summaryrefslogtreecommitdiffstats
path: root/block
Commit message (Collapse)AuthorAgeFilesLines
* vfs: do bulk POLL* -> EPOLL* replacementLinus Torvalds2018-02-111-2/+2
| | | | | | | | | | | | | | | | | | | | | | | This is the mindless scripted replacement of kernel use of POLL* variables as described by Al, done by this script: for V in IN OUT PRI ERR RDNORM RDBAND WRNORM WRBAND HUP RDHUP NVAL MSG; do L=`git grep -l -w POLL$V | grep -v '^t' | grep -v /um/ | grep -v '^sa' | grep -v '/poll.h$'|grep -v '^D'` for f in $L; do sed -i "-es/^\([^\"]*\)\(\<POLL$V\>\)/\\1E\\2/" $f; done done with de-mangling cleanups yet to come. NOTE! On almost all architectures, the EPOLL* constants have the same values as the POLL* constants do. But they keyword here is "almost". For various bad reasons they aren't the same, and epoll() doesn't actually work quite correctly in some cases due to this on Sparc et al. The next patch from Al will sort out the final differences, and we should be all done. Scripted-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* block, bfq: add requeue-request hookPaolo Valente2018-02-071-25/+82
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Commit 'a6a252e64914 ("blk-mq-sched: decide how to handle flush rq via RQF_FLUSH_SEQ")' makes all non-flush re-prepared requests for a device be re-inserted into the active I/O scheduler for that device. As a consequence, I/O schedulers may get the same request inserted again, even several times, without a finish_request invoked on that request before each re-insertion. This fact is the cause of the failure reported in [1]. For an I/O scheduler, every re-insertion of the same re-prepared request is equivalent to the insertion of a new request. For schedulers like mq-deadline or kyber, this fact causes no harm. In contrast, it confuses a stateful scheduler like BFQ, which keeps state for an I/O request, until the finish_request hook is invoked on the request. In particular, BFQ may get stuck, waiting forever for the number of request dispatches, of the same request, to be balanced by an equal number of request completions (while there will be one completion for that request). In this state, BFQ may refuse to serve I/O requests from other bfq_queues. The hang reported in [1] then follows. However, the above re-prepared requests undergo a requeue, thus the requeue_request hook of the active elevator is invoked for these requests, if set. This commit then addresses the above issue by properly implementing the hook requeue_request in BFQ. [1] https://marc.info/?l=linux-block&m=151211117608676 Reported-by: Ivan Kozik <ivan@ludios.org> Reported-by: Alban Browaeys <alban.browaeys@gmail.com> Tested-by: Mike Galbraith <efault@gmx.de> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Serena Ziviani <ziviani.serena@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
* block: Add should_fail_bio() for bpf error injectionHoward McLauchlan2018-02-061-1/+10
| | | | | | | | | | | | | | | The classic error injection mechanism, should_fail_request() does not support use cases where more information is required (from the entire struct bio, for example). To that end, this patch introduces should_fail_bio(), which calls should_fail_request() under the hood but provides a convenient place for kprobes to hook into if they require the entire struct bio. This patch also replaces some existing calls to should_fail_request() with should_fail_bio() with no degradation in performance. Signed-off-by: Howard McLauchlan <hmclauchlan@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
* blk-wbt: account flush requests correctlyJens Axboe2018-02-061-1/+9
| | | | | | | | | | | | | | | | | | Mikulas reported a workload that saw bad performance, and figured out what it was due to various other types of requests being accounted as reads. Flush requests, for instance. Due to the high latency of those, we heavily throttle the writes to keep the latencies in balance. But they really should be accounted as writes. Fix this by checking the exact type of the request. If it's a read, account as a read, if it's a write or a flush, account as a write. Any other request we disregard. Previously everything would have been mistakenly accounted as reads. Reported-by: Mikulas Patocka <mpatocka@redhat.com> Cc: stable@vger.kernel.org # v4.12+ Signed-off-by: Jens Axboe <axboe@kernel.dk>
* Merge tag 'for-linus-20180204' of git://git.kernel.dk/linux-blockLinus Torvalds2018-02-044-7/+47
|\ | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Pull more block updates from Jens Axboe: "Most of this is fixes and not new code/features: - skd fix from Arnd, fixing a build error dependent on sla allocator type. - blk-mq scheduler discard merging fixes, one from me and one from Keith. This fixes a segment miscalculation for blk-mq-sched, where we mistakenly think two segments are physically contigious even though the request isn't carrying real data. Also fixes a bio-to-rq merge case. - Don't re-set a bit on the buffer_head flags, if it's already set. This can cause scalability concerns on bigger machines and workloads. From Kemi Wang. - Add BLK_STS_DEV_RESOURCE return value to blk-mq, allowing us to distuingish between a local (device related) resource starvation and a global one. The latter might happen without IO being in flight, so it has to be handled a bit differently. From Ming" * tag 'for-linus-20180204' of git://git.kernel.dk/linux-block: block: skd: fix incorrect linux/slab_def.h inclusion buffer: Avoid setting buffer bits that are already set blk-mq-sched: Enable merging discard bio into request blk-mq: fix discard merge with scheduler attached blk-mq: introduce BLK_STS_DEV_RESOURCE
| * blk-mq-sched: Enable merging discard bio into requestKeith Busch2018-02-011-0/+2
| | | | | | | | | | Signed-off-by: Keith Busch <keith.busch@intel.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * blk-mq: fix discard merge with scheduler attachedJens Axboe2018-02-012-3/+28
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | I ran into an issue on my laptop that triggered a bug on the discard path: WARNING: CPU: 2 PID: 207 at drivers/nvme/host/core.c:527 nvme_setup_cmd+0x3d3/0x430 Modules linked in: rfcomm fuse ctr ccm bnep arc4 binfmt_misc snd_hda_codec_hdmi nls_iso8859_1 nls_cp437 vfat snd_hda_codec_conexant fat snd_hda_codec_generic iwlmvm snd_hda_intel snd_hda_codec snd_hwdep mac80211 snd_hda_core snd_pcm snd_seq_midi snd_seq_midi_event snd_rawmidi snd_seq x86_pkg_temp_thermal intel_powerclamp kvm_intel uvcvideo iwlwifi btusb snd_seq_device videobuf2_vmalloc btintel videobuf2_memops kvm snd_timer videobuf2_v4l2 bluetooth irqbypass videobuf2_core aesni_intel aes_x86_64 crypto_simd cryptd snd glue_helper videodev cfg80211 ecdh_generic soundcore hid_generic usbhid hid i915 psmouse e1000e ptp pps_core xhci_pci xhci_hcd intel_gtt CPU: 2 PID: 207 Comm: jbd2/nvme0n1p7- Tainted: G U 4.15.0+ #176 Hardware name: LENOVO 20FBCTO1WW/20FBCTO1WW, BIOS N1FET59W (1.33 ) 12/19/2017 RIP: 0010:nvme_setup_cmd+0x3d3/0x430 RSP: 0018:ffff880423e9f838 EFLAGS: 00010217 RAX: 0000000000000000 RBX: ffff880423e9f8c8 RCX: 0000000000010000 RDX: ffff88022b200010 RSI: 0000000000000002 RDI: 00000000327f0000 RBP: ffff880421251400 R08: ffff88022b200000 R09: 0000000000000009 R10: 0000000000000000 R11: 0000000000000000 R12: 000000000000ffff R13: ffff88042341e280 R14: 000000000000ffff R15: ffff880421251440 FS: 0000000000000000(0000) GS:ffff880441500000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 000055b684795030 CR3: 0000000002e09006 CR4: 00000000001606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: nvme_queue_rq+0x40/0xa00 ? __sbitmap_queue_get+0x24/0x90 ? blk_mq_get_tag+0xa3/0x250 ? wait_woken+0x80/0x80 ? blk_mq_get_driver_tag+0x97/0xf0 blk_mq_dispatch_rq_list+0x7b/0x4a0 ? deadline_remove_request+0x49/0xb0 blk_mq_do_dispatch_sched+0x4f/0xc0 blk_mq_sched_dispatch_requests+0x106/0x170 __blk_mq_run_hw_queue+0x53/0xa0 __blk_mq_delay_run_hw_queue+0x83/0xa0 blk_mq_run_hw_queue+0x6c/0xd0 blk_mq_sched_insert_request+0x96/0x140 __blk_mq_try_issue_directly+0x3d/0x190 blk_mq_try_issue_directly+0x30/0x70 blk_mq_make_request+0x1a4/0x6a0 generic_make_request+0xfd/0x2f0 ? submit_bio+0x5c/0x110 submit_bio+0x5c/0x110 ? __blkdev_issue_discard+0x152/0x200 submit_bio_wait+0x43/0x60 ext4_process_freed_data+0x1cd/0x440 ? account_page_dirtied+0xe2/0x1a0 ext4_journal_commit_callback+0x4a/0xc0 jbd2_journal_commit_transaction+0x17e2/0x19e0 ? kjournald2+0xb0/0x250 kjournald2+0xb0/0x250 ? wait_woken+0x80/0x80 ? commit_timeout+0x10/0x10 kthread+0x111/0x130 ? kthread_create_worker_on_cpu+0x50/0x50 ? do_group_exit+0x3a/0xa0 ret_from_fork+0x1f/0x30 Code: 73 89 c1 83 ce 10 c1 e1 10 09 ca 83 f8 04 0f 87 0f ff ff ff 8b 4d 20 48 8b 7d 00 c1 e9 09 48 01 8c c7 00 08 00 00 e9 f8 fe ff ff <0f> ff 4c 89 c7 41 bc 0a 00 00 00 e8 0d 78 d6 ff e9 a1 fc ff ff ---[ end trace 50d361cc444506c8 ]--- print_req_error: I/O error, dev nvme0n1, sector 847167488 Decoding the assembly, the request claims to have 0xffff segments, while nvme counts two. This turns out to be because we don't check for a data carrying request on the mq scheduler path, and since blk_phys_contig_segment() returns true for a non-data request, we decrement the initial segment count of 0 and end up with 0xffff in the unsigned short. There are a few issues here: 1) We should initialize the segment count for a discard to 1. 2) The discard merging is currently using the data limits for segments and sectors. Fix this up by having attempt_merge() correctly identify the request, and by initializing the segment count correctly for discards. This can only be triggered with mq-deadline on discard capable devices right now, which isn't a common configuration. Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * blk-mq: introduce BLK_STS_DEV_RESOURCEMing Lei2018-01-302-4/+17
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This status is returned from driver to block layer if device related resource is unavailable, but driver can guarantee that IO dispatch will be triggered in future when the resource is available. Convert some drivers to return BLK_STS_DEV_RESOURCE. Also, if driver returns BLK_STS_RESOURCE and SCHED_RESTART is set, rerun queue after a delay (BLK_MQ_DELAY_QUEUE) to avoid IO stalls. BLK_MQ_DELAY_QUEUE is 3 ms because both scsi-mq and nvmefc are using that magic value. If a driver can make sure there is in-flight IO, it is safe to return BLK_STS_DEV_RESOURCE because: 1) If all in-flight IOs complete before examining SCHED_RESTART in blk_mq_dispatch_rq_list(), SCHED_RESTART must be cleared, so queue is run immediately in this case by blk_mq_dispatch_rq_list(); 2) if there is any in-flight IO after/when examining SCHED_RESTART in blk_mq_dispatch_rq_list(): - if SCHED_RESTART isn't set, queue is run immediately as handled in 1) - otherwise, this request will be dispatched after any in-flight IO is completed via blk_mq_sched_restart() 3) if SCHED_RESTART is set concurently in context because of BLK_STS_RESOURCE, blk_mq_delay_run_hw_queue() will cover the above two cases and make sure IO hang can be avoided. One invariant is that queue will be rerun if SCHED_RESTART is set. Suggested-by: Jens Axboe <axboe@kernel.dk> Tested-by: Laurence Oberman <loberman@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
* | Merge branch 'misc.poll' of ↵Linus Torvalds2018-01-301-2/+2
|\ \ | |/ |/| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs Pull poll annotations from Al Viro: "This introduces a __bitwise type for POLL### bitmap, and propagates the annotations through the tree. Most of that stuff is as simple as 'make ->poll() instances return __poll_t and do the same to local variables used to hold the future return value'. Some of the obvious brainos found in process are fixed (e.g. POLLIN misspelled as POLL_IN). At that point the amount of sparse warnings is low and most of them are for genuine bugs - e.g. ->poll() instance deciding to return -EINVAL instead of a bitmap. I hadn't touched those in this series - it's large enough as it is. Another problem it has caught was eventpoll() ABI mess; select.c and eventpoll.c assumed that corresponding POLL### and EPOLL### were equal. That's true for some, but not all of them - EPOLL### are arch-independent, but POLL### are not. The last commit in this series separates userland POLL### values from the (now arch-independent) kernel-side ones, converting between them in the few places where they are copied to/from userland. AFAICS, this is the least disruptive fix preserving poll(2) ABI and making epoll() work on all architectures. As it is, it's simply broken on sparc - try to give it EPOLLWRNORM and it will trigger only on what would've triggered EPOLLWRBAND on other architectures. EPOLLWRBAND and EPOLLRDHUP, OTOH, are never triggered at all on sparc. With this patch they should work consistently on all architectures" * 'misc.poll' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (37 commits) make kernel-side POLL... arch-independent eventpoll: no need to mask the result of epi_item_poll() again eventpoll: constify struct epoll_event pointers debugging printk in sg_poll() uses %x to print POLL... bitmap annotate poll(2) guts 9p: untangle ->poll() mess ->si_band gets POLL... bitmap stored into a user-visible long field ring_buffer_poll_wait() return value used as return value of ->poll() the rest of drivers/*: annotate ->poll() instances media: annotate ->poll() instances fs: annotate ->poll() instances ipc, kernel, mm: annotate ->poll() instances net: annotate ->poll() instances apparmor: annotate ->poll() instances tomoyo: annotate ->poll() instances sound: annotate ->poll() instances acpi: annotate ->poll() instances crypto: annotate ->poll() instances block: annotate ->poll() instances x86: annotate ->poll() instances ...
| * block: annotate ->poll() instancesAl Viro2017-11-271-2/+2
| | | | | | | | Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
* | Merge branch 'for-4.16/block' of git://git.kernel.dk/linux-blockLinus Torvalds2018-01-2932-684/+1506
|\ \ | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Pull block updates from Jens Axboe: "This is the main pull request for block IO related changes for the 4.16 kernel. Nothing major in this pull request, but a good amount of improvements and fixes all over the map. This contains: - BFQ improvements, fixes, and cleanups from Angelo, Chiara, and Paolo. - Support for SMR zones for deadline and mq-deadline from Damien and Christoph. - Set of fixes for bcache by way of Michael Lyle, including fixes from himself, Kent, Rui, Tang, and Coly. - Series from Matias for lightnvm with fixes from Hans Holmberg, Javier, and Matias. Mostly centered around pblk, and the removing rrpc 1.2 in preparation for supporting 2.0. - A couple of NVMe pull requests from Christoph. Nothing major in here, just fixes and cleanups, and support for command tracing from Johannes. - Support for blk-throttle for tracking reads and writes separately. From Joseph Qi. A few cleanups/fixes also for blk-throttle from Weiping. - Series from Mike Snitzer that enables dm to register its queue more logically, something that's alwways been problematic on dm since it's a stacked device. - Series from Ming cleaning up some of the bio accessor use, in preparation for supporting multipage bvecs. - Various fixes from Ming closing up holes around queue mapping and quiescing. - BSD partition fix from Richard Narron, fixing a problem where we can't mount newer (10/11) FreeBSD partitions. - Series from Tejun reworking blk-mq timeout handling. The previous scheme relied on atomic bits, but it had races where we would think a request had timed out if it to reused at the wrong time. - null_blk now supports faking timeouts, to enable us to better exercise and test that functionality separately. From me. - Kill the separate atomic poll bit in the request struct. After this, we don't use the atomic bits on blk-mq anymore at all. From me. - sgl_alloc/free helpers from Bart. - Heavily contended tag case scalability improvement from me. - Various little fixes and cleanups from Arnd, Bart, Corentin, Douglas, Eryu, Goldwyn, and myself" * 'for-4.16/block' of git://git.kernel.dk/linux-block: (186 commits) block: remove smart1,2.h nvme: add tracepoint for nvme_complete_rq nvme: add tracepoint for nvme_setup_cmd nvme-pci: introduce RECONNECTING state to mark initializing procedure nvme-rdma: remove redundant boolean for inline_data nvme: don't free uuid pointer before printing it nvme-pci: Suspend queues after deleting them bsg: use pr_debug instead of hand crafted macros blk-mq-debugfs: don't allow write on attributes with seq_operations set nvme-pci: Fix queue double allocations block: Set BIO_TRACE_COMPLETION on new bio during split blk-throttle: use queue_is_rq_based block: Remove kblockd_schedule_delayed_work{,_on}() blk-mq: Avoid that blk_mq_delay_run_hw_queue() introduces unintended delays blk-mq: Rename blk_mq_request_direct_issue() into blk_mq_request_issue_directly() lib/scatterlist: Fix chaining support in sgl_alloc_order() blk-throttle: track read and write request individually block: add bdev_read_only() checks to common helpers block: fail op_is_write() requests to read-only partitions blk-throttle: export io_serviced_recursive, io_service_bytes_recursive ...
| * | bsg: use pr_debug instead of hand crafted macrosJohannes Thumshirn2018-01-241-23/+17
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Use pr_debug instead of hand crafted macros. This way it is not needed to re-compile the kernel to enable bsg debug outputs and it's possible to selectively enable specific prints. Cc: Joe Perches <joe@perches.com> Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq-debugfs: don't allow write on attributes with seq_operations setEryu Guan2018-01-241-1/+5
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Attributes that only implement .seq_ops are read-only, any write to them should be rejected. But currently kernel would crash when writing to such debugfs entries, e.g. chmod +w /sys/kernel/debug/block/<dev>/requeue_list echo 0 > /sys/kernel/debug/block/<dev>/requeue_list chmod -w /sys/kernel/debug/block/<dev>/requeue_list Fix it by returning -EPERM in blk_mq_debugfs_write() when writing to such attributes. Cc: Ming Lei <ming.lei@redhat.com> Signed-off-by: Eryu Guan <eguan@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: Set BIO_TRACE_COMPLETION on new bio during splitGoldwyn Rodrigues2018-01-231-1/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | We inadvertently set it again on the source bio, but we need to set it on the new split bio instead. Fixes: fbbaf700e7b1 ("block: trace completion of all bios.") Signed-off-by: Goldwyn Rodrigues <rgoldwyn@suse.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-throttle: use queue_is_rq_basedweiping zhang2018-01-191-1/+1
| | | | | | | | | | | | | | | | | | | | | use queue_is_rq_based instead of open code. Signed-off-by: weiping zhang <zhangweiping@didichuxing.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: Remove kblockd_schedule_delayed_work{,_on}()Bart Van Assche2018-01-191-14/+0
| | | | | | | | | | | | | | | | | | | | | | | | | | | The previous patch removed all users of these two functions. Hence also remove the functions themselves. Reviewed-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: Avoid that blk_mq_delay_run_hw_queue() introduces unintended delaysBart Van Assche2018-01-191-4/+3
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Make sure that calling blk_mq_run_hw_queue() or blk_mq_kick_requeue_list() triggers a queue run without delay even if blk_mq_delay_run_hw_queue() has been called recently and if its delay has not yet expired. Reviewed-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: Rename blk_mq_request_direct_issue() into ↵Bart Van Assche2018-01-193-4/+4
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | blk_mq_request_issue_directly() Most blk-mq functions have a name that follows the pattern blk_mq_${action}. However, the function name blk_mq_request_direct_issue is an exception. Hence rename this function. This patch does not change any functionality. Reviewed-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-throttle: track read and write request individuallyJoseph Qi2018-01-181-55/+79
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | In mixed read/write workload on SSD, write latency is much lower than read. But now we only track and record read latency and then use it as threshold base for both read and write io latency accounting. As a result, write io latency will always be considered as good and bad_bio_cnt is much smaller than 20% of bio_cnt. That is to mean, the tg to be checked will be treated as idle most of the time and still let others dispatch more ios, even it is truly running under low limit and wants its low limit to be guaranteed, which is not we expected in fact. So track read and write request individually, which can bring more precise latency control for low limit idle detection. Signed-off-by: Joseph Qi <qijiang.qj@alibaba-inc.com> Reviewed-by: Shaohua Li <shli@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: add bdev_read_only() checks to common helpersIlya Dryomov2018-01-181-0/+12
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Similar to blkdev_write_iter(), return -EPERM if the partition is read-only. This covers ioctl(), fallocate() and most in-kernel users but isn't meant to be exhaustive -- everything else will be caught in generic_make_request_checks(), fail with -EIO and can be fixed later. Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Ilya Dryomov <idryomov@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: fail op_is_write() requests to read-only partitionsIlya Dryomov2018-01-181-18/+38
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Regular block device writes go through blkdev_write_iter(), which does bdev_read_only(), while zeroout/discard/etc requests are never checked, both userspace- and kernel-triggered. Add a generic catch-all check to generic_make_request_checks() to actually enforce ioctl(BLKROSET) and set_disk_ro(), which is used by quite a few drivers for things like snapshots, read-only backing files/images, etc. Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Ilya Dryomov <idryomov@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-throttle: export io_serviced_recursive, io_service_bytes_recursiveweiping zhang2018-01-181-0/+10
| | | | | | | | | | | | | | | | | | | | | | | | export these two interface for cgroup-v1. Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: weiping zhang <zhangweiping@didichuxing.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: Protect less code with sysfs_lock in blk_{un,}register_queue()Bart Van Assche2018-01-181-9/+28
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The __blk_mq_register_dev(), blk_mq_unregister_dev(), elv_register_queue() and elv_unregister_queue() calls need to be protected with sysfs_lock but other code in these functions not. Hence protect only this code with sysfs_lock. This patch fixes a locking inversion issue in blk_unregister_queue() and also in an error path of blk_register_queue(): it is not allowed to hold sysfs_lock around the kobject_del(&q->kobj) call. Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: Document scheduler modification locking requirementsBart Van Assche2018-01-181-0/+8
| | | | | | | | | | | | | | | | | | | | | | | | This patch does not change any functionality. Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: Unexport elv_register_queue() and elv_unregister_queue()Bart Van Assche2018-01-182-2/+3
| | | | | | | | | | | | | | | | | | | | | | | | | | | These two functions are only called from inside the block layer so unexport them. Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block, bfq: limit sectors served with interactive weight raisingPaolo Valente2018-01-183-9/+80
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | To maximise responsiveness, BFQ raises the weight, and performs device idling, for bfq_queues associated with processes deemed as interactive. In particular, weight raising has a maximum duration, equal to the time needed to start a large application. If a weight-raised process goes on doing I/O beyond this maximum duration, it loses weight-raising. This mechanism is evidently vulnerable to the following false positives: I/O-bound applications that will go on doing I/O for much longer than the duration of weight-raising. These applications have basically no benefit from being weight-raised at the beginning of their I/O. On the opposite end, while being weight-raised, these applications a) unjustly steal throughput to applications that may truly need low latency; b) make BFQ uselessly perform device idling; device idling results in loss of device throughput with most flash-based storage, and may increase latencies when used purposelessly. This commit adds a countermeasure to reduce both the above problems. To introduce this countermeasure, we provide the following extra piece of information (full details in the comments added by this commit). During the start-up of the large application used as a reference to set the duration of weight-raising, involved processes transfer at most ~110K sectors each. Accordingly, a process initially deemed as interactive has no right to be weight-raised any longer, once transferred 110K sectors or more. Basing on this consideration, this commit early-ends weight-raising for a bfq_queue if the latter happens to have received an amount of service at least equal to 110K sectors (actually, a little bit more, to keep a safety margin). I/O-bound applications that reach a high throughput, such as file copy, get to this threshold much before the allowed weight-raising period finishes. Thus this early ending of weight-raising reduces the amount of time during which these applications cause the problems described above. Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block, bfq: limit tags for writes and async I/OPaolo Valente2018-01-182-0/+89
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Asynchronous I/O can easily starve synchronous I/O (both sync reads and sync writes), by consuming all request tags. Similarly, storms of synchronous writes, such as those that sync(2) may trigger, can starve synchronous reads. In their turn, these two problems may also cause BFQ to loose control on latency for interactive and soft real-time applications. For example, on a PLEXTOR PX-256M5S SSD, LibreOffice Writer takes 0.6 seconds to start if the device is idle, but it takes more than 45 seconds (!) if there are sequential writes in the background. This commit addresses this issue by limiting the maximum percentage of tags that asynchronous I/O requests and synchronous write requests can consume. In particular, this commit grants a higher threshold to synchronous writes, to prevent the latter from being starved by asynchronous I/O. According to the above test, LibreOffice Writer now starts in about 1.2 seconds on average, regardless of the background workload, and apart from some rare outlier. To check this improvement, run, e.g., sudo ./comm_startup_lat.sh bfq 5 5 seq 10 "lowriter --terminate_after_init" for the comm_startup_lat benchmark in the S suite [1]. [1] https://github.com/Algodev-github/S Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: don't dispatch request in blk_mq_request_direct_issue if queue is busyMing Lei2018-01-171-12/+10
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | If we run into blk_mq_request_direct_issue(), when queue is busy, we don't want to dispatch this request into hctx->dispatch_list, and what we need to do is to return the queue busy info to caller, so that caller can deal with it well. Fixes: 396eaf21ee ("blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback") Reported-by: Laurence Oberman <loberman@redhat.com> Reviewed-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: Fix __bio_integrity_endio() documentationBart Van Assche2018-01-171-1/+0
| | | | | | | | | | | | | | | | | | | | | Fixes: 4246a0b63bd8 ("block: add a bi_error field to struct bio") Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq-sched: remove unused 'can_block' arg from blk_mq_sched_insert_requestMike Snitzer2018-01-174-12/+10
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | After commit: 923218f6166a ("blk-mq: don't allocate driver tag upfront for flush rq") we no longer use the 'can_block' argument in blk_mq_sched_insert_request(). Kill it. Signed-off-by: Mike Snitzer <snitzer@redhat.com> Added actual commit message as to why it's being removed. Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedbackMing Lei2018-01-173-10/+33
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | blk_insert_cloned_request() is called in the fast path of a dm-rq driver (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses blk_mq_request_bypass_insert() to directly append the request to the blk-mq hctx->dispatch_list of the underlying queue. 1) This way isn't efficient enough because the hctx spinlock is always used. 2) With blk_insert_cloned_request(), we completely bypass underlying queue's elevator and depend on the upper-level dm-rq driver's elevator to schedule IO. But dm-rq currently can't get the underlying queue's dispatch feedback at all. Without knowing whether a request was issued or not (e.g. due to underlying queue being busy) the dm-rq elevator will not be able to provide effective IO merging (as a side-effect of dm-rq currently blindly destaging a request from its elevator only to requeue it after a delay, which kills any opportunity for merging). This obviously causes very bad sequential IO performance. Fix this by updating blk_insert_cloned_request() to use blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a request to be issued directly to the underlying queue and returns the dispatch feedback (blk_status_t). If blk_mq_request_direct_issue() returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE to _not_ destage the request. Whereby preserving the opportunity to merge IO. With this, request-based DM's blk-mq sequential IO performance is vastly improved (as much as 3X in mpath/virtio-scsi testing). Signed-off-by: Ming Lei <ming.lei@redhat.com> [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but they were refactored to make them less fragile and easier to read/review] Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: factor out a few helpers from __blk_mq_try_issue_directlyMike Snitzer2018-01-171-27/+52
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | No functional change. Just makes code flow more logically. In following commit, __blk_mq_try_issue_directly() will be used to return the dispatch result (blk_status_t) to DM. DM needs this information to improve IO merging. Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: turn WARN_ON in __blk_mq_run_hw_queue into printkMing Lei2018-01-171-2/+20
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We know this WARN_ON is harmless and in reality it may be trigged, so convert it to printk() and dump_stack() to avoid to confusing people. Also add comment about two releated races here. Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Stefan Haberland <sth@linux.vnet.ibm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: make sure hctx->next_cpu is set correctlyMing Lei2018-01-171-2/+28
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | When hctx->next_cpu is set from possible online CPUs, there is one race in which hctx->next_cpu may be set as >= nr_cpu_ids, and finally break workqueue. The race can be triggered in the following two sitations: 1) when one CPU is becoming DEAD, blk_mq_hctx_notify_dead() is called to dispatch requests from the DEAD cpu context, but at that time, this DEAD CPU has been cleared from 'cpu_online_mask', so all CPUs in hctx->cpumask may become offline, and cause hctx->next_cpu set a bad value. 2) blk_mq_delay_run_hw_queue() is called from CPU B, and found the queue should be run on the other CPU A, then CPU A may become offline at the same time and all CPUs in hctx->cpumask become offline. This patch deals with this issue by re-selecting next CPU, and making sure it is set correctly. Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Stefan Haberland <sth@linux.vnet.ibm.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Thomas Gleixner <tglx@linutronix.de> Reported-by: "jianchao.wang" <jianchao.w.wang@oracle.com> Tested-by: "jianchao.wang" <jianchao.w.wang@oracle.com> Fixes: 20e4d81393 ("blk-mq: simplify queue mapping & schedule with each possisble CPU") Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk_rq_map_user_iov: fix error overrideDouglas Gilbert2018-01-151-2/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | During stress tests by syzkaller on the sg driver the block layer infrequently returns EINVAL. Closer inspection shows the block layer was trying to return ENOMEM (which is much more understandable) but for some reason overroad that useful error. Patch below does not show this (unchanged) line: ret =__blk_rq_map_user_iov(rq, map_data, &i, gfp_mask, copy); That 'ret' was being overridden when that function failed. Signed-off-by: Douglas Gilbert <dgilbert@interlog.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: allow gendisk's request_queue registration to be deferredMike Snitzer2018-01-152-3/+22
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Since I can remember DM has forced the block layer to allow the allocation and initialization of the request_queue to be distinct operations. Reason for this is block/genhd.c:add_disk() has requires that the request_queue (and associated bdi) be tied to the gendisk before add_disk() is called -- because add_disk() also deals with exposing the request_queue via blk_register_queue(). DM's dynamic creation of arbitrary device types (and associated request_queue types) requires the DM device's gendisk be available so that DM table loads can establish a master/slave relationship with subordinate devices that are referenced by loaded DM tables -- using bd_link_disk_holder(). But until these DM tables, and their associated subordinate devices, are known DM cannot know what type of request_queue it needs -- nor what its queue_limits should be. This chicken and egg scenario has created all manner of problems for DM and, at times, the block layer. Summary of changes: - Add device_add_disk_no_queue_reg() and add_disk_no_queue_reg() variant that drivers may use to add a disk without also calling blk_register_queue(). Driver must call blk_register_queue() once its request_queue is fully initialized. - Return early from blk_unregister_queue() if QUEUE_FLAG_REGISTERED is not set. It won't be set if driver used add_disk_no_queue_reg() but driver encounters an error and must del_gendisk() before calling blk_register_queue(). - Export blk_register_queue(). These changes allow DM to use add_disk_no_queue_reg() to anchor its gendisk as the "master" for master/slave relationships DM must establish with subordinate devices referenced in DM tables that get loaded. Once all "slave" devices for a DM device are known its request_queue can be properly initialized and then advertised via sysfs -- important improvement being that no request_queue resource initialization performed by blk_register_queue() is missed for DM devices anymore. Signed-off-by: Mike Snitzer <snitzer@redhat.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: properly protect the 'queue' kobj in blk_unregister_queueMike Snitzer2018-01-152-11/+11
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The original commit e9a823fb34a8b (block: fix warning when I/O elevator is changed as request_queue is being removed) is pretty conflated. "conflated" because the resource being protected by q->sysfs_lock isn't the queue_flags (it is the 'queue' kobj). q->sysfs_lock serializes __elevator_change() (via elv_iosched_store) from racing with blk_unregister_queue(): 1) By holding q->sysfs_lock first, __elevator_change() can complete before a racing blk_unregister_queue(). 2) Conversely, __elevator_change() is testing for QUEUE_FLAG_REGISTERED in case elv_iosched_store() loses the race with blk_unregister_queue(), it needs a way to know the 'queue' kobj isn't there. Expand the scope of blk_unregister_queue()'s q->sysfs_lock use so it is held until after the 'queue' kobj is removed. To do so blk_mq_unregister_dev() must not also take q->sysfs_lock. So rename __blk_mq_unregister_dev() to blk_mq_unregister_dev(). Also, blk_unregister_queue() should use q->queue_lock to protect against any concurrent writes to q->queue_flags -- even though chances are the queue is being cleaned up so no concurrent writes are likely. Fixes: e9a823fb34a8b ("block: fix warning when I/O elevator is changed as request_queue is being removed") Signed-off-by: Mike Snitzer <snitzer@redhat.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: only bdi_unregister() in del_gendisk() if !GENHD_FL_HIDDENMike Snitzer2018-01-151-1/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | device_add_disk() will only call bdi_register_owner() if !GENHD_FL_HIDDEN, so it follows that del_gendisk() should only call bdi_unregister() if !GENHD_FL_HIDDEN. Found with code inspection. bdi_unregister() won't do any harm if bdi_register_owner() wasn't used but best to avoid the unnecessary call to bdi_unregister(). Fixes: 8ddcd65325 ("block: introduce GENHD_FL_HIDDEN") Signed-off-by: Mike Snitzer <snitzer@redhat.com> Reviewed-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Hannes Reinecke <hare@suse.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: fix bad clear of RQF_MQ_INFLIGHT in blk_mq_ct_ctx_init()Jens Axboe2018-01-141-2/+3
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | A previous commit moved the clearing of rq->rq_flags later, but we may have already set RQF_MQ_INFLIGHT when that happens. Ensure that we correctly initialize rq->rq_flags to the right value. This is based on an original fix by Ming, just rewritten to not require a conditional. Fixes: 7c3fb70f0341 ("block: rearrange a few request fields for better cache layout") Reviewed-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: add missing RQF_STARTED to debugfsJens Axboe2018-01-121-0/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Looking at debug output, we see: ./000000009ddfa913/requeue_list:000000009646711c {.op=READ, .state=idle, gen=0x1 18, abort_gen=0x0, .cmd_flags=, .rq_flags=SORTED|1|SOFTBARRIER|IO_STAT, complete =0, .tag=-1, .internal_tag=217} Note the '1' between SORTED and SOFTBARRIER - that's because no name as defined for RQF_STARTED. Fixed that. Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: simplify queue mapping & schedule with each possisble CPUChristoph Hellwig2018-01-121-11/+8
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The previous patch assigns interrupt vectors to all possible CPUs, so now hctx can be mapped to possible CPUs, this patch applies this fact to simplify queue mapping & schedule so that we don't need to handle CPU hotplug for dealing with physical CPU plug & unplug. With this simplication, we can work well on physical CPU plug & unplug, which is a normal use case for VM at least. Make sure we allocate blk_mq_ctx structures for all possible CPUs, and set hctx->numa_node for possible CPUs which are mapped to this hctx. And only choose the online CPUs for schedule. Reported-by: Christian Borntraeger <borntraeger@de.ibm.com> Tested-by: Christian Borntraeger <borntraeger@de.ibm.com> Tested-by: Stefan Haberland <sth@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Christoph Hellwig <hch@lst.de> Fixes: 4b855ad37194 ("blk-mq: Create hctx for each present CPU") (merged the three into one because any single one may not work, and fix selecting online CPUs for scheduler) Signed-off-by: Ming Lei <ming.lei@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: Reduce the number of if-statements in blk_mq_mark_tag_wait()Bart Van Assche2018-01-111-34/+35
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This patch does not change any functionality but makes the blk_mq_mark_tag_wait() code slightly easier to read. Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Omar Sandoval <osandov@fb.com> Cc: Hannes Reinecke <hare@suse.de> Cc: Johannes Thumshirn <jthumshirn@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: Add locking annotations to hctx_lock() and hctx_unlock()Bart Van Assche2018-01-101-0/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This patch avoids that sparse reports the following: block/blk-mq.c:637:33: warning: context imbalance in 'hctx_unlock' - unexpected unlock block/blk-mq.c:642:9: warning: context imbalance in 'hctx_lock' - wrong count at exit Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Cc: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: silently forbid sending any ioctl to a partitionPaolo Bonzini2018-01-101-29/+0
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | After the first few months, the message has not led to many bug reports. It's been almost five years now, and in practice the main source of it seems to be MTIOCGET that someone is using to detect tape devices. While we could whitelist it just like CDROM_GET_CAPABILITY, this patch just removes the message altogether. The patch also removes the "safe but not very useful" ioctl whitelist, as suggested by Christoph. I doubt anything is using most of those ioctls _in general_, let alone on a partition. Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: rearrange a few request fields for better cache layoutJens Axboe2018-01-101-9/+9
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Move completion related items (like the call single data) near the end of the struct, instead of mixing them in with the initial queueing related fields. Move queuelist below the bio structures. Then we have all queueing related bits in the first cache line. This yields a 1.5-2% increase in IOPS for a null_blk test, both for sync and for high thread count access. Sync test goes form 975K to 992K, 32-thread case from 20.8M to 21.2M IOPS. Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com> Reviewed-by: Omar Sandoval <osandov@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: convert REQ_ATOM_COMPLETE to stealing rq->__deadline bitJens Axboe2018-01-104-20/+12
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We only have one atomic flag left. Instead of using an entire unsigned long for that, steal the bottom bit of the deadline field that we already reserved. Remove ->atomic_flags, since it's now unused. Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com> Reviewed-by: Omar Sandoval <osandov@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: add accessors for setting/querying request deadlineJens Axboe2018-01-103-7/+24
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We reduce the resolution of request expiry, but since we're already using jiffies for this where resolution depends on the kernel configuration and since the timeout resolution is coarse anyway, that should be fine. Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com> Reviewed-by: Omar Sandoval <osandov@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | block: remove REQ_ATOM_POLL_SLEPTJens Axboe2018-01-103-6/+3
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | We don't need this to be an atomic flag, it can be a regular flag. We either end up on the same CPU for the polling, in which case the state is sane, or we did the sleep which would imply the needed barrier to ensure we see the right state. Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com> Reviewed-by: Omar Sandoval <osandov@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: add a few missing debugfs RQF_ flagsJens Axboe2018-01-101-0/+2
| | | | | | | | | | | | | | | | | | | | | We are missing ZONE_WRITE_LOCKED and MQ_TIMEOUT_EXPIRED, add them so the debugfs bits can decode them. Signed-off-by: Jens Axboe <axboe@kernel.dk>
| * | blk-mq: Explain when 'active_queues' is decrementedBart Van Assche2018-01-101-0/+6
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | It is nontrivial to derive from the blk-mq source code when blk_mq_tags.active_queues is decremented. Hence add a comment that explains this. Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Hannes Reinecke <hare@suse.de> Cc: Johannes Thumshirn <jthumshirn@suse.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
OpenPOWER on IntegriCloud