Hello, Currently block layer shares a single request_list (@q->rq) for all IOs regardless of their blkcg associations. This means that once the shared pool is exhausted, blkcg limits don't mean much. Whoever grabs the requests being freed the first grabs the next IO slot. This priority inversion can be easily demonstrated by creating a blkio cgroup w/ very low weight, put a program which can issue a lot of random direct IOs there and running a sequential IO from a different cgroup. As soon as the request pool is used up, the sequential IO bandwidth crashes. This patchset implements per-blkg request allocation so that each blkcg-request_queue pair has its own request pool to allocate from. This isolates different blkcgs in terms of request allocation. Most changes are straight-forward; unfortunately, bdi isn't blkcg-aware yet so it currently just propagates the congestion state from root blkcg. As writeback currently is always on the root blkcg, this kinda works for write congestion but readahead may behave non-optimally under congestion for now. This needs to be improved but the situation is still way better than blkcg completely collapsing. 0001-blkcg-fix-blkg_alloc-failure-path.patch 0002-blkcg-__blkg_lookup_create-doesn-t-have-to-fail-on-r.patch 0003-blkcg-make-root-blkcg-allocation-use-GFP_KERNEL.patch 0004-mempool-add-gfp_mask-to-mempool_create_node.patch 0005-block-drop-custom-queue-draining-used-by-scsi_transp.patch 0006-block-refactor-get_request-_wait.patch 0007-block-allocate-io_context-upfront.patch 0008-blkcg-inline-bio_blkcg-and-friends.patch 0009-block-add-q-nr_rqs-and-move-q-rq.elvpriv-to-q-nr_rqs.patch 0010-block-prepare-for-multiple-request_lists.patch 0011-blkcg-implement-per-blkg-request-allocation.patch 0001-0003 are assorted fixes / improvements which can be separated from this patchset. Just sending as part of this series for convenience. 0004 adds @gfp_mask to mempool_create_node(). This is necessary because blkg allocation is on the IO path and now blkg contains mempool for request_list. Note that blkg allocation failure doesn't lead to catastrophic failure. It just hinders blkcg enforcement. 0005 drops custom queue draining which I dont't think is necessary and hinders with further updates. 0006-0010 are prep patches and 0011 implements per-blkg request allocation. This patchset is on top of, block/for-3.5/core bd1a68b59c "vmsplice: relax alignement requireme..." + [1] blkcg: tg_stats_alloc_lock is an irq lock and is also available in the following git branch. git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blkcg-rl Jens, I still can't reproduce the boot failure you were seeing on block/for-3.5/core, so am just basing this series on top. Once we figure that one out, we can resequence the patches. Thanks. block/blk-cgroup.c | 147 ++++++++++++++++---------- block/blk-cgroup.h | 121 +++++++++++++++++++++ block/blk-core.c | 200 ++++++++++++++++++------------------ block/blk-sysfs.c | 34 +++--- block/blk-throttle.c | 3 block/blk.h | 3 block/bsg-lib.c | 53 --------- drivers/scsi/scsi_transport_fc.c | 38 ------ drivers/scsi/scsi_transport_iscsi.c | 2 include/linux/blkdev.h | 53 +++++---- include/linux/bsg-lib.h | 1 include/linux/mempool.h | 3 mm/mempool.c | 12 +- 13 files changed, 379 insertions(+), 291 deletions(-) -- tejun [1] http://article.gmane.org/gmane.linux.kernel/1288400 _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/containers
