From: Hugh Dickins Date: Sun, 6 Apr 2008 22:56:57 +0000 (+0100) Subject: scsi: fix sense_slab/bio swapping livelock X-Git-Tag: v2.6.25-rc9~60 X-Git-Url: http://pilppa.com/gitweb/?a=commitdiff_plain;h=164fc5dcd6a1026fc713f5c63fad899aa484888c;p=linux-2.6-omap-h63xx.git scsi: fix sense_slab/bio swapping livelock Since 2.6.25-rc7, I've been seeing an occasional livelock on one x86_64 machine, copying kernel trees to tmpfs, paging out to swap. Signature: 6000 pages under writeback but never getting written; most tasks of interest trying to reclaim, but each get_swap_bio waiting for a bio in mempool_alloc's io_schedule_timeout(5*HZ); every five seconds an atomic page allocation failure report from kblockd failing to allocate a sense_buffer in __scsi_get_command. __scsi_get_command has a (one item) free_list to protect against this, but rc1's [SCSI] use dynamically allocated sense buffer de25deb18016f66dcdede165d07654559bb332bc upset that slightly. When it fails to allocate from the separate sense_slab, instead of giving up, it must fall back to the command free_list, which is sure to have a sense_buffer attached. Either my earlier -rc testing missed this, or there's some recent contributory factor. One very significant factor is SLUB, which merges slab caches when it can, and on 64-bit happens to merge both bio cache and sense_slab cache into kmalloc's 128-byte cache: so that under this swapping load, bios above are liable to gobble up all the slots needed for scsi_cmnd sense_buffers below. That's disturbing behaviour, and I tried a few things to fix it. Adding a no-op constructor to the sense_slab inhibits SLUB from merging it, and stops all the allocation failures I was seeing; but it's rather a hack, and perhaps in different configurations we have other caches on the swapout path which are ill-merged. Another alternative is to revert the separate sense_slab, using cache-line-aligned sense_buffer allocated beyond scsi_cmnd from the one kmem_cache; but that might waste more memory, and is only a way of diverting around the known problem. While I don't like seeing the allocation failures, and hate the idea of all those bios piled up above a scsi host working one by one, it does seem to emerge fairly soon with the livelock fix. So lacking better ideas, stick with that one clear fix for now. Signed-off-by: Hugh Dickins Cc: James Bottomley Cc: Andrew Morton Cc: FUJITA Tomonori Cc: Jens Axboe Cc: Christoph Lameter Cc: Pekka Enberg Cc: Peter Zijlstra Cc: Rafael J. Wysocki Signed-off-by: Linus Torvalds --- diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c index e5c6f6af876..c78b836f59d 100644 --- a/drivers/scsi/scsi.c +++ b/drivers/scsi/scsi.c @@ -181,6 +181,18 @@ struct scsi_cmnd *__scsi_get_command(struct Scsi_Host *shost, gfp_t gfp_mask) cmd = kmem_cache_alloc(shost->cmd_pool->cmd_slab, gfp_mask | shost->cmd_pool->gfp_mask); + if (likely(cmd)) { + buf = kmem_cache_alloc(shost->cmd_pool->sense_slab, + gfp_mask | shost->cmd_pool->gfp_mask); + if (likely(buf)) { + memset(cmd, 0, sizeof(*cmd)); + cmd->sense_buffer = buf; + } else { + kmem_cache_free(shost->cmd_pool->cmd_slab, cmd); + cmd = NULL; + } + } + if (unlikely(!cmd)) { unsigned long flags; @@ -197,16 +209,6 @@ struct scsi_cmnd *__scsi_get_command(struct Scsi_Host *shost, gfp_t gfp_mask) memset(cmd, 0, sizeof(*cmd)); cmd->sense_buffer = buf; } - } else { - buf = kmem_cache_alloc(shost->cmd_pool->sense_slab, - gfp_mask | shost->cmd_pool->gfp_mask); - if (likely(buf)) { - memset(cmd, 0, sizeof(*cmd)); - cmd->sense_buffer = buf; - } else { - kmem_cache_free(shost->cmd_pool->cmd_slab, cmd); - cmd = NULL; - } } return cmd;