blk-mq: only iterate over inflight requests in blk_mq_tagset_busy_iter
We already check for started commands in all callbacks, but we should
also protect against already completed commands. Do this by taking
the checks to common code.
Kevin Vigor [Wed, 30 May 2018 16:45:11 +0000 (10:45 -0600)]
nbd: clear DISCONNECT_REQUESTED flag once disconnection occurs.
When a userspace client requests a NBD device be disconnected, the
DISCONNECT_REQUESTED flag is set. While this flag is set, the driver
will not inform userspace when a connection is closed.
Unfortunately the flag was never cleared, so once a disconnect was
requested the driver would thereafter never tell userspace about a
closed connection. Thus when connections failed due to timeout, no
attempt to reconnect was made and eventually the device would fail.
Fix by clearing the DISCONNECT_REQUESTED flag (and setting the
DISCONNECTED flag) once all connections are closed.
Jianchao Wang [Wed, 30 May 2018 16:47:40 +0000 (10:47 -0600)]
block: kyber: make kyber more friendly with merging
Currently, kyber is very unfriendly with merging. kyber depends
on ctx rq_list to do merging, however, most of time, it will not
leave any requests in ctx rq_list. This is because even if tokens
of one domain is used up, kyber will try to dispatch requests
from other domain and flush the rq_list there.
To improve this, we setup kyber_ctx_queue (kcq) which is similar
with ctx, but it has rq_lists for different domain and build same
mapping between kcq and khd as the ctx & hctx. Then we could merge,
insert and dispatch for different domains separately. At the same
time, only flush the rq_list of kcq when get domain token successfully.
Then if one domain token is used up, the requests could be left in
the rq_list of that domain and maybe merged with following io.
Following is my test result on machine with 8 cores and NVMe card
INTEL SSDPEKKR128G7
fio size=256m ioengine=libaio iodepth=64 direct=1 numjobs=8
seq/random
+------+---------------------------------------------------------------+
|patch?| bw(MB/s) | iops | slat(usec) | clat(usec) | merge |
+----------------------------------------------------------------------+
| w/o | 606/612 | 151k/153k | 6.89/7.03 | 3349.21/3305.40 | 0/0 |
+----------------------------------------------------------------------+
| w/ | 1083/616 | 277k/154k | 4.93/6.95 | 1830.62/3279.95 | 223k/3k |
+----------------------------------------------------------------------+
When set numjobs to 16, the bw and iops could reach 1662MB/s and 425k
on my platform.
block: remove parent device reference from struct bsg_class_device
Bsg holding a reference to the parent device may result in a crash if a
bsg file handle is closed after the parent device driver has unloaded.
Holding a reference is not really needed: the parent device must exist
between bsg_register_queue and bsg_unregister_queue. Before the device
goes away the caller does blk_cleanup_queue so that all in-flight
requests to the device are gone and all new requests cannot pass beyond
the queue. The queue itself is a refcounted object and it will stay
alive with a bsg file.
Based on analysis, previous patch and changelog from Anatoliy Glagolev.
Jens Axboe [Tue, 29 May 2018 18:56:20 +0000 (12:56 -0600)]
Merge branch 'nvme-4.18-2' of git://git.infradead.org/nvme into for-4.18/block
Pull NVMe changes from Christoph:
"Here is the current batch of nvme updates for 4.18, we have a few more
patches in the queue, but I'd like to get this pile into your tree
and linux-next ASAP.
The biggest item is support for file-backed namespaces in the NVMe
target from Chaitanya, in addition to that we mostly small fixes from
all the usual suspects."
* 'nvme-4.18-2' of git://git.infradead.org/nvme:
nvme: fixup memory leak in nvme_init_identify()
nvme: fix KASAN warning when parsing host nqn
nvmet-loop: use nr_phys_segments when map rq to sgl
nvmet-fc: increase LS buffer count per fc port
nvmet: add simple file backed ns support
nvmet: remove duplicate NULL initialization for req->ns
nvmet: make a few error messages more generic
nvme-fabrics: allow duplicate connections to the discovery controller
nvme-fabrics: centralize discovery controller defaults
nvme-fabrics: remove unnecessary controller subnqn validation
nvme-fc: remove setting DNR on exception conditions
nvme-rdma: stop admin queue before freeing it
nvme-pci: Fix AER reset handling
nvme-pci: set nvmeq->cq_vector after alloc cq/sq
nvme: host: core: fix precedence of ternary operator
nvme: fix lockdep warning in nvme_mpath_clear_current_path
Jens Axboe [Tue, 29 May 2018 14:47:57 +0000 (08:47 -0600)]
block: move ->timeout request member
After the recent timeout handling changes, we have two holes in
the struct. Move the timeout near the deadline, killing both,
and moving related members closer together. On my config on
x86-64, this shrinks struct request from 312 to 304 bytes.
libiscsi is the only SCSI code that return BLK_EH_HANDLED, thus trying to
bypass the normal SCSI EH code. We are going to remove this return value
at the block layer, and at least from a quick look it doesn't look too
harmful to try to send an abort for these cases, especially as the first
one should not actually be possible. If this doesn't work out iscsi
will probably need its own eh_strategy_handler instead to just do the
right thing.
By completing the request entirely in the driver we can remove the
BLK_EH_HANDLED return value and thus the split responsibility between the
driver and the block layer that has been causing trouble.
[While this keeps existing behavior it seems to mismatch the comment,
maintainers please chime in!]
scsi_transport_fc: complete requests from ->timeout
By completing the request entirely in the driver we can remove the
BLK_EH_HANDLED return value and thus the split responsibility between the
driver and the block layer that has been causing trouble.
By completing the request entirely in the driver we can remove the
BLK_EH_HANDLED return value and thus the split responsibility between the
driver and the block layer that has been causing trouble.
By completing the request entirely in the driver we can remove the
BLK_EH_HANDLED return value and thus the split responsibility between the
driver and the block layer that has been causing trouble.
By completing the request entirely in the driver we can remove the
BLK_EH_HANDLED return value and thus the split responsibility between the
driver and the block layer that has been causing trouble.
NVMe always completes the request before returning from ->timeout, either
by polling for it, or by disabling the controller. Return BLK_EH_DONE so
that the block layer doesn't even try to complete it again.
The BLK_EH_NOT_HANDLED implies nothing happen, but very often that
is not what is happening - instead the driver already completed the
command. Fix the symbolic name to reflect that a little better.
Keith Busch [Tue, 29 May 2018 13:52:28 +0000 (15:52 +0200)]
blk-mq: Remove generation seqeunce
This patch simplifies the timeout handling by relying on the request
reference counting to ensure the iterator is operating on an inflight
and truly timed out request. Since the reference counting prevents the
tag from being reallocated, the block layer no longer needs to prevent
drivers from completing their requests while the timeout handler is
operating on it: a driver completing a request is allowed to proceed to
the next state without additional syncronization with the block layer.
This also removes any need for generation sequence numbers since the
request lifetime is prevented from being reallocated as a new sequence
while timeout handling is operating on it.
To enables this a refcount is added to struct request so that request
users can be sure they're operating on the same request without it
changing while they're processing it. The request's tag won't be
released for reuse until both the timeout handler and the completion
are done with it.
Signed-off-by: Keith Busch <[email protected]>
[hch: slight cleanups, added back submission side hctx lock, use cmpxchg
for completions] Signed-off-by: Christoph Hellwig <[email protected]> Signed-off-by: Jens Axboe <[email protected]>
Keith Busch [Tue, 29 May 2018 13:52:27 +0000 (15:52 +0200)]
blk-mq: Fix timeout and state order
The block layer had been setting the state to in-flight prior to updating
the timer. This is the wrong order since the timeout handler could observe
the in-flight state with the older timeout, believing the request had
expired when in fact it is just getting started.
As far as I can tell this function can't even be called any more, given
that ATA implements its own eh_strategy_handler with ata_scsi_error, which
never calls ->eh_timed_out.
Andy Shevchenko [Mon, 28 May 2018 07:37:44 +0000 (15:37 +0800)]
bcache: Replace bch_read_string_list() by __sysfs_match_string()
Kernel library has a common function to match user input from sysfs
against an array of strings. Thus, replace bch_read_string_list() by
__sysfs_match_string().
Coly Li [Mon, 28 May 2018 07:37:41 +0000 (15:37 +0800)]
bcache: stop bcache device when backing device is offline
Currently bcache does not handle backing device failure, if backing
device is offline and disconnected from system, its bcache device can still
be accessible. If the bcache device is in writeback mode, I/O requests even
can success if the requests hit on cache device. That is to say, when and
how bcache handles offline backing device is undefined.
This patch tries to handle backing device offline in a rather simple way,
- Add cached_dev->status_update_thread kernel thread to update backing
device status in every 1 second.
- Add cached_dev->offline_seconds to record how many seconds the backing
device is observed to be offline. If the backing device is offline for
BACKING_DEV_OFFLINE_TIMEOUT (30) seconds, set dc->io_disable to 1 and
call bcache_device_stop() to stop the bache device which linked to the
offline backing device.
Now if a backing device is offline for BACKING_DEV_OFFLINE_TIMEOUT seconds,
its bcache device will be removed, then user space application writing on
it will get error immediately, and handler the device failure in time.
This patch is quite simple, does not handle more complicated situations.
Once the bcache device is stopped, users need to recovery the backing
device, register and attach it manually.
Changelog:
v3: call wait_for_kthread_stop() before exits kernel thread.
v2: remove "bcache: " prefix when calling pr_warn().
v1: initial version.
nvmet-loop: use nr_phys_segments when map rq to sgl
Use blk_rq_nr_phys_segments() instead of blk_rq_payload_bytes() to check
if a command contains data to me mapped. This fixes the case where
a struct requests contains LBAs, but no data will actually be send,
e.g. the pending Write Zeroes support.
James Smart [Mon, 21 May 2018 23:27:42 +0000 (16:27 -0700)]
nvmet-fc: increase LS buffer count per fc port
Todays limit on concurrent LS's is very small - 4 buffers. With large
subsystem counts or large numbers of initiators connecting, the limit
may be exceeded.
This patch adds simple file backed namespace support for NVMeOF target.
The new file io-cmd-file.c is responsible for handling the code for I/O
commands when ns is file backed. Also, we introduce mempools based slow
path using sync I/Os for file backed ns to ensure forward progress under
reclaim.
The old block device based implementation is moved to io-cmd-bdev.c and
use a "nvmet_bdev_" symbol prefix. The enable/disable calls are also
move into the respective files.
Signed-off-by: Chaitanya Kulkarni <[email protected]>
[hch: updated changelog, fixed double req->ns lookup in bdev case] Signed-off-by: Christoph Hellwig <[email protected]>
nvmet: remove duplicate NULL initialization for req->ns
Remove the duplicate NULL initialization for req->ns. req->ns is always
initialized to NULL in nvmet_req_init(), so there is no need to reset
it later on failures unless we have previously assigned a value to it.
Hannes Reinecke [Thu, 24 May 2018 14:18:17 +0000 (16:18 +0200)]
nvme-fabrics: allow duplicate connections to the discovery controller
The whole point of the discovery controller is that it can accept
multiple connections. Additionally the cmic field is not even defined for
the discovery controller identify page.
After creating the nvme controller, nvmf_create_ctrl() validates
the newly created subsysnqn vs the one specified by the options.
In general, this is an unnecessary check as the Connect message
should implicitly ensure this value matches.
With the change to the FC transport to do an asynchronous connect
for the first association create, the transport will return to
nvmf_create_ctrl() before that first association has been established,
thus the subnqn will not yet be set.
Jianchao Wang [Thu, 24 May 2018 01:27:38 +0000 (09:27 +0800)]
nvme-rdma: stop admin queue before freeing it
For any failure after nvme_rdma_start_queue in
nvme_rdma_configure_admin_queue, the admin queue will be freed with the
NVME_RDMA_Q_LIVE flag still set. Once nvme_rdma_stop_queue is invoked,
that will cause a use-after-free.
BUG: KASAN: use-after-free in rdma_disconnect+0x1f/0xe0 [rdma_cm]
To fix it, call nvme_rdma_stop_queue for all the failed cases after
nvme_rdma_start_queue.
Keith Busch [Thu, 24 May 2018 22:16:04 +0000 (16:16 -0600)]
nvme-pci: Fix AER reset handling
The nvme timeout handling doesn't do anything if the pci channel is
offline, which is the case when recovering from PCI error event, so it
was a bad idea to sync the controller reset in this state. This patch
flushes the reset work in the error_resume callback instead when the
channel is back to online. This keeps AER handling serialized and
can recover from timeouts.
Jianchao Wang [Thu, 24 May 2018 09:51:33 +0000 (17:51 +0800)]
nvme-pci: set nvmeq->cq_vector after alloc cq/sq
Set cq_vector after alloc cq/sq, otherwise nvme_suspend_queue will invoke
free_irq for it and cause a 'Trying to free already-free IRQ xxx'
warning if the create CQ/SQ command times out.
Ming Lei [Thu, 24 May 2018 17:00:39 +0000 (11:00 -0600)]
blk-mq: avoid starving tag allocation after allocating process migrates
When the allocation process is scheduled back and the mapped hw queue is
changed, fake one extra wake up on previous queue for compensating wake
up miss, so other allocations on the previous queue won't be starved.
This patch fixes one request allocation hang issue, which can be
triggered easily in case of very low nr_request.
The race is as follows:
1) 2 hw queues, nr_requests are 2, and wake_batch is one
2) there are 3 waiters on hw queue 0
3) two in-flight requests in hw queue 0 are completed, and only two
waiters of 3 are waken up because of wake_batch, but both the two
waiters can be scheduled to another CPU and cause to switch to hw
queue 1
4) then the 3rd waiter will wait for ever, since no in-flight request
is in hw queue 0 any more.
5) this patch fixes it by the fake wakeup when waiter is scheduled to
another hw queue
cgwb_release() punts the actual release to cgwb_release_workfn() on
system_wq. Depending on the number of cgroups or block devices, there
can be a lot of cgwb_release_workfn() in flight at the same time.
We're periodically seeing close to 256 kworkers getting stuck with the
following stack trace and overtime the entire system gets stuck.
1. A lot of cgwb_release_workfn() is queued at the same time and all
system_wq kworkers are assigned to execute them.
2. They all end up calling synchronize_rcu_expedited(). One of them
wins and tries to perform the expedited synchronization.
3. However, that invovles queueing rcu_exp_work to system_wq and
waiting for it. Because #1 is holding all available kworkers on
system_wq, rcu_exp_work can't be executed. cgwb_release_workfn()
is waiting for synchronize_rcu_expedited() which in turn is waiting
for cgwb_release_workfn() to free up some of the kworkers.
We shouldn't be scheduling hundreds of cgwb_release_workfn() at the
same time. There's nothing to be gained from that. This patch
updates cgwb release path to use a dedicated percpu workqueue with
@max_active of 1.
While this resolves the problem at hand, it might be a good idea to
isolate rcu_exp_work to its own workqueue too as it can be used from
various paths and is prone to this sort of indirect A-A deadlocks.
Josef Bacik [Wed, 23 May 2018 17:35:59 +0000 (13:35 -0400)]
nbd: set discard granularity properly
For some reason we had discard granularity set to 512 always even when
discards were disabled. Fix this by having the default be 0, and then
if we turn it on set the discard granularity to the blocksize.
The only caller of nvme_mpath_clear_current_path() is nvme_ns_remove()
which holds the subsys lock so it's likely a false positive, but when
using rcu_access_pointer(), we're telling rcu and lockdep that we're
only after the pointer falue.
Fixes: 32acab3181c7 ("nvme: implement multipath access to nvme subsystems") Signed-off-by: Johannes Thumshirn <[email protected]> Suggested-by: Paul E. McKenney <[email protected]> Reviewed-by: Christoph Hellwig <[email protected]> Signed-off-by: Keith Busch <[email protected]>
Jens Axboe [Mon, 21 May 2018 14:41:52 +0000 (08:41 -0600)]
nvme-pci: fix race between poll and IRQ completions
If polling completions are racing with the IRQ triggered by a
completion, the IRQ handler will find no work and return IRQ_NONE.
This can trigger complaints about spurious interrupts:
A previous commit removed ->cqe_seen that was handling this case,
but we need to handle this a bit differently due to completions
now running outside the queue lock. Return IRQ_HANDLED from the
IRQ handler, if the completion ring head was moved since we last
saw it.
Jens Axboe [Mon, 21 May 2018 14:33:37 +0000 (08:33 -0600)]
Merge branch 'nvme-4.18' of git://git.infradead.org/nvme into for-4.18/block
Pull NVMe changes from Keith:
"This is just the first nvme pull request for 4.18. There are several
fabrics and target patches that I missed, so there will be more to
come."
* 'nvme-4.18' of git://git.infradead.org/nvme:
nvme-pci: drop IRQ disabling on submission queue lock
nvme-pci: split the nvme queue lock into submission and completion locks
nvme-pci: handle completions outside of the queue lock
nvme-pci: move ->cq_vector == -1 check outside of ->q_lock
nvme-pci: remove cq check after submission
nvme-pci: simplify nvme_cqe_valid
nvme: mark the result argument to nvme_complete_async_event volatile
nvme/pci: Sync controller reset for AER slot_reset
nvme/pci: Hold controller reference during async probe
nvme: only reconfigure discard if necessary
nvme/pci: Use async_schedule for initial reset work
nvme: lightnvm: add granby support
NVMe: Add Quirk Delay before CHK RDY for Seagate Nytro Flash Storage
nvme: change order of qid and cmdid in completion trace
nvme: fc: provide a descriptive error
Jens Axboe [Thu, 17 May 2018 16:31:50 +0000 (18:31 +0200)]
nvme-pci: handle completions outside of the queue lock
Split the completion of events into a two part process:
1) Reap the events inside the queue lock
2) Complete the events outside the queue lock
Since we never wrap the queue, we can access it locklessly after we've
updated the completion queue head. This patch started off with batching
events on the stack, but with this trick we don't have to. Keith Busch
<[email protected]> came up with that idea.
Note that this kills the ->cqe_seen as well. I haven't been able to
trigger any ill effects of this. If we do race with polling every so
often, it should be rare enough NOT to trigger any issues.
Jens Axboe [Thu, 17 May 2018 16:31:48 +0000 (18:31 +0200)]
nvme-pci: remove cq check after submission
We always check the completion queue after submitting, but in my testing
this isn't a win even on DRAM/xpoint devices. In some cases it's
actually worse. Kill it.
huhai [Fri, 18 May 2018 14:32:30 +0000 (08:32 -0600)]
blk-mq: clear hctx->dispatch_from when mappings change
When the number of hardware queues is changed, the drivers will call
blk_mq_update_nr_hw_queues() to remap hardware queues. This changes
the ctx mappings, but the current code doesn't clear the
->dispatch_from hint. This can result in dispatch_from pointing to
a ctx that isn't mapped to the hctx anymore.
Fixes: b347689ffbca ("blk-mq-sched: improve dispatching from sw queue") Signed-off-by: huhai <[email protected]> Reviewed-by: Ming Lei <[email protected]>
Moved the placement of the clearing to where we clear other items
pertaining to the existing mapping, added Fixes line, and reworded
the commit message.
Josef Bacik [Wed, 16 May 2018 18:51:22 +0000 (14:51 -0400)]
nbd: call nbd_bdev_reset instead of bd_set_size on disconnect
We need to make sure we don't just set the size of the bdev to 0 while
it's being used by a file system. We have the appropriate check in
nbd_bdev_reset, simply use that helper instead.
Josef Bacik [Wed, 16 May 2018 18:51:21 +0000 (14:51 -0400)]
nbd: fix how we set bd_invalidated
bd_invalidated is kind of a pain wrt partitions as it really only
triggers the partition rescan if it is set after bd_ops->open() runs, so
setting it when we reset the device isn't useful. We also sporadically
would still have partitions left over in some disconnect cases, so fix
this by always setting bd_invalidated on open if there's no
configuration or if we've had a disconnect action happen, that way the
partition table gets invalidated and rescanned properly.
Josef Bacik [Wed, 16 May 2018 18:51:20 +0000 (14:51 -0400)]
nbd: clear_sock on netlink disconnect
This is what the ioctl based nbd disconnect does as well. Without this
the device will just sit there and wait for the connection to go away
(or IO to occur) before the device gets torn down. Instead clear
everything up on our end so the configuration goes away as quickly as
possible.
Josef Bacik [Wed, 16 May 2018 18:51:19 +0000 (14:51 -0400)]
nbd: use bd_set_size when updating disk size
When we stopped relying on the bdev everywhere I broke updating the
block device size on the fly, which ceph relies on. We can't just do
set_capacity, we also have to do bd_set_size so things like parted will
notice the device size change.
Josef Bacik [Wed, 16 May 2018 18:51:18 +0000 (14:51 -0400)]
nbd: update size when connected
I messed up changing the size of an NBD device while it was connected by
not actually updating the device or doing the uevent. Fix this by
updating everything if we're connected and we change the size.
Josef Bacik [Wed, 16 May 2018 18:51:17 +0000 (14:51 -0400)]
nbd: fix nbd device deletion
This fixes a use after free bug, we shouldn't be doing disk->queue right
after we do del_gendisk(disk). Save the queue and do the cleanup after
the del_gendisk.
Josef Bacik [Wed, 16 May 2018 18:36:01 +0000 (14:36 -0400)]
block: fix MAINTAINERS email for nbd
I've been missing stuff because it's been going into my work email which
is a black hole. Update to the email I actually use so I stop missing
patches and bug reports.
huhai [Wed, 16 May 2018 14:21:21 +0000 (08:21 -0600)]
blk-mq: remove redundant insert case in blk_mq_make_request()
We can use blk_mq_sched_insert_request() even if we don't have
an IO scheduler attached, since that case will end up being
exactly the same as what blk_mq_queue_io() was doing now.
Kent Overstreet [Wed, 9 May 2018 01:33:56 +0000 (21:33 -0400)]
block: Add warning for bi_next not NULL in bio_endio()
Recently found a bug where a driver left bi_next not NULL and then
called bio_endio(), and then the submitter of the bio used
bio_copy_data() which was treating src and dst as lists of bios.
Fixed that bug by splitting out bio_list_copy_data(), but in case other
things are depending on bi_next in weird ways, add a warning to help
avoid more bugs like that in the future.
Kent Overstreet [Wed, 9 May 2018 01:33:54 +0000 (21:33 -0400)]
block: Split out bio_list_copy_data()
Found a bug (with ASAN) where we were passing a bio to bio_copy_data()
with bi_next not NULL, when it should have been - a driver had left
bi_next set to something after calling bio_endio().
Since the normal case is only copying single bios, split out
bio_list_copy_data() to avoid more bugs like this in the future.
Jens Axboe [Mon, 14 May 2018 18:17:31 +0000 (12:17 -0600)]
sbitmap: fix race in wait batch accounting
If we have multiple callers of sbq_wake_up(), we can end up in a
situation where the wait_cnt will continually go more and more
negative. Consider the case where our wake batch is 1, hence
wait_cnt will start out as 1.
This ends up with wait_cnt being 0, we'll wakeup immediately
next time. Going through the same loop as above again, and
we'll have wait_cnt -1.
For the case where we have a larger wake batch, the only
difference is that the starting point will be higher. We'll
still end up with continually smaller batch wakeups, which
defeats the purpose of the rolling wakeups.
Always reset the wait_cnt to the batch value. Then it doesn't
matter who wins the race. But ensure that whomever does win
the race is the one that increments the ws index and wakes up
our batch count, loser gets to call __sbq_wake_up() again to
account his wakeups towards the next active wait state index.
Fixes: 6c0ca7ae292a ("sbitmap: fix wakeup hang after sbq resize") Reviewed-by: Omar Sandoval <[email protected]> Signed-off-by: Jens Axboe <[email protected]>
projects / linux.git / log