]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
zsmalloc: return -EBUSY for zspage migration lock contention
authorteawater <zhuhui@kylinos.cn>
Thu, 19 Mar 2026 06:59:24 +0000 (14:59 +0800)
committerAndrew Morton <akpm@linux-foundation.org>
Sun, 5 Apr 2026 20:53:33 +0000 (13:53 -0700)
movable_operations::migrate_page() should return an appropriate error code
for temporary migration failures so the migration core can handle them
correctly.

zs_page_migrate() currently returns -EINVAL when zspage_write_trylock()
fails.  That path reflects transient lock contention, not invalid input,
so -EINVAL is clearly wrong.

However, -EAGAIN is also inappropriate here: the zspage's reader-lock
owner may hold the lock for an unbounded duration due to slow
decompression or reader-lock owner preemption.  Since migration retries
are bounded by NR_MAX_MIGRATE_PAGES_RETRY and performed with virtually no
delay between attempts, there is no guarantee the lock will be released in
time for a retry to succeed.  -EAGAIN implies "try again soon", which does
not hold in this case.

Return -EBUSY instead, which more accurately conveys that the resource is
occupied and migration cannot proceed at this time.

Link: https://lkml.kernel.org/r/20260319065924.69337-1-hui.zhu@linux.dev
Signed-off-by: teawater <zhuhui@kylinos.cn>
Acked-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/zsmalloc.c

index 2c1430bf8d57cf4f0203ad5375bf8334b6d2886f..e7417ece1c12ebe058486ceeb9aa3f265b1e6583 100644 (file)
@@ -1727,7 +1727,19 @@ static int zs_page_migrate(struct page *newpage, struct page *page,
        if (!zspage_write_trylock(zspage)) {
                spin_unlock(&class->lock);
                write_unlock(&pool->lock);
-               return -EINVAL;
+               /*
+                * Return -EBUSY but not -EAGAIN: the zspage's reader-lock
+                * owner may hold the lock for an unbounded duration due to a
+                * slow decompression or reader-lock owner preemption.
+                * Since migration retries are bounded by
+                * NR_MAX_MIGRATE_PAGES_RETRY and performed with virtually no
+                * delay between attempts, there is no guarantee the lock will
+                * be released in time for a retry to succeed.
+                * -EAGAIN implies "try again soon", which does not hold here.
+                * -EBUSY more accurately conveys "resource is occupied,
+                * migration cannot proceed".
+                */
+               return -EBUSY;
        }
 
        /* We're committed, tell the world that this is a Zsmalloc page. */