X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosc%2Fosc_cache.c;h=658e8f419fbc56f352e77b833510eebca2ac74ce;hp=ff0652fd317e48fe49d7fff401168495b1864fce;hb=a4fbe7341b;hpb=a7ff5d050ee7db0e80baac5fb3848ffcfa04dea6

diff --git a/lustre/osc/osc_cache.c b/lustre/osc/osc_cache.c
index ff0652f..658e8f4 100644
--- a/lustre/osc/osc_cache.c
+++ b/lustre/osc/osc_cache.c
@@ -28,7 +28,6 @@
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
  *
  * osc cache management.
  *
@@ -38,6 +37,7 @@
 #define DEBUG_SUBSYSTEM S_OSC
 
 #include <lustre_osc.h>
+#include <lustre_dlm.h>
 
 #include "osc_internal.h"
 
@@ -74,7 +74,7 @@ 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';
@@ -94,14 +94,9 @@ 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[] = {
+static const char *const oes_strings[] = {
 	"inv", "active", "cache", "locking", "lockdone", "rpc", "trunc", NULL };
 
 #define OSC_EXTENT_DUMP_WITH_LOC(file, func, line, mask, extent, fmt, ...) do {\
@@ -115,7 +110,7 @@ static const char *oes_strings[] = {
 		/* ----- extent part 0 ----- */				      \
 		__ext, EXTPARA(__ext),					      \
 		/* ----- part 1 ----- */				      \
-		atomic_read(&__ext->oe_refc),				      \
+		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),	      \
@@ -151,10 +146,7 @@ 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)
@@ -162,7 +154,7 @@ 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));
 }
 
@@ -171,7 +163,7 @@ 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));
 }
 
@@ -189,16 +181,15 @@ 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) {
@@ -286,6 +277,11 @@ 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.
@@ -295,7 +291,7 @@ 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;
@@ -303,8 +299,7 @@ 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;
@@ -312,15 +307,15 @@ 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)
@@ -334,7 +329,7 @@ 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;
@@ -345,35 +340,50 @@ 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);
 	}
 }
 
@@ -384,9 +394,9 @@ 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);
 }
 
 /**
@@ -399,7 +409,7 @@ 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) {
@@ -436,9 +446,9 @@ 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;
@@ -453,17 +463,16 @@ 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);
 	}
@@ -473,7 +482,7 @@ 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);
@@ -486,7 +495,7 @@ 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);
@@ -517,16 +526,25 @@ 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;
 
-	if (victim->oe_state != OES_CACHE || victim->oe_fsync_wait)
+	if (victim->oe_state != OES_INV &&
+	    (victim->oe_state != OES_CACHE || victim->oe_fsync_wait))
 		return -EBUSY;
 
 	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;
@@ -552,7 +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);
@@ -566,11 +583,10 @@ static int osc_extent_merge(const struct lu_env *env, struct osc_extent *cur,
 /**
  * Drop user count of osc_extent, and unplug IO asynchronously.
  */
-int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
+void osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 {
 	struct osc_object *obj = ext->oe_obj;
 	struct client_obd *cli = osc_cli(obj);
-	int rc = 0;
 	ENTRY;
 
 	LASSERT(atomic_read(&ext->oe_users) > 0);
@@ -599,7 +615,10 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 			if (osc_extent_merge(env, ext, next_extent(ext)) == 0)
 				grant += cli->cl_grant_extent_tax;
 
-			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) {
@@ -614,12 +633,8 @@ int osc_extent_release(const struct lu_env *env, struct osc_extent *ext)
 		osc_io_unplug_async(env, cli, obj);
 	}
 	osc_extent_put(env, 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);
+	RETURN_EXIT;
 }
 
 /**
@@ -682,7 +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);
@@ -699,11 +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)
@@ -714,15 +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
@@ -747,68 +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);
@@ -822,7 +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;
 
@@ -914,17 +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.
  */
@@ -953,14 +912,16 @@ 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 = wait_event_idle_timeout(ext->oe_waitq, extent_wait_cb(ext, state),
+	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);
 
-		wait_event_idle(ext->oe_waitq, extent_wait_cb(ext, state));
+		wait_event_idle(ext->oe_waitq,
+				smp_load_acquire(&ext->oe_state) == state);
 	}
 	if (ext->oe_rc < 0)
 		rc = ext->oe_rc;
@@ -1149,7 +1110,8 @@ static int osc_extent_make_ready(const struct lu_env *env,
 	 * the size of file. */
 	if (!(last->oap_async_flags & ASYNC_COUNT_STABLE)) {
 		int last_oap_count = osc_refresh_count(env, last, OBD_BRW_WRITE);
-		LASSERT(last_oap_count > 0);
+		LASSERTF(last_oap_count > 0,
+			 "last_oap_count %d\n", last_oap_count);
 		LASSERT(last->oap_page_off + last_oap_count <= PAGE_SIZE);
 		last->oap_count = last_oap_count;
 		spin_lock(&last->oap_lock);
@@ -1327,7 +1289,7 @@ 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;
 }
@@ -1408,7 +1370,6 @@ static void osc_consume_write_grant(struct client_obd *cli,
 	pga->flag |= OBD_BRW_FROM_GRANT;
 	CDEBUG(D_CACHE, "using %lu grant credits for brw %p page %p\n",
 	       PAGE_SIZE, pga, pga->pg);
-	osc_update_next_shrink(cli);
 }
 
 /* the companion to osc_consume_write_grant, called when a brw has completed.
@@ -1564,15 +1525,26 @@ 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
@@ -1583,10 +1555,22 @@ 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;
-	int			 rc = -EDQUOT;
+	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);
+
 	ENTRY;
 
 	OSC_DUMP_GRANT(D_CACHE, cli, "need:%d\n", bytes);
@@ -1602,83 +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 (list_empty(&cli->cl_cache_waiters) &&
-	    osc_enter_cache_try(cli, oap, bytes)) {
-		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 = wait_event_idle_timeout(ocw.ocw_waitq,
-					     ocw_granted(cli, &ocw),
-					     cfs_time_seconds(AT_OFF ?
-							      obd_timeout :
-							      at_max));
-
-		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);
-			if (rc == 0)
-				rc = -ETIMEDOUT;
-			break;
-		}
-		LASSERT(list_empty(&ocw.ocw_entry));
-		rc = ocw.ocw_rc;
-
-		if (rc != -EDQUOT)
-			break;
-		if (osc_enter_cache_try(cli, oap, bytes)) {
-			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:
@@ -1686,36 +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);
-
-		ocw->ocw_rc = -EDQUOT;
-
-		if (osc_enter_cache_try(cli, ocw->ocw_oap, ocw->ocw_grant))
-			ocw->ocw_rc = 0;
-
-		if (ocw->ocw_rc == 0 ||
-		    !(cli->cl_dirty_pages > 0 || cli->cl_w_in_flight > 0)) {
-			list_del_init(&ocw->ocw_entry);
-			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);
@@ -1755,8 +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);
 		}
@@ -1885,7 +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);
@@ -1918,6 +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
@@ -1929,9 +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),
@@ -1960,30 +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--;
@@ -2020,11 +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);
@@ -2055,19 +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;
 }
@@ -2086,7 +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)));
@@ -2164,7 +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))
@@ -2210,8 +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));
 
@@ -2242,7 +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;
 		}
@@ -2319,10 +2237,11 @@ int osc_io_unplug0(const struct lu_env *env, struct client_obd *cli,
 EXPORT_SYMBOL(osc_io_unplug0);
 
 int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
-			struct page *page, loff_t offset)
+			struct cl_page *page, loff_t offset)
 {
 	struct obd_export     *exp = osc_export(osc);
 	struct osc_async_page *oap = &ops->ops_oap;
+	struct page	      *vmpage = page->cp_vmpage;
 	ENTRY;
 
 	if (!page)
@@ -2332,16 +2251,24 @@ int osc_prep_async_page(struct osc_object *osc, struct osc_page *ops,
 	oap->oap_cli = &exp->exp_obd->u.cli;
 	oap->oap_obj = osc;
 
-	oap->oap_page = page;
+	oap->oap_page = vmpage;
 	oap->oap_obj_off = offset;
 	LASSERT(!(offset & ~PAGE_MASK));
 
+	/* Count of transient (direct i/o) pages is always stable by the time
+	 * they're submitted.  Setting this here lets us avoid calling
+	 * cl_page_clip later to set this.
+	 */
+	if (page->cp_type == CPT_TRANSIENT)
+		oap->oap_async_flags |= ASYNC_COUNT_STABLE|ASYNC_URGENT|
+					ASYNC_READY;
+
 	INIT_LIST_HEAD(&oap->oap_pending_item);
 	INIT_LIST_HEAD(&oap->oap_rpc_item);
 
 	spin_lock_init(&oap->oap_lock);
-	CDEBUG(D_INFO, "oap %p page %p obj off %llu\n",
-	       oap, page, oap->oap_obj_off);
+	CDEBUG(D_INFO, "oap %p vmpage %p obj off %llu\n",
+	       oap, vmpage, oap->oap_obj_off);
 	RETURN(0);
 }
 EXPORT_SYMBOL(osc_prep_async_page);
@@ -2376,13 +2303,13 @@ 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 (oio->oi_cap_sys_resource) {
+	if (oio->oi_cap_sys_resource || io->ci_noquota) {
 		brw_flags |= OBD_BRW_NOQUOTA;
 		cmd |= OBD_BRW_NOQUOTA;
 	}
 
 	/* check if the file's owner/group is over quota */
-	if (!(cmd & OBD_BRW_NOQUOTA)) {
+	if (!io->ci_noquota) {
 		struct cl_object *obj;
 		struct cl_attr   *attr;
 		unsigned int qid[LL_MAXQUOTAS];
@@ -2397,7 +2324,7 @@ 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);
@@ -2405,7 +2332,7 @@ 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;
@@ -2619,7 +2546,7 @@ 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;
@@ -2641,72 +2568,7 @@ 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,
+int osc_queue_sync_pages(const struct lu_env *env, struct cl_io *io,
 			 struct osc_object *obj, struct list_head *list,
 			 int brw_flags)
 {
@@ -2731,7 +2593,8 @@ 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;
 	}
 
@@ -2755,6 +2618,40 @@ 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);
+
+		CDEBUG(D_CACHE, "requesting %d bytes grant\n", grants);
+		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;
+		} else {
+			/* We cannot report ENOSPC correctly if we do parallel
+			 * DIO (async RPC submission), so turn off parallel dio
+			 * if there is not sufficient grant available.  This
+			 * makes individual RPCs synchronous.
+			 */
+			io->ci_parallel_dio = false;
+			CDEBUG(D_CACHE,
+			"not enough grant available, switching to sync for this i/o\n");
+		}
+		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);
@@ -2764,7 +2661,21 @@ 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);
@@ -3029,7 +2940,7 @@ 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;
 				}
@@ -3126,18 +3037,14 @@ 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;
@@ -3146,7 +3053,7 @@ 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;
 
@@ -3193,11 +3100,10 @@ int osc_page_gang_lookup(const struct lu_env *env, struct cl_io *io,
 		spin_unlock(&osc->oo_tree_lock);
 		tree_lock = false;
 
+		res = (*cb)(env, io, pvec, j, cbdata);
+
 		for (i = 0; i < j; ++i) {
 			ops = pvec[i];
-			if (res == CLP_GANG_OKAY)
-				res = (*cb)(env, io, ops, cbdata);
-
 			page = ops->ops_cl.cpl_page;
 			lu_ref_del(&page->cp_reference, "gang_lookup", current);
 			cl_pagevec_put(env, page, pagevec);
@@ -3207,11 +3113,20 @@ 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 (io->ci_type == CIT_MISC &&
+		    io->u.ci_misc.lm_next_rpc_time &&
+		    ktime_get_seconds() > io->u.ci_misc.lm_next_rpc_time) {
+			osc_send_empty_rpc(osc, idx << PAGE_SHIFT);
+			io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
+							 5 * obd_timeout / 16;
+		}
+
+		if (need_resched())
+			cond_resched();
+
 		spin_lock(&osc->oo_tree_lock);
 		tree_lock = true;
 	}
@@ -3224,66 +3139,105 @@ 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,
-				struct osc_page *ops, void *cbdata)
+static bool check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
+				 void **pvec, int count, void *cbdata)
 {
 	struct osc_thread_info *info = osc_env_info(env);
 	struct osc_object *osc = cbdata;
-	pgoff_t index;
+	int i;
 
-	index = osc_index(ops);
-	if (index >= info->oti_fn_index) {
-		struct ldlm_lock *tmp;
+	for (i = 0; i < count; i++) {
+		struct osc_page *ops = pvec[i];
 		struct cl_page *page = ops->ops_cl.cpl_page;
+		pgoff_t index = osc_index(ops);
+		bool discard = false;
+
+		/* negative lock caching */
+		if (index < info->oti_ng_index) {
+			discard = true;
+		} else if (index >= info->oti_fn_index) {
+			struct ldlm_lock *tmp;
+			/* refresh non-overlapped index */
+			tmp = osc_dlmlock_at_pgoff(env, osc, index,
+					OSC_DAP_FL_TEST_LOCK |
+					OSC_DAP_FL_AST |
+					OSC_DAP_FL_RIGHT);
+			if (tmp != NULL) {
+				__u64 end =
+					tmp->l_policy_data.l_extent.end;
+				__u64 start =
+					tmp->l_policy_data.l_extent.start;
+
+				/* no lock covering this page */
+				if (index < cl_index(osc2cl(osc), start)) {
+					/* no lock at @index,
+					 * first lock at @start
+					 */
+					info->oti_ng_index =
+						cl_index(osc2cl(osc), start);
+					discard = true;
+				} else {
+					/* Cache the first-non-overlapped
+					 * index so as to skip all pages
+					 * within [index, oti_fn_index).
+					 * This is safe because if tmp lock
+					 * is canceled, it will discard these
+					 * pages.
+					 */
+					info->oti_fn_index =
+						cl_index(osc2cl(osc), end + 1);
+					if (end == OBD_OBJECT_EOF)
+						info->oti_fn_index =
+							CL_PAGE_EOF;
+				}
+				LDLM_LOCK_PUT(tmp);
+			} else {
+				info->oti_ng_index = CL_PAGE_EOF;
+				discard = true;
+			}
+		}
 
-		/* refresh non-overlapped index */
-		tmp = osc_dlmlock_at_pgoff(env, osc, index,
-					   OSC_DAP_FL_TEST_LOCK);
-		if (tmp != NULL) {
-			__u64 end = tmp->l_policy_data.l_extent.end;
-			/* Cache the first-non-overlapped index so as to skip
-			 * all pages within [index, oti_fn_index). This is safe
-			 * because if tmp lock is canceled, it will discard
-			 * these pages. */
-			info->oti_fn_index = cl_index(osc2cl(osc), end + 1);
-			if (end == OBD_OBJECT_EOF)
-				info->oti_fn_index = CL_PAGE_EOF;
-			LDLM_LOCK_PUT(tmp);
-		} else if (cl_page_own(env, io, page) == 0) {
-			/* discard the page */
-			cl_page_discard(env, io, page);
-			cl_page_disown(env, io, page);
-		} else {
-			LASSERT(page->cp_state == CPS_FREEING);
+		if (discard) {
+			if (cl_page_own(env, io, page) == 0) {
+				cl_page_discard(env, io, page);
+				cl_page_disown(env, io, page);
+			} else {
+				LASSERT(page->cp_state == CPS_FREEING);
+			}
 		}
-	}
 
-	info->oti_next_index = index + 1;
-	return CLP_GANG_OKAY;
+		info->oti_next_index = index + 1;
+	}
+	return true;
 }
 
-int osc_discard_cb(const struct lu_env *env, struct cl_io *io,
-		   struct osc_page *ops, void *cbdata)
+bool osc_discard_cb(const struct lu_env *env, struct cl_io *io,
+		    void **pvec, int count, void *cbdata)
 {
 	struct osc_thread_info *info = osc_env_info(env);
-	struct cl_page *page = ops->ops_cl.cpl_page;
-
-	/* page is top page. */
-	info->oti_next_index = osc_index(ops) + 1;
-	if (cl_page_own(env, io, page) == 0) {
-		if (!ergo(page->cp_type == CPT_CACHEABLE,
-			  !PageDirty(cl_page_vmpage(page))))
-			CL_PAGE_DEBUG(D_ERROR, env, page,
-					"discard dirty page?\n");
-
-		/* discard the page */
-		cl_page_discard(env, io, page);
-		cl_page_disown(env, io, page);
-	} else {
-		LASSERT(page->cp_state == CPS_FREEING);
+	int i;
+
+	for (i = 0; i < count; i++) {
+		struct osc_page *ops = pvec[i];
+		struct cl_page *page = ops->ops_cl.cpl_page;
+
+		/* page is top page. */
+		info->oti_next_index = osc_index(ops) + 1;
+		if (cl_page_own(env, io, page) == 0) {
+			if (!ergo(page->cp_type == CPT_CACHEABLE,
+				  !PageDirty(cl_page_vmpage(page))))
+				CL_PAGE_DEBUG(D_ERROR, env, page,
+					      "discard dirty page?\n");
+
+			/* discard the page */
+			cl_page_discard(env, io, page);
+			cl_page_disown(env, io, page);
+		} else {
+			LASSERT(page->cp_state == CPS_FREEING);
+		}
 	}
 
-	return CLP_GANG_OKAY;
+	return true;
 }
 EXPORT_SYMBOL(osc_discard_cb);
 
@@ -3301,28 +3255,24 @@ 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;
 
 	io->ci_obj = cl_object_top(osc2cl(osc));
 	io->ci_ignore_layout = 1;
+	io->u.ci_misc.lm_next_rpc_time = ktime_get_seconds() +
+					 5 * obd_timeout / 16;
 	result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
 	if (result != 0)
 		GOTO(out, result);
 
 	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;
+	info->oti_ng_index = 0;
 
-		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);