mm: hwpoison: don't drop slab caches for offlining non-LRU page
authorYang Shi <shy828301@gmail.com>
Thu, 2 Sep 2021 21:58:31 +0000 (14:58 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 3 Sep 2021 16:58:15 +0000 (09:58 -0700)
In the current implementation of soft offline, if non-LRU page is met,
all the slab caches will be dropped to free the page then offline.  But
if the page is not slab page all the effort is wasted in vain.  Even
though it is a slab page, it is not guaranteed the page could be freed
at all.

However the side effect and cost is quite high.  It does not only drop
the slab caches, but also may drop a significant amount of page caches
which are associated with inode caches.  It could make the most
workingset gone in order to just offline a page.  And the offline is not
guaranteed to succeed at all, actually I really doubt the success rate
for real life workload.

Furthermore the worse consequence is the system may be locked up and
unusable since the page cache release may incur huge amount of works
queued for memcg release.

Actually we ran into such unpleasant case in our production environment.
Firstly, the workqueue of memory_failure_work_func is locked up as
below:

    BUG: workqueue lockup - pool cpus=1 node=0 flags=0x0 nice=0 stuck for 53s!
    Showing busy workqueues and worker pools:
    workqueue events: flags=0x0
     pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=14/256 refcnt=15
      in-flight: 409271:memory_failure_work_func
      pending: kfree_rcu_work, kfree_rcu_monitor, kfree_rcu_work, rht_deferred_worker, rht_deferred_worker, rht_deferred_worker, rht_deferred_worker, kfree_rcu_work, kfree_rcu_work, kfree_rcu_work, kfree_rcu_work, drain_local_stock, kfree_rcu_work
    workqueue mm_percpu_wq: flags=0x8
     pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/256 refcnt=2
      pending: vmstat_update
    workqueue cgroup_destroy: flags=0x0
      pwq 2: cpus=1 node=0 flags=0x0 nice=0 active=1/1 refcnt=12072
        pending: css_release_work_fn

There were over 12K css_release_work_fn queued, and this caused a few
lockups due to the contention of worker pool lock with IRQ disabled, for
example:

    NMI watchdog: Watchdog detected hard LOCKUP on cpu 1
    Modules linked in: amd64_edac_mod edac_mce_amd crct10dif_pclmul crc32_pclmul ghash_clmulni_intel xt_DSCP iptable_mangle kvm_amd bpfilter vfat fat acpi_ipmi i2c_piix4 usb_storage ipmi_si k10temp i2c_core ipmi_devintf ipmi_msghandler acpi_cpufreq sch_fq_codel xfs libcrc32c crc32c_intel mlx5_core mlxfw nvme xhci_pci ptp nvme_core pps_core xhci_hcd
    CPU: 1 PID: 205500 Comm: kworker/1:0 Tainted: G             L    5.10.32-t1.el7.twitter.x86_64 #1
    Hardware name: TYAN F5AMT /z        /S8026GM2NRE-CGN, BIOS V8.030 03/30/2021
    Workqueue: events memory_failure_work_func
    RIP: 0010:queued_spin_lock_slowpath+0x41/0x1a0
    Code: 41 f0 0f ba 2f 08 0f 92 c0 0f b6 c0 c1 e0 08 89 c2 8b 07 30 e4 09 d0 a9 00 01 ff ff 75 1b 85 c0 74 0e 8b 07 84 c0 74 08 f3 90 <8b> 07 84 c0 75 f8 b8 01 00 00 00 66 89 07 c3 f6 c4 01 75 04 c6 47
    RSP: 0018:ffff9b2ac278f900 EFLAGS: 00000002
    RAX: 0000000000480101 RBX: ffff8ce98ce71800 RCX: 0000000000000084
    RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff8ce98ce6a140
    RBP: 00000000000284c8 R08: ffffd7248dcb6808 R09: 0000000000000000
    R10: 0000000000000003 R11: ffff9b2ac278f9b0 R12: 0000000000000001
    R13: ffff8cb44dab9c00 R14: ffffffffbd1ce6a0 R15: ffff8cacaa37f068
    FS:  0000000000000000(0000) GS:ffff8ce98ce40000(0000) knlGS:0000000000000000
    CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
    CR2: 00007fcf6e8cb000 CR3: 0000000a0c60a000 CR4: 0000000000350ee0
    Call Trace:
     __queue_work+0xd6/0x3c0
     queue_work_on+0x1c/0x30
     uncharge_batch+0x10e/0x110
     mem_cgroup_uncharge_list+0x6d/0x80
     release_pages+0x37f/0x3f0
     __pagevec_release+0x1c/0x50
     __invalidate_mapping_pages+0x348/0x380
     inode_lru_isolate+0x10a/0x160
     __list_lru_walk_one+0x7b/0x170
     list_lru_walk_one+0x4a/0x60
     prune_icache_sb+0x37/0x50
     super_cache_scan+0x123/0x1a0
     do_shrink_slab+0x10c/0x2c0
     shrink_slab+0x1f1/0x290
     drop_slab_node+0x4d/0x70
     soft_offline_page+0x1ac/0x5b0
     memory_failure_work_func+0x6a/0x90
     process_one_work+0x19e/0x340
     worker_thread+0x30/0x360
     kthread+0x116/0x130

The lockup made the machine is quite unusable.  And it also made the
most workingset gone, the reclaimabled slab caches were reduced from 12G
to 300MB, the page caches were decreased from 17G to 4G.

But the most disappointing thing is all the effort doesn't make the page
offline, it just returns:

    soft_offline: 0x1469f2: unknown non LRU page type 5ffff0000000000 ()

It seems the aggressive behavior for non-LRU page didn't pay back, so it
doesn't make too much sense to keep it considering the terrible side
effect.

Link: https://lkml.kernel.org/r/20210819054116.266126-1-shy828301@gmail.com
Signed-off-by: Yang Shi <shy828301@gmail.com>
Reported-by: David Mackey <tdmackey@twitter.com>
Acked-by: David Hildenbrand <david@redhat.com>
Acked-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/mm.h
mm/hwpoison-inject.c
mm/memory-failure.c

index 35bbac3..11c3855 100644 (file)
@@ -3110,7 +3110,7 @@ extern void memory_failure_queue_kick(int cpu);
 extern int unpoison_memory(unsigned long pfn);
 extern int sysctl_memory_failure_early_kill;
 extern int sysctl_memory_failure_recovery;
-extern void shake_page(struct page *p, int access);
+extern void shake_page(struct page *p);
 extern atomic_long_t num_poisoned_pages __read_mostly;
 extern int soft_offline_page(unsigned long pfn, int flags);
 
index 1ae1ebc..aff4d27 100644 (file)
@@ -30,7 +30,7 @@ static int hwpoison_inject(void *data, u64 val)
        if (!hwpoison_filter_enable)
                goto inject;
 
-       shake_page(hpage, 0);
+       shake_page(hpage);
        /*
         * This implies unable to support non-LRU pages.
         */
index f83a2af..5decacb 100644 (file)
@@ -282,9 +282,9 @@ static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
 
 /*
  * Unknown page type encountered. Try to check whether it can turn PageLRU by
- * lru_add_drain_all, or a free page by reclaiming slabs when possible.
+ * lru_add_drain_all.
  */
-void shake_page(struct page *p, int access)
+void shake_page(struct page *p)
 {
        if (PageHuge(p))
                return;
@@ -296,11 +296,9 @@ void shake_page(struct page *p, int access)
        }
 
        /*
-        * Only call drop_slab_node here (which would also shrink
-        * other caches) if access is not potentially fatal.
+        * TODO: Could shrink slab caches here if a lightweight range-based
+        * shrinker will be available.
         */
-       if (access)
-               drop_slab_node(page_to_nid(p));
 }
 EXPORT_SYMBOL_GPL(shake_page);
 
@@ -1205,7 +1203,7 @@ try_again:
                         * page, retry.
                         */
                        if (pass++ < 3) {
-                               shake_page(p, 1);
+                               shake_page(p);
                                goto try_again;
                        }
                        ret = -EIO;
@@ -1222,7 +1220,7 @@ try_again:
                 */
                if (pass++ < 3) {
                        put_page(p);
-                       shake_page(p, 1);
+                       shake_page(p);
                        count_increased = false;
                        goto try_again;
                }
@@ -1369,7 +1367,7 @@ static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
         * shake_page() again to ensure that it's flushed.
         */
        if (mlocked)
-               shake_page(hpage, 0);
+               shake_page(hpage);
 
        /*
         * Now that the dirty bit has been propagated to the
@@ -1723,7 +1721,7 @@ try_again:
         * The check (unnecessarily) ignores LRU pages being isolated and
         * walked by the page reclaim code, however that's not a big loss.
         */
-       shake_page(p, 0);
+       shake_page(p);
 
        lock_page(p);