When boxes are run near (or to) OOM, we have a problem with the discard
page allocation in sd. If we fail allocating the special page, we return
busy, and it'll get retried. But since ordering is honored for dispatch
requests, we can keep retrying this same IO and failing. Behind that IO
could be requests that want to free memory, but they never get the
chance. This means you get repeated spews of traces like this:
[1201401.625972] Call Trace:
[1201401.631748] dump_stack+0x4d/0x65
[1201401.639445] warn_alloc+0xec/0x190
[1201401.647335] __alloc_pages_slowpath+0xe84/0xf30
[1201401.657722] ? get_page_from_freelist+0x11b/0xb10
[1201401.668475] ? __alloc_pages_slowpath+0x2e/0xf30
[1201401.679054] __alloc_pages_nodemask+0x1f9/0x210
[1201401.689424] alloc_pages_current+0x8c/0x110
[1201401.699025] sd_setup_write_same16_cmnd+0x51/0x150
[1201401.709987] sd_init_command+0x49c/0xb70
[1201401.719029] scsi_setup_cmnd+0x9c/0x160
[1201401.727877] scsi_queue_rq+0x4d9/0x610
[1201401.736535] blk_mq_dispatch_rq_list+0x19a/0x360
[1201401.747113] blk_mq_sched_dispatch_requests+0xff/0x190
[1201401.758844] __blk_mq_run_hw_queue+0x95/0xa0
[1201401.768653] blk_mq_run_work_fn+0x2c/0x30
[1201401.777886] process_one_work+0x14b/0x400
[1201401.787119] worker_thread+0x4b/0x470
[1201401.795586] kthread+0x110/0x150
[1201401.803089] ? rescuer_thread+0x320/0x320
[1201401.812322] ? kthread_park+0x90/0x90
[1201401.820787] ? do_syscall_64+0x53/0x150
[1201401.829635] ret_from_fork+0x29/0x40
Ensure that the discard page allocation has a mempool backing, so we
know we can make progress.
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>
---
We actually hit this in production, it's not a theoretical issue.
diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
index 4a6ed2fc8c71..a1a44f52e0e8 100644
--- a/drivers/scsi/sd.c
+++ b/drivers/scsi/sd.c
@@ -133,6 +133,7 @@ static DEFINE_MUTEX(sd_ref_mutex);
static struct kmem_cache *sd_cdb_cache;
static mempool_t *sd_cdb_pool;
+static mempool_t *sd_page_pool;
static const char *sd_cache_types[] = {
"write through", "none", "write back",
@@ -759,9 +760,10 @@ static blk_status_t sd_setup_unmap_cmnd(struct scsi_cmnd *cmd)
unsigned int data_len = 24;
char *buf;
- rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
+ rq->special_vec.bv_page = mempool_alloc(sd_page_pool, GFP_ATOMIC);
if (!rq->special_vec.bv_page)
return BLK_STS_RESOURCE;
+ clear_highpage(rq->special_vec.bv_page);
rq->special_vec.bv_offset = 0;
rq->special_vec.bv_len = data_len;
rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
@@ -793,9 +795,10 @@ static blk_status_t sd_setup_write_same16_cmnd(struct scsi_cmnd *cmd,
u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
u32 data_len = sdp->sector_size;
- rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
+ rq->special_vec.bv_page = mempool_alloc(sd_page_pool, GFP_ATOMIC);
if (!rq->special_vec.bv_page)
return BLK_STS_RESOURCE;
+ clear_highpage(rq->special_vec.bv_page);
rq->special_vec.bv_offset = 0;
rq->special_vec.bv_len = data_len;
rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
@@ -824,9 +827,10 @@ static blk_status_t sd_setup_write_same10_cmnd(struct scsi_cmnd *cmd,
u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
u32 data_len = sdp->sector_size;
- rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
+ rq->special_vec.bv_page = mempool_alloc(sd_page_pool, GFP_ATOMIC);
if (!rq->special_vec.bv_page)
return BLK_STS_RESOURCE;
+ clear_highpage(rq->special_vec.bv_page);
rq->special_vec.bv_offset = 0;
rq->special_vec.bv_len = data_len;
rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
@@ -1277,7 +1281,7 @@ static void sd_uninit_command(struct scsi_cmnd *SCpnt)
u8 *cmnd;
if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
- __free_page(rq->special_vec.bv_page);
+ mempool_free(rq->special_vec.bv_page, sd_page_pool);
if (SCpnt->cmnd != scsi_req(rq)->cmd) {
cmnd = SCpnt->cmnd;
@@ -3614,6 +3618,13 @@ static int __init init_sd(void)
goto err_out_cache;
}
+ sd_page_pool = mempool_create_page_pool(SD_MEMPOOL_SIZE, 0);
+ if (!sd_page_pool) {
+ printk(KERN_ERR "sd: can't init discard page pool\n");
+ err = -ENOMEM;
+ goto err_out_ppool;
+ }
+
err = scsi_register_driver(&sd_template.gendrv);
if (err)
goto err_out_driver;
@@ -3621,6 +3632,9 @@ static int __init init_sd(void)
return 0;
err_out_driver:
+ mempool_destroy(sd_page_pool);
+
+err_out_ppool:
mempool_destroy(sd_cdb_pool);
err_out_cache:
@@ -3647,6 +3661,7 @@ static void __exit exit_sd(void)
scsi_unregister_driver(&sd_template.gendrv);
mempool_destroy(sd_cdb_pool);
+ mempool_destroy(sd_page_pool);
kmem_cache_destroy(sd_cdb_cache);
class_unregister(&sd_disk_class);
--
Jens Axboe
