On Wed, May 7, 2025 at 2:49 PM Uday Shankar ushankar@purestorage.com wrote:
In update_alloc_hint_after_get, we wrap the new hint back to 0 one bit too early. This breaks round robin tag allocation (BLK_MQ_F_TAG_RR) - some tags get skipped, so we don't get round robin tags even in the simple case of single-threaded load on a single hctx. Fix the off-by-one in the wrapping condition so that round robin tag allocation works properly.
The same pattern occurs in __sbitmap_get_word, so fix it there too.
Should this have a Fixes tag? Looks like the off-by-one wrapping has existed since 4bb659b15699 ("blk-mq: implement new and more efficient tagging scheme"), but it's only a correctness issue with round-robin tag allocation, which was added in 24391c0dc57c ("blk-mq: add tag allocation policy").
I don't have much background on blk-mq's round-robin tag allocation, but FWIW,
Reviewed-by: Caleb Sander Mateos csander@purestorage.com
Signed-off-by: Uday Shankar ushankar@purestorage.com
lib/sbitmap.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/lib/sbitmap.c b/lib/sbitmap.c index d3412984170c03dc6600bbe53f130404b765ac5a..aa1cec78b9649f1f3e8ef2d617dd7ee724391a8c 100644 --- a/lib/sbitmap.c +++ b/lib/sbitmap.c @@ -51,7 +51,7 @@ static inline void update_alloc_hint_after_get(struct sbitmap *sb, } else if (nr == hint || unlikely(sb->round_robin)) { /* Only update the hint if we used it. */ hint = nr + 1;
if (hint >= depth - 1)
if (hint >= depth) hint = 0; this_cpu_write(*sb->alloc_hint, hint); }@@ -182,7 +182,7 @@ static int __sbitmap_get_word(unsigned long *word, unsigned long depth, break;
hint = nr + 1;
if (hint >= depth - 1)
if (hint >= depth) hint = 0; }-- 2.34.1