summaryrefslogtreecommitdiff
path: root/block/bfq-iosched.c
AgeCommit message (Collapse)AuthorFilesLines
2019-07-26Merge tag 'for-linus-20190726' of git://git.kernel.dk/linux-blockLinus Torvalds1-24/+43
Pull block fixes from Jens Axboe: - Several io_uring fixes/improvements: - Blocking fix for O_DIRECT (me) - Latter page slowness for registered buffers (me) - Fix poll hang under certain conditions (me) - Defer sequence check fix for wrapped rings (Zhengyuan) - Mismatch in async inc/dec accounting (Zhengyuan) - Memory ordering issue that could cause stall (Zhengyuan) - Track sequential defer in bytes, not pages (Zhengyuan) - NVMe pull request from Christoph - Set of hang fixes for wbt (Josef) - Redundant error message kill for libahci (Ding) - Remove unused blk_mq_sched_started_request() and related ops (Marcos) - drbd dynamic alloc shash descriptor to reduce stack use (Arnd) - blkcg ->pd_stat() non-debug print (Tejun) - bcache memory leak fix (Wei) - Comment fix (Akinobu) - BFQ perf regression fix (Paolo) * tag 'for-linus-20190726' of git://git.kernel.dk/linux-block: (24 commits) io_uring: ensure ->list is initialized for poll commands Revert "nvme-pci: don't create a read hctx mapping without read queues" nvme: fix multipath crash when ANA is deactivated nvme: fix memory leak caused by incorrect subsystem free nvme: ignore subnqn for ADATA SX6000LNP drbd: dynamically allocate shash descriptor block: blk-mq: Remove blk_mq_sched_started_request and started_request bcache: fix possible memory leak in bch_cached_dev_run() io_uring: track io length in async_list based on bytes io_uring: don't use iov_iter_advance() for fixed buffers block: properly handle IOCB_NOWAIT for async O_DIRECT IO blk-mq: allow REQ_NOWAIT to return an error inline io_uring: add a memory barrier before atomic_read rq-qos: use a mb for got_token rq-qos: set ourself TASK_UNINTERRUPTIBLE after we schedule rq-qos: don't reset has_sleepers on spurious wakeups rq-qos: fix missed wake-ups in rq_qos_throttle wait: add wq_has_single_sleeper helper block, bfq: check also in-flight I/O in dispatch plugging block: fix sysfs module parameters directory path in comment ...
2019-07-18block, bfq: check also in-flight I/O in dispatch pluggingPaolo Valente1-24/+43
Consider a sync bfq_queue Q that remains empty while in service, and suppose that, when this happens, there is a fair amount of already in-flight I/O not belonging to Q. In such a situation, I/O dispatching may need to be plugged (until new I/O arrives for Q), for the following reason. The drive may decide to serve in-flight non-Q's I/O requests before Q's ones, thereby delaying the arrival of new I/O requests for Q (recall that Q is sync). If I/O-dispatching is not plugged, then, while Q remains empty, a basically uncontrolled amount of I/O from other queues may be dispatched too, possibly causing the service of Q's I/O to be delayed even longer in the drive. This problem gets more and more serious as the speed and the queue depth of the drive grow, because, as these two quantities grow, the probability to find no queue busy but many requests in flight grows too. If Q has the same weight and priority as the other queues, then the above delay is unlikely to cause any issue, because all queues tend to undergo the same treatment. So, since not plugging I/O dispatching is convenient for throughput, it is better not to plug. Things change in case Q has a higher weight or priority than some other queue, because Q's service guarantees may simply be violated. For this reason, commit 1de0c4cd9ea6 ("block, bfq: reduce idling only in symmetric scenarios") does plug I/O in such an asymmetric scenario. Plugging minimizes the delay induced by already in-flight I/O, and enables Q to recover the bandwidth it may lose because of this delay. Yet the above commit does not cover the case of weight-raised queues, for efficiency concerns. For weight-raised queues, I/O-dispatch plugging is activated simply if not all bfq_queues are weight-raised. But this check does not handle the case of in-flight requests, because a bfq_queue may become non busy *before* all its in-flight requests are completed. This commit performs I/O-dispatch plugging for weight-raised queues if there are some in-flight requests. As a practical example of the resulting recover of control, under write load on a Samsung SSD 970 PRO, gnome-terminal starts in 1.5 seconds after this fix, against 15 seconds before the fix (as a reference, gnome-terminal takes about 35 seconds to start with any of the other I/O schedulers). Fixes: 1de0c4cd9ea6 ("block, bfq: reduce idling only in symmetric scenarios") Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-07-15docs: block: convert to ReSTMauro Carvalho Chehab1-1/+1
Rename the block documentation files to ReST, add an index for them and adjust in order to produce a nice html output via the Sphinx build system. At its new index.rst, let's add a :orphan: while this is not linked to the main index.rst file, in order to avoid build warnings. Signed-off-by: Mauro Carvalho Chehab <mchehab+samsung@kernel.org>
2019-07-09Merge tag 'for-5.3/block-20190708' of git://git.kernel.dk/linux-blockLinus Torvalds1-296/+671
Pull block updates from Jens Axboe: "This is the main block updates for 5.3. Nothing earth shattering or major in here, just fixes, additions, and improvements all over the map. This contains: - Series of documentation fixes (Bart) - Optimization of the blk-mq ctx get/put (Bart) - null_blk removal race condition fix (Bob) - req/bio_op() cleanups (Chaitanya) - Series cleaning up the segment accounting, and request/bio mapping (Christoph) - Series cleaning up the page getting/putting for bios (Christoph) - block cgroup cleanups and moving it to where it is used (Christoph) - block cgroup fixes (Tejun) - Series of fixes and improvements to bcache, most notably a write deadlock fix (Coly) - blk-iolatency STS_AGAIN and accounting fixes (Dennis) - Series of improvements and fixes to BFQ (Douglas, Paolo) - debugfs_create() return value check removal for drbd (Greg) - Use struct_size(), where appropriate (Gustavo) - Two lighnvm fixes (Heiner, Geert) - MD fixes, including a read balance and corruption fix (Guoqing, Marcos, Xiao, Yufen) - block opal shadow mbr additions (Jonas, Revanth) - sbitmap compare-and-exhange improvemnts (Pavel) - Fix for potential bio->bi_size overflow (Ming) - NVMe pull requests: - improved PCIe suspent support (Keith Busch) - error injection support for the admin queue (Akinobu Mita) - Fibre Channel discovery improvements (James Smart) - tracing improvements including nvmetc tracing support (Minwoo Im) - misc fixes and cleanups (Anton Eidelman, Minwoo Im, Chaitanya Kulkarni)" - Various little fixes and improvements to drivers and core" * tag 'for-5.3/block-20190708' of git://git.kernel.dk/linux-block: (153 commits) blk-iolatency: fix STS_AGAIN handling block: nr_phys_segments needs to be zero for REQ_OP_WRITE_ZEROES blk-mq: simplify blk_mq_make_request() blk-mq: remove blk_mq_put_ctx() sbitmap: Replace cmpxchg with xchg block: fix .bi_size overflow block: sed-opal: check size of shadow mbr block: sed-opal: ioctl for writing to shadow mbr block: sed-opal: add ioctl for done-mark of shadow mbr block: never take page references for ITER_BVEC direct-io: use bio_release_pages in dio_bio_complete block_dev: use bio_release_pages in bio_unmap_user block_dev: use bio_release_pages in blkdev_bio_end_io iomap: use bio_release_pages in iomap_dio_bio_end_io block: use bio_release_pages in bio_map_user_iov block: use bio_release_pages in bio_unmap_user block: optionally mark pages dirty in bio_release_pages block: move the BIO_NO_PAGE_REF check into bio_release_pages block: skd_main.c: Remove call to memset after dma_alloc_coherent block: mtip32xx: Remove call to memset after dma_alloc_coherent ...
2019-06-28block, bfq: NULL out the bic when it's no longer validDouglas Anderson1-0/+1
In reboot tests on several devices we were seeing a "use after free" when slub_debug or KASAN was enabled. The kernel complained about: Unable to handle kernel paging request at virtual address 6b6b6c2b ...which is a classic sign of use after free under slub_debug. The stack crawl in kgdb looked like: 0 test_bit (addr=<optimized out>, nr=<optimized out>) 1 bfq_bfqq_busy (bfqq=<optimized out>) 2 bfq_select_queue (bfqd=<optimized out>) 3 __bfq_dispatch_request (hctx=<optimized out>) 4 bfq_dispatch_request (hctx=<optimized out>) 5 0xc056ef00 in blk_mq_do_dispatch_sched (hctx=0xed249440) 6 0xc056f728 in blk_mq_sched_dispatch_requests (hctx=0xed249440) 7 0xc0568d24 in __blk_mq_run_hw_queue (hctx=0xed249440) 8 0xc0568d94 in blk_mq_run_work_fn (work=<optimized out>) 9 0xc024c5c4 in process_one_work (worker=0xec6d4640, work=0xed249480) 10 0xc024cff4 in worker_thread (__worker=0xec6d4640) Digging in kgdb, it could be found that, though bfqq looked fine, bfqq->bic had been freed. Through further digging, I postulated that perhaps it is illegal to access a "bic" (AKA an "icq") after bfq_exit_icq() had been called because the "bic" can be freed at some point in time after this call is made. I confirmed that there certainly were cases where the exact crashing code path would access the "bic" after bfq_exit_icq() had been called. Sspecifically I set the "bfqq->bic" to (void *)0x7 and saw that the bic was 0x7 at the time of the crash. To understand a bit more about why this crash was fairly uncommon (I saw it only once in a few hundred reboots), you can see that much of the time bfq_exit_icq_fbqq() fully frees the bfqq and thus it can't access the ->bic anymore. The only case it doesn't is if bfq_put_queue() sees a reference still held. However, even in the case when bfqq isn't freed, the crash is still rare. Why? I tracked what happened to the "bic" after the exit routine. It doesn't get freed right away. Rather, put_io_context_active() eventually called put_io_context() which queued up freeing on a workqueue. The freeing then actually happened later than that through call_rcu(). Despite all these delays, some extra debugging showed that all the hoops could be jumped through in time and the memory could be freed causing the original crash. Phew! To make a long story short, assuming it truly is illegal to access an icq after the "exit_icq" callback is finished, this patch is needed. Cc: stable@vger.kernel.org Reviewed-by: Paolo Valente <paolo.valente@unimore.it> Signed-off-by: Douglas Anderson <dianders@chromium.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-27block, bfq: Init saved_wr_start_at_switch_to_srt in unlikely caseDouglas Anderson1-0/+1
Some debug code suggested by Paolo was tripping when I did reboot stress tests. Specifically in bfq_bfqq_resume_state() "bic->saved_wr_start_at_switch_to_srt" was later than the current value of "jiffies". A bit of debugging showed that "bic->saved_wr_start_at_switch_to_srt" was actually 0 and a bit more debugging showed that was because we had run through the "unlikely" case in the bfq_bfqq_save_state() function. Let's init "saved_wr_start_at_switch_to_srt" in the unlikely case to something sane. NOTE: this fixes no known real-world errors. Reviewed-by: Paolo Valente <paolo.valente@linaro.org> Reviewed-by: Guenter Roeck <groeck@chromium.org> Signed-off-by: Douglas Anderson <dianders@chromium.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: fix operator in BFQQ_TOTALLY_SEEKYPaolo Valente1-1/+1
By mistake, there is a '&' instead of a '==' in the definition of the macro BFQQ_TOTALLY_SEEKY. This commit replaces the wrong operator with the correct one. Fixes: 7074f076ff15 ("block, bfq: do not tag totally seeky queues as soft rt") Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: re-schedule empty queues if they deserve I/O pluggingPaolo Valente1-184/+203
Consider, on one side, a bfq_queue Q that remains empty while in service, and, on the other side, the pending I/O of bfq_queues that, according to their timestamps, have to be served after Q. If an uncontrolled amount of I/O from the latter bfq_queues were dispatched while Q is waiting for its new I/O to arrive, then Q's bandwidth guarantees would be violated. To prevent this, I/O dispatch is plugged until Q receives new I/O (except for a properly controlled amount of injected I/O). Unfortunately, preemption breaks I/O-dispatch plugging, for the following reason. Preemption is performed in two steps. First, Q is expired and re-scheduled. Second, the new bfq_queue to serve is chosen. The first step is needed by the second, as the second can be performed only after Q's timestamps have been properly updated (done in the expiration step), and Q has been re-queued for service. This dependency is a consequence of the way how BFQ's scheduling algorithm is currently implemented. But Q is not re-scheduled at all in the first step, because Q is empty. As a consequence, an uncontrolled amount of I/O may be dispatched until Q becomes non empty again. This breaks Q's service guarantees. This commit addresses this issue by re-scheduling Q even if it is empty. This in turn breaks the assumption that all scheduled queues are non empty. Then a few extra checks are now needed. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: preempt lower-weight or lower-priority queuesPaolo Valente1-20/+75
BFQ enqueues the I/O coming from each process into a separate bfq_queue, and serves bfq_queues one at a time. Each bfq_queue may be served for at most timeout_sync milliseconds (default: 125 ms). This service scheme is prone to the following inaccuracy. While a bfq_queue Q1 is in service, some empty bfq_queue Q2 may receive I/O, and, according to BFQ's scheduling policy, may become the right bfq_queue to serve, in place of the currently in-service bfq_queue. In this respect, postponing the service of Q2 to after the service of Q1 finishes may delay the completion of Q2's I/O, compared with an ideal service in which all non-empty bfq_queues are served in parallel, and every non-empty bfq_queue is served at a rate proportional to the bfq_queue's weight. This additional delay is equal at most to the time Q1 may unjustly remain in service before switching to Q2. If Q1 and Q2 have the same weight, then this time is most likely negligible compared with the completion time to be guaranteed to Q2's I/O. In addition, first, one of the reasons why BFQ may want to serve Q1 for a while is that this boosts throughput and, second, serving Q1 longer reduces BFQ's overhead. As a conclusion, it is usually better not to preempt Q1 if both Q1 and Q2 have the same weight. In contrast, as Q2's weight or priority becomes higher and higher compared with that of Q1, the above delay becomes larger and larger, compared with the I/O completion times that have to be guaranteed to Q2 according to Q2's weight. So reducing this delay may be more important than avoiding the costs of preempting Q1. Accordingly, this commit preempts Q1 if Q2 has a higher weight or a higher priority than Q1. Preemption causes Q1 to be re-scheduled, and triggers a new choice of the next bfq_queue to serve. If Q2 really is the next bfq_queue to serve, then Q2 will be set in service immediately. This change reduces the component of the I/O latency caused by the above delay by about 80%. For example, on an (old) PLEXTOR PX-256M5 SSD, the maximum latency reported by fio drops from 15.1 to 3.2 ms for a process doing sporadic random reads while another process is doing continuous sequential reads. Signed-off-by: Nicola Bottura <bottura.nicola95@gmail.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: detect wakers and unconditionally inject their I/OPaolo Valente1-33/+237
A bfq_queue Q may happen to be synchronized with another bfq_queue Q2, i.e., the I/O of Q2 may need to be completed for Q to receive new I/O. We call Q2 "waker queue". If I/O plugging is being performed for Q, and Q is not receiving any more I/O because of the above synchronization, then, thanks to BFQ's injection mechanism, the waker queue is likely to get served before the I/O-plugging timeout fires. Unfortunately, this fact may not be sufficient to guarantee a high throughput during the I/O plugging, because the inject limit for Q may be too low to guarantee a lot of injected I/O. In addition, the duration of the plugging, i.e., the time before Q finally receives new I/O, may not be minimized, because the waker queue may happen to be served only after other queues. To address these issues, this commit introduces the explicit detection of the waker queue, and the unconditional injection of a pending I/O request of the waker queue on each invocation of bfq_dispatch_request(). One may be concerned that this systematic injection of I/O from the waker queue delays the service of Q's I/O. Fortunately, it doesn't. On the contrary, next Q's I/O is brought forward dramatically, for it is not blocked for milliseconds. Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: bring forward seek&think time updatePaolo Valente1-10/+4
Until the base value for request service times gets finally computed for a bfq_queue, the inject limit for that queue does depend on the think-time state (short|long) of the queue. A timely update of the think time then guarantees a quicker activation or deactivation of the injection. Fortunately, the think time of a bfq_queue is updated in the same code path as the inject limit; but after the inject limit. This commits moves the update of the think time before the update of the inject limit. For coherence, it moves the update of the seek time too. Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: update base request service times when possiblePaolo Valente1-1/+12
I/O injection gets reduced if it increases the request service times of the victim queue beyond a certain threshold. The threshold, in its turn, is computed as a function of the base service time enjoyed by the queue when it undergoes no injection. As a consequence, for injection to work properly, the above base value has to be accurate. In this respect, such a value may vary over time. For example, it varies if the size or the spatial locality of the I/O requests in the queue change. It is then important to update this value whenever possible. This commit performs this update. Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: fix rq_in_driver check in bfq_update_inject_limitPaolo Valente1-1/+7
One of the cases where the parameters for injection may be updated is when there are no more in-flight I/O requests. The number of in-flight requests is stored in the field bfqd->rq_in_driver of the descriptor bfqd of the device. So, the controlled condition is bfqd->rq_in_driver == 0. Unfortunately, this is wrong because, the instruction that checks this condition is in the code path that handles the completion of a request, and, in particular, the instruction is executed before bfqd->rq_in_driver is decremented in such a code path. This commit fixes this issue by just replacing 0 with 1 in the comparison. Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-25block, bfq: reset inject limit when think-time state changesPaolo Valente1-68/+151
Until the base value of the request service times gets finally computed for a bfq_queue, the inject limit does depend on the think-time state (short|long). The limit must be 0 or 1 if the think time is deemed, respectively, as short or long. However, such a check and possible limit update is performed only periodically, once per second. So, to make the injection mechanism much more reactive, this commit performs the update also every time the think-time state changes. In addition, in the following special case, this commit lets the inject limit of a bfq_queue bfqq remain equal to 1 even if bfqq's think time is short: bfqq's I/O is synchronized with that of some other queue, i.e., bfqq may receive new I/O only after the I/O of the other queue is completed. Keeping the inject limit to 1 allows the blocking I/O to be served while bfqq is in service. And this is very convenient both for bfqq and for the total throughput, as explained in detail in the comments in bfq_update_has_short_ttime(). Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-20block: rename CONFIG_DEBUG_BLK_CGROUP to CONFIG_BFQ_CGROUP_DEBUGChristoph Hellwig1-4/+4
This option is entirely bfq specific, give it an appropinquate name. Also make it depend on CONFIG_BFQ_GROUP_IOSCHED in Kconfig, as all the functionality already does so anyway. Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-06-20block: remove the bi_phys_segments field in struct bioChristoph Hellwig1-2/+3
We only need the number of segments in the blk-mq submission path. Remove the field from struct bio, and return it from a variant of blk_queue_split instead of that it can passed as an argument to those functions that need the value. This also means we stop recounting segments except for cloning and partial segments. To keep the number of arguments in this how path down remove pointless struct request_queue arguments from any of the functions that had it and grew a nr_segs argument. Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-05-01block: switch all files cleared marked as GPLv2 or later to SPDX tagsChristoph Hellwig1-10/+1
All these files have some form of the usual GPLv2 or later boilerplate. Switch them to use SPDX tags instead. Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-22Merge tag 'v5.1-rc6' into for-5.2/blockJens Axboe1-9/+14
Pull in v5.1-rc6 to resolve two conflicts. One is in BFQ, in just a comment, and is trivial. The other one is a conflict due to a later fix in the bio multi-page work, and needs a bit more care. * tag 'v5.1-rc6': (770 commits) Linux 5.1-rc6 block: make sure that bvec length can't be overflow block: kill all_q_node in request_queue x86/cpu/intel: Lower the "ENERGY_PERF_BIAS: Set to normal" message's log priority coredump: fix race condition between mmget_not_zero()/get_task_mm() and core dumping mm/kmemleak.c: fix unused-function warning init: initialize jump labels before command line option parsing kernel/watchdog_hld.c: hard lockup message should end with a newline kcov: improve CONFIG_ARCH_HAS_KCOV help text mm: fix inactive list balancing between NUMA nodes and cgroups mm/hotplug: treat CMA pages as unmovable proc: fixup proc-pid-vm test proc: fix map_files test on F29 mm/vmstat.c: fix /proc/vmstat format for CONFIG_DEBUG_TLBFLUSH=y CONFIG_SMP=n mm/memory_hotplug: do not unlock after failing to take the device_hotplug_lock mm: swapoff: shmem_unuse() stop eviction without igrab() mm: swapoff: take notice of completion sooner mm: swapoff: remove too limiting SWAP_UNUSE_MAX_TRIES mm: swapoff: shmem_find_swap_entries() filter out other types slab: store tagged freelist for off-slab slabmgmt ... Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-14bfq: update internal depth state when queue depth changesJens Axboe1-1/+7
A previous commit moved the shallow depth and BFQ depth map calculations to be done at init time, moving it outside of the hotter IO path. This potentially causes hangs if the users changes the depth of the scheduler map, by writing to the 'nr_requests' sysfs file for that device. Add a blk-mq-sched hook that allows blk-mq to inform the scheduler if the depth changes, so that the scheduler can update its internal state. Tested-by: Kai Krakow <kai@kaishome.de> Reported-by: Paolo Valente <paolo.valente@linaro.org> Fixes: f0635b8a416e ("bfq: calculate shallow depths at init time") Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-10block, bfq: fix use after free in bfq_bfqq_expirePaolo Valente1-8/+7
The function bfq_bfqq_expire() invokes the function __bfq_bfqq_expire(), and the latter may free the in-service bfq-queue. If this happens, then no other instruction of bfq_bfqq_expire() must be executed, or a use-after-free will occur. Basing on the assumption that __bfq_bfqq_expire() invokes bfq_put_queue() on the in-service bfq-queue exactly once, the queue is assumed to be freed if its refcounter is equal to one right before invoking __bfq_bfqq_expire(). But, since commit 9dee8b3b057e ("block, bfq: fix queue removal from weights tree") this assumption is false. __bfq_bfqq_expire() may also invoke bfq_weights_tree_remove() and, since commit 9dee8b3b057e ("block, bfq: fix queue removal from weights tree"), also the latter function may invoke bfq_put_queue(). So __bfq_bfqq_expire() may invoke bfq_put_queue() twice, and this is the actual case where the in-service queue may happen to be freed. To address this issue, this commit moves the check on the refcounter of the queue right around the last bfq_put_queue() that may be invoked on the queue. Fixes: 9dee8b3b057e ("block, bfq: fix queue removal from weights tree") Reported-by: Dmitrii Tcvetkov <demfloro@demfloro.ru> Reported-by: Douglas Anderson <dianders@chromium.org> Tested-by: Dmitrii Tcvetkov <demfloro@demfloro.ru> Tested-by: Douglas Anderson <dianders@chromium.org> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-08block, bfq: fix some typos in commentsAngelo Ruocco1-8/+8
Some of the comments in the bfq files had typos. This patch fixes them. Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: save & resume weight on a queue merge/splitFrancesco Pollicino1-0/+2
bfq saves the state of a queue each time a merge occurs, to be able to resume such a state when the queue is associated again with its original process, on a split. Unfortunately bfq does not save & restore also the weight of the queue. If the weight is not correctly resumed when the queue is recycled, then the weight of the recycled queue could differ from the weight of the original queue. This commit adds the missing save & resume of the weight. Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Signed-off-by: Francesco Pollicino <fra.fra.800@gmail.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: print SHARED instead of pid for shared queues in logsFrancesco Pollicino1-0/+10
The function "bfq_log_bfqq" prints the pid of the process associated with the queue passed as input. Unfortunately, if the queue is shared, then more than one process is associated with the queue. The pid that gets printed in this case is the pid of one of the associated processes. Which process gets printed depends on the exact sequence of merge events the queue underwent. So printing such a pid is rather useless and above all is often rather confusing because it reports a random pid between those of the associated processes. This commit addresses this issue by printing SHARED instead of a pid if the queue is shared. Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Signed-off-by: Francesco Pollicino <fra.fra.800@gmail.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: always protect newly-created queues from existing active queuesPaolo Valente1-13/+51
If many bfq_queues belonging to the same group happen to be created shortly after each other, then the processes associated with these queues have typically a common goal. In particular, bursts of queue creations are usually caused by services or applications that spawn many parallel threads/processes. Examples are systemd during boot, or git grep. If there are no other active queues, then, to help these processes get their job done as soon as possible, the best thing to do is to reach a high throughput. To this goal, it is usually better to not grant either weight-raising or device idling to the queues associated with these processes. And this is exactly what BFQ currently does. There is however a drawback: if, in contrast, some other queues are already active, then the newly created queues must be protected from the I/O flowing through the already existing queues. In this case, the best thing to do is the opposite as in the other case: it is much better to grant weight-raising and device idling to the newly-created queues, if they deserve it. This commit addresses this issue by doing so if there are already other active queues. This change also helps eliminating false positives, which occur when the newly-created queues do not belong to an actual large burst of creations, but some background task (e.g., a service) happens to trigger the creation of new queues in the middle, i.e., very close to when the victim queues are created. These false positive may cause total loss of control on process latencies. Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: do not tag totally seeky queues as soft rtPaolo Valente1-0/+14
Sync random I/O is likely to be confused with soft real-time I/O, because it is characterized by limited throughput and apparently isochronous arrival pattern. To avoid false positives, this commits prevents bfq_queues containing only random (seeky) I/O from being tagged as soft real-time. Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: do not merge queues on flash storage with queueingPaolo Valente1-5/+68
To boost throughput with a set of processes doing interleaved I/O (i.e., a set of processes whose individual I/O is random, but whose merged cumulative I/O is sequential), BFQ merges the queues associated with these processes, i.e., redirects the I/O of these processes into a common, shared queue. In the shared queue, I/O requests are ordered by their position on the medium, thus sequential I/O gets dispatched to the device when the shared queue is served. Queue merging costs execution time, because, to detect which queues to merge, BFQ must maintain a list of the head I/O requests of active queues, ordered by request positions. Measurements showed that this costs about 10% of BFQ's total per-request processing time. Request processing time becomes more and more critical as the speed of the underlying storage device grows. Yet, fortunately, queue merging is basically useless on the very devices that are so fast to make request processing time critical. To reach a high throughput, these devices must have many requests queued at the same time. But, in this configuration, the internal scheduling algorithms of these devices do also the job of queue merging: they reorder requests so as to obtain as much as possible a sequential I/O pattern. As a consequence, with processes doing interleaved I/O, the throughput reached by one such device is likely to be the same, with and without queue merging. In view of this fact, this commit disables queue merging, and all related housekeeping, for non-rotational devices with internal queueing. The total, single-lock-protected, per-request processing time of BFQ drops to, e.g., 1.9 us on an Intel Core i7-2760QM@2.40GHz (time measured with simple code instrumentation, and using the throughput-sync.sh script of the S suite [1], in performance-profiling mode). To put this result into context, the total, single-lock-protected, per-request execution time of the lightest I/O scheduler available in blk-mq, mq-deadline, is 0.7 us (mq-deadline is ~800 LOC, against ~10500 LOC for BFQ). Disabling merging provides a further, remarkable benefit in terms of throughput. Merging tends to make many workloads artificially more uneven, mainly because of shared queues remaining non empty for incomparably more time than normal queues. So, if, e.g., one of the queues in a set of merged queues has a higher weight than a normal queue, then the shared queue may inherit such a high weight and, by staying almost always active, may force BFQ to perform I/O plugging most of the time. This evidently makes it harder for BFQ to let the device reach a high throughput. As a practical example of this problem, and of the benefits of this commit, we measured again the throughput in the nasty scenario considered in previous commit messages: dbench test (in the Phoronix suite), with 6 clients, on a filesystem with journaling, and with the journaling daemon enjoying a higher weight than normal processes. With this commit, the throughput grows from ~150 MB/s to ~200 MB/s on a PLEXTOR PX-256M5 SSD. This is the same peak throughput reached by any of the other I/O schedulers. As such, this is also likely to be the maximum possible throughput reachable with this workload on this device, because I/O is mostly random, and the other schedulers basically just pass I/O requests to the drive as fast as possible. [1] https://github.com/Algodev-github/S Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Tested-by: Francesco Pollicino <fra.fra.800@gmail.com> Signed-off-by: Alessio Masola <alessio.masola@gmail.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: tune service injection basing on request service timesPaolo Valente1-36/+381
The processes associated with a bfq_queue, say Q, may happen to generate their cumulative I/O at a lower rate than the rate at which the device could serve the same I/O. This is rather probable, e.g., if only one process is associated with Q and the device is an SSD. It results in Q becoming often empty while in service. If BFQ is not allowed to switch to another queue when Q becomes empty, then, during the service of Q, there will be frequent "service holes", i.e., time intervals during which Q gets empty and the device can only consume the I/O already queued in its hardware queues. This easily causes considerable losses of throughput. To counter this problem, BFQ implements a request injection mechanism, which tries to fill the above service holes with I/O requests taken from other bfq_queues. The hard part in this mechanism is finding the right amount of I/O to inject, so as to both boost throughput and not break Q's bandwidth and latency guarantees. To this goal, the current version of this mechanism measures the bandwidth enjoyed by Q while it is being served, and tries to inject the maximum possible amount of extra service that does not cause Q's bandwidth to decrease too much. This solution has an important shortcoming. For bandwidth measurements to be stable and reliable, Q must remain in service for a much longer time than that needed to serve a single I/O request. Unfortunately, this does not hold with many workloads. This commit addresses this issue by changing the way the amount of injection allowed is dynamically computed. It tunes injection as a function of the service times of single I/O requests of Q, instead of Q's bandwidth. Single-request service times are evidently meaningful even if Q gets very few I/O requests completed while it is in service. As a testbed for this new solution, we measured the throughput reached by BFQ for one of the nastiest workloads and configurations for this scheduler: the workload generated by the dbench test (in the Phoronix suite), with 6 clients, on a filesystem with journaling, and with the journaling daemon enjoying a higher weight than normal processes. With this commit, the throughput grows from ~100 MB/s to ~150 MB/s on a PLEXTOR PX-256M5. Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Tested-by: Francesco Pollicino <fra.fra.800@gmail.com> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: do not idle for lowest-weight queuesPaolo Valente1-90/+114
In most cases, it is detrimental for throughput to plug I/O dispatch when the in-service bfq_queue becomes temporarily empty (plugging is performed to wait for the possible arrival, soon, of new I/O from the in-service queue). There is however a case where plugging is needed for service guarantees. If a bfq_queue, say Q, has a higher weight than some other active bfq_queue, and is sync, i.e., contains sync I/O, then, to guarantee that Q does receive a higher share of the throughput than other lower-weight queues, it is necessary to plug I/O dispatch when Q remains temporarily empty while being served. For this reason, BFQ performs I/O plugging when some active bfq_queue has a higher weight than some other active bfq_queue. But this is unnecessarily overkill. In fact, if the in-service bfq_queue actually has a weight lower than or equal to the other queues, then the queue *must not* be guaranteed a higher share of the throughput than the other queues. So, not plugging I/O cannot cause any harm to the queue. And can boost throughput. Taking advantage of this fact, this commit does not plug I/O for sync bfq_queues with a weight lower than or equal to the weights of the other queues. Here is an example of the resulting throughput boost with the dbench workload, which is particularly nasty for BFQ. With the dbench test in the Phoronix suite, BFQ reaches its lowest total throughput with 6 clients on a filesystem with journaling, in case the journaling daemon has a higher weight than normal processes. Before this commit, the total throughput was ~80 MB/sec on a PLEXTOR PX-256M5, after this commit it is ~100 MB/sec. Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block, bfq: increase idling for weight-raised queuesPaolo Valente1-0/+2
If a sync bfq_queue has a higher weight than some other queue, and remains temporarily empty while in service, then, to preserve the bandwidth share of the queue, it is necessary to plug I/O dispatching until a new request arrives for the queue. In addition, a timeout needs to be set, to avoid waiting for ever if the process associated with the queue has actually finished its I/O. Even with the above timeout, the device is however not fed with new I/O for a while, if the process has finished its I/O. If this happens often, then throughput drops and latencies grow. For this reason, the timeout is kept rather low: 8 ms is the current default. Unfortunately, such a low value may cause, on the opposite end, a violation of bandwidth guarantees for a process that happens to issue new I/O too late. The higher the system load, the higher the probability that this happens to some process. This is a problem in scenarios where service guarantees matter more than throughput. One important case are weight-raised queues, which need to be granted a very high fraction of the bandwidth. To address this issue, this commit lower-bounds the plugging timeout for weight-raised queues to 20 ms. This simple change provides relevant benefits. For example, on a PLEXTOR PX-256M5S, with which gnome-terminal starts in 0.6 seconds if there is no other I/O in progress, the same applications starts in - 0.8 seconds, instead of 1.2 seconds, if ten files are being read sequentially in parallel - 1 second, instead of 2 seconds, if, in parallel, five files are being read sequentially, and five more files are being written sequentially Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com> Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01block/bfq: fix ifdef for CONFIG_BFQ_GROUP_IOSCHED=yKonstantin Khlebnikov1-1/+1
Replace BFQ_GROUP_IOSCHED_ENABLED with CONFIG_BFQ_GROUP_IOSCHED. Code under these ifdefs never worked, something might be broken. Fixes: 0471559c2fbd ("block, bfq: add/remove entity weights correctly") Fixes: 73d58118498b ("block, bfq: consider also ioprio classes in symmetry detection") Reviewed-by: Holger Hoffstätte <holger@applied-asynchrony.com> Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: fix in-service-queue check for queue mergingPaolo Valente1-1/+4
When a new I/O request arrives for a bfq_queue, say Q, bfq checks whether that request is close to (a) the head request of some other queue waiting to be served, or (b) the last request dispatched for the in-service queue (in case Q itself is not the in-service queue) If a queue, say Q2, is found for which the above condition holds, then bfq merges Q and Q2, to hopefully get a more sequential I/O in the resulting merged queue, and thus a possibly higher throughput. Case (b) is checked by comparing the new request for Q with the last request dispatched, assuming that the latter necessarily belonged to the in-service queue. Unfortunately, this assumption is no longer always correct, since commit d0edc2473be9 ("block, bfq: inject other-queue I/O into seeky idle queues on NCQ flash"). When the assumption does not hold, queues that must not be merged may be merged, causing unexpected loss of control on per-queue service guarantees. This commit solves this problem by adding an extra field, which stores the actual last request dispatched for the in-service queue, and by using this new field to correctly check case (b). Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: do not overcharge writes in asymmetric scenariosPaolo Valente1-1/+2
Writes tend to starve reads. bfq counters this problem by overcharging writes with an inflated service w.r.t. the actual service (number of sector written) they receive. Yet his overcharging is useless, and actually causes unfairness in the opposite direction, when bfq happens to be enforcing strong I/O control. bfq does this enforcing when the scenario is asymmetric, i.e., when some bfq_queue or group of bfq_queues is to be granted a different bandwidth than some other bfq_queue or group of bfq_queues. So, in such a scenario, this commit disables write overcharging. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: port commit "cfq-iosched: improve hw_tag detection"Paolo Valente1-0/+13
The original commit is commit 1a1238a7dd48 ("cfq-iosched: improve hw_tag detection") and has the following commit message: If active queue hasn't enough requests and idle window opens, cfq will not dispatch sufficient requests to hardware. In such situation, current code will zero hw_tag. But this is because cfq doesn't dispatch enough requests instead of hardware queue doesn't work. Don't zero hw_tag in such case. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: reduce threshold for detecting command queueingPaolo Valente1-2/+2
bfq simple heuristic from cfq for detecting whether the drive performs command queueing: check whether the average number of in-flight requests is above a given threshold. Unfortunately this heuristic does fail to detect queueing (on drives with queueing) if processes doing I/O are few and issue I/O with a low depth. To reduce false negatives, this commit lowers the threshold. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: fix queue removal from weights treePaolo Valente1-4/+13
bfq maintains an ordered list, through a red-black tree, of unique weights of active bfq_queues. This list is used to detect whether there are active queues with differentiated weights. The weight of a queue is removed from the list when both the following two conditions become true: (1) the bfq_queue is flagged as inactive (2) the has no in-flight request any longer; Unfortunately, in the rare cases where condition (2) becomes true before condition (1), the removal fails, because the function to remove the weight of the queue (bfq_weights_tree_remove) is rightly invoked in the path that deactivates the bfq_queue, but mistakenly invoked *before* the function that actually performs the deactivation (bfq_deactivate_bfqq). This commits moves the invocation of bfq_weights_tree_remove for condition (1) to after bfq_deactivate_bfqq. As a consequence of this move, it is necessary to add a further reference to the queue when the weight of a queue is added, because the queue might otherwise be freed before bfq_weights_tree_remove is invoked. This commit adds this reference and makes all related modifications. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: fix sequential rq detection in rate estimationPaolo Valente1-5/+7
In bfq_update_peak_rate, to check whether an I/O request rq is sequential, only the seek distance of rq w.r.t. the last request dispatched is controlled. This is not sufficient for non-rotational storage, where the size of rq is at least as relevant. This commit adds the missing control. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: unconditionally plug I/O in asymmetric scenariosPaolo Valente1-181/+165
bfq detects the creation of multiple bfq_queues shortly after each other, namely a burst of queue creations in the terminology used in the code. If the burst is large, then no queue in the burst is granted - either I/O-dispatch plugging when the queue remains temporarily idle while in service; - or weight raising, because it causes even longer plugging. In fact, such a plugging tends to lower throughput, while these bursts are typically due to applications or services that spawn multiple processes, to reach a common goal as soon as possible. Examples are a "git grep" or the booting of a system. Unfortunately, disabling plugging may cause a loss of service guarantees in asymmetric scenarios, i.e., if queue weights are differentiated or if more than one group is active. This commit addresses this issue by no longer disabling I/O-dispatch plugging for queues in large bursts. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: do not plug I/O of in-service queue when harmfulPaolo Valente1-14/+17
If the in-service bfq_queue is sync and remains temporarily idle, then I/O dispatching (from other queues) may be plugged. It may be dome for two reasons: either to boost throughput, or to preserve the bandwidth share of the in-service queue. In the first case, if the I/O of the in-service queue, when it finally arrives, consists only of one small I/O request, then it makes sense to plug even the I/O of the in-service queue. In fact, serving such a small request immediately is likely to lower throughput instead of boosting it, whereas waiting a little bit is likely to let that request grow, thanks to request merging, and become more profitable in terms of throughput (this is likely to happen exactly because the I/O of the queue has been detected to boost throughput). On the opposite end, if I/O dispatching is being plugged only to preserve the bandwidth of the in-service queue, then it would be better not to plug also the I/O of the in-service queue, because such a plugging is likely to cause only loss of bandwidth for the queue. Unfortunately, no distinction is made between the two cases, and the I/O of the in-service queue is always plugged in case just a small I/O request arrives. This commit draws this missing distinction and does not perform harmful plugging. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: split function bfq_better_to_idlePaolo Valente1-73/+82
This is a preparatory commit for commits that need to check only one of the two main reasons for idling. This change should also improve the quality of the code a little bit, by splitting a function that contains very long, non-trivial and little related comments. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: consider also ioprio classes in symmetry detectionPaolo Valente1-41/+45
In asymmetric scenarios, i.e., when some bfq_queue or bfq_group needs to be guaranteed a different bandwidth than other bfq_queues or bfq_groups, these service guaranteed can be provided only by plugging I/O dispatch, completely or partially, when the queue in service remains temporarily empty. A case where asymmetry is particularly strong is when some active bfq_queues belong to a higher-priority class than some other active bfq_queues. Unfortunately, this important case is not considered at all in the code for detecting asymmetric scenarios. This commit adds the missing logic. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: remove case of redirected bic from insert_requestPaolo Valente1-2/+0
Before commit 18e5a57d7987 ("block, bfq: postpone rq preparation to insert or merge"), the destination queue for a request was chosen by a different hook than the one that then inserted the request. So, between the execution of the two hooks, the bic of the process generating the request could happen to be redirected to a different bfq_queue. As a consequence, the destination bfq_queue stored in the request could be wrong. Such an event does not need to ba handled any longer. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: make sure queue budgets are not below service receivedPaolo Valente1-2/+4
With some unlucky sequences of events, the function bfq_updated_next_req updates the current budget of a bfq_queue to a lower value than the service received by the queue using such a budget. Unfortunately, if this happens, then the return value of the function bfq_bfqq_budget_left becomes inconsistent. This commit solves this problem by lower-bounding the budget computed in bfq_updated_next_req to the service currently charged to the queue. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: avoid selecting a queue w/o budgetPaolo Valente1-1/+9
To boost throughput on devices with internal queueing and in scenarios where device idling is not strictly needed, bfq immediately starts serving a new bfq_queue if the in-service bfq_queue remains without pending I/O, even if new I/O may arrive soon for the latter queue. Then, if such I/O actually arrives soon, bfq preempts the new in-service bfq_queue so as to give the previous queue a chance to go on being served (in case the previous queue should actually be the one to be served, according to its timestamps). However, the in-service bfq_queue, say Q, may also be without further budget when it remains also pending I/O. Since bfq changes budgets dynamically to fit the needs of bfq_queues, this happens more often than one may expect. If this happens, then there is no point in trying to go on serving Q when new I/O arrives for it soon: Q would be expired immediately after being selected for service. This would only cause useless overhead. This commit avoids such a useless selection. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31block, bfq: do not consider interactive queues in srt filteringPaolo Valente1-10/+29
The speed at which a bfq_queue receives I/O is one of the parameters by which bfq decides whether the queue is soft real-time (i.e., whether the queue contains the I/O of a soft real-time application). In particular, when a bfq_queue remains without outstanding I/O requests, bfq computes the minimum time instant, named soft_rt_next_start, at which the next request of the queue may arrive for the queue to be deemed as soft real time. Unfortunately this filtering may cause problems with a queue in interactive weight raising. In fact, such a queue may be conveying the I/O needed to load a soft real-time application. The latter will actually exhibit a soft real-time I/O pattern after it finally starts doing its job. But, if soft_rt_next_start is updated for an interactive bfq_queue, and the queue has received a lot of service before remaining with no outstanding request (likely to happen on a fast device), then soft_rt_next_start is assigned such a high value that, for a very long time, the queue is prevented from being possibly considered as soft real time. This commit removes the updating of soft_rt_next_start for bfq_queues in interactive weight raising. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-12-10Merge tag 'v4.20-rc6' into for-4.21/blockJens Axboe1-22/+54
Pull in v4.20-rc6 to resolve the conflict in NVMe, but also to get the two corruption fixes. We're going to be overhauling the direct dispatch path, and we need to do that on top of the changes we made for that in mainline. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-12-08blkcg: fix ref count issue with bio_blkcg() using task_cssDennis Zhou1-1/+1
The bio_blkcg() function turns out to be inconsistent and consequently dangerous to use. The first part returns a blkcg where a reference is owned by the bio meaning it does not need to be rcu protected. However, the third case, the last line, is problematic: return css_to_blkcg(task_css(current, io_cgrp_id)); This can race against task migration and the cgroup dying. It is also semantically different as it must be called rcu protected and is susceptible to failure when trying to get a reference to it. This patch adds association ahead of calling bio_blkcg() rather than after. This makes association a required and explicit step along the code paths for calling bio_blkcg(). In blk-iolatency, association is moved above the bio_blkcg() call to ensure it will not return %NULL. BFQ uses the old bio_blkcg() function, but I do not want to address it in this series due to the complexity. I have created a private version documenting the inconsistency and noting not to use it. Signed-off-by: Dennis Zhou <dennis@kernel.org> Acked-by: Tejun Heo <tj@kernel.org> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-12-07block, bfq: fix decrement of num_active_groupsPaolo Valente1-22/+54
Since commit '2d29c9f89fcd ("block, bfq: improve asymmetric scenarios detection")', if there are process groups with I/O requests waiting for completion, then BFQ tags the scenario as 'asymmetric'. This detection is needed for preserving service guarantees (for details, see comments on the computation * of the variable asymmetric_scenario in the function bfq_better_to_idle). Unfortunately, commit '2d29c9f89fcd ("block, bfq: improve asymmetric scenarios detection")' contains an error exactly in the updating of the number of groups with I/O requests waiting for completion: if a group has more than one descendant process, then the above number of groups, which is renamed from num_active_groups to a more appropriate num_groups_with_pending_reqs by this commit, may happen to be wrongly decremented multiple times, namely every time one of the descendant processes gets all its pending I/O requests completed. A correct, complete solution should work as follows. Consider a group that is inactive, i.e., that has no descendant process with pending I/O inside BFQ queues. Then suppose that num_groups_with_pending_reqs is still accounting for this group, because the group still has some descendant process with some I/O request still in flight. num_groups_with_pending_reqs should be decremented when the in-flight request of the last descendant process is finally completed (assuming that nothing else has changed for the group in the meantime, in terms of composition of the group and active/inactive state of child groups and processes). To accomplish this, an additional pending-request counter must be added to entities, and must be updated correctly. To avoid this additional field and operations, this commit resorts to the following tradeoff between simplicity and accuracy: for an inactive group that is still counted in num_groups_with_pending_reqs, this commit decrements num_groups_with_pending_reqs when the first descendant process of the group remains with no request waiting for completion. This simplified scheme provides a fix to the unbalanced decrements introduced by 2d29c9f89fcd. Since this error was also caused by lack of comments on this non-trivial issue, this commit also adds related comments. Fixes: 2d29c9f89fcd ("block, bfq: improve asymmetric scenarios detection") Reported-by: Steven Barrett <steven@liquorix.net> Tested-by: Steven Barrett <steven@liquorix.net> Tested-by: Lucjan Lucjanov <lucjan.lucjanov@gmail.com> Reviewed-by: Federico Motta <federico@willer.it> Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-11-15block: remove the queue_lock indirectionChristoph Hellwig1-8/+8
With the legacy request path gone there is no good reason to keep queue_lock as a pointer, we can always use the embedded lock now. Reviewed-by: Hannes Reinecke <hare@suse.com> Signed-off-by: Christoph Hellwig <hch@lst.de> Fixed floppy and blk-cgroup missing conversions and half done edits. Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-11-07block: get rid of MQ scheduler ops unionJens Axboe1-1/+1
This is a remnant of when we had ops for both SQ and MQ schedulers. Now it's just MQ, so get rid of the union. Reviewed-by: Omar Sandoval <osandov@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-11-07block: remove dead elevator codeJens Axboe1-1/+0
This removes a bunch of core and elevator related code. On the core front, we remove anything related to queue running, draining, initialization, plugging, and congestions. We also kill anything related to request allocation, merging, retrieval, and completion. Remove any checking for single queue IO schedulers, as they no longer exist. This means we can also delete a bunch of code related to request issue, adding, completion, etc - and all the SQ related ops and helpers. Also kill the load_default_modules(), as all that did was provide for a way to load the default single queue elevator. Tested-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Omar Sandoval <osandov@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>