]> git.baikalelectronics.ru Git - kernel.git/commitdiff
mm: fs: invalidate BH LRU during page migration
authorMinchan Kim <minchan@kernel.org>
Wed, 5 May 2021 01:37:00 +0000 (18:37 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 5 May 2021 18:27:24 +0000 (11:27 -0700)
Pages containing buffer_heads that are in one of the per-CPU buffer_head
LRU caches will be pinned and thus cannot be migrated.  This can prevent
CMA allocations from succeeding, which are often used on platforms with
co-processors (such as a DSP) that can only use physically contiguous
memory.  It can also prevent memory hot-unplugging from succeeding,
which involves migrating at least MIN_MEMORY_BLOCK_SIZE bytes of memory,
which ranges from 8 MiB to 1 GiB based on the architecture in use.

Correspondingly, invalidate the BH LRU caches before a migration starts
and stop any buffer_head from being cached in the LRU caches, until
migration has finished.

Link: https://lkml.kernel.org/r/20210319175127.886124-3-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Reported-by: Chris Goldsworthy <cgoldswo@codeaurora.org>
Reported-by: Laura Abbott <labbott@kernel.org>
Tested-by: Oliver Sang <oliver.sang@intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: John Dias <joaodias@google.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/buffer.c
include/linux/buffer_head.h
mm/swap.c

index 0cb7ffd4977c9283aba3794f4da84ecb1758ec47..e9872d0dcbf1724a00db6f265b4becf491f9bffc 100644 (file)
@@ -1264,6 +1264,15 @@ static void bh_lru_install(struct buffer_head *bh)
        int i;
 
        check_irqs_on();
+       /*
+        * the refcount of buffer_head in bh_lru prevents dropping the
+        * attached page(i.e., try_to_free_buffers) so it could cause
+        * failing page migration.
+        * Skip putting upcoming bh into bh_lru until migration is done.
+        */
+       if (lru_cache_disabled())
+               return;
+
        bh_lru_lock();
 
        b = this_cpu_ptr(&bh_lrus);
@@ -1404,6 +1413,15 @@ __bread_gfp(struct block_device *bdev, sector_t block,
 }
 EXPORT_SYMBOL(__bread_gfp);
 
+static void __invalidate_bh_lrus(struct bh_lru *b)
+{
+       int i;
+
+       for (i = 0; i < BH_LRU_SIZE; i++) {
+               brelse(b->bhs[i]);
+               b->bhs[i] = NULL;
+       }
+}
 /*
  * invalidate_bh_lrus() is called rarely - but not only at unmount.
  * This doesn't race because it runs in each cpu either in irq
@@ -1412,16 +1430,12 @@ EXPORT_SYMBOL(__bread_gfp);
 static void invalidate_bh_lru(void *arg)
 {
        struct bh_lru *b = &get_cpu_var(bh_lrus);
-       int i;
 
-       for (i = 0; i < BH_LRU_SIZE; i++) {
-               brelse(b->bhs[i]);
-               b->bhs[i] = NULL;
-       }
+       __invalidate_bh_lrus(b);
        put_cpu_var(bh_lrus);
 }
 
-static bool has_bh_in_lru(int cpu, void *dummy)
+bool has_bh_in_lru(int cpu, void *dummy)
 {
        struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
        int i;
@@ -1440,6 +1454,16 @@ void invalidate_bh_lrus(void)
 }
 EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
 
+void invalidate_bh_lrus_cpu(int cpu)
+{
+       struct bh_lru *b;
+
+       bh_lru_lock();
+       b = per_cpu_ptr(&bh_lrus, cpu);
+       __invalidate_bh_lrus(b);
+       bh_lru_unlock();
+}
+
 void set_bh_page(struct buffer_head *bh,
                struct page *page, unsigned long offset)
 {
index 6b47f94378c5ad9638f0dddde0682f426aab96d3..e7e99da31349633b791a63cc21e0262e63cb32b6 100644 (file)
@@ -194,6 +194,8 @@ void __breadahead_gfp(struct block_device *, sector_t block, unsigned int size,
 struct buffer_head *__bread_gfp(struct block_device *,
                                sector_t block, unsigned size, gfp_t gfp);
 void invalidate_bh_lrus(void);
+void invalidate_bh_lrus_cpu(int cpu);
+bool has_bh_in_lru(int cpu, void *dummy);
 struct buffer_head *alloc_buffer_head(gfp_t gfp_flags);
 void free_buffer_head(struct buffer_head * bh);
 void unlock_buffer(struct buffer_head *bh);
@@ -406,6 +408,8 @@ static inline int inode_has_buffers(struct inode *inode) { return 0; }
 static inline void invalidate_inode_buffers(struct inode *inode) {}
 static inline int remove_inode_buffers(struct inode *inode) { return 1; }
 static inline int sync_mapping_buffers(struct address_space *mapping) { return 0; }
+static inline void invalidate_bh_lrus_cpu(int cpu) {}
+static inline bool has_bh_in_lru(int cpu, void *dummy) { return 0; }
 #define buffer_heads_over_limit 0
 
 #endif /* CONFIG_BLOCK */
index c94f55e7b6493e695d5e9d2105512a4c2c53286c..a75a8265302b57d86f9a7e551f1cd43922dbeefe 100644 (file)
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -36,6 +36,7 @@
 #include <linux/hugetlb.h>
 #include <linux/page_idle.h>
 #include <linux/local_lock.h>
+#include <linux/buffer_head.h>
 
 #include "internal.h"
 
@@ -641,6 +642,7 @@ void lru_add_drain_cpu(int cpu)
                pagevec_lru_move_fn(pvec, lru_lazyfree_fn);
 
        activate_page_drain(cpu);
+       invalidate_bh_lrus_cpu(cpu);
 }
 
 /**
@@ -828,7 +830,8 @@ inline void __lru_add_drain_all(bool force_all_cpus)
                    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate_file, cpu)) ||
                    pagevec_count(&per_cpu(lru_pvecs.lru_deactivate, cpu)) ||
                    pagevec_count(&per_cpu(lru_pvecs.lru_lazyfree, cpu)) ||
-                   need_activate_page_drain(cpu)) {
+                   need_activate_page_drain(cpu) ||
+                   has_bh_in_lru(cpu, NULL)) {
                        INIT_WORK(work, lru_add_drain_per_cpu);
                        queue_work_on(cpu, mm_percpu_wq, work);
                        __cpumask_set_cpu(cpu, &has_work);