Re: [PATCH v3 3/3] fsnotify: allow sleepable child flag update

[Stephen Brennan <stephen.s.brennan@oracle.com>]

Amir Goldstein <amir73il@gmail.com> writes:

> On Fri, Oct 28, 2022 at 3:10 AM Stephen Brennan
> <stephen.s.brennan@oracle.com> wrote:
>>
>> With very large d_subdirs lists, iteration can take a long time. Since
>> iteration needs to hold parent->d_lock, this can trigger soft lockups.
>> It would be best to make this iteration sleepable. Since we have the
>> inode locked exclusive, we can drop the parent->d_lock and sleep,
>> holding a reference to a child dentry, and continue iteration once we
>> wake.
>>
>> Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com>
>> ---
>>
>
> Reviewed-by: Amir Goldstein <amir73il@gmail.com>
>
> some comment nits and one fortify suggestion
>
>> Notes:
>>     v3:
>>     - removed if statements around dput()
>>     v2:
>>     - added a check for child->d_parent != alias and retry logic
>>
>>  fs/notify/fsnotify.c | 36 ++++++++++++++++++++++++++++++++----
>>  1 file changed, 32 insertions(+), 4 deletions(-)
>>
>> diff --git a/fs/notify/fsnotify.c b/fs/notify/fsnotify.c
>> index ccb8a3a6c522..34e5d18235a7 100644
>> --- a/fs/notify/fsnotify.c
>> +++ b/fs/notify/fsnotify.c
>> @@ -102,10 +102,12 @@ void fsnotify_sb_delete(struct super_block *sb)
>>   * on a child we run all of our children and set a dentry flag saying that the
>>   * parent cares.  Thus when an event happens on a child it can quickly tell
>>   * if there is a need to find a parent and send the event to the parent.
>> + *
>> + * Context: inode locked exclusive
>>   */
>>  static bool __fsnotify_update_children_dentry_flags(struct inode *inode)
>>  {
>> -       struct dentry *alias, *child;
>> +       struct dentry *child, *alias, *last_ref = NULL;
>>         int watched;
>>
>>         if (!S_ISDIR(inode->i_mode))
>> @@ -120,12 +122,37 @@ static bool __fsnotify_update_children_dentry_flags(struct inode *inode)
>>         alias = d_find_any_alias(inode);
>>
>>         /*
>> -        * run all of the children of the original inode and fix their
>> -        * d_flags to indicate parental interest (their parent is the
>> -        * original inode)
>> +        * These lists can get very long, so we may need to sleep during
>> +        * iteration. Normally this would be impossible without a cursor,
>> +        * but since we have the inode locked exclusive, we're guaranteed
>> +        * that the directory won't be modified, so whichever dentry we
>> +        * pick to sleep on won't get moved. So, start a manual iteration
>> +        * over d_subdirs which will allow us to sleep.
>>          */
>>         spin_lock(&alias->d_lock);
>> +retry:
>>         list_for_each_entry(child, &alias->d_subdirs, d_child) {
>> +               if (need_resched()) {
>> +                       /*
>> +                        * We need to hold a reference while we sleep. But when
>> +                        * we wake, dput() could free the dentry, invalidating
>> +                        * the list pointers. We can't look at the list pointers
>> +                        * until we re-lock the parent, and we can't dput() once
>> +                        * we have the parent locked. So the solution is to hold
>> +                        * onto our reference and free it the *next* time we drop
>> +                        * alias->d_lock: either at the end of the function, or
>> +                        * at the time of the next sleep.
>> +                        */
>
> My personal preference would be to move this above if (needed_reschd())
> it is not any less clear when this comment is above the condition
> and less indented will read nicer.

Definitely.

>> +                       dget(child);
>> +                       spin_unlock(&alias->d_lock);
>> +                       dput(last_ref);
>> +                       last_ref = child;
>> +                       cond_resched();
>> +                       spin_lock(&alias->d_lock);
>> +                       if (child->d_parent != alias)
>> +                               goto retry;
>
> Is this expected? If not, then we need a WARN_ON_ONCE().
> Also, I wonder if it would be better to break out and leave
> dentry flags as they are instead of risking some endless
> or very long retry loop?
>
> And how about asserting on unexpected !list_empty(&child->d_child)
> to avoid an endless loop in list_for_each_entry()?

I was trying to think this through as I prepared v3, and ended I up
leaving it as-is, out of a hope that it was doing something helpful. But
I'm pretty sure Al would ask why I believe that this could happen (e.g.
what scenario am I guarding against happening?). I didn't have a clear
idea of what I was guarding against here when I wrote it, which is my
fault. What's necessary is an audit of the places where d_child is
modified, so we can understand the risky places.

What we're doing here is dropping parent->d_lock and going to sleep,
while still having parent->d_inode->i_rwsem held in write mode. The risk
we have is that this particular dentry is removed from this list while
we slept. Checking d_parent seems particularly unhelpful for that, since
what we care far more about is the d_child list pointers. I did a code
audit of all locations where d_child is modified. Here's all I could
identify:

1. dentry_unlist() in fs/dcache.c
2. __d_move() in fs/dcache.c
3. Initialization code in d_alloc() family
4. Cursor code in fs/libfs.c for dcache_readdir()

For case #1, dentry_unlist() is a helper of dentry_kill(). The real guts
of dentry_kill() must happen with parent->d_lock held. Since we get a
reference to the child with parent->d_lock held, we can be confident
that any dentry_kill() which could have been happening concurrently will
bail out after seeing that we got a reference. At least, that's my
reading of the code.

For case #2, __d_move() can (should?) NOT happen, because the caller
must hold the i_mutex (read i_rwsem in write mode), but we are holding
that ourselves.

For case #3, it doesn't matter... it's initialization time, and we
couldn't have gotten a reference to the dentry until the dentry is put
into the tree.

For case #4, the simple_readdir() code should hold the inode rwsem in
read mode, so we have mutual exclusion from this code too. While it's
possible that we could go to sleep holding a reference to a cursor, I
believe that the cursor could not move without holding i_rwsem in read
mode. So we could never wake up to find we've skipped part of the list.

So to summarize, I don't think there's a case where we could actually
expect that d_child pointers get updated while we sleep with a reference
held and the parent i_rwsem held exclusive. Consequently, no place where
d_parent would change from under our feet. So I will remove this check
form the patch.

Thanks,
Stephen