Re: [PATCH 1/2] mm: Retry migration earlier upon refcount mismatch

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 8/19/24 12:28, Huang, Ying wrote:
Dev Jain <dev.jain@xxxxxxx> writes:

On 8/13/24 12:52, Dev Jain wrote:
On 8/13/24 10:30, Dev Jain wrote:
On 8/12/24 17:38, Dev Jain wrote:
On 8/12/24 13:01, Huang, Ying wrote:
Dev Jain <dev.jain@xxxxxxx> writes:

On 8/12/24 11:45, Huang, Ying wrote:
Dev Jain <dev.jain@xxxxxxx> writes:

On 8/12/24 11:04, Huang, Ying wrote:
Hi, Dev,

Dev Jain <dev.jain@xxxxxxx> writes:

As already being done in __migrate_folio(), wherein we
backoff if the
folio refcount is wrong, make this check during the
unmapping phase, upon
the failure of which, the original state of the PTEs will be
restored and
the folio lock will be dropped via migrate_folio_undo_src(),
any racing
thread will make progress and migration will be retried.

Signed-off-by: Dev Jain <dev.jain@xxxxxxx>
---
     mm/migrate.c | 9 +++++++++
     1 file changed, 9 insertions(+)

diff --git a/mm/migrate.c b/mm/migrate.c
index e7296c0fb5d5..477acf996951 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1250,6 +1250,15 @@ static int
migrate_folio_unmap(new_folio_t get_new_folio,
         }
           if (!folio_mapped(src)) {
+        /*
+         * Someone may have changed the refcount and maybe
sleeping
+         * on the folio lock. In case of refcount mismatch,
bail out,
+         * let the system make progress and retry.
+         */
+        struct address_space *mapping = folio_mapping(src);
+
+        if (folio_ref_count(src) !=
folio_expected_refs(mapping, src))
+            goto out;
             __migrate_folio_record(dst, old_page_state,
anon_vma);
             return MIGRATEPAGE_UNMAP;
         }
Do you have some test results for this?  For example, after
applying the
patch, the migration success rate increased XX%, etc.
I'll get back to you on this.

My understanding for this issue is that the migration success
rate can
increase if we undo all changes before retrying. This is the
current
behavior for sync migration, but not for async migration.  If
so, we can
use migrate_pages_sync() for async migration too to increase
success
rate?  Of course, we need to change the function name and
comments.
As per my understanding, this is not the current behaviour for sync
migration. After successful unmapping, we fail in
migrate_folio_move()
with -EAGAIN, we do not call undo src+dst (rendering the loop
around
migrate_folio_move() futile), we do not push the failed folio
onto the
ret_folios list, therefore, in _sync(), _batch() is never
tried again.
In migrate_pages_sync(), migrate_pages_batch(,MIGRATE_ASYNC) will be
called first, if failed, the folio will be restored to the original
state (unlocked).  Then migrate_pages_batch(,_SYNC*) is called
again.
So, we unlock once.  If it's necessary, we can unlock more times via
another level of loop.
Yes, that's my point. We need to undo src+dst and retry.
For sync migration, we undo src+dst and retry now, but only once.  You
have shown that more retrying increases success rate.

We will have
to decide where we want this retrying to be; do we want to change the
return value, end up in the while loop wrapped around _sync(),
and retry
there by adding another level of loop, or do we want to make use
of the
existing retry loops, one of which is wrapped around _unmap();
the latter
is my approach. The utility I see for the former approach is
that, in case
of a large number of page migrations (which should usually be
the case),
we are giving more time for the folio to get retried. The latter
does not
give much time and discards the folio if it did not succeed
under 7 times.
Because it's a race, I guess that most folios will be migrated
successfully in the first pass.

My concerns of your method are that it deal with just one case
specially.  While retrying after undoing all appears more general.

Makes sense. Also, please ignore my "change the return value"
thing, I got confused between unmap_folios, ret_folios, etc.
Now I think I understood what the lists are doing :)

If it's really important to retry after undoing all, we can either
convert two retying loops of migrate_pages_batch() into one loop, or
remove retry loop in migrate_pages_batch() and retry in its caller
instead.
And if I implemented this correctly, the following makes the test
pass always:
https://www.codedump.xyz/diff/Zrn7EdxzNXmXyNXe

Okay, I did mess up with the implementation, leading to a false
positive. Let me try again :)

Hopefully this should do the job:
https://www.codedump.xyz/diff/ZrsIV8JSOPYx5V_u

But the result is worse than the patch proposed; I rarely hit
a 3 digit number of successes of move_pages(). But, on a
base kernel without any changes, when I apply David's
suggestion to change the test, if I choose 7 as the number
of retries (= NR_MAX_MIGRATE_SYNC_RETRY) in the test, I
can touch even 4 digits. I am puzzled.
We can also try merging the for loops of unmap and move...

If people are okay with this change, I guess I can send it as
a v2? I concur with your assessment that my initial approach
is solving a specific case; the above approach does give me
a slight improvement on arm64 and should be an improvement
in general, since it makes sense to defer retrying the failed folio
as much as we can.
We need to deal with something else before a formal v2,

- stats need to be fixed, please check result processing for the first
   loop of migrate_pages_sync().

Sorry, can you point out where do they need to be fixed exactly?
The change I did is inside the while(!list_empty(from)) block,
and there is no stat computation being done there already.


- Do we need something similar for async migration.

- Can we add another level of explicit loop for the second loop of
   migrate_pages_sync()?  That is to improve code readability.  Or, add a
   function to dot that?

- Is it good to remove retry loop in migrate_pages_batch()?  And do
   retry in the caller?

I am personally in favour of leaving the retry loop, and async
migration, as it is. Since async version is basically minimal-effort
migration, it won't make sense to "optimize" it, given the code churn
it would create, including the change we will have to then do in
"if (mode == MIGRATE_ASYNC) => migrate_pages_batch(ASYNC)" inside
migrate_pages().

Sorry, what do you mean by "another level of explicit loop"?


--
Best Regards,
Huang, Ying




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux