- 19 Jan, 2018 1 commit
-
-
Joseph Qi authored
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>
-
- 18 Jan, 2018 9 commits
-
-
Ilya Dryomov authored
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:
-
Ilya Dryomov authored
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:
-
weiping zhang authored
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:
-
Bart Van Assche authored
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:
-
Bart Van Assche authored
This patch does not change any functionality. Reviewed-by:
-
Bart Van Assche authored
These two functions are only called from inside the block layer so unexport them. Reviewed-by:
-
Paolo Valente authored
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:
-
Paolo Valente authored
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/STested-by:
-
Ming Lei authored
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: 396eaf21 ("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:
-
- 17 Jan, 2018 7 commits
-
-
Bart Van Assche authored
Fixes: 4246a0b6 ("block: add a bi_error field to struct bio") Reviewed-by:
Martin K. Petersen <martin.petersen@oracle.com> Signed-off-by:
-
Mike Snitzer authored
After commit: 923218f6 ("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:
-
Ming Lei authored
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:
-
Mike Snitzer authored
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:
-
Ming Lei authored
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 authored
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:
-
Tina Ruchandani authored
'struct frame' uses two variables to store the sent timestamp - 'struct timeval' and jiffies. jiffies is used to avoid discrepancies caused by updates to system time. 'struct timeval' is deprecated because it uses 32-bit representation for seconds which will overflow in year 2038. This patch does the following: - Replace the use of 'struct timeval' and jiffies with ktime_t, which is the recommended type for timestamping - ktime_t provides both long range (like jiffies) and high resolution (like timeval). Using ktime_get (monotonic time) instead of wall-clock time prevents any discprepancies caused by updates to system time. [updates by Arnd below] The original patch from Tina never went anywhere as we discussed how to keep the impact on performance minimal. I've started over now but arrived at basically the same patch that she had originally, except for an slightly improved tsince_hr() function. I'm making it more robust against overflows, and also optimize explicitly for the common case in which a frame is less than 4.2 seconds old, using only a 32-bit division in that case. This should make the new version more efficient than the old code, since we replace the existing two 32-bit division in do_gettimeofday() plus one multiplication with a single single 32-bit division in tsince_hr() and drop the double bookkeeping. It's also more efficient than the ktime_get_us() API we discussed before, since that would also rely on multiple divisions. Link: https://lists.linaro.org/pipermail/y2038/2015-May/000276.htmlSigned-off-by:
Tina Ruchandani <ruchandani.tina@gmail.com> Cc: Ed Cashin <ed.cashin@acm.org> Signed-off-by:
Arnd Bergmann <arnd@arndb.de> Signed-off-by:
-
- 16 Jan, 2018 1 commit
-
-
Arnd Bergmann authored
Some older compilers (gcc-4.4 through 4.6 in particular) struggle with the way that blkg_rwstat_read() returns a structure, leading to excessive stack usage and rather inefficient code: block/blk-cgroup.c: In function 'blkg_destroy': block/blk-cgroup.c:354:1: error: the frame size of 1296 bytes is larger than 1024 bytes [-Werror=frame-larger-than=] block/cfq-iosched.c: In function 'cfqg_stats_add_aux': block/cfq-iosched.c:753:1: error: the frame size of 1928 bytes is larger than 1024 bytes [-Werror=frame-larger-than=] block/bfq-cgroup.c: In function 'bfqg_stats_add_aux': block/bfq-cgroup.c:299:1: error: the frame size of 1928 bytes is larger than 1024 bytes [-Werror=frame-larger-than=] I also notice that there is no point in using atomic accesses for the local variables, so storing the temporaries in simple 'u64' variables not only avoids the stack usage on older compilers but also improves the object code on modern versions. Fixes: e6269c44 ("blkcg: add blkg_[rw]stat->aux_cnt and replace cfq_group->dead_stats with it") Acked-by:
-
- 15 Jan, 2018 5 commits
-
-
Mike Snitzer authored
DM is no longer prone to having its request_queue be improperly initialized. Summary of changes: - defer DM's blk_register_queue() from add_disk()-time until dm_setup_md_queue() by using add_disk_no_queue_reg() in alloc_dev(). - dm_setup_md_queue() is updated to fully initialize DM's request_queue (_after_ all table loads have occurred and the request_queue's type, features and limits are known). A very welcome side-effect of these changes is DM no longer needs to: 1) backfill the "mq" sysfs entry (because historically DM didn't initialize the request_queue to use blk-mq until _after_ blk_register_queue() was called via add_disk()). 2) call elv_register_queue() to get .request_fn request-based DM device's "iosched" exposed in syfs. In addition, blk-mq debugfs support is now made available because request-based DM's blk-mq request_queue is now properly initialized before dm_setup_md_queue() calls blk_register_queue(). These changes also stave off the need to introduce new DM-specific workarounds in block core, e.g. this proposal: https://patchwork.kernel.org/patch/10067961/ In the end DM devices should be less unicorn in nature (relative to initialization and availability of block core infrastructure provided by the request_queue). Signed-off-by:
-
Douglas Gilbert authored
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:
-
Mike Snitzer authored
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 authored
The original commit e9a823fb (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: e9a823fb ("block: fix warning when I/O elevator is changed as request_queue is being removed") Signed-off-by:
-
Mike Snitzer authored
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: 8ddcd653 ("block: introduce GENHD_FL_HIDDEN") Signed-off-by:
Hannes Reinecke <hare@suse.com> Signed-off-by:
-
- 14 Jan, 2018 1 commit
-
-
Jens Axboe authored
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: 7c3fb70f ("block: rearrange a few request fields for better cache layout") Reviewed-by:
-
- 12 Jan, 2018 3 commits
-
-
Jens Axboe authored
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: -
Christoph Hellwig authored
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:
Stefan Haberland <sth@linux.vnet.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by:
-
Christoph Hellwig authored
Currently we assign managed interrupt vectors to all present CPUs. This works fine for systems were we only online/offline CPUs. But in case of systems that support physical CPU hotplug (or the virtualized version of it) this means the additional CPUs covered for in the ACPI tables or on the command line are not catered for. To fix this we'd either need to introduce new hotplug CPU states just for this case, or we can start assining vectors to possible but not present CPUs. Reported-by:
-
- 11 Jan, 2018 2 commits
-
-
Bart Van Assche authored
This patch does not change any functionality but makes the blk_mq_mark_tag_wait() code slightly easier to read. Signed-off-by:
-
Arnd Bergmann authored
Selecting FAULT_INJECTION causes a Kconfig warning when CONFIG_DEBUG_KERNEL is not set: warning: (BLK_DEV_NULL_BLK && DRM_I915_SELFTEST) selects FAULT_INJECTION which has unmet direct dependencies (DEBUG_KERNEL) The other drivers that use FAULT_INJECTION tend to have a separate Kconfig symbol for turning on that feature, so let's do the same thing here. This may add a bit more complexity than we like, but it avoids the warning and is more consistent with the rest of the kernel. Fixes: 93b57046 ("null_blk: add option for managing IO timeouts") Signed-off-by:
-
- 10 Jan, 2018 11 commits
-
-
Bart Van Assche authored
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:
-
Paolo Bonzini authored
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:
Paolo Bonzini <pbonzini@redhat.com> Signed-off-by:
-
Jens Axboe authored
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:
Omar Sandoval <osandov@fb.com> Signed-off-by:
-
Jens Axboe authored
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:
-
Jens Axboe authored
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:
-
Jens Axboe authored
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:
-
Jens Axboe authored
We are missing ZONE_WRITE_LOCKED and MQ_TIMEOUT_EXPIRED, add them so the debugfs bits can decode them. Signed-off-by: -
Keith Busch authored
Uses common code for determining if an error should be retried on alternate path. Acked-by:
Keith Busch <keith.busch@intel.com> Reviewed-by:
Johannes Thumshirn <jthumshirn@suse.de> Reviewed-by:
-
Keith Busch authored
Uses common code for determining if an error should be retried on alternate path. Acked-by:
-
Keith Busch authored
This patch provides a common decoder for block status path related errors that may be retried so various entities wishing to consult this do not have to duplicate this decision. Acked-by:
-
Keith Busch authored
This removes nvme multipath's specific status decoding to see if failover is needed, using the generic blk_status_t that was decoded earlier. This abstraction from the raw NVMe status means all status decoding exists in one place. Acked-by:
-
