mm: swapoff: shmem_unuse() stop eviction without igrab()
authorHugh Dickins <hughd@google.com>
Fri, 19 Apr 2019 00:50:13 +0000 (17:50 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 19 Apr 2019 16:46:04 +0000 (09:46 -0700)
The igrab() in shmem_unuse() looks good, but we forgot that it gives no
protection against concurrent unmounting: a point made by Konstantin
Khlebnikov eight years ago, and then fixed in 2.6.39 by 778dd893ae78
("tmpfs: fix race between umount and swapoff").  The current 5.1-rc
swapoff is liable to hit "VFS: Busy inodes after unmount of tmpfs.
Self-destruct in 5 seconds.  Have a nice day..." followed by GPF.

Once again, give up on using igrab(); but don't go back to making such
heavy-handed use of shmem_swaplist_mutex as last time: that would spoil
the new design, and I expect could deadlock inside shmem_swapin_page().

Instead, shmem_unuse() just raise a "stop_eviction" count in the shmem-
specific inode, and shmem_evict_inode() wait for that to go down to 0.
Call it "stop_eviction" rather than "swapoff_busy" because it can be put
to use for others later (huge tmpfs patches expect to use it).

That simplifies shmem_unuse(), protecting it from both unlink and
unmount; and in practice lets it locate all the swap in its first try.
But do not rely on that: there's still a theoretical case, when
shmem_writepage() might have been preempted after its get_swap_page(),
before making the swap entry visible to swapoff.

[hughd@google.com: remove incorrect list_del()]
Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1904091133570.1898@eggly.anvils
Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1904081259400.1523@eggly.anvils
Fixes: b56a2d8af914 ("mm: rid swapoff of quadratic complexity")
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: "Alex Xu (Hello71)" <alex_y_xu@yahoo.ca>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Kelley Nielsen <kelleynnn@gmail.com>
Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Cc: Rik van Riel <riel@surriel.com>
Cc: Vineeth Pillai <vpillai@digitalocean.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/shmem_fs.h
mm/shmem.c
mm/swapfile.c

index f3fb1edb3526ddc0c582f0ad32017ab7eaf21dd3..20d815a331454f93e7a66d808a5a5f84601e9a58 100644 (file)
@@ -21,6 +21,7 @@ struct shmem_inode_info {
        struct list_head        swaplist;       /* chain of maybes on swap */
        struct shared_policy    policy;         /* NUMA memory alloc policy */
        struct simple_xattrs    xattrs;         /* list of xattrs */
+       atomic_t                stop_eviction;  /* hold when working on inode */
        struct inode            vfs_inode;
 };
 
index 859e8628071f19c51a6eb65d2197809f43044ab1..2275a0ff7c3051d9674ee59b153451c6078741c3 100644 (file)
@@ -1081,9 +1081,14 @@ static void shmem_evict_inode(struct inode *inode)
                        }
                        spin_unlock(&sbinfo->shrinklist_lock);
                }
-               if (!list_empty(&info->swaplist)) {
+               while (!list_empty(&info->swaplist)) {
+                       /* Wait while shmem_unuse() is scanning this inode... */
+                       wait_var_event(&info->stop_eviction,
+                                      !atomic_read(&info->stop_eviction));
                        mutex_lock(&shmem_swaplist_mutex);
-                       list_del_init(&info->swaplist);
+                       /* ...but beware of the race if we peeked too early */
+                       if (!atomic_read(&info->stop_eviction))
+                               list_del_init(&info->swaplist);
                        mutex_unlock(&shmem_swaplist_mutex);
                }
        }
@@ -1227,36 +1232,27 @@ int shmem_unuse(unsigned int type, bool frontswap,
                unsigned long *fs_pages_to_unuse)
 {
        struct shmem_inode_info *info, *next;
-       struct inode *inode;
-       struct inode *prev_inode = NULL;
        int error = 0;
 
        if (list_empty(&shmem_swaplist))
                return 0;
 
        mutex_lock(&shmem_swaplist_mutex);
-
-       /*
-        * The extra refcount on the inode is necessary to safely dereference
-        * p->next after re-acquiring the lock. New shmem inodes with swap
-        * get added to the end of the list and we will scan them all.
-        */
        list_for_each_entry_safe(info, next, &shmem_swaplist, swaplist) {
                if (!info->swapped) {
                        list_del_init(&info->swaplist);
                        continue;
                }
-
-               inode = igrab(&info->vfs_inode);
-               if (!inode)
-                       continue;
-
+               /*
+                * Drop the swaplist mutex while searching the inode for swap;
+                * but before doing so, make sure shmem_evict_inode() will not
+                * remove placeholder inode from swaplist, nor let it be freed
+                * (igrab() would protect from unlink, but not from unmount).
+                */
+               atomic_inc(&info->stop_eviction);
                mutex_unlock(&shmem_swaplist_mutex);
-               if (prev_inode)
-                       iput(prev_inode);
-               prev_inode = inode;
 
-               error = shmem_unuse_inode(inode, type, frontswap,
+               error = shmem_unuse_inode(&info->vfs_inode, type, frontswap,
                                          fs_pages_to_unuse);
                cond_resched();
 
@@ -1264,14 +1260,13 @@ int shmem_unuse(unsigned int type, bool frontswap,
                next = list_next_entry(info, swaplist);
                if (!info->swapped)
                        list_del_init(&info->swaplist);
+               if (atomic_dec_and_test(&info->stop_eviction))
+                       wake_up_var(&info->stop_eviction);
                if (error)
                        break;
        }
        mutex_unlock(&shmem_swaplist_mutex);
 
-       if (prev_inode)
-               iput(prev_inode);
-
        return error;
 }
 
@@ -2238,6 +2233,7 @@ static struct inode *shmem_get_inode(struct super_block *sb, const struct inode
                info = SHMEM_I(inode);
                memset(info, 0, (char *)inode - (char *)info);
                spin_lock_init(&info->lock);
+               atomic_set(&info->stop_eviction, 0);
                info->seals = F_SEAL_SEAL;
                info->flags = flags & VM_NORESERVE;
                INIT_LIST_HEAD(&info->shrinklist);
index 71383625a582c47dd87281655329b6eeada87121..cf63b5f01adf7da9d1def93f8763b50243adc698 100644 (file)
@@ -2116,12 +2116,11 @@ retry:
         * Under global memory pressure, swap entries can be reinserted back
         * into process space after the mmlist loop above passes over them.
         *
-        * Limit the number of retries? No: when shmem_unuse()'s igrab() fails,
-        * a shmem inode using swap is being evicted; and when mmget_not_zero()
-        * above fails, that mm is likely to be freeing swap from exit_mmap().
-        * Both proceed at their own independent pace: we could move them to
-        * separate lists, and wait for those lists to be emptied; but it's
-        * easier and more robust (though cpu-intensive) just to keep retrying.
+        * Limit the number of retries? No: when mmget_not_zero() above fails,
+        * that mm is likely to be freeing swap from exit_mmap(), which proceeds
+        * at its own independent pace; and even shmem_writepage() could have
+        * been preempted after get_swap_page(), temporarily hiding that swap.
+        * It's easy and robust (though cpu-intensive) just to keep retrying.
         */
        if (si->inuse_pages) {
                if (!signal_pending(current))