1 files changed, 265 insertions, 279 deletions

diff --git a/drivers/staging/lustre/lustre/osc/osc_page.c b/drivers/staging/lustre/lustre/osc/osc_page.c
index ce9ddd515f64..c29c2eabe39c 100644
--- a/drivers/staging/lustre/lustre/osc/osc_page.c
+++ b/drivers/staging/lustre/lustre/osc/osc_page.c
@@ -36,14 +36,15 @@
  * Implementation of cl_page for OSC layer.
  *
  *   Author: Nikita Danilov <nikita.danilov@sun.com>
+ *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  */
 
 #define DEBUG_SUBSYSTEM S_OSC
 
 #include "osc_cl_internal.h"
 
-static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del);
-static void osc_lru_add(struct client_obd *cli, struct osc_page *opg);
+static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
+static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 			   struct osc_page *opg);
 
@@ -63,18 +64,9 @@ static int osc_page_protected(const struct lu_env *env,
  * Page operations.
  *
  */
-static void osc_page_fini(const struct lu_env *env,
-			  struct cl_page_slice *slice)
-{
-	struct osc_page *opg = cl2osc_page(slice);
-
-	CDEBUG(D_TRACE, "%p\n", opg);
-	LASSERT(!opg->ops_lock);
-}
-
 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 {
-	struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
+	struct cl_page *page = opg->ops_cl.cpl_page;
 
 	LASSERT(!opg->ops_transfer_pinned);
 	cl_page_get(page);
@@ -85,11 +77,11 @@ static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 static void osc_page_transfer_put(const struct lu_env *env,
 				  struct osc_page *opg)
 {
-	struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
+	struct cl_page *page = opg->ops_cl.cpl_page;
 
 	if (opg->ops_transfer_pinned) {
-		lu_ref_del(&page->cp_reference, "transfer", page);
 		opg->ops_transfer_pinned = 0;
+		lu_ref_del(&page->cp_reference, "transfer", page);
 		cl_page_put(env, page);
 	}
 }
@@ -104,10 +96,7 @@ static void osc_page_transfer_add(const struct lu_env *env,
 {
 	struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 
-	/* ops_lru and ops_inflight share the same field, so take it from LRU
-	 * first and then use it as inflight.
-	 */
-	osc_lru_del(osc_cli(obj), opg, false);
+	osc_lru_use(osc_cli(obj), opg);
 
 	spin_lock(&obj->oo_seatbelt);
 	list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
@@ -115,11 +104,9 @@ static void osc_page_transfer_add(const struct lu_env *env,
 	spin_unlock(&obj->oo_seatbelt);
 }
 
-static int osc_page_cache_add(const struct lu_env *env,
-			      const struct cl_page_slice *slice,
-			      struct cl_io *io)
+int osc_page_cache_add(const struct lu_env *env,
+		       const struct cl_page_slice *slice, struct cl_io *io)
 {
-	struct osc_io *oio = osc_env_io(env);
 	struct osc_page *opg = cl2osc_page(slice);
 	int result;
 
@@ -132,17 +119,6 @@ static int osc_page_cache_add(const struct lu_env *env,
 	else
 		osc_page_transfer_add(env, opg, CRT_WRITE);
 
-	/* for sync write, kernel will wait for this page to be flushed before
-	 * osc_io_end() is called, so release it earlier.
-	 * for mkwrite(), it's known there is no further pages.
-	 */
-	if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) {
-		if (oio->oi_active) {
-			osc_extent_release(env, oio->oi_active);
-			oio->oi_active = NULL;
-		}
-	}
-
 	return result;
 }
 
@@ -154,102 +130,25 @@ void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 	policy->l_extent.end = cl_offset(obj, end + 1) - 1;
 }
 
-static int osc_page_addref_lock(const struct lu_env *env,
-				struct osc_page *opg,
-				struct cl_lock *lock)
-{
-	struct osc_lock *olock;
-	int rc;
-
-	LASSERT(!opg->ops_lock);
-
-	olock = osc_lock_at(lock);
-	if (atomic_inc_return(&olock->ols_pageref) <= 0) {
-		atomic_dec(&olock->ols_pageref);
-		rc = -ENODATA;
-	} else {
-		cl_lock_get(lock);
-		opg->ops_lock = lock;
-		rc = 0;
-	}
-	return rc;
-}
-
-static void osc_page_putref_lock(const struct lu_env *env,
-				 struct osc_page *opg)
-{
-	struct cl_lock *lock = opg->ops_lock;
-	struct osc_lock *olock;
-
-	LASSERT(lock);
-	olock = osc_lock_at(lock);
-
-	atomic_dec(&olock->ols_pageref);
-	opg->ops_lock = NULL;
-
-	cl_lock_put(env, lock);
-}
-
 static int osc_page_is_under_lock(const struct lu_env *env,
 				  const struct cl_page_slice *slice,
-				  struct cl_io *unused)
+				  struct cl_io *unused, pgoff_t *max_index)
 {
-	struct cl_lock *lock;
+	struct osc_page *opg = cl2osc_page(slice);
+	struct ldlm_lock *dlmlock;
 	int result = -ENODATA;
 
-	lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
-			       NULL, 1, 0);
-	if (lock) {
-		if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
-			result = -EBUSY;
-		cl_lock_put(env, lock);
+	dlmlock = osc_dlmlock_at_pgoff(env, cl2osc(slice->cpl_obj),
+				       osc_index(opg), 1, 0);
+	if (dlmlock) {
+		*max_index = cl_index(slice->cpl_obj,
+				      dlmlock->l_policy_data.l_extent.end);
+		LDLM_LOCK_PUT(dlmlock);
+		result = 0;
 	}
 	return result;
 }
 
-static void osc_page_disown(const struct lu_env *env,
-			    const struct cl_page_slice *slice,
-			    struct cl_io *io)
-{
-	struct osc_page *opg = cl2osc_page(slice);
-
-	if (unlikely(opg->ops_lock))
-		osc_page_putref_lock(env, opg);
-}
-
-static void osc_page_completion_read(const struct lu_env *env,
-				     const struct cl_page_slice *slice,
-				     int ioret)
-{
-	struct osc_page *opg = cl2osc_page(slice);
-	struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
-
-	if (likely(opg->ops_lock))
-		osc_page_putref_lock(env, opg);
-	osc_lru_add(osc_cli(obj), opg);
-}
-
-static void osc_page_completion_write(const struct lu_env *env,
-				      const struct cl_page_slice *slice,
-				      int ioret)
-{
-	struct osc_page *opg = cl2osc_page(slice);
-	struct osc_object *obj = cl2osc(slice->cpl_obj);
-
-	osc_lru_add(osc_cli(obj), opg);
-}
-
-static int osc_page_fail(const struct lu_env *env,
-			 const struct cl_page_slice *slice,
-			 struct cl_io *unused)
-{
-	/*
-	 * Cached read?
-	 */
-	LBUG();
-	return 0;
-}
-
 static const char *osc_list(struct list_head *head)
 {
 	return list_empty(head) ? "-" : "+";
@@ -272,8 +171,8 @@ static int osc_page_print(const struct lu_env *env,
 	struct osc_object *obj = cl2osc(slice->cpl_obj);
 	struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 
-	return (*printer)(env, cookie, LUSTRE_OSC_NAME "-page@%p: 1< %#x %d %u %s %s > 2< %llu %u %u %#x %#x | %p %p %p > 3< %s %p %d %lu %d > 4< %d %d %d %lu %s | %s %s %s %s > 5< %s %s %s %s | %d %s | %d %s %s>\n",
-			  opg,
+	return (*printer)(env, cookie, LUSTRE_OSC_NAME "-page@%p %lu: 1< %#x %d %u %s %s > 2< %llu %u %u %#x %#x | %p %p %p > 3< %s %p %d %lu %d > 4< %d %d %d %lu %s | %s %s %s %s > 5< %s %s %s %s | %d %s | %d %s %s>\n",
+			  opg, osc_index(opg),
 			  /* 1 */
 			  oap->oap_magic, oap->oap_cmd,
 			  oap->oap_interrupted,
@@ -321,7 +220,7 @@ static void osc_page_delete(const struct lu_env *env,
 	osc_page_transfer_put(env, opg);
 	rc = osc_teardown_async_page(env, obj, opg);
 	if (rc) {
-		CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
+		CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 			      "Trying to teardown failed: %d\n", rc);
 		LASSERT(0);
 	}
@@ -334,7 +233,19 @@ static void osc_page_delete(const struct lu_env *env,
 	}
 	spin_unlock(&obj->oo_seatbelt);
 
-	osc_lru_del(osc_cli(obj), opg, true);
+	osc_lru_del(osc_cli(obj), opg);
+
+	if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
+		void *value;
+
+		spin_lock(&obj->oo_tree_lock);
+		value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
+		if (value)
+			--obj->oo_npages;
+		spin_unlock(&obj->oo_tree_lock);
+
+		LASSERT(ergo(value, value == opg));
+	}
 }
 
 static void osc_page_clip(const struct lu_env *env,
@@ -382,28 +293,16 @@ static int osc_page_flush(const struct lu_env *env,
 }
 
 static const struct cl_page_operations osc_page_ops = {
-	.cpo_fini	  = osc_page_fini,
 	.cpo_print	 = osc_page_print,
 	.cpo_delete	= osc_page_delete,
 	.cpo_is_under_lock = osc_page_is_under_lock,
-	.cpo_disown	= osc_page_disown,
-	.io = {
-		[CRT_READ] = {
-			.cpo_cache_add  = osc_page_fail,
-			.cpo_completion = osc_page_completion_read
-		},
-		[CRT_WRITE] = {
-			.cpo_cache_add  = osc_page_cache_add,
-			.cpo_completion = osc_page_completion_write
-		}
-	},
 	.cpo_clip	   = osc_page_clip,
 	.cpo_cancel	 = osc_page_cancel,
 	.cpo_flush	  = osc_page_flush
 };
 
 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
-		  struct cl_page *page, struct page *vmpage)
+		  struct cl_page *page, pgoff_t index)
 {
 	struct osc_object *osc = cl2osc(obj);
 	struct osc_page *opg = cl_object_page_slice(obj, page);
@@ -412,13 +311,14 @@ int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 	opg->ops_from = 0;
 	opg->ops_to = PAGE_SIZE;
 
-	result = osc_prep_async_page(osc, opg, vmpage,
-				     cl_offset(obj, page->cp_index));
+	result = osc_prep_async_page(osc, opg, page->cp_vmpage,
+				     cl_offset(obj, index));
 	if (result == 0) {
 		struct osc_io *oio = osc_env_io(env);
 
 		opg->ops_srvlock = osc_io_srvlock(oio);
-		cl_page_slice_add(page, &opg->ops_cl, obj, &osc_page_ops);
+		cl_page_slice_add(page, &opg->ops_cl, obj, index,
+				  &osc_page_ops);
 	}
 	/*
 	 * Cannot assert osc_page_protected() here as read-ahead
@@ -431,12 +331,47 @@ int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 	INIT_LIST_HEAD(&opg->ops_lru);
 
 	/* reserve an LRU space for this page */
-	if (page->cp_type == CPT_CACHEABLE && result == 0)
+	if (page->cp_type == CPT_CACHEABLE && result == 0) {
 		result = osc_lru_reserve(env, osc, opg);
+		if (result == 0) {
+			spin_lock(&osc->oo_tree_lock);
+			result = radix_tree_insert(&osc->oo_tree, index, opg);
+			if (result == 0)
+				++osc->oo_npages;
+			spin_unlock(&osc->oo_tree_lock);
+			LASSERT(result == 0);
+		}
+	}
 
 	return result;
 }
 
+int osc_over_unstable_soft_limit(struct client_obd *cli)
+{
+	long obd_upages, obd_dpages, osc_upages;
+
+	/* Can't check cli->cl_unstable_count, therefore, no soft limit */
+	if (!cli)
+		return 0;
+
+	obd_upages = atomic_read(&obd_unstable_pages);
+	obd_dpages = atomic_read(&obd_dirty_pages);
+
+	osc_upages = atomic_read(&cli->cl_unstable_count);
+
+	/*
+	 * obd_max_dirty_pages is the max number of (dirty + unstable)
+	 * pages allowed at any given time. To simulate an unstable page
+	 * only limit, we subtract the current number of dirty pages
+	 * from this max. This difference is roughly the amount of pages
+	 * currently available for unstable pages. Thus, the soft limit
+	 * is half of that difference. Check osc_upages to ensure we don't
+	 * set SOFT_SYNC for OSCs without any outstanding unstable pages.
+	 */
+	return osc_upages &&
+	       obd_upages >= (obd_max_dirty_pages - obd_dpages) / 2;
+}
+
 /**
  * Helper function called by osc_io_submit() for every page in an immediate
  * transfer (i.e., transferred synchronously).
@@ -460,6 +395,9 @@ void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 	oap->oap_count = opg->ops_to - opg->ops_from;
 	oap->oap_brw_flags = brw_flags | OBD_BRW_SYNC;
 
+	if (osc_over_unstable_soft_limit(oap->oap_cli))
+		oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
+
 	if (!client_is_remote(osc_export(obj)) &&
 	    capable(CFS_CAP_SYS_RESOURCE)) {
 		oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
@@ -483,13 +421,12 @@ void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
  */
 
 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
-static atomic_t osc_lru_waiters = ATOMIC_INIT(0);
 /* LRU pages are freed in batch mode. OSC should at least free this
  * number of pages to avoid running out of LRU budget, and..
  */
 static const int lru_shrink_min = 2 << (20 - PAGE_SHIFT);  /* 2M */
 /* free this number at most otherwise it will take too long time to finish. */
-static const int lru_shrink_max = 32 << (20 - PAGE_SHIFT); /* 32M */
+static const int lru_shrink_max = 8 << (20 - PAGE_SHIFT); /* 8M */
 
 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
  * we should free slots aggressively. In this way, slots are freed in a steady
@@ -500,65 +437,142 @@ static const int lru_shrink_max = 32 << (20 - PAGE_SHIFT); /* 32M */
 static int osc_cache_too_much(struct client_obd *cli)
 {
 	struct cl_client_cache *cache = cli->cl_cache;
-	int pages = atomic_read(&cli->cl_lru_in_list) >> 1;
+	int pages = atomic_read(&cli->cl_lru_in_list);
+	unsigned long budget;
 
-	if (atomic_read(&osc_lru_waiters) > 0 &&
-	    atomic_read(cli->cl_lru_left) < lru_shrink_max)
-		/* drop lru pages aggressively */
-		return min(pages, lru_shrink_max);
+	budget = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
 
 	/* if it's going to run out LRU slots, we should free some, but not
 	 * too much to maintain fairness among OSCs.
 	 */
 	if (atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
-		unsigned long tmp;
+		if (pages >= budget)
+			return lru_shrink_max;
+		else if (pages >= budget / 2)
+			return lru_shrink_min;
+	} else if (pages >= budget * 2) {
+		return lru_shrink_min;
+	}
+	return 0;
+}
 
-		tmp = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
-		if (pages > tmp)
-			return min(pages, lru_shrink_max);
+int lru_queue_work(const struct lu_env *env, void *data)
+{
+	struct client_obd *cli = data;
 
-		return pages > lru_shrink_min ? lru_shrink_min : 0;
-	}
+	CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
+
+	if (osc_cache_too_much(cli))
+		osc_lru_shrink(env, cli, lru_shrink_max, true);
 
 	return 0;
 }
 
-/* Return how many pages are not discarded in @pvec. */
-static int discard_pagevec(const struct lu_env *env, struct cl_io *io,
-			   struct cl_page **pvec, int max_index)
+void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
+{
+	LIST_HEAD(lru);
+	struct osc_async_page *oap;
+	int npages = 0;
+
+	list_for_each_entry(oap, plist, oap_pending_item) {
+		struct osc_page *opg = oap2osc_page(oap);
+
+		if (!opg->ops_in_lru)
+			continue;
+
+		++npages;
+		LASSERT(list_empty(&opg->ops_lru));
+		list_add(&opg->ops_lru, &lru);
+	}
+
+	if (npages > 0) {
+		spin_lock(&cli->cl_lru_list_lock);
+		list_splice_tail(&lru, &cli->cl_lru_list);
+		atomic_sub(npages, &cli->cl_lru_busy);
+		atomic_add(npages, &cli->cl_lru_in_list);
+		spin_unlock(&cli->cl_lru_list_lock);
+
+		/* XXX: May set force to be true for better performance */
+		if (osc_cache_too_much(cli))
+			(void)ptlrpcd_queue_work(cli->cl_lru_work);
+	}
+}
+
+static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
+{
+	LASSERT(atomic_read(&cli->cl_lru_in_list) > 0);
+	list_del_init(&opg->ops_lru);
+	atomic_dec(&cli->cl_lru_in_list);
+}
+
+/**
+ * Page is being destroyed. The page may be not in LRU list, if the transfer
+ * has never finished(error occurred).
+ */
+static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
+{
+	if (opg->ops_in_lru) {
+		spin_lock(&cli->cl_lru_list_lock);
+		if (!list_empty(&opg->ops_lru)) {
+			__osc_lru_del(cli, opg);
+		} else {
+			LASSERT(atomic_read(&cli->cl_lru_busy) > 0);
+			atomic_dec(&cli->cl_lru_busy);
+		}
+		spin_unlock(&cli->cl_lru_list_lock);
+
+		atomic_inc(cli->cl_lru_left);
+		/* this is a great place to release more LRU pages if
+		 * this osc occupies too many LRU pages and kernel is
+		 * stealing one of them.
+		 */
+		if (!memory_pressure_get())
+			(void)ptlrpcd_queue_work(cli->cl_lru_work);
+		wake_up(&osc_lru_waitq);
+	} else {
+		LASSERT(list_empty(&opg->ops_lru));
+	}
+}
+
+/**
+ * Delete page from LRUlist for redirty.
+ */
+static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
+{
+	/* If page is being transferred for the first time,
+	 * ops_lru should be empty
+	 */
+	if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
+		spin_lock(&cli->cl_lru_list_lock);
+		__osc_lru_del(cli, opg);
+		spin_unlock(&cli->cl_lru_list_lock);
+		atomic_inc(&cli->cl_lru_busy);
+	}
+}
+
+static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
+			    struct cl_page **pvec, int max_index)
 {
-	int count;
 	int i;
 
-	for (count = 0, i = 0; i < max_index; i++) {
+	for (i = 0; i < max_index; i++) {
 		struct cl_page *page = pvec[i];
 
-		if (cl_page_own_try(env, io, page) == 0) {
-			/* free LRU page only if nobody is using it.
-			 * This check is necessary to avoid freeing the pages
-			 * having already been removed from LRU and pinned
-			 * for IO.
-			 */
-			if (!cl_page_in_use(page)) {
-				cl_page_unmap(env, io, page);
-				cl_page_discard(env, io, page);
-				++count;
-			}
-			cl_page_disown(env, io, page);
-		}
+		LASSERT(cl_page_is_owned(page, io));
+		cl_page_discard(env, io, page);
+		cl_page_disown(env, io, page);
 		cl_page_put(env, page);
+
 		pvec[i] = NULL;
 	}
-	return max_index - count;
 }
 
 /**
  * Drop @target of pages from LRU at most.
  */
-int osc_lru_shrink(struct client_obd *cli, int target)
+int osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
+		   int target, bool force)
 {
-	struct cl_env_nest nest;
-	struct lu_env *env;
 	struct cl_io *io;
 	struct cl_object *clobj = NULL;
 	struct cl_page **pvec;
@@ -573,23 +587,31 @@ int osc_lru_shrink(struct client_obd *cli, int target)
 	if (atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 		return 0;
 
-	env = cl_env_nested_get(&nest);
-	if (IS_ERR(env))
-		return PTR_ERR(env);
+	if (!force) {
+		if (atomic_read(&cli->cl_lru_shrinkers) > 0)
+			return -EBUSY;
 
-	pvec = osc_env_info(env)->oti_pvec;
+		if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
+			atomic_dec(&cli->cl_lru_shrinkers);
+			return -EBUSY;
+		}
+	} else {
+		atomic_inc(&cli->cl_lru_shrinkers);
+	}
+
+	pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 	io = &osc_env_info(env)->oti_io;
 
-	client_obd_list_lock(&cli->cl_lru_list_lock);
-	atomic_inc(&cli->cl_lru_shrinkers);
+	spin_lock(&cli->cl_lru_list_lock);
 	maxscan = min(target << 1, atomic_read(&cli->cl_lru_in_list));
 	list_for_each_entry_safe(opg, temp, &cli->cl_lru_list, ops_lru) {
 		struct cl_page *page;
+		bool will_free = false;
 
 		if (--maxscan < 0)
 			break;
 
-		page = cl_page_top(opg->ops_cl.cpl_page);
+		page = opg->ops_cl.cpl_page;
 		if (cl_page_in_use_noref(page)) {
 			list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 			continue;
@@ -600,10 +622,10 @@ int osc_lru_shrink(struct client_obd *cli, int target)
 			struct cl_object *tmp = page->cp_obj;
 
 			cl_object_get(tmp);
-			client_obd_list_unlock(&cli->cl_lru_list_lock);
+			spin_unlock(&cli->cl_lru_list_lock);
 
 			if (clobj) {
-				count -= discard_pagevec(env, io, pvec, index);
+				discard_pagevec(env, io, pvec, index);
 				index = 0;
 
 				cl_io_fini(env, io);
@@ -616,7 +638,7 @@ int osc_lru_shrink(struct client_obd *cli, int target)
 			io->ci_ignore_layout = 1;
 			rc = cl_io_init(env, io, CIT_MISC, clobj);
 
-			client_obd_list_lock(&cli->cl_lru_list_lock);
+			spin_lock(&cli->cl_lru_list_lock);
 
 			if (rc != 0)
 				break;
@@ -625,98 +647,54 @@ int osc_lru_shrink(struct client_obd *cli, int target)
 			continue;
 		}
 
-		/* move this page to the end of list as it will be discarded
-		 * soon. The page will be finally removed from LRU list in
-		 * osc_page_delete().
-		 */
-		list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
+		if (cl_page_own_try(env, io, page) == 0) {
+			if (!cl_page_in_use_noref(page)) {
+				/* remove it from lru list earlier to avoid
+				 * lock contention
+				 */
+				__osc_lru_del(cli, opg);
+				opg->ops_in_lru = 0; /* will be discarded */
+
+				cl_page_get(page);
+				will_free = true;
+			} else {
+				cl_page_disown(env, io, page);
+			}
+		}
 
-		/* it's okay to grab a refcount here w/o holding lock because
-		 * it has to grab cl_lru_list_lock to delete the page.
-		 */
-		cl_page_get(page);
-		pvec[index++] = page;
-		if (++count >= target)
-			break;
+		if (!will_free) {
+			list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
+			continue;
+		}
 
+		/* Don't discard and free the page with cl_lru_list held */
+		pvec[index++] = page;
 		if (unlikely(index == OTI_PVEC_SIZE)) {
-			client_obd_list_unlock(&cli->cl_lru_list_lock);
-			count -= discard_pagevec(env, io, pvec, index);
+			spin_unlock(&cli->cl_lru_list_lock);
+			discard_pagevec(env, io, pvec, index);
 			index = 0;
 
-			client_obd_list_lock(&cli->cl_lru_list_lock);
+			spin_lock(&cli->cl_lru_list_lock);
 		}
+
+		if (++count >= target)
+			break;
 	}
-	client_obd_list_unlock(&cli->cl_lru_list_lock);
+	spin_unlock(&cli->cl_lru_list_lock);
 
 	if (clobj) {
-		count -= discard_pagevec(env, io, pvec, index);
+		discard_pagevec(env, io, pvec, index);
 
 		cl_io_fini(env, io);
 		cl_object_put(env, clobj);
 	}
-	cl_env_nested_put(&nest, env);
 
 	atomic_dec(&cli->cl_lru_shrinkers);
-	return count > 0 ? count : rc;
-}
-
-static void osc_lru_add(struct client_obd *cli, struct osc_page *opg)
-{
-	bool wakeup = false;
-
-	if (!opg->ops_in_lru)
-		return;
-
-	atomic_dec(&cli->cl_lru_busy);
-	client_obd_list_lock(&cli->cl_lru_list_lock);
-	if (list_empty(&opg->ops_lru)) {
-		list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
-		atomic_inc_return(&cli->cl_lru_in_list);
-		wakeup = atomic_read(&osc_lru_waiters) > 0;
-	}
-	client_obd_list_unlock(&cli->cl_lru_list_lock);
-
-	if (wakeup) {
-		osc_lru_shrink(cli, osc_cache_too_much(cli));
+	if (count > 0) {
+		atomic_add(count, cli->cl_lru_left);
 		wake_up_all(&osc_lru_waitq);
 	}
-}
-
-/* delete page from LRUlist. The page can be deleted from LRUlist for two
- * reasons: redirtied or deleted from page cache.
- */
-static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del)
-{
-	if (opg->ops_in_lru) {
-		client_obd_list_lock(&cli->cl_lru_list_lock);
-		if (!list_empty(&opg->ops_lru)) {
-			LASSERT(atomic_read(&cli->cl_lru_in_list) > 0);
-			list_del_init(&opg->ops_lru);
-			atomic_dec(&cli->cl_lru_in_list);
-			if (!del)
-				atomic_inc(&cli->cl_lru_busy);
-		} else if (del) {
-			LASSERT(atomic_read(&cli->cl_lru_busy) > 0);
-			atomic_dec(&cli->cl_lru_busy);
-		}
-		client_obd_list_unlock(&cli->cl_lru_list_lock);
-		if (del) {
-			atomic_inc(cli->cl_lru_left);
-			/* this is a great place to release more LRU pages if
-			 * this osc occupies too many LRU pages and kernel is
-			 * stealing one of them.
-			 * cl_lru_shrinkers is to avoid recursive call in case
-			 * we're already in the context of osc_lru_shrink().
-			 */
-			if (atomic_read(&cli->cl_lru_shrinkers) == 0 &&
-			    !memory_pressure_get())
-				osc_lru_shrink(cli, osc_cache_too_much(cli));
-			wake_up(&osc_lru_waitq);
-		}
-	} else {
-		LASSERT(list_empty(&opg->ops_lru));
-	}
+	return count > 0 ? count : rc;
 }
 
 static inline int max_to_shrink(struct client_obd *cli)
@@ -724,19 +702,28 @@ static inline int max_to_shrink(struct client_obd *cli)
 	return min(atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 }
 
-static int osc_lru_reclaim(struct client_obd *cli)
+int osc_lru_reclaim(struct client_obd *cli)
 {
+	struct cl_env_nest nest;
+	struct lu_env *env;
 	struct cl_client_cache *cache = cli->cl_cache;
 	int max_scans;
-	int rc;
+	int rc = 0;
 
 	LASSERT(cache);
 
-	rc = osc_lru_shrink(cli, lru_shrink_min);
+	env = cl_env_nested_get(&nest);
+	if (IS_ERR(env))
+		return 0;
+
+	rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 	if (rc != 0) {
+		if (rc == -EBUSY)
+			rc = 0;
+
 		CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 		       cli->cl_import->imp_obd->obd_name, rc, cli);
-		return rc;
+		goto out;
 	}
 
 	CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
@@ -764,10 +751,11 @@ static int osc_lru_reclaim(struct client_obd *cli)
 		       atomic_read(&cli->cl_lru_busy));
 
 		list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
-		if (atomic_read(&cli->cl_lru_in_list) > 0) {
+		if (osc_cache_too_much(cli) > 0) {
 			spin_unlock(&cache->ccc_lru_lock);
 
-			rc = osc_lru_shrink(cli, max_to_shrink(cli));
+			rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
+					    true);
 			spin_lock(&cache->ccc_lru_lock);
 			if (rc != 0)
 				break;
@@ -775,6 +763,8 @@ static int osc_lru_reclaim(struct client_obd *cli)
 	}
 	spin_unlock(&cache->ccc_lru_lock);
 
+out:
+	cl_env_nested_put(&nest, env);
 	CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 	       cli->cl_import->imp_obd->obd_name, cli, rc);
 	return rc;
@@ -784,16 +774,20 @@ static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 			   struct osc_page *opg)
 {
 	struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
+	struct osc_io *oio = osc_env_io(env);
 	struct client_obd *cli = osc_cli(obj);
 	int rc = 0;
 
 	if (!cli->cl_cache) /* shall not be in LRU */
 		return 0;
 
+	if (oio->oi_lru_reserved > 0) {
+		--oio->oi_lru_reserved;
+		goto out;
+	}
+
 	LASSERT(atomic_read(cli->cl_lru_left) >= 0);
 	while (!atomic_add_unless(cli->cl_lru_left, -1, 0)) {
-		int gen;
-
 		/* run out of LRU spaces, try to drop some by itself */
 		rc = osc_lru_reclaim(cli);
 		if (rc < 0)
@@ -803,23 +797,15 @@ static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 
 		cond_resched();
 
-		/* slowest case, all of caching pages are busy, notifying
-		 * other OSCs that we're lack of LRU slots.
-		 */
-		atomic_inc(&osc_lru_waiters);
-
-		gen = atomic_read(&cli->cl_lru_in_list);
 		rc = l_wait_event(osc_lru_waitq,
-				  atomic_read(cli->cl_lru_left) > 0 ||
-				  (atomic_read(&cli->cl_lru_in_list) > 0 &&
-				   gen != atomic_read(&cli->cl_lru_in_list)),
+				  atomic_read(cli->cl_lru_left) > 0,
 				  &lwi);
 
-		atomic_dec(&osc_lru_waiters);
 		if (rc < 0)
 			break;
 	}
 
+out:
 	if (rc >= 0) {
 		atomic_inc(&cli->cl_lru_busy);
 		opg->ops_in_lru = 1;