openwrt/staging/blogic.git
6 years agoblock: move initialization of elevator-related fields to blk_alloc_queue_node
Christoph Hellwig [Thu, 31 May 2018 17:11:36 +0000 (19:11 +0200)]
block: move initialization of elevator-related fields to blk_alloc_queue_node

No point in doing this in elevator_init.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Damien Le Moal <Damien.LeMoal@wdc.com>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Tested-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock, bfq: prevent soft_rt_next_start from being stuck at infinity
Davide Sapienza [Thu, 31 May 2018 14:45:08 +0000 (16:45 +0200)]
block, bfq: prevent soft_rt_next_start from being stuck at infinity

BFQ can deem a bfq_queue as soft real-time only if the queue
- periodically becomes completely idle, i.e., empty and with
  no still-outstanding I/O request;
- after becoming idle, gets new I/O only after a special reference
  time soft_rt_next_start.

In this respect, after commit "block, bfq: consider also past I/O in
soft real-time detection", the value of soft_rt_next_start can never
decrease. This causes a problem with the following special updating
case for soft_rt_next_start: to prevent queues that are not completely
idle to be wrongly detected as soft real-time (when they become
non-empty again), soft_rt_next_start is temporarily set to infinity
for empty queues with still outstanding I/O requests. But, if such an
update is actually performed, then, because of the above commit,
soft_rt_next_start will be stuck at infinity forever, and the queue
will have no more chance to be considered soft real-time.

On slow systems, this problem does cause actual soft real-time
applications to be occasionally not detected as such.

This commit addresses this issue by eliminating the pushing of
soft_rt_next_start to infinity, and by changing the way non-empty
queues are prevented from being wrongly detected as soft
real-time. Simply, a queue that becomes non-empty again can now be
detected as soft real-time only if it has no outstanding I/O request.

Signed-off-by: Davide Sapienza <sapienza.dav@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock, bfq: increase weight-raising duration for interactive apps
Davide Sapienza [Thu, 31 May 2018 14:45:07 +0000 (16:45 +0200)]
block, bfq: increase weight-raising duration for interactive apps

The maximum possible duration of the weight-raising period for
interactive applications is limited to 13 seconds, as this is the time
needed to load the largest application that we considered when tuning
weight raising. Unfortunately, in such an evaluation, we did not
consider the case of very slow virtual machines.

For example, on a QEMU/KVM virtual machine
- running in a slow PC;
- with a virtual disk stacked on a slow low-end 5400rpm HDD;
- serving a heavy I/O workload, such as the sequential reading of
several files;
mplayer takes 23 seconds to start, if constantly weight-raised.

To address this issue, this commit conservatively sets the upper limit
for weight-raising duration to 25 seconds.

Signed-off-by: Davide Sapienza <sapienza.dav@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock, bfq: remove slow-system class
Paolo Valente [Thu, 31 May 2018 14:45:06 +0000 (16:45 +0200)]
block, bfq: remove slow-system class

BFQ computes the duration of weight raising for interactive
applications automatically, using some reference parameters. In
particular, BFQ uses the best durations (see comments in the code for
how these durations have been assessed) for two classes of systems:
slow and fast ones. Examples of slow systems are old phones or systems
using micro HDDs. Fast systems are all the remaining ones. Using these
parameters, BFQ computes the actual duration of the weight raising,
for the system at hand, as a function of the relative speed of the
system w.r.t. the speed of a reference system, belonging to the same
class of systems as the system at hand.

This slow vs fast differentiation proved to be useful in the past, but
happens to have little meaning with current hardware. Even worse, it
does cause problems in virtual systems, where the speed of the system
can vary frequently, and so widely to just confuse the class-detection
mechanism, and, as we have verified experimentally, to cause BFQ to
compute non-sensical weight-raising durations.

This commit addresses this issue by removing the slow class and the
class-detection mechanism.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock, bfq: add description of weight-raising heuristics
Paolo Valente [Thu, 31 May 2018 14:45:05 +0000 (16:45 +0200)]
block, bfq: add description of weight-raising heuristics

A description of how weight raising works is missing in BFQ
sources. In addition, the code for handling weight raising is
scattered across a few functions. This makes it rather hard to
understand the mechanism and its rationale. This commits adds such a
description at the beginning of the main source file.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock, bfq: remove the removal of 'next' rq in bfq_requests_merged
Filippo Muzzini [Thu, 31 May 2018 13:23:13 +0000 (15:23 +0200)]
block, bfq: remove the removal of 'next' rq in bfq_requests_merged

Since bfq_finish_request() is always called on the request 'next',
after bfq_requests_merged() is finished, and bfq_finish_request()
removes 'next' from its bfq_queue if needed, it isn't necessary to do
such a removal in advance in bfq_merged_requests().

This commit removes such a useless 'next' removal.

Signed-off-by: Filippo Muzzini <filippo.muzzini@outlook.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock, bfq: remove wrong check in bfq_requests_merged
Paolo Valente [Thu, 31 May 2018 14:48:05 +0000 (08:48 -0600)]
block, bfq: remove wrong check in bfq_requests_merged

The request rq passed to the function bfq_requests_merged is always in
a bfq_queue, so the check !RB_EMPTY_NODE(&rq->rb_node) at the
beginning of bfq_requests_merged always succeeds, and the control
flow systematically skips to the end of the function.  This implies
that the body of the function is never executed, i.e., the
repositioning of rq is never performed.

On the opposite end, a control is missing in the body of the function:
'next' must be removed only if it is inside a bfq_queue.

This commit removes the wrong check on rq, and adds the missing check
on 'next'. In addition, this commit adds comments on
bfq_requests_merged.

Signed-off-by: Filippo Muzzini <filippo.muzzini@outlook.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock, bfq: remove wrong lock in bfq_requests_merged
Filippo Muzzini [Thu, 31 May 2018 13:23:11 +0000 (15:23 +0200)]
block, bfq: remove wrong lock in bfq_requests_merged

In bfq_requests_merged(), there is a deadlock because the lock on
bfqq->bfqd->lock is held by the calling function, but the code of
this function tries to grab the lock again.

This deadlock is currently hidden by another bug (fixed by next commit
for this source file), which causes the body of bfq_requests_merged()
to be never executed.

This commit removes the deadlock by removing the lock/unlock pair.

Signed-off-by: Filippo Muzzini <filippo.muzzini@outlook.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: fixup bioset_integrity_create() call
Jens Axboe [Thu, 31 May 2018 00:51:21 +0000 (18:51 -0600)]
block: fixup bioset_integrity_create() call

Missed converting the bioset_integrity_create() bounce bio set
call.

Fixes: 338aa96d5661 ("block: convert bounce, q->bio_split to bioset_init()/mempool_init()")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Drop bioset_create()
Kent Overstreet [Sun, 20 May 2018 22:25:58 +0000 (18:25 -0400)]
block: Drop bioset_create()

All users have been converted to bioset_init(), kill off the
old API.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoxfs: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:57 +0000 (18:25 -0400)]
xfs: convert to bioset_init()/mempool_init()

Convert XFS to embedded bio sets.

Acked-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agobtrfs: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:56 +0000 (18:25 -0400)]
btrfs: convert to bioset_init()/mempool_init()

Convert btrfs to embedded bio sets.

Acked-by: Chris Mason <clm@fb.com>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agofs: convert block_dev.c to bioset_init()
Kent Overstreet [Sun, 20 May 2018 22:25:55 +0000 (18:25 -0400)]
fs: convert block_dev.c to bioset_init()

Convert block DIO code to embedded bio sets.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agotarget: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:54 +0000 (18:25 -0400)]
target: convert to bioset_init()/mempool_init()

Convert the target code to embedded bio sets.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agodm: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:53 +0000 (18:25 -0400)]
dm: convert to bioset_init()/mempool_init()

Convert dm to embedded bio sets.

Acked-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agomd: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:52 +0000 (18:25 -0400)]
md: convert to bioset_init()/mempool_init()

Convert md to embedded bio sets.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agobcache: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:51 +0000 (18:25 -0400)]
bcache: convert to bioset_init()/mempool_init()

Convert bcache to embedded bio sets.

Reviewed-by: Coly Li <colyli@suse.de>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agolightnvm: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:50 +0000 (18:25 -0400)]
lightnvm: convert to bioset_init()/mempool_init()

Convert lightnvm to embedded bio sets.

Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agopktcdvd: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:49 +0000 (18:25 -0400)]
pktcdvd: convert to bioset_init()/mempool_init()

Convert pktcdvd to embedded bio sets.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agodrbd: convert to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:48 +0000 (18:25 -0400)]
drbd: convert to bioset_init()/mempool_init()

Convert drbd to embedded bio sets and mempools.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: convert bounce, q->bio_split to bioset_init()/mempool_init()
Kent Overstreet [Sun, 20 May 2018 22:25:47 +0000 (18:25 -0400)]
block: convert bounce, q->bio_split to bioset_init()/mempool_init()

Convert the core block functionality to embedded bio sets.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-throttle: return proper bool type to caller instead of 0/1
Chengguang Xu [Tue, 29 May 2018 10:32:44 +0000 (18:32 +0800)]
blk-throttle: return proper bool type to caller instead of 0/1

Change to return true/false only for bool type return code.

Signed-off-by: Chengguang Xu <cgxu519@gmx.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: only iterate over inflight requests in blk_mq_tagset_busy_iter
Christoph Hellwig [Wed, 30 May 2018 16:51:00 +0000 (18:51 +0200)]
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.

Acked-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: clear DISCONNECT_REQUESTED flag once disconnection occurs.
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.

Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Kevin Vigor <kvigor@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-throttle: fix potential NULL pointer dereference in throtl_select_dispatch
Liu Bo [Tue, 29 May 2018 08:29:12 +0000 (16:29 +0800)]
blk-throttle: fix potential NULL pointer dereference in throtl_select_dispatch

tg in throtl_select_dispatch is used first and then do check. Since tg
may be NULL, it has potential NULL pointer dereference risk. So fix
it.

Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>
Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: kyber: make kyber more friendly with merging
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.

Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: abstract out blk-mq-sched rq list iteration bio merge helper
Jens Axboe [Wed, 30 May 2018 07:26:07 +0000 (15:26 +0800)]
blk-mq: abstract out blk-mq-sched rq list iteration bio merge helper

No functional changes in this patch, just a prep patch for utilizing
this in an IO scheduler.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Omar Sandoval <osandov@fb.com>
6 years agoblock: remove parent device reference from struct bsg_class_device
Christoph Hellwig [Tue, 29 May 2018 06:40:23 +0000 (08:40 +0200)]
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.

Reported-by: Anatoliy Glagolev <glagolig@gmail.com>
Reviewed-by: James E.J. Bottomley <jejb@linux.vnet.ibm.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoMerge branch 'nvme-4.18-2' of git://git.infradead.org/nvme into for-4.18/block
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

6 years agoblock: don't print a message when the device went away
Christoph Hellwig [Tue, 29 May 2018 14:42:59 +0000 (16:42 +0200)]
block: don't print a message when the device went away

The information about a size change in this case just creates confusion.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: unexport check_disk_size_change
Christoph Hellwig [Tue, 29 May 2018 14:42:58 +0000 (16:42 +0200)]
block: unexport check_disk_size_change

Only used in block_dev.c and the partitions code, and it should remain
that way..

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: move ->timeout request member
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.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: simplify blk_mq_rq_timed_out
Christoph Hellwig [Tue, 29 May 2018 13:52:39 +0000 (15:52 +0200)]
blk-mq: simplify blk_mq_rq_timed_out

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: document the blk_eh_timer_return values
Christoph Hellwig [Tue, 29 May 2018 13:52:38 +0000 (15:52 +0200)]
block: document the blk_eh_timer_return values

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: remove BLK_EH_HANDLED
Christoph Hellwig [Tue, 29 May 2018 13:52:37 +0000 (15:52 +0200)]
block: remove BLK_EH_HANDLED

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agolibiscsi: don't try to bypass SCSI EH
Christoph Hellwig [Tue, 29 May 2018 13:52:36 +0000 (15:52 +0200)]
libiscsi: don't try to bypass SCSI EH

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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agommc: complete requests from ->timeout
Christoph Hellwig [Tue, 29 May 2018 13:52:35 +0000 (15:52 +0200)]
mmc: 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.

[While this keeps existing behavior it seems to mismatch the comment,
 maintainers please chime in!]

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoscsi_transport_fc: complete requests from ->timeout
Christoph Hellwig [Tue, 29 May 2018 13:52:34 +0000 (15:52 +0200)]
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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonull_blk: complete requests from ->timeout
Christoph Hellwig [Tue, 29 May 2018 13:52:33 +0000 (15:52 +0200)]
null_blk: 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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agomtip32xx: complete requests from ->timeout
Christoph Hellwig [Tue, 29 May 2018 13:52:32 +0000 (15:52 +0200)]
mtip32xx: 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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: complete requests from ->timeout
Christoph Hellwig [Tue, 29 May 2018 13:52:31 +0000 (15:52 +0200)]
nbd: 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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonvme: return BLK_EH_DONE from ->timeout
Christoph Hellwig [Tue, 29 May 2018 13:52:30 +0000 (15:52 +0200)]
nvme: return BLK_EH_DONE from ->timeout

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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: rename BLK_EH_NOT_HANDLED to BLK_EH_DONE
Christoph Hellwig [Tue, 29 May 2018 13:52:29 +0000 (15:52 +0200)]
block: rename BLK_EH_NOT_HANDLED to BLK_EH_DONE

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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: Remove generation seqeunce
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 <keith.busch@intel.com>
[hch: slight cleanups, added back submission side hctx lock, use cmpxchg
 for completions]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: Fix timeout and state order
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.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agolibata: remove ata_scsi_timed_out
Christoph Hellwig [Tue, 29 May 2018 13:52:26 +0000 (15:52 +0200)]
libata: remove ata_scsi_timed_out

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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agobcache: Replace bch_read_string_list() by __sysfs_match_string()
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().

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agobcache: Move couple of functions to sysfs.c
Andy Shevchenko [Mon, 28 May 2018 07:37:43 +0000 (15:37 +0800)]
bcache: Move couple of functions to sysfs.c

There is couple of functions that are used exclusively in sysfs.c.
Move it to there and make them static.

Besides above, it will allow further clean up.

No functional change intended.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agobcache: Move couple of string arrays to sysfs.c
Andy Shevchenko [Mon, 28 May 2018 07:37:42 +0000 (15:37 +0800)]
bcache: Move couple of string arrays to sysfs.c

There is couple of string arrays that are used exclusively in sysfs.c.
Move it to there and make them static.

Besides above, it will allow further clean up.

No functional change intended.

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agobcache: stop bcache device when backing device is offline
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.

Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Cc: Michael Lyle <mlyle@lyle.org>
Cc: Junhui Tang <tang.junhui@zte.com.cn>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonull_blk: add blocking description and remove lightnvm
Liu Bo [Fri, 25 May 2018 14:40:04 +0000 (22:40 +0800)]
null_blk: add blocking description and remove lightnvm

- The description of 'blocking' is missing in null_blk.txt

- The 'lightnvm' parameter has been removed in null_blk.c

This updates both in null_blk.txt.

Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonvme: fixup memory leak in nvme_init_identify()
Hannes Reinecke [Fri, 25 May 2018 09:06:27 +0000 (11:06 +0200)]
nvme: fixup memory leak in nvme_init_identify()

If nvme_get_effects_log() failed the 'id' buffer from the previous
nvme_identify_ctrl() call will never be freed.

Signed-off-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme: fix KASAN warning when parsing host nqn
Hannes Reinecke [Fri, 25 May 2018 09:04:03 +0000 (11:04 +0200)]
nvme: fix KASAN warning when parsing host nqn

The host nqn actually is smaller than the space reserved for it,
so we should be using strlcpy to keep KASAN happy.

Signed-off-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvmet-loop: use nr_phys_segments when map rq to sgl
Chaitanya Kulkarni [Fri, 11 May 2018 06:38:15 +0000 (02:38 -0400)]
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.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvmet-fc: increase LS buffer count per fc port
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.

Raise the LS buffer count to 256.

Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvmet: add simple file backed ns support
Chaitanya Kulkarni [Wed, 23 May 2018 04:34:39 +0000 (00:34 -0400)]
nvmet: add simple file backed ns support

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 <chaitanya.kulkarni@wdc.com>
[hch: updated changelog, fixed double req->ns lookup in bdev case]
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvmet: remove duplicate NULL initialization for req->ns
Chaitanya Kulkarni [Thu, 10 May 2018 06:46:30 +0000 (02:46 -0400)]
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.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvmet: make a few error messages more generic
Chaitanya Kulkarni [Tue, 8 May 2018 06:02:33 +0000 (02:02 -0400)]
nvmet: make a few error messages more generic

"nvmet_check_ctrl_status()" is called from admin-cmd.c along
with io-cmd.c, make the error message more generic.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-fabrics: allow duplicate connections to the discovery controller
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.

Signed-off-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-fabrics: centralize discovery controller defaults
Hannes Reinecke [Thu, 24 May 2018 14:18:15 +0000 (16:18 +0200)]
nvme-fabrics: centralize discovery controller defaults

When connecting to the discovery controller we have certain defaults
to observe, so centralize them to avoid inconsistencies due to argument
ordering.

Signed-off-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-fabrics: remove unnecessary controller subnqn validation
James Smart [Sat, 12 May 2018 00:50:22 +0000 (17:50 -0700)]
nvme-fabrics: remove unnecessary controller subnqn validation

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.

Remove the unnecessary validation.

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-fc: remove setting DNR on exception conditions
James Smart [Sat, 12 May 2018 00:50:24 +0000 (17:50 -0700)]
nvme-fc: remove setting DNR on exception conditions

Current code will set DNR if the controller is deleting or there is
an error during controller init. None of this is necessary.

Remove the code that sets DNR

Signed-off-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-rdma: stop admin queue before freeing it
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.

Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Suggested-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-pci: Fix AER reset handling
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.

Link: https://bugzilla.kernel.org/show_bug.cgi?id=199757
Fixes: cc1d5e749a2e ("nvme/pci: Sync controller reset for AER slot_reset")
Reported-by: Alex Gagniuc <mr.nuke.me@gmail.com>
Tested-by: Alex Gagniuc <mr.nuke.me@gmail.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-pci: set nvmeq->cq_vector after alloc cq/sq
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.

Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
[hch: fixed to pass a s16 and clean up the comment]
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agoblock drivers/block: Use octal not symbolic permissions
Joe Perches [Thu, 24 May 2018 19:38:59 +0000 (13:38 -0600)]
block drivers/block: Use octal not symbolic permissions

Convert the S_<FOO> symbolic permissions to their octal equivalents as
using octal and not symbolic permissions is preferred by many as more
readable.

see: https://lkml.org/lkml/2016/8/2/1945

Done with automated conversion via:
$ ./scripts/checkpatch.pl -f --types=SYMBOLIC_PERMS --fix-inplace <files...>

Miscellanea:

o Wrapped modified multi-line calls to a single line where appropriate
o Realign modified multi-line calls to open parenthesis

Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: avoid starving tag allocation after allocating process migrates
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

Cc: <stable@vger.kernel.org>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Modified commit message to make it clearer, and make it apply on
top of the 4.18 branch.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agobdi: Move cgroup bdi_writeback to a dedicated low concurrency workqueue
Tejun Heo [Wed, 23 May 2018 17:56:32 +0000 (10:56 -0700)]
bdi: Move cgroup bdi_writeback to a dedicated low concurrency workqueue

From 0aa2e9b921d6db71150633ff290199554f0842a8 Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj@kernel.org>
Date: Wed, 23 May 2018 10:29:00 -0700

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.

  [<ffffffff810ee40c>] _synchronize_rcu_expedited.constprop.72+0x2fc/0x330
  [<ffffffff810ee634>] synchronize_rcu_expedited+0x24/0x30
  [<ffffffff811ccf23>] bdi_unregister+0x53/0x290
  [<ffffffff811cd1e9>] release_bdi+0x89/0xc0
  [<ffffffff811cd645>] wb_exit+0x85/0xa0
  [<ffffffff811cdc84>] cgwb_release_workfn+0x54/0xb0
  [<ffffffff810a68d0>] process_one_work+0x150/0x410
  [<ffffffff810a71fd>] worker_thread+0x6d/0x520
  [<ffffffff810ad3dc>] kthread+0x12c/0x160
  [<ffffffff81969019>] ret_from_fork+0x29/0x40
  [<ffffffffffffffff>] 0xffffffffffffffff

The events leading to the lockup are...

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.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: set discard granularity properly
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.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonvme: host: core: fix precedence of ternary operator
Ivan Bornyakov [Wed, 23 May 2018 14:56:11 +0000 (17:56 +0300)]
nvme: host: core: fix precedence of ternary operator

Ternary operator have lower precedence then bitwise or, so 'cdw10' was
calculated wrong.

Signed-off-by: Ivan Bornyakov <brnkv.i1@gmail.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
6 years agonvme: fix lockdep warning in nvme_mpath_clear_current_path
Johannes Thumshirn [Thu, 17 May 2018 11:52:50 +0000 (13:52 +0200)]
nvme: fix lockdep warning in nvme_mpath_clear_current_path

When running blktest's nvme/005 with a lockdep enabled kernel the test
case fails due to the following lockdep splat in dmesg:

 =============================
 WARNING: suspicious RCU usage
 4.17.0-rc5 #881 Not tainted
 -----------------------------
 drivers/nvme/host/nvme.h:457 suspicious rcu_dereference_check() usage!

 other info that might help us debug this:

 rcu_scheduler_active = 2, debug_locks = 1
 3 locks held by kworker/u32:5/1102:
  #0:         (ptrval) ((wq_completion)"nvme-wq"){+.+.}, at: process_one_work+0x152/0x5c0
  #1:         (ptrval) ((work_completion)(&ctrl->scan_work)){+.+.}, at: process_one_work+0x152/0x5c0
  #2:         (ptrval) (&subsys->lock#2){+.+.}, at: nvme_ns_remove+0x43/0x1c0 [nvme_core]

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 <jthumshirn@suse.de>
Suggested-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Keith Busch <keith.busch@intel.com>
6 years agoblkdev_report_zones_ioctl(): Use vmalloc() to allocate large buffers
Bart Van Assche [Tue, 22 May 2018 15:27:22 +0000 (08:27 -0700)]
blkdev_report_zones_ioctl(): Use vmalloc() to allocate large buffers

Avoid that complaints similar to the following appear in the kernel log
if the number of zones is sufficiently large:

  fio: page allocation failure: order:9, mode:0x140c0c0(GFP_KERNEL|__GFP_COMP|__GFP_ZERO), nodemask=(null)
  Call Trace:
  dump_stack+0x63/0x88
  warn_alloc+0xf5/0x190
  __alloc_pages_slowpath+0x8f0/0xb0d
  __alloc_pages_nodemask+0x242/0x260
  alloc_pages_current+0x6a/0xb0
  kmalloc_order+0x18/0x50
  kmalloc_order_trace+0x26/0xb0
  __kmalloc+0x20e/0x220
  blkdev_report_zones_ioctl+0xa5/0x1a0
  blkdev_ioctl+0x1ba/0x930
  block_ioctl+0x41/0x50
  do_vfs_ioctl+0xaa/0x610
  SyS_ioctl+0x79/0x90
  do_syscall_64+0x79/0x1b0
  entry_SYSCALL_64_after_hwframe+0x3d/0xa2

Fixes: 3ed05a987e0f ("blk-zoned: implement ioctls")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Shaun Tancheff <shaun.tancheff@seagate.com>
Cc: Damien Le Moal <damien.lemoal@hgst.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Martin K. Petersen <martin.petersen@oracle.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock/ndb: add WQ_UNBOUND to the knbd-recv workqueue
Dan Melnic [Mon, 18 Sep 2017 20:08:51 +0000 (13:08 -0700)]
block/ndb: add WQ_UNBOUND to the knbd-recv workqueue

Add WQ_UNBOUND to the knbd-recv workqueue so we're not bound
to a single CPU that is selected at device creation time.

Signed-off-by: Dan Melnic <dmm@fb.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: remove wrong 'unlikely' check
huhai [Tue, 22 May 2018 09:39:34 +0000 (17:39 +0800)]
blk-mq: remove wrong 'unlikely' check

When dispatch_rq_from_ctx is called, in the vast majority of cases
the ctx->rq_list is not empty.

Signed-off-by: huhai <huhai@kylinos.cn>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonvme-pci: fix race between poll and IRQ completions
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:

[  560.169153] irq 630: nobody cared (try booting with the "irqpoll" option)
[  560.175988] CPU: 40 PID: 0 Comm: swapper/40 Not tainted 4.17.0-rc2+ #65
[  560.175990] Hardware name: Intel Corporation S2600STB/S2600STB, BIOS SE5C620.86B.00.01.0010.010920180151 01/09/2018
[  560.175991] Call Trace:
[  560.175994]  <IRQ>
[  560.176005]  dump_stack+0x5c/0x7b
[  560.176010]  __report_bad_irq+0x30/0xc0
[  560.176013]  note_interrupt+0x235/0x280
[  560.176020]  handle_irq_event_percpu+0x51/0x70
[  560.176023]  handle_irq_event+0x27/0x50
[  560.176026]  handle_edge_irq+0x6d/0x180
[  560.176031]  handle_irq+0xa5/0x110
[  560.176036]  do_IRQ+0x41/0xc0
[  560.176042]  common_interrupt+0xf/0xf
[  560.176043]  </IRQ>
[  560.176050] RIP: 0010:cpuidle_enter_state+0x9b/0x2b0
[  560.176052] RSP: 0018:ffffa0ed4659fe98 EFLAGS: 00000246 ORIG_RAX: ffffffffffffffdd
[  560.176055] RAX: ffff9527beb20a80 RBX: 000000826caee491 RCX: 000000000000001f
[  560.176056] RDX: 000000826caee491 RSI: 00000000335206ee RDI: 0000000000000000
[  560.176057] RBP: 0000000000000001 R08: 00000000ffffffff R09: 0000000000000008
[  560.176059] R10: ffffa0ed4659fe78 R11: 0000000000000001 R12: ffff9527beb29358
[  560.176060] R13: ffffffffa235d4b8 R14: 0000000000000000 R15: 000000826caed593
[  560.176065]  ? cpuidle_enter_state+0x8b/0x2b0
[  560.176071]  do_idle+0x1f4/0x260
[  560.176075]  cpu_startup_entry+0x6f/0x80
[  560.176080]  start_secondary+0x184/0x1d0
[  560.176085]  secondary_startup_64+0xa5/0xb0
[  560.176088] handlers:
[  560.178387] [<00000000efb612be>] nvme_irq [nvme]
[  560.183019] Disabling IRQ #630

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.

Fixes: 5cb525c8315f ("nvme-pci: handle completions outside of the queue lock")
Reported-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Tested-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoMerge branch 'nvme-4.18' of git://git.infradead.org/nvme into for-4.18/block
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

6 years agonvme-pci: drop IRQ disabling on submission queue lock
Jens Axboe [Thu, 17 May 2018 16:31:52 +0000 (18:31 +0200)]
nvme-pci: drop IRQ disabling on submission queue lock

Since we aren't sharing the lock for completions now, we don't
have to make it IRQ safe.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-pci: split the nvme queue lock into submission and completion locks
Jens Axboe [Thu, 17 May 2018 16:31:51 +0000 (18:31 +0200)]
nvme-pci: split the nvme queue lock into submission and completion locks

This is now feasible. We protect the submission queue ring with
->sq_lock, and the completion side with ->cq_lock.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-pci: handle completions outside of the queue lock
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
<keith.busch@intel.com> 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.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Keith Busch <keith.busch@intel.com>
[hch: refactored, restored poll early exit optimization]
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-pci: move ->cq_vector == -1 check outside of ->q_lock
Jens Axboe [Thu, 17 May 2018 16:31:49 +0000 (18:31 +0200)]
nvme-pci: move ->cq_vector == -1 check outside of ->q_lock

We only clear it dynamically in nvme_suspend_queue(). When we do, ensure
to do a full flush so that any nvme_queue_rq() invocation will see it.

Ideally we'd kill this check completely, but we're using it to flush
requests on a dying queue.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-pci: remove cq check after submission
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.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agonvme-pci: simplify nvme_cqe_valid
Christoph Hellwig [Fri, 18 May 2018 14:37:04 +0000 (08:37 -0600)]
nvme-pci: simplify nvme_cqe_valid

We always look at the current CQ head and phase, so don't pass these
as separate arguments, and rename the function to nvme_cqe_pending.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonvme: mark the result argument to nvme_complete_async_event volatile
Christoph Hellwig [Thu, 17 May 2018 16:31:46 +0000 (18:31 +0200)]
nvme: mark the result argument to nvme_complete_async_event volatile

We'll need that in the PCIe driver soon as we'll read it straight off the
CQ.

Signed-off-by: Christoph Hellwig <hch@lst.de>
6 years agoblk-mq: clear hctx->dispatch_from when mappings change
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 <huhai@kylinos.cn>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
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.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: call nbd_bdev_reset instead of bd_set_size on disconnect
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.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: fix how we set bd_invalidated
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.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: clear_sock on netlink disconnect
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.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: use bd_set_size when updating disk size
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.

Fixes: 29eaadc ("nbd: stop using the bdev everywhere")
cc: stable@vger.kernel.org
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: update size when connected
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.

cc: stable@vger.kernel.org
Fixes: 639812a ("nbd: don't set the device size until we're connected")
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agonbd: fix nbd device deletion
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.

Fixes: c6a4759ea0c9 ("nbd: add device refcounting")
cc: stable@vger.kernel.org
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: fix MAINTAINERS email for nbd
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.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblk-mq: remove redundant insert case in blk_mq_make_request()
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.

Signed-off-by: huhai <huhai@kylinos.cn>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoRemove jsflash driver
Jens Axboe [Tue, 15 May 2018 19:54:11 +0000 (13:54 -0600)]
Remove jsflash driver

Nobody is using it anymore, and it's been abandoned. Since David
is fine with removing it, kill it.

Suggested-by: Christoph Hellwig <hch@lst.de>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Add sysfs entry for fua support
Kent Overstreet [Wed, 9 May 2018 01:33:58 +0000 (21:33 -0400)]
block: Add sysfs entry for fua support

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Export bio check/set pages_dirty
Kent Overstreet [Wed, 9 May 2018 01:33:57 +0000 (21:33 -0400)]
block: Export bio check/set pages_dirty

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Add warning for bi_next not NULL in bio_endio()
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.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Add missing flush_dcache_page() call
Kent Overstreet [Wed, 9 May 2018 01:33:55 +0000 (21:33 -0400)]
block: Add missing flush_dcache_page() call

Since a bio can point to userspace pages (e.g. direct IO), this is
generally necessary.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Split out bio_list_copy_data()
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.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Add bio_copy_data_iter(), zero_fill_bio_iter()
Kent Overstreet [Wed, 9 May 2018 01:33:53 +0000 (21:33 -0400)]
block: Add bio_copy_data_iter(), zero_fill_bio_iter()

Add versions that take bvec_iter args instead of using bio->bi_iter - to
be used by bcachefs.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
6 years agoblock: Use bioset_init() for fs_bio_set
Kent Overstreet [Wed, 9 May 2018 01:33:52 +0000 (21:33 -0400)]
block: Use bioset_init() for fs_bio_set

Minor optimization - remove a pointer indirection when using fs_bio_set.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>