On 08.04.2019 9:05, Hugh Dickins wrote:
On Fri, 5 Apr 2019, Konstantin Khlebnikov wrote:
On 05.04.2019 5:12, Hugh Dickins wrote:
Hi Alex, could you please give the patch below a try? It fixes a
problem, but I'm not sure that it's your problem - please let us know.
I've not yet written up the commit description, and this should end up
as 4/4 in a series fixing several new swapoff issues: I'll wait to post
the finished series until heard back from you.
I did first try following the suggestion Konstantin had made back then,
for a similar shmem_writepage() case: atomic_inc_not_zero(&sb->s_active).
But it turned out to be difficult to get right in shmem_unuse(), because
of the way that relies on the inode as a cursor in the list - problem
when you've acquired an s_active reference, but fail to acquire inode
reference, and cannot safely release the s_active reference while still
holding the swaplist mutex.
If VFS offered an isgrab(inode), like igrab() but acquiring s_active
reference while holding i_lock, that would drop very easily into the
current shmem_unuse() as a replacement there for igrab(). But the rest
of the world has managed without that for years, so I'm disinclined to
add it just for this. And the patch below seems good enough without it.
Thanks,
Hugh
---
include/linux/shmem_fs.h | 1 +
mm/shmem.c | 39 ++++++++++++++++++---------------------
2 files changed, 19 insertions(+), 21 deletions(-)
--- 5.1-rc3/include/linux/shmem_fs.h 2019-03-17 16:18:15.181820820 -0700
+++ linux/include/linux/shmem_fs.h 2019-04-04 16:18:08.193512968 -0700
@@ -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;
};
--- 5.1-rc3/mm/shmem.c 2019-03-17 16:18:15.701823872 -0700
+++ linux/mm/shmem.c 2019-04-04 16:18:08.193512968 -0700
@@ -1081,9 +1081,15 @@ static void shmem_evict_inode(struct ino
}
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 +1233,27 @@ int shmem_unuse(unsigned int type, bool
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;
This seems too ad hoc solution.
I see what you mean by "ad hoc", but disagree with "too" ad hoc:
it's an appropriate solution, and a general one - I didn't invent it
for this, but for the huge tmpfs recoveries work items four years ago;
just changed the name from "info->recoveries" to "info->stop_eviction"
to let it be generalized to this swapoff case.
I prefer mine, since it simplifies shmem_unuse() (no igrab!), and has
the nice (but admittedly not essential) property of letting swapoff
proceed without delay and without unnecessary locking on unmounting
filesystems and evicting inodes. (Would I prefer to use the s_umount
technique for my recoveries case? I think not.) >
But yours should work too, with a slight change - see comments below,
where I've inlined yours. I'd better get on and post my four fixes
tomorrow, whether or not they fix Alex's case; then if people prefer
yours to my 4/4, yours can be swapped in instead.
Ok. But both swapoff and tmpfs umount does not look like
operations that should be concurrent by any cost.
shmem: fix race between shmem_unuse and umount
From: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
Function shmem_unuse could race with generic_shutdown_super.
Inode reference is not enough for preventing umount and freeing superblock.
Signed-off-by: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
---
mm/shmem.c | 24 +++++++++++++++++++++++-
1 file changed, 23 insertions(+), 1 deletion(-)
diff --git a/mm/shmem.c b/mm/shmem.c
index b3db3779a30a..2018a9a96bb7 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1218,6 +1218,10 @@ static int shmem_unuse_inode(struct inode *inode, unsigned int type,
return ret;
}
+static void shmem_synchronize_umount(struct super_block *sb, void *arg)
+{
+}
+
I think this can go away, see below.
/*
* Read all the shared memory data that resides in the swap
* device 'type' back into memory, so the swap device can be
@@ -1229,6 +1233,7 @@ int shmem_unuse(unsigned int type, bool frontswap,
struct shmem_inode_info *info, *next;
struct inode *inode;
struct inode *prev_inode = NULL;
+ struct super_block *sb;
int error = 0;
if (list_empty(&shmem_swaplist))
@@ -1247,9 +1252,22 @@ int shmem_unuse(unsigned int type, bool frontswap,
continue;
}
+ /*
+ * Lock superblock to prevent umount and freeing it under us.
+ * If umount in progress it will free swap enties.
+ *
+ * Must be done before grabbing inode reference, otherwise
+ * generic_shutdown_super() will complain about busy inodes.
+ */
+ sb = info->vfs_inode.i_sb;
+ if (!trylock_super(sb))
Right, trylock important there.
+ continue;
+
inode = igrab(&info->vfs_inode);
- if (!inode)
+ if (!inode) {
+ up_read(&sb->s_umount);
Yes, that indeed avoids the difficulty I had with when to call
deactivate_super(), that put me off trying to use s_active.
continue;
+ }
mutex_unlock(&shmem_swaplist_mutex);
if (prev_inode)
@@ -1258,6 +1276,7 @@ int shmem_unuse(unsigned int type, bool frontswap,
error = shmem_unuse_inode(inode, type, frontswap,
fs_pages_to_unuse);
+ up_read(&sb->s_umount);
No, not here. I think you have to note prev_sb, and then only
up_read(&prev_sb->s_umount) after each iput(prev_inode): otherwise
there's still a risk of "Self-destruct in 5 seconds", isn't there?
Oh yes. So, this code have to swap sb locks above with this monster
if (sb != info->vfs_inode.i_sb) {
if (sb)
up_read(&sb->s_umount);
sb = NULL;
if (!trylock_super(info->vfs_inode.i_sb))
continue;
sb = info->vfs_inode.i_sb
}
Locking shmem_swaplist_mutex under s_umount should be fine.
Also I looking into idea of treating swapoff like reverse-writeback:
-> iterate over superblocks
-> lock s_umount with normal down_read
-> iterate over inodes
-> iterate over inode tags
-> ...
Whole code will be more natural in this way.
cond_resched();
mutex_lock(&shmem_swaplist_mutex);
@@ -1272,6 +1291,9 @@ int shmem_unuse(unsigned int type, bool frontswap,
if (prev_inode)
iput(prev_inode);
+ /* Wait for umounts, this grabs s_umount for each superblock. */
+ iterate_supers_type(&shmem_fs_type, shmem_synchronize_umount, NULL);
+
I guess that's an attempt to compensate for the somewhat unsatisfactory
trylock above (bearing in mind the SWAP_UNUSE_MAX_TRIES 3, but I remove
that in my 2/4). Nice idea, and if it had the effect of never needing to
retry shmem_unuse(), I'd say yes; but since you're still passing over
un-igrab()-able inodes without an equivalent synchronization, I think
this odd iterate_supers_type() just delays swapoff without buying any
guarantee: better just deleted to keep your patch simpler.
Yep, robust algorithm is better than try-3-times-and-give-up =)
(could hide bugs for ages)
I suppose your solution will wait for wakeup from shmem_evict_inode()?
That should work. In more general design this could be something like
__wait_on_freeing_inode(), but with killable wait.
return error;
}
Thanks,
Hugh
