
-		list_del_init(&ocw->ocw_entry);
-
-		ocw->ocw_rc = -EDQUOT;
-		/* we can't dirty more */
-		if ((cli->cl_dirty_pages  >= cli->cl_dirty_max_pages) ||
-		    (1 + atomic_long_read(&obd_dirty_pages) >
-		     obd_max_dirty_pages)) {
-			CDEBUG(D_CACHE, "no dirty room: dirty: %ld "
-			       "osc max %ld, sys max %ld\n",
-			       cli->cl_dirty_pages, cli->cl_dirty_max_pages,
-			       obd_max_dirty_pages);
-			goto wakeup;
-		}
-
-		if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant, 0))
-			ocw->ocw_rc = 0;
-wakeup:
-		CDEBUG(D_CACHE, "wake up %p for oap %p, avail grant %ld, %d\n",
-		       ocw, ocw->ocw_oap, cli->cl_avail_grant, ocw->ocw_rc);
-
-		wake_up(&ocw->ocw_waitq);
-	}
-
-	EXIT;
-}
-EXPORT_SYMBOL(osc_wake_cache_waiters);
-
 static int osc_max_rpc_in_flight(struct client_obd *cli, struct osc_object *osc)
 {
 	int hprpc = !!list_empty(&osc->oo_hp_exts);
@@ -1752,8 +1666,9 @@ static int osc_makes_rpc(struct client_obd *cli, struct osc_object *osc,
 		}
 		/* trigger a write rpc stream as long as there are dirtiers
 		 * waiting for space.  as they're waiting, they're not going to
-		 * create more pages to coalesce with what's waiting.. */
-		if (!list_empty(&cli->cl_cache_waiters)) {
+		 * create more pages to coalesce with what's waiting..
+		 */
+		if (waitqueue_active(&cli->cl_cache_waiters)) {
 			CDEBUG(D_CACHE, "cache waiters forcing RPC\n");
 			RETURN(1);
 		}
@@ -1882,7 +1797,6 @@ static void osc_ap_completion(const struct lu_env *env, struct client_obd *cli,
 	spin_lock(&oap->oap_lock);
 	oap->oap_async_flags = 0;
 	spin_unlock(&oap->oap_lock);
-	oap->oap_interrupted = 0;
 
 	if (oap->oap_cmd & OBD_BRW_WRITE && xid > 0) {
 		spin_lock(&cli->cl_loi_list_lock);
@@ -1915,6 +1829,34 @@ static inline unsigned osc_extent_chunks(const struct osc_extent *ext)
 	return (ext->oe_end >> ppc_bits) - (ext->oe_start >> ppc_bits) + 1;
 }
 
+static inline bool
+can_merge(const struct osc_extent *ext, const struct osc_extent *in_rpc)
+{
+	if (ext->oe_no_merge || in_rpc->oe_no_merge)
+		return false;
+
+	if (ext->oe_srvlock != in_rpc->oe_srvlock)
+		return false;
+
+	if (ext->oe_ndelay != in_rpc->oe_ndelay)
+		return false;
+
+	if (!ext->oe_grants != !in_rpc->oe_grants)
+		return false;
+
+	if (ext->oe_dio != in_rpc->oe_dio)
+		return false;
+
+	/* It's possible to have overlap on DIO */
+	if (in_rpc->oe_dio && overlapped(ext, in_rpc))
+		return false;
+
+	if (ext->oe_is_rdma_only != in_rpc->oe_is_rdma_only)
+		return false;
+
+	return true;
+}
+
 /**
  * Try to add extent to one RPC. We need to think about the following things:
  * - # of pages must not be over max_pages_per_rpc
@@ -1926,9 +1868,6 @@ static int try_to_add_extent_for_io(struct client_obd *cli,
 {
 	struct osc_extent *tmp;
 	unsigned int chunk_count;
-	struct osc_async_page *oap = list_first_entry(&ext->oe_pages,
-						      struct osc_async_page,
-						      oap_pending_item);
 	ENTRY;
 
 	EASSERT((ext->oe_state == OES_CACHE || ext->oe_state == OES_LOCK_DONE),
@@ -1957,30 +1896,10 @@ static int try_to_add_extent_for_io(struct client_obd *cli,
 		RETURN(0);
 
 	list_for_each_entry(tmp, data->erd_rpc_list, oe_link) {
-		struct osc_async_page *oap2;
-		oap2 = list_first_entry(&tmp->oe_pages, struct osc_async_page,
-					oap_pending_item);
 		EASSERT(tmp->oe_owner == current, tmp);
-#if 0
-		if (overlapped(tmp, ext)) {
-			OSC_EXTENT_DUMP(D_ERROR, tmp, "overlapped %p.\n", ext);
-			EASSERT(0, ext);
-		}
-#endif
-		if (oap2cl_page(oap)->cp_type != oap2cl_page(oap2)->cp_type) {
-			CDEBUG(D_CACHE, "Do not permit different types of IO "
-			       "in one RPC\n");
-			RETURN(0);
-		}
 
-		if (tmp->oe_srvlock != ext->oe_srvlock ||
-		    !tmp->oe_grants != !ext->oe_grants ||
-		    tmp->oe_ndelay != ext->oe_ndelay ||
-		    tmp->oe_no_merge || ext->oe_no_merge)
+		if (!can_merge(ext, tmp))
 			RETURN(0);
-
-		/* remove break for strict check */
-		break;
 	}
 
 	data->erd_max_extents--;
@@ -1991,36 +1910,6 @@ static int try_to_add_extent_for_io(struct client_obd *cli,
 	RETURN(1);
 }
 
-static inline unsigned osc_max_write_chunks(const struct client_obd *cli)
-{
-	/*
-	 * LU-8135:
-	 *
-	 * The maximum size of a single transaction is about 64MB in ZFS.
-	 * #define DMU_MAX_ACCESS (64 * 1024 * 1024)
-	 *
-	 * Since ZFS is a copy-on-write file system, a single dirty page in
-	 * a chunk will result in the rewrite of the whole chunk, therefore
-	 * an RPC shouldn't be allowed to contain too many chunks otherwise
-	 * it will make transaction size much bigger than 64MB, especially
-	 * with big block size for ZFS.
-	 *
-	 * This piece of code is to make sure that OSC won't send write RPCs
-	 * with too many chunks. The maximum chunk size that an RPC can cover
-	 * is set to PTLRPC_MAX_BRW_SIZE, which is defined to 16MB. Ideally
-	 * OST should tell the client what the biggest transaction size is,
-	 * but it's good enough for now.
-	 *
-	 * This limitation doesn't apply to ldiskfs, which allows as many
-	 * chunks in one RPC as we want. However, it won't have any benefits
-	 * to have too many discontiguous pages in one RPC.
-	 *
-	 * An osc_extent won't cover over a RPC size, so the chunks in an
-	 * osc_extent won't bigger than PTLRPC_MAX_BRW_SIZE >> chunkbits.
-	 */
-	return PTLRPC_MAX_BRW_SIZE >> cli->cl_chunkbits;
-}
-
 /**
  * In order to prevent multiple ptlrpcd from breaking contiguous extents,
  * get_write_extent() takes all appropriate extents in atomic.
@@ -2047,11 +1936,10 @@ static unsigned int get_write_extents(struct osc_object *obj,
 		.erd_max_extents = 256,
 	};
 
-	LASSERT(osc_object_is_locked(obj));
+	assert_osc_object_is_locked(obj);
 	while (!list_empty(&obj->oo_hp_exts)) {
 		ext = list_entry(obj->oo_hp_exts.next, struct osc_extent,
 				 oe_link);
-		LASSERT(ext->oe_state == OES_CACHE);
 		if (!try_to_add_extent_for_io(cli, ext, &data))
 			return data.erd_page_count;
 		EASSERT(ext->oe_nr_pages <= data.erd_max_pages, ext);
@@ -2082,19 +1970,16 @@ static unsigned int get_write_extents(struct osc_object *obj,
 	if (data.erd_page_count == data.erd_max_pages)
 		return data.erd_page_count;
 
-	ext = first_extent(obj);
-	while (ext != NULL) {
+	for (ext = first_extent(obj);
+	     ext;
+	     ext = next_extent(ext)) {
 		if ((ext->oe_state != OES_CACHE) ||
 		    /* this extent may be already in current rpclist */
-		    (!list_empty(&ext->oe_link) && ext->oe_owner != NULL)) {
-			ext = next_extent(ext);
+		    (!list_empty(&ext->oe_link) && ext->oe_owner))
 			continue;
-		}
 
 		if (!try_to_add_extent_for_io(cli, ext, &data))
 			return data.erd_page_count;
-
-		ext = next_extent(ext);
 	}
 	return data.erd_page_count;
 }
@@ -2104,7 +1989,7 @@ osc_send_write_rpc(const struct lu_env *env, struct client_obd *cli,
 		   struct osc_object *osc)
 __must_hold(osc)
 {
-	struct list_head   rpclist = LIST_HEAD_INIT(rpclist);
+	LIST_HEAD(rpclist);
 	struct osc_extent *ext;
 	struct osc_extent *tmp;
 	struct osc_extent *first = NULL;
@@ -2113,7 +1998,7 @@ __must_hold(osc)
 	int rc = 0;
 	ENTRY;
 
-	LASSERT(osc_object_is_locked(osc));
+	assert_osc_object_is_locked(osc);
 
 	page_count = get_write_extents(osc, &rpclist);
 	LASSERT(equi(page_count == 0, list_empty(&rpclist)));
@@ -2180,7 +2065,7 @@ __must_hold(osc)
 {
 	struct osc_extent *ext;
 	struct osc_extent *next;
-	struct list_head rpclist = LIST_HEAD_INIT(rpclist);
+	LIST_HEAD(rpclist);
 	struct extent_rpc_data data = {
 		.erd_rpc_list	= &rpclist,
 		.erd_page_count	= 0,
@@ -2191,7 +2076,7 @@ __must_hold(osc)
 	int rc = 0;
 	ENTRY;
 
-	LASSERT(osc_object_is_locked(osc));
+	assert_osc_object_is_locked(osc);
 	list_for_each_entry_safe(ext, next, &osc->oo_reading_exts, oe_link) {
 		EASSERT(ext->oe_state == OES_LOCK_DONE, ext);
 		if (!try_to_add_extent_for_io(cli, ext, &data))
@@ -2237,8 +2122,9 @@ static struct osc_object *osc_next_obj(struct client_obd *cli)
 	/* then if we have cache waiters, return all objects with queued
 	 * writes.  This is especially important when many small files
 	 * have filled up the cache and not been fired into rpcs because
-	 * they don't pass the nr_pending/object threshhold */
-	if (!list_empty(&cli->cl_cache_waiters) &&
+	 * they don't pass the nr_pending/object threshhold
+	 */
+	if (waitqueue_active(&cli->cl_cache_waiters) &&
 	    !list_empty(&cli->cl_loi_write_list))
 		RETURN(list_to_obj(&cli->cl_loi_write_list, write_item));
 
@@ -2269,7 +2155,12 @@ __must_hold(&cli->cl_loi_list_lock)
 
 		OSC_IO_DEBUG(osc, "%lu in flight\n", rpcs_in_flight(cli));
 
-		if (osc_max_rpc_in_flight(cli, osc)) {
+		/* even if we have reached our max in flight RPCs, we still
+		 * allow all high-priority RPCs through to prevent their
+		 * starvation and leading to server evicting us for not
+		 * writing out pages in a timely manner LU-13131 */
+		if (osc_max_rpc_in_flight(cli, osc) &&
+		    list_empty(&osc->oo_hp_exts)) {
 			__osc_list_maint(cli, osc);
 			break;
 		}
@@ -2363,9 +2254,6 @@ int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
 	oap->oap_obj_off = offset;
 	LASSERT(!(offset & ~PAGE_MASK));
 
-	if (cfs_capable(CFS_CAP_SYS_RESOURCE))
-		oap->oap_brw_flags = OBD_BRW_NOQUOTA;
-
 	INIT_LIST_HEAD(&oap->oap_pending_item);
 	INIT_LIST_HEAD(&oap->oap_rpc_item);
 
@@ -2377,13 +2265,14 @@ int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
 EXPORT_SYMBOL(osc_prep_async_page);
 
 int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
-		       struct osc_page *ops)
+		       struct osc_page *ops, cl_commit_cbt cb)
 {
 	struct osc_io *oio = osc_env_io(env);
 	struct osc_extent     *ext = NULL;
 	struct osc_async_page *oap = &ops->ops_oap;
 	struct client_obd     *cli = oap->oap_cli;
 	struct osc_object     *osc = oap->oap_obj;
+	struct pagevec        *pvec = &osc_env_info(env)->oti_pagevec;
 	pgoff_t index;
 	unsigned int tmp;
 	unsigned int grants = 0;
@@ -2405,7 +2294,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 
 	/* Set the OBD_BRW_SRVLOCK before the page is queued. */
 	brw_flags |= ops->ops_srvlock ? OBD_BRW_SRVLOCK : 0;
-	if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
+	if (oio->oi_cap_sys_resource || io->ci_noquota) {
 		brw_flags |= OBD_BRW_NOQUOTA;
 		cmd |= OBD_BRW_NOQUOTA;
 	}
@@ -2426,7 +2315,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 		qid[USRQUOTA] = attr->cat_uid;
 		qid[GRPQUOTA] = attr->cat_gid;
 		qid[PRJQUOTA] = attr->cat_projid;
-		if (rc == 0 && osc_quota_chkdq(cli, qid) == NO_QUOTA)
+		if (rc == 0 && osc_quota_chkdq(cli, qid) == -EDQUOT)
 			rc = -EDQUOT;
 		if (rc)
 			RETURN(rc);
@@ -2434,7 +2323,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 
 	oap->oap_cmd = cmd;
 	oap->oap_page_off = ops->ops_from;
-	oap->oap_count = ops->ops_to - ops->ops_from;
+	oap->oap_count = ops->ops_to - ops->ops_from + 1;
 	/* No need to hold a lock here,
 	 * since this page is not in any list yet. */
 	oap->oap_async_flags = 0;
@@ -2461,8 +2350,7 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 
 		/* it doesn't need any grant to dirty this page */
 		spin_lock(&cli->cl_loi_list_lock);
-		rc = osc_enter_cache_try(cli, oap, grants, 0);
-		spin_unlock(&cli->cl_loi_list_lock);
+		rc = osc_enter_cache_try(cli, oap, grants);
 		if (rc == 0) { /* try failed */
 			grants = 0;
 			need_release = 1;
@@ -2476,10 +2364,11 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 			} else {
 				OSC_EXTENT_DUMP(D_CACHE, ext,
 						"expanded for %lu.\n", index);
-				osc_unreserve_grant(cli, grants, tmp);
+				osc_unreserve_grant_nolock(cli, grants, tmp);
 				grants = 0;
 			}
 		}
+		spin_unlock(&cli->cl_loi_list_lock);
 		rc = 0;
 	} else if (ext != NULL) {
 		/* index is located outside of active extent */
@@ -2502,7 +2391,14 @@ int osc_queue_async_io(const struct lu_env *env, struct cl_io *io,
 
 		rc = 0;
 		if (grants == 0) {
-			/* we haven't allocated grant for this page. */
+			/* We haven't allocated grant for this page, and we
+			 * must not hold a page lock while we do enter_cache,
+			 * so we must mark dirty & unlock any pages in the
+			 * write commit pagevec. */
+			if (pagevec_count(pvec)) {
+				cb(env, io, pvec);
+				pagevec_reinit(pvec);
+			}
 			rc = osc_enter_cache(env, cli, oap, tmp);
 			if (rc == 0)
 				grants = tmp;
@@ -2641,7 +2537,7 @@ int osc_flush_async_page(const struct lu_env *env, struct cl_io *io,
 	oap->oap_async_flags |= ASYNC_READY|ASYNC_URGENT;
 	spin_unlock(&oap->oap_lock);
 
-	if (memory_pressure_get())
+	if (current->flags & PF_MEMALLOC)
 		ext->oe_memalloc = 1;
 
 	ext->oe_urgent = 1;
@@ -2663,71 +2559,6 @@ out:
 	return rc;
 }
 
-/**
- * this is called when a sync waiter receives an interruption.  Its job is to
- * get the caller woken as soon as possible.  If its page hasn't been put in an
- * rpc yet it can dequeue immediately.  Otherwise it has to mark the rpc as
- * desiring interruption which will forcefully complete the rpc once the rpc
- * has timed out.
- */
-int osc_cancel_async_page(const struct lu_env *env, struct osc_page *ops)
-{
-	struct osc_async_page *oap = &ops->ops_oap;
-	struct osc_object     *obj = oap->oap_obj;
-	struct client_obd     *cli = osc_cli(obj);
-	struct osc_extent     *ext;
-	struct osc_extent     *found = NULL;
-	struct list_head            *plist;
-	pgoff_t index = osc_index(ops);
-	int     rc = -EBUSY;
-	int     cmd;
-	ENTRY;
-
-	LASSERT(!oap->oap_interrupted);
-	oap->oap_interrupted = 1;
-
-	/* Find out the caching extent */
-	osc_object_lock(obj);
-	if (oap->oap_cmd & OBD_BRW_WRITE) {
-		plist = &obj->oo_urgent_exts;
-		cmd   = OBD_BRW_WRITE;
-	} else {
-		plist = &obj->oo_reading_exts;
-		cmd   = OBD_BRW_READ;
-	}
-	list_for_each_entry(ext, plist, oe_link) {
-		if (ext->oe_start <= index && ext->oe_end >= index) {
-			LASSERT(ext->oe_state == OES_LOCK_DONE);
-			/* For OES_LOCK_DONE state extent, it has already held
-			 * a refcount for RPC. */
-			found = osc_extent_get(ext);
-			break;
-		}
-	}
-	if (found != NULL) {
-		list_del_init(&found->oe_link);
-		osc_update_pending(obj, cmd, -found->oe_nr_pages);
-		osc_object_unlock(obj);
-
-		osc_extent_finish(env, found, 0, -EINTR);
-		osc_extent_put(env, found);
-		rc = 0;
-	} else {
-		osc_object_unlock(obj);
-		/* ok, it's been put in an rpc. only one oap gets a request
-		 * reference */
-		if (oap->oap_request != NULL) {
-			ptlrpc_mark_interrupted(oap->oap_request);
-			ptlrpcd_wake(oap->oap_request);
-			ptlrpc_req_finished(oap->oap_request);
-			oap->oap_request = NULL;
-		}
-	}
-
-	osc_list_maint(cli, obj);
-	RETURN(rc);
-}
-
 int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
 			 struct osc_object *obj, struct list_head *list,
 			 int brw_flags)
@@ -2753,7 +2584,8 @@ int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
 		++page_count;
 		mppr <<= (page_count > mppr);
 
-		if (unlikely(opg->ops_from > 0 || opg->ops_to < PAGE_SIZE))
+		if (unlikely(opg->ops_from > 0 ||
+			     opg->ops_to < PAGE_SIZE - 1))
 			can_merge = false;
 	}
 
@@ -2777,6 +2609,30 @@ int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
 	ext->oe_obj = obj;
 	ext->oe_srvlock = !!(brw_flags & OBD_BRW_SRVLOCK);
 	ext->oe_ndelay = !!(brw_flags & OBD_BRW_NDELAY);
+	ext->oe_dio = !!(brw_flags & OBD_BRW_NOCACHE);
+	if (ext->oe_dio && !ext->oe_rw) { /* direct io write */
+		int grants;
+		int ppc;
+
+		ppc = 1 << (cli->cl_chunkbits - PAGE_SHIFT);
+		grants = cli->cl_grant_extent_tax;
+		grants += (1 << cli->cl_chunkbits) *
+			((page_count + ppc - 1) / ppc);
+
+		spin_lock(&cli->cl_loi_list_lock);
+		if (osc_reserve_grant(cli, grants) == 0) {
+			list_for_each_entry(oap, list, oap_pending_item) {
+				osc_consume_write_grant(cli,
+							&oap->oap_brw_page);
+				atomic_long_inc(&obd_dirty_pages);
+			}
+			osc_unreserve_grant_nolock(cli, grants, 0);
+			ext->oe_grants = grants;
+		}
+		spin_unlock(&cli->cl_loi_list_lock);
+	}
+
+	ext->oe_is_rdma_only = !!(brw_flags & OBD_BRW_RDMA_ONLY);
 	ext->oe_nr_pages = page_count;
 	ext->oe_mppr = mppr;
 	list_splice_init(list, &ext->oe_pages);
@@ -2786,7 +2642,21 @@ int osc_queue_sync_pages(const struct lu_env *env, const struct cl_io *io,
 	/* Reuse the initial refcount for RPC, don't drop it */
 	osc_extent_state_set(ext, OES_LOCK_DONE);
 	if (!ext->oe_rw) { /* write */
-		list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
+		if (!ext->oe_srvlock && !ext->oe_dio) {
+			/* The most likely case here is from lack of grants
+			 * so we are either out of quota or out of space.
+			 * Since this means we are holding locks across
+			 * potentially multi-striped IO, we must send out
+			 * everything out instantly to avoid prolonged
+			 * waits resulting in lock eviction (likely since
+			 * the extended wait in osc_cache_enter() did not
+			 * yield any additional grant due to a timeout.
+			 * LU-13131 */
+			ext->oe_hp = 1;
+			list_add_tail(&ext->oe_link, &obj->oo_hp_exts);
+		} else {
+			list_add_tail(&ext->oe_link, &obj->oo_urgent_exts);
+		}
 		osc_update_pending(obj, OBD_BRW_WRITE, page_count);
 	} else {
 		list_add_tail(&ext->oe_link, &obj->oo_reading_exts);
@@ -2808,7 +2678,7 @@ int osc_cache_truncate_start(const struct lu_env *env, struct osc_object *obj,
 	struct osc_extent *ext;
 	struct osc_extent *waiting = NULL;
 	pgoff_t index;
-	struct list_head list = LIST_HEAD_INIT(list);
+	LIST_HEAD(list);
 	int result = 0;
 	bool partial;
 	ENTRY;
@@ -3026,7 +2896,7 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
 			      pgoff_t start, pgoff_t end, int hp, int discard)
 {
 	struct osc_extent *ext;
-	struct list_head discard_list = LIST_HEAD_INIT(discard_list);
+	LIST_HEAD(discard_list);
 	bool unplug = false;
 	int result = 0;
 	ENTRY;
@@ -3051,7 +2921,7 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
 					EASSERT(!ext->oe_hp, ext);
 					ext->oe_hp = 1;
 					list = &obj->oo_hp_exts;
-				} else if (!ext->oe_urgent) {
+				} else if (!ext->oe_urgent && !ext->oe_hp) {
 					ext->oe_urgent = 1;
 					list = &obj->oo_urgent_exts;
 				}
@@ -3059,10 +2929,25 @@ int osc_cache_writeback_range(const struct lu_env *env, struct osc_object *obj,
 					list_move_tail(&ext->oe_link, list);
 				unplug = true;
 			} else {
+				struct client_obd *cli = osc_cli(obj);
+				int pcc_bits = cli->cl_chunkbits - PAGE_SHIFT;
+				pgoff_t align_by = (1 << pcc_bits);
+				pgoff_t a_start = round_down(start, align_by);
+				pgoff_t a_end = round_up(end, align_by);
+
+				/* overflow case */
+				if (end && !a_end)
+					a_end = CL_PAGE_EOF;
 				/* the only discarder is lock cancelling, so
-				 * [start, end] must contain this extent */
-				EASSERT(ext->oe_start >= start &&
-					ext->oe_max_end <= end, ext);
+				 * [start, end], aligned by chunk size, must
+				 * contain this extent */
+				LASSERTF(ext->oe_start >= a_start &&
+					 ext->oe_end <= a_end,
+					 "ext [%lu, %lu] reg [%lu, %lu] "
+					 "orig [%lu %lu] align %lu bits "
+					 "%d\n", ext->oe_start, ext->oe_end,
+					 a_start, a_end, start, end,
+					 align_by, pcc_bits);
 				osc_extent_state_set(ext, OES_LOCKING);
 				ext->oe_owner = current;
 				list_move_tail(&ext->oe_link,
@@ -3133,18 +3018,14 @@ EXPORT_SYMBOL(osc_cache_writeback_range);
 /**
  * Returns a list of pages by a given [start, end] of \a obj.
  *
- * \param resched If not NULL, then we give up before hogging CPU for too
- * long and set *resched = 1, in that case caller should implement a retry
- * logic.
- *
  * Gang tree lookup (radix_tree_gang_lookup()) optimization is absolutely
  * crucial in the face of [offset, EOF] locks.
  *
  * Return at least one page in @queue unless there is no covered page.
  */
-int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
-			struct osc_object *osc, pgoff_t start, pgoff_t end,
-			osc_page_gang_cbt cb, void *cbdata)
+bool osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
+			  struct osc_object *osc, pgoff_t start, pgoff_t end,
+			  osc_page_gang_cbt cb, void *cbdata)
 {
 	struct osc_page *ops;
 	struct pagevec	*pagevec;
@@ -3153,7 +3034,7 @@ int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
 	unsigned int    nr;
 	unsigned int    i;
 	unsigned int    j;
-	int             res = CLP_GANG_OKAY;
+	bool            res = true;
 	bool            tree_lock = true;
 	ENTRY;
 
@@ -3202,7 +3083,7 @@ int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
 
 		for (i = 0; i < j; ++i) {
 			ops = pvec[i];
-			if (res == CLP_GANG_OKAY)
+			if (res)
 				res = (*cb)(env, io, ops, cbdata);
 
 			page = ops->ops_cl.cpl_page;
@@ -3214,10 +3095,10 @@ int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
 		if (nr < OTI_PVEC_SIZE || end_of_region)
 			break;
 
-		if (res == CLP_GANG_OKAY && need_resched())
-			res = CLP_GANG_RESCHED;
-		if (res != CLP_GANG_OKAY)
+		if (!res)
 			break;
+		if (need_resched())
+			cond_resched();
 
 		spin_lock(&osc->oo_tree_lock);
 		tree_lock = true;
@@ -3231,7 +3112,7 @@ EXPORT_SYMBOL(osc_page_gang_lookup);
 /**
  * Check if page @page is covered by an extra lock or discard it.
  */
-static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
+static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
 				struct osc_page *ops, void *cbdata)
 {
 	struct osc_thread_info *info = osc_env_info(env);
@@ -3266,10 +3147,10 @@ static int check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
 	}
 
 	info->oti_next_index = index + 1;
-	return CLP_GANG_OKAY;
+	return true;
 }
 
-int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
+bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
 		   struct osc_page *ops, void *cbdata)
 {
 	struct osc_thread_info *info = osc_env_info(env);
@@ -3290,7 +3171,7 @@ int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
 		LASSERT(page->cp_state == CPS_FREEING);
 	}
 
-	return CLP_GANG_OKAY;
+	return true;
 }
 EXPORT_SYMBOL(osc_discard_cb);
 
@@ -3308,7 +3189,6 @@ int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
 	struct osc_thread_info *info = osc_env_info(env);
 	struct cl_io *io = osc_env_thread_io(env);
 	osc_page_gang_cbt cb;
-	int res;
 	int result;
 
 	ENTRY;
@@ -3321,15 +3201,9 @@ int osc_lock_discard_pages(const struct lu_env *env, struct osc_object *osc,
 
 	cb = discard ? osc_discard_cb : check_and_discard_cb;
 	info->oti_fn_index = info->oti_next_index = start;
-	do {
-		res = osc_page_gang_lookup(env, io, osc,
-					   info->oti_next_index, end, cb, osc);
-		if (info->oti_next_index > end)
-			break;
 
-		if (res == CLP_GANG_RESCHED)
-			cond_resched();
-	} while (res != CLP_GANG_OKAY);
+	osc_page_gang_lookup(env, io, osc,
+			     info->oti_next_index, end, cb, osc);
 out:
 	cl_io_fini(env, io);
 	RETURN(result);