We ran into an issue in production where reading an NVMe drive log
randomly fails. The request is a big one, 1056K, and the application
(nvme-cli) nicely aligns the buffer. This means the kernel will attempt
to map the pages in for zero-copy DMA, but we ultimately fail adding
enough pages to the request to satisfy the size requirement as we ran
out of segments supported by the hardware.
If we fail a user map that attempts to map pages in directly, retry with
copy == true set.
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
---
diff --git a/block/blk-map.c b/block/blk-map.c
index db9373bd31ac..2dc5286baec6 100644
--- a/block/blk-map.c
+++ b/block/blk-map.c
@@ -131,6 +131,7 @@ int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
else if (queue_virt_boundary(q))
copy = queue_virt_boundary(q) & iov_iter_gap_alignment(iter);
+retry:
i = *iter;
do {
ret =__blk_rq_map_user_iov(rq, map_data, &i, gfp_mask, copy);
@@ -148,6 +149,10 @@ int blk_rq_map_user_iov(struct request_queue *q, struct request *rq,
__blk_rq_unmap_user(bio);
fail:
rq->bio = NULL;
+ if (ret && !copy) {
+ copy = true;
+ goto retry;
+ }
return ret;
}
EXPORT_SYMBOL(blk_rq_map_user_iov);
--
Jens Axboe
