Age | Commit message (Collapse) | Author |
|
This reverts commit 3a366e614d0837d9fc23f78cdb1a1186ebc3387f.
Wanlong Gao reports that it causes a kernel panic on his machine several
minutes after boot. Reverting it removes the panic.
Jens says:
"It's not quite clear why that is yet, so I think we should just revert
the commit for 3.9 final (which I'm assuming is pretty close).
The wifi is crap at the LSF hotel, so sending this email instead of
queueing up a revert and pull request."
Reported-by: Wanlong Gao <gaowanlong@cn.fujitsu.com>
Requested-by: Jens Axboe <axboe@kernel.dk>
Cc: Tejun Heo <tj@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This reverts commit 8761a3dc1f07b163414e2215a2cadbb4cfe2a107.
There are situations where the destruction path is called
with the bdev->bd_mutex already held, which then deadlocks in
loop_clr_fd(). The normal partition cleanup does a trylock()
on the mutex, but it'd be nice to have a more bullet proof
method in loop. So punt this more involved fix to the next
merge window, and just back out this buggy fix for now.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
As found by gcc-4.8, the QUEUE_SYSFS_BIT_FNS macro creates functions
that use a value generated by queue_var_store independent of whether
that value was set or not.
block/blk-sysfs.c: In function 'queue_store_nonrot':
block/blk-sysfs.c:244:385: warning: 'val' may be used uninitialized in this function [-Wmaybe-uninitialized]
Unlike most other such warnings, this one is not a false positive,
writing any non-number string into the sysfs files indeed has
an undefined result, rather than returning an error.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Fixed code indent should use tabs where possible.
Signed-off-by: Alice Ferrazzi <alice.ferrazzi@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Any partitions added by user space to the loop device were being
left in place after detaching the loop device. This was because
the detach path issued a BLKRRPART to clean up partitions if
LO_FLAGS_PARTSCAN was set, meaning that the partitions were auto
scanned on attach. Replace this BLKRRPART with code that
unconditionally cleans up partitions on detach instead.
Signed-off-by: Phillip Susi <psusi@ubuntu.com>
Modified by Jens to export delete_partition().
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Pull block IO core bits from Jens Axboe:
"Below are the core block IO bits for 3.9. It was delayed a few days
since my workstation kept crashing every 2-8h after pulling it into
current -git, but turns out it is a bug in the new pstate code (divide
by zero, will report separately). In any case, it contains:
- The big cfq/blkcg update from Tejun and and Vivek.
- Additional block and writeback tracepoints from Tejun.
- Improvement of the should sort (based on queues) logic in the plug
flushing.
- _io() variants of the wait_for_completion() interface, using
io_schedule() instead of schedule() to contribute to io wait
properly.
- Various little fixes.
You'll get two trivial merge conflicts, which should be easy enough to
fix up"
Fix up the trivial conflicts due to hlist traversal cleanups (commit
b67bfe0d42ca: "hlist: drop the node parameter from iterators").
* 'for-3.9/core' of git://git.kernel.dk/linux-block: (39 commits)
block: remove redundant check to bd_openers()
block: use i_size_write() in bd_set_size()
cfq: fix lock imbalance with failed allocations
drivers/block/swim3.c: fix null pointer dereference
block: don't select PERCPU_RWSEM
block: account iowait time when waiting for completion of IO request
sched: add wait_for_completion_io[_timeout]
writeback: add more tracepoints
block: add block_{touch|dirty}_buffer tracepoint
buffer: make touch_buffer() an exported function
block: add @req to bio_{front|back}_merge tracepoints
block: add missing block_bio_complete() tracepoint
block: Remove should_sort judgement when flush blk_plug
block,elevator: use new hashtable implementation
cfq-iosched: add hierarchical cfq_group statistics
cfq-iosched: collect stats from dead cfqgs
cfq-iosched: separate out cfqg_stats_reset() from cfq_pd_reset_stats()
blkcg: make blkcg_print_blkgs() grab q locks instead of blkcg lock
block: RCU free request_queue
blkcg: implement blkg_[rw]stat_recursive_sum() and blkg_[rw]stat_merge()
...
|
|
I'm not sure why, but the hlist for each entry iterators were conceived
list_for_each_entry(pos, head, member)
The hlist ones were greedy and wanted an extra parameter:
hlist_for_each_entry(tpos, pos, head, member)
Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.
Besides the semantic patch, there was some manual work required:
- Fix up the actual hlist iterators in linux/list.h
- Fix up the declaration of other iterators based on the hlist ones.
- A very small amount of places were using the 'node' parameter, this
was modified to use 'obj->member' instead.
- Coccinelle didn't handle the hlist_for_each_entry_safe iterator
properly, so those had to be fixed up manually.
The semantic patch which is mostly the work of Peter Senna Tschudin is here:
@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;
type T;
expression a,c,d,e;
identifier b;
statement S;
@@
-T b;
<+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
...+>
[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: Peter Senna Tschudin <peter.senna@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Currently, sizeof(struct parsed_partitions) may be 64KB in 32bit arch, so
it is easy to trigger page allocation failure by check_partition,
especially in hotplug block device situation(such as, USB mass storage,
MMC card, ...), and Felipe Balbi has observed the failure.
This patch does below optimizations on the allocation of struct
parsed_partitions to try to address the issue:
- make parsed_partitions.parts as pointer so that the pointed memory can
fit in 32KB buffer, then approximate 32KB memory can be saved
- vmalloc the buffer pointed by parsed_partitions.parts because 32KB is
still a bit big for kmalloc
- given that many devices have the partition count limit, so only
allocate disk_max_parts() partitions instead of 256 partitions always
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Reported-by: Felipe Balbi <balbi@ti.com>
Cc: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
It isn't necessary to read the information of partitions whose number is
equal and more than state->limit since only maximum state->limit
partitions will be added inside rescan_partitions().
That is also what other kind of partitions are doing.
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
the structure.
UEFI 2.3.1D will include a change to the spec language mandating that a
GPT header must be greater than *or equal to* the size of the defined
structure. While verifying that this would work on Linux, I discovered
that we're not actually checking the minimum bound at all.
The result of this is that when we verify the checksum, it's possible that
on a malformed header (with header_size of 0), we won't actually verify
any data.
[akpm@linux-foundation.org: fix printk warning]
Signed-off-by: Peter Jones <pjones@redhat.com>
Acked-by: Matt Fleming <matt.fleming@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Stephen Warren <swarren@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
AIX formatted disks do not always have the MSDOS 55aa signature.
This happens e.g. for unbootable AIX disks.
Up to now, such disks were not recognized as AIX disks, because of the
missing 55aa. Fix that by inverting the two tests. Let's first
check for the AIX magic strings, and only if that fails check for
the MSDOS magic word.
Signed-off-by: Philippe De Muyter <phdm@macqel.be>
Cc: Andreas Mohr <andi@lisas.de>
Cc: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Olaf Hering <olh@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Convert to the much saner new idr interface. Both bsg and genhd
protect idr w/ mutex making preloading unnecessary.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
idr allocation in blk_alloc_devt() wasn't synchronized against lookup
and removal, and its limit check was off by one - 1 << MINORBITS is
the number of minors allowed, not the maximum allowed minor.
Add locking and rename MAX_EXT_DEVT to NR_EXT_DEVT and fix limit
checking.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Jens Axboe <axboe@kernel.dk>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
While adding and removing a lot of disks disks and partitions this
sometimes shows up:
WARNING: at fs/sysfs/dir.c:512 sysfs_add_one+0xc9/0x130() (Not tainted)
Hardware name:
sysfs: cannot create duplicate filename '/dev/block/259:751'
Modules linked in: raid1 autofs4 bnx2fc cnic uio fcoe libfcoe libfc 8021q scsi_transport_fc scsi_tgt garp stp llc sunrpc cpufreq_ondemand powernow_k8 freq_table mperf ipv6 dm_mirror dm_region_hash dm_log power_meter microcode dcdbas serio_raw amd64_edac_mod edac_core edac_mce_amd i2c_piix4 i2c_core k10temp bnx2 sg ixgbe dca mdio ext4 mbcache jbd2 dm_round_robin sr_mod cdrom sd_mod crc_t10dif ata_generic pata_acpi pata_atiixp ahci mptsas mptscsih mptbase scsi_transport_sas dm_multipath dm_mod [last unloaded: scsi_wait_scan]
Pid: 44103, comm: async/16 Not tainted 2.6.32-195.el6.x86_64 #1
Call Trace:
warn_slowpath_common+0x87/0xc0
warn_slowpath_fmt+0x46/0x50
sysfs_add_one+0xc9/0x130
sysfs_do_create_link+0x12b/0x170
sysfs_create_link+0x13/0x20
device_add+0x317/0x650
idr_get_new+0x13/0x50
add_partition+0x21c/0x390
rescan_partitions+0x32b/0x470
sd_open+0x81/0x1f0 [sd_mod]
__blkdev_get+0x1b6/0x3c0
blkdev_get+0x10/0x20
register_disk+0x155/0x170
add_disk+0xa6/0x160
sd_probe_async+0x13b/0x210 [sd_mod]
add_wait_queue+0x46/0x60
async_thread+0x102/0x250
default_wake_function+0x0/0x20
async_thread+0x0/0x250
kthread+0x96/0xa0
child_rip+0xa/0x20
kthread+0x0/0xa0
child_rip+0x0/0x20
This most likely happens because dev_t is freed while the number is
still used and idr_get_new() is not protected on every use. The fix
adds a mutex where it wasn't before and moves the dev_t free function so
it is called after device del.
Signed-off-by: Tomas Henzl <thenzl@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Apply the introduced pm_runtime_set_memalloc_noio on block device so
that PM core will teach mm to not allocate memory with GFP_IOFS when
calling the runtime_resume and runtime_suspend callback for block
devices and its ancestors.
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Oliver Neukum <oneukum@suse.de>
Cc: Jiri Kosina <jiri.kosina@suse.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Greg KH <greg@kroah.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Decotigny <david.decotigny@google.com>
Cc: Tom Herbert <therbert@google.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
While stress-running very-small container scenarios with the Kernel Memory
Controller, I've run into a lockdep-detected lock imbalance in
cfq-iosched.c.
I'll apologize beforehand for not posting a backlog: I didn't anticipate
it would be so hard to reproduce, so I didn't save my serial output and
went directly on debugging. Turns out that it did not happen again in
more than 20 runs, making it a quite rare pattern.
But here is my analysis:
When we are in very low-memory situations, we will arrive at
cfq_find_alloc_queue and may not find a queue, having to resort to the oom
queue, in an rcu-locked condition:
if (!cfqq || cfqq == &cfqd->oom_cfqq)
[ ... ]
Next, we will release the rcu lock, and try to allocate a queue, retrying
if we succeed:
rcu_read_unlock();
spin_unlock_irq(cfqd->queue->queue_lock);
new_cfqq = kmem_cache_alloc_node(cfq_pool,
gfp_mask | __GFP_ZERO,
cfqd->queue->node);
spin_lock_irq(cfqd->queue->queue_lock);
if (new_cfqq)
goto retry;
We are unlocked at this point, but it should be fine, since we will
reacquire the rcu_read_lock when we retry.
Except of course, that we may not retry: the allocation may very well fail
and we'll keep on going through the flow:
The next branch is:
if (cfqq) {
[ ... ]
} else
cfqq = &cfqd->oom_cfqq;
And right before exiting, we'll issue rcu_read_unlock().
Being already unlocked, this is the likely source of our imbalance. Since
cfqq is either already NULL or made NULL in the first statement of the
outter branch, the only viable alternative here seems to be to return the
oom queue right away in case of allocation failure.
Please review the following patch and apply if you agree with my analysis.
Signed-off-by: Glauber Costa <glommer@parallels.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
The block device doesn't use percpu rw-semaphore anymore, so don't select
it for compilation.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
This provides a band-aid to provide stable page writes on jbd without
needing to backport the fixed locking and page writeback bit handling
schemes of jbd2. The band-aid works by using bounce buffers to snapshot
page contents instead of waiting.
For those wondering about the ext3 bandage -- fixing the jbd locking
(which was done as part of ext4dev years ago) is a lot of surgery, and
setting PG_writeback on data pages when we actually hold the page lock
dropped ext3 performance by nearly an order of magnitude. If we're
going to migrate iscsi and raid to use stable page writes, the
complaints about high latency will likely return. We might as well
centralize their page snapshotting thing to one place.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Tested-by: Andy Lutomirski <luto@amacapital.net>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Artem Bityutskiy <dedekind1@gmail.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Mark Fasheh <mfasheh@suse.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Eric Van Hensbergen <ericvh@gmail.com>
Cc: Ron Minnich <rminnich@sandia.gov>
Cc: Latchesar Ionkov <lucho@ionkov.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
This patchset ("stable page writes, part 2") makes some key
modifications to the original 'stable page writes' patchset. First, it
provides creators (devices and filesystems) of a backing_dev_info a flag
that declares whether or not it is necessary to ensure that page
contents cannot change during writeout. It is no longer assumed that
this is true of all devices (which was never true anyway). Second, the
flag is used to relaxed the wait_on_page_writeback calls so that wait
only occurs if the device needs it. Third, it fixes up the remaining
disk-backed filesystems to use this improved conditional-wait logic to
provide stable page writes on those filesystems.
It is hoped that (for people not using checksumming devices, anyway)
this patchset will give back unnecessary performance decreases since the
original stable page write patchset went into 3.0. Sorry about not
fixing it sooner.
Complaints were registered by several people about the long write
latencies introduced by the original stable page write patchset.
Generally speaking, the kernel ought to allocate as little extra memory
as possible to facilitate writeout, but for people who simply cannot
wait, a second page stability strategy is (re)introduced: snapshotting
page contents. The waiting behavior is still the default strategy; to
enable page snapshotting, a superblock flag (MS_SNAP_STABLE) must be
set. This flag is used to bandaid^Henable stable page writeback on
ext3[1], and is not used anywhere else.
Given that there are already a few storage devices and network FSes that
have rolled their own page stability wait/page snapshot code, it would
be nice to move towards consolidating all of these. It seems possible
that iscsi and raid5 may wish to use the new stable page write support
to enable zero-copy writeout.
Thank you to Jan Kara for helping fix a couple more filesystems.
Per Andrew Morton's request, here are the result of using dbench to measure
latencies on ext2:
3.8.0-rc3:
Operation Count AvgLat MaxLat
----------------------------------------
WriteX 109347 0.028 59.817
ReadX 347180 0.004 3.391
Flush 15514 29.828 287.283
Throughput 57.429 MB/sec 4 clients 4 procs max_latency=287.290 ms
3.8.0-rc3 + patches:
WriteX 105556 0.029 4.273
ReadX 335004 0.005 4.112
Flush 14982 30.540 298.634
Throughput 55.4496 MB/sec 4 clients 4 procs max_latency=298.650 ms
As you can see, for ext2 the maximum write latency decreases from ~60ms
on a laptop hard disk to ~4ms. I'm not sure why the flush latencies
increase, though I suspect that being able to dirty pages faster gives
the flusher more work to do.
On ext4, the average write latency decreases as well as all the maximum
latencies:
3.8.0-rc3:
WriteX 85624 0.152 33.078
ReadX 272090 0.010 61.210
Flush 12129 36.219 168.260
Throughput 44.8618 MB/sec 4 clients 4 procs max_latency=168.276 ms
3.8.0-rc3 + patches:
WriteX 86082 0.141 30.928
ReadX 273358 0.010 36.124
Flush 12214 34.800 165.689
Throughput 44.9941 MB/sec 4 clients 4 procs max_latency=165.722 ms
XFS seems to exhibit similar latency improvements as ext2:
3.8.0-rc3:
WriteX 125739 0.028 104.343
ReadX 399070 0.005 4.115
Flush 17851 25.004 131.390
Throughput 66.0024 MB/sec 4 clients 4 procs max_latency=131.406 ms
3.8.0-rc3 + patches:
WriteX 123529 0.028 6.299
ReadX 392434 0.005 4.287
Flush 17549 25.120 188.687
Throughput 64.9113 MB/sec 4 clients 4 procs max_latency=188.704 ms
...and btrfs, just to round things out, also shows some latency
decreases:
3.8.0-rc3:
WriteX 67122 0.083 82.355
ReadX 212719 0.005 2.828
Flush 9547 47.561 147.418
Throughput 35.3391 MB/sec 4 clients 4 procs max_latency=147.433 ms
3.8.0-rc3 + patches:
WriteX 64898 0.101 71.631
ReadX 206673 0.005 7.123
Flush 9190 47.963 219.034
Throughput 34.0795 MB/sec 4 clients 4 procs max_latency=219.044 ms
Before this patchset, all filesystems would block, regardless of whether
or not it was necessary. ext3 would wait, but still generate occasional
checksum errors. The network filesystems were left to do their own
thing, so they'd wait too.
After this patchset, all the disk filesystems except ext3 and btrfs will
wait only if the hardware requires it. ext3 (if necessary) snapshots
pages instead of blocking, and btrfs provides its own bdi so the mm will
never wait. Network filesystems haven't been touched, so either they
provide their own wait code, or they don't block at all. The blocking
behavior is back to what it was before 3.0 if you don't have a disk
requiring stable page writes.
This patchset has been tested on 3.8.0-rc3 on x64 with ext3, ext4, and
xfs. I've spot-checked 3.8.0-rc4 and seem to be getting the same
results as -rc3.
[1] The alternative fixes to ext3 include fixing the locking order and
page bit handling like we did for ext4 (but then why not just use
ext4?), or setting PG_writeback so early that ext3 becomes extremely
slow. I tried that, but the number of write()s I could initiate dropped
by nearly an order of magnitude. That was a bit much even for the
author of the stable page series! :)
This patch:
Creates a per-backing-device flag that tracks whether or not pages must
be held immutable during writeout. Eventually it will be used to waive
wait_for_page_writeback() if nothing requires stable pages.
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Artem Bityutskiy <dedekind1@gmail.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Mark Fasheh <mfasheh@suse.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Eric Van Hensbergen <ericvh@gmail.com>
Cc: Ron Minnich <rminnich@sandia.gov>
Cc: Latchesar Ionkov <lucho@ionkov.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq
Pull async changes from Tejun Heo:
"These are followups for the earlier deadlock issue involving async
ending up waiting for itself through block requesting module[1]. The
following changes are made by these commits.
- Instead of requesting default elevator on each request_queue init,
block now requests it once early during boot.
- Kmod triggers warning if invoked from an async worker.
- Async synchronization implementation has been reimplemented. It's
a lot simpler now."
* 'for-3.9-async' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
async: initialise list heads to fix crash
async: replace list of active domains with global list of pending items
async: keep pending tasks on async_domain and remove async_pending
async: use ULLONG_MAX for infinity cookie value
async: bring sanity to the use of words domain and running
async, kmod: warn on synchronous request_module() from async workers
block: don't request module during elevator init
init, block: try to load default elevator module early during boot
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull scheduler changes from Ingo Molnar:
"Main changes:
- scheduler side full-dynticks (user-space execution is undisturbed
and receives no timer IRQs) preparation changes that convert the
cputime accounting code to be full-dynticks ready, from Frederic
Weisbecker.
- Initial sched.h split-up changes, by Clark Williams
- select_idle_sibling() performance improvement by Mike Galbraith:
" 1 tbench pair (worst case) in a 10 core + SMT package:
pre 15.22 MB/sec 1 procs
post 252.01 MB/sec 1 procs "
- sched_rr_get_interval() ABI fix/change. We think this detail is not
used by apps (so it's not an ABI in practice), but lets keep it
under observation.
- misc RT scheduling cleanups, optimizations"
* 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (24 commits)
sched/rt: Add <linux/sched/rt.h> header to <linux/init_task.h>
cputime: Remove irqsave from seqlock readers
sched, powerpc: Fix sched.h split-up build failure
cputime: Restore CPU_ACCOUNTING config defaults for PPC64
sched/rt: Move rt specific bits into new header file
sched/rt: Add a tuning knob to allow changing SCHED_RR timeslice
sched: Move sched.h sysctl bits into separate header
sched: Fix signedness bug in yield_to()
sched: Fix select_idle_sibling() bouncing cow syndrome
sched/rt: Further simplify pick_rt_task()
sched/rt: Do not account zero delta_exec in update_curr_rt()
cputime: Safely read cputime of full dynticks CPUs
kvm: Prepare to add generic guest entry/exit callbacks
cputime: Use accessors to read task cputime stats
cputime: Allow dynamic switch between tick/virtual based cputime accounting
cputime: Generic on-demand virtual cputime accounting
cputime: Move default nsecs_to_cputime() to jiffies based cputime file
cputime: Librarize per nsecs resolution cputime definitions
cputime: Avoid multiplication overflow on utime scaling
context_tracking: Export context state for generic vtime
...
Fix up conflict in kernel/context_tracking.c due to comment additions.
|
|
Using wait_for_completion() for waiting for a IO request to be executed
results in wrong iowait time accounting. For example, a system having
the only task doing write() and fdatasync() on a block device can be
reported being idle instead of iowaiting as it should because
blkdev_issue_flush() calls wait_for_completion() which in turn calls
schedule() that does not increment the iowait proc counter and thus does
not turn on iowait time accounting.
The patch makes block layer use wait_for_completion_io() instead of
wait_for_completion() where appropriate to account iowait time
correctly.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Move the sysctl-related bits from include/linux/sched.h into
a new file: include/linux/sched/sysctl.h. Then update source
files requiring access to those bits by including the new
header file.
Signed-off-by: Clark Williams <williams@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Link: http://lkml.kernel.org/r/20130207094659.06dced96@riff.lan
Signed-off-by: Ingo Molnar <mingo@kernel.org>
|
|
Block layer allows selecting an elevator which is built as a module to
be selected as system default via kernel param "elevator=". This is
achieved by automatically invoking request_module() whenever a new
block device is initialized and the elevator is not available.
This led to an interesting deadlock problem involving async and module
init. Block device probing running off an async job invokes
request_module(). While the module is being loaded, it performs
async_synchronize_full() which ends up waiting for the async job which
is already waiting for request_module() to finish, leading to
deadlock.
Invoking request_module() from deep in block device init path is
already nasty in itself. It seems best to avoid these situations from
the beginning by moving on-demand module loading out of block init
path.
The previous patch made sure that the default elevator module is
loaded early during boot if available. This patch removes on-demand
loading of the default elevator from elevator init path. As the
module would have been loaded during boot, userland-visible behavior
difference should be minimal.
For more details, please refer to the following thread.
http://thread.gmane.org/gmane.linux.kernel/1420814
v2: The bool parameter was named @request_module which conflicted with
request_module(). This built okay w/ CONFIG_MODULES because
request_module() was defined as a macro. W/o CONFIG_MODULES, it
causes build breakage. Rename the parameter to @try_loading.
Reported by Fengguang.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Alex Riesen <raa.lkml@gmail.com>
Cc: Fengguang Wu <fengguang.wu@intel.com>
|
|
This patch adds default module loading and uses it to load the default
block elevator. During boot, it's called right after initramfs or
initrd is made available and right before control is passed to
userland. This ensures that as long as the modules are available in
the usual places in initramfs, initrd or the root filesystem, the
default modules are loaded as soon as possible.
This will replace the on-demand elevator module loading from elevator
init path.
v2: Fixed build breakage when !CONFIG_BLOCK. Reported by kbuild test
robot.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Alex Riesen <raa.lkml@gmail.com>
Cc: Fengguang We <fengguang.wu@intel.com>
|
|
bio_{front|back}_merge tracepoints report a bio merging into an
existing request but didn't specify which request the bio is being
merged into. Add @req to it. This makes it impossible to share the
event template with block_bio_queue - split it out.
@req isn't used or exported to userland at this point and there is no
userland visible behavior change. Later changes will make use of the
extra parameter.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
bio completion didn't kick block_bio_complete TP. Only dm was
explicitly triggering the TP on IO completion. This makes
block_bio_complete TP useless for tracers which want to know about
bios, and all other bio based drivers skip generating blktrace
completion events.
This patch makes all bio completions via bio_endio() generate
block_bio_complete TP.
* Explicit trace_block_bio_complete() invocation removed from dm and
the trace point is unexported.
* @rq dropped from trace_block_bio_complete(). bios may fly around
w/o queue associated. Verifying and accessing the assocaited queue
belongs to TP probes.
* blktrace now gets both request and bio completions. Make it ignore
bio completions if request completion path is happening.
This makes all bio based drivers generate blktrace completion events
properly and makes the block_bio_complete TP actually useful.
v2: With this change, block_bio_complete TP could be invoked on sg
commands which have bio's with %NULL bi_bdev. Update TP
assignment code to check whether bio->bi_bdev is %NULL before
dereferencing.
Signed-off-by: Tejun Heo <tj@kernel.org>
Original-patch-by: Namhyung Kim <namhyung@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Alasdair Kergon <agk@redhat.com>
Cc: dm-devel@redhat.com
Cc: Neil Brown <neilb@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup into for-3.9/core
Tejun writes:
Hello, Jens.
Please consider pulling from the following branch to receive cfq blkcg
hierarchy support. The branch is based on top of v3.8-rc2.
git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup.git blkcg-cfq-hierarchy
The patchset was reviewd in the following thread.
http://thread.gmane.org/gmane.linux.kernel.cgroups/5571
|
|
In commit 975927b942c932,it add blk_rq_pos to sort rq when flushing.
Although this commit was used for the situation which blk_plug handled
multi devices on the same time like md device.
I think there must be some situations like this but only single
device.
So remove the should_sort judgement.
Because the parameter should_sort is only for this purpose,it can delete
should_sort from blk_plug.
CC: Shaohua Li <shli@kernel.org>
Signed-off-by: Jianpeng Ma <majianpeng@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Switch elevator to use the new hashtable implementation. This reduces the
amount of generic unrelated code in the elevator.
This also removes the dymanic allocation of the hash table. The size of the table is
constant so there's no point in paying the price of an extra dereference when accessing
it.
This patch depends on d9b482c ("hashtable: introduce a small and naive
hashtable") which was merged in v3.6.
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
|
|
Unfortunately, at this point, there's no way to make the existing
statistics hierarchical without creating nasty surprises for the
existing users. Just create recursive counterpart of the existing
stats.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
To support hierarchical stats, it's necessary to remember stats from
dead children. Add cfqg->dead_stats and make a dying cfqg transfer
its stats to the parent's dead-stats.
The transfer happens form ->pd_offline_fn() and it is possible that
there are some residual IOs completing afterwards. Currently, we lose
these stats. Given that cgroup removal isn't a very high frequency
operation and the amount of residual IOs on offline are likely to be
nil or small, this shouldn't be a big deal and the complexity needed
to handle residual IOs - another callback and rather elaborate
synchronization to reach and lock the matching q - doesn't seem
justified.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Separate out cfqg_stats_reset() which takes struct cfqg_stats * from
cfq_pd_reset_stats() and move the latter to where other pd methods are
defined. cfqg_stats_reset() will be used to implement hierarchical
stats.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Instead of holding blkcg->lock while walking ->blkg_list and executing
prfill(), RCU walk ->blkg_list and hold the blkg's queue lock while
executing prfill(). This makes prfill() implementations easier as
stats are mostly protected by queue lock.
This will be used to implement hierarchical stats.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
RCU free request_queue so that blkcg_gq->q can be dereferenced under
RCU lock. This will be used to implement hierarchical stats.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Implement blkg_[rw]stat_recursive_sum() and blkg_[rw]stat_merge().
The former two collect the [rw]stats designated by the target policy
data and offset from the pd's subtree. The latter two add one
[rw]stat to another.
Note that the recursive sum functions require the queue lock to be
held on entry to make blkg online test reliable. This is necessary to
properly handle stats of a dying blkg.
These will be used to implement hierarchical stats.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Hierarchical stats for cfq-iosched will need __blkg_prfill_rwstat().
Export it.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
|
|
Rename blkg_rwstat_sum() to blkg_rwstat_total(). sum will be used for
summing up stats from multiple blkgs.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Add two blkcg_policy methods, ->online_pd_fn() and ->offline_pd_fn(),
which are invoked as the policy_data gets activated and deactivated
while holding both blkcg and q locks.
Also, add blkcg_gq->online bool, which is set and cleared as the
blkcg_gq gets activated and deactivated. This flag also is toggled
while holding both blkcg and q locks.
These will be used to implement hierarchical stats.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Add pd->plid so that the policy a pd belongs to can be identified
easily. This will be used to implement hierarchical blkg_[rw]stats.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
With the previous two patches, all cfqg scheduling decisions are based
on vfraction and ready for hierarchy support. The only thing which
keeps the behavior flat is cfqg_flat_parent() which makes vfraction
calculation consider all non-root cfqgs children of the root cfqg.
Replace it with cfqg_parent() which returns the real parent. This
enables full blkcg hierarchy support for cfq-iosched. For example,
consider the following hierarchy.
root
/ \
A:500 B:250
/ \
AA:500 AB:1000
For simplicity, let's say all the leaf nodes have active tasks and are
on service tree. For each leaf node, vfraction would be
AA: (500 / 1500) * (500 / 750) =~ 0.2222
AB: (1000 / 1500) * (500 / 750) =~ 0.4444
B: (250 / 750) =~ 0.3333
and vdisktime will be distributed accordingly. For more detail,
please refer to Documentation/block/cfq-iosched.txt.
v2: cfq-iosched.txt updated to describe group scheduling as suggested
by Vivek.
v3: blkio-controller.txt updated.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
cfq_group_slice() calculates slice by taking a fraction of
cfq_target_latency according to the ratio of cfqg->weight against
service_tree->total_weight. This currently works only because all
cfqgs are treated to be at the same level.
To prepare for proper hierarchy support, convert cfq_group_slice() to
base the calculation on cfqg->vfraction. As cfqg->vfraction is always
a fraction of 1 and represents the fraction allocated to the cfqg with
hierarchy considered, the slice can be simply calculated by
multiplying cfqg->vfraction to cfq_target_latency (with fixed point
shift factored in).
As vfraction calculation currently treats all non-root cfqgs as
children of the root cfqg, this patch doesn't introduce noticeable
behavior difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Currently, cfqg charges are scaled directly according to cfqg->weight.
Regardless of the number of active cfqgs or the amount of active
weights, a given weight value always scales charge the same way. This
works fine as long as all cfqgs are treated equally regardless of
their positions in the hierarchy, which is what cfq currently
implements. It can't work in hierarchical settings because the
interpretation of a given weight value depends on where the weight is
located in the hierarchy.
This patch reimplements cfqg charge scaling so that it can be used to
support hierarchy properly. The scheme is fairly simple and
light-weight.
* When a cfqg is added to the service tree, v(disktime)weight is
calculated. It walks up the tree to root calculating the fraction
it has in the hierarchy. At each level, the fraction can be
calculated as
cfqg->weight / parent->level_weight
By compounding these, the global fraction of vdisktime the cfqg has
claim to - vfraction - can be determined.
* When the cfqg needs to be charged, the charge is scaled inversely
proportionally to the vfraction.
The new scaling scheme uses the same CFQ_SERVICE_SHIFT for fixed point
representation as before; however, the smallest scaling factor is now
1 (ie. 1 << CFQ_SERVICE_SHIFT). This is different from before where 1
was for CFQ_WEIGHT_DEFAULT and higher weight would result in smaller
scaling factor.
While this shifts the global scale of vdisktime a bit, it doesn't
change the relative relationships among cfqgs and the scheduling
result isn't different.
cfq_group_notify_queue_add uses fixed CFQ_IDLE_DELAY when appending
new cfqg to the service tree. The specific value of CFQ_IDLE_DELAY
didn't have any relevance to vdisktime before and is unlikely to cause
any visible behavior difference now especially as the scale shift
isn't that large.
As the new scheme now makes proper distinction between cfqg->weight
and ->leaf_weight, reverse the weight aliasing for root cfqgs. For
root, both weights are now mapped to ->leaf_weight instead of the
other way around.
Because we're still using cfqg_flat_parent(), this patch shouldn't
change the scheduling behavior in any noticeable way.
v2: Beefed up comments on vfraction as requested by Vivek.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
To prepare for blkcg hierarchy support, add cfqg->nr_active and
->children_weight. cfqg->nr_active counts the number of active cfqgs
at the cfqg's level and ->children_weight is sum of weights of those
cfqgs. The level covers itself (cfqg->leaf_weight) and immediate
children.
The two values are updated when a cfqg enters and leaves the group
service tree. Unless the hierarchy is very deep, the added overhead
should be negligible.
Currently, the parent is determined using cfqg_flat_parent() which
makes the root cfqg the parent of all other cfqgs. This is to make
the transition to hierarchy-aware scheduling gradual. Scheduling
logic will be converted to use cfqg->children_weight without actually
changing the behavior. When everything is ready,
blkcg_weight_parent() will be replaced with proper parent function.
This patch doesn't introduce any behavior chagne.
v2: s/cfqg->level_weight/cfqg->children_weight/ as per Vivek.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
cfq blkcg is about to grow proper hierarchy handling, where a child
blkg's weight would nest inside the parent's. This makes tasks in a
blkg to compete against both tasks in the sibling blkgs and the tasks
of child blkgs.
We're gonna use the existing weight as the group weight which decides
the blkg's weight against its siblings. This patch introduces a new
weight - leaf_weight - which decides the weight of a blkg against the
child blkgs.
It's named leaf_weight because another way to look at it is that each
internal blkg nodes have a hidden child leaf node which contains all
its tasks and leaf_weight is the weight of the leaf node and handled
the same as the weight of the child blkgs.
This patch only adds leaf_weight fields and exposes it to userland.
The new weight isn't actually used anywhere yet. Note that
cfq-iosched currently offcially supports only single level hierarchy
and root blkgs compete with the first level blkgs - ie. root weight is
basically being used as leaf_weight. For root blkgs, the two weights
are kept in sync for backward compatibility.
v2: cfqd->root_group->leaf_weight initialization was missing from
cfq_init_queue() causing divide by zero when
!CONFIG_CFQ_GROUP_SCHED. Fix it. Reported by Fengguang.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Fengguang Wu <fengguang.wu@intel.com>
|
|
Currently a child blkg (blkcg_gq) can be created even if its parent
doesn't exist. ie. Given a blkg, it's not guaranteed that its
ancestors will exist. This makes it difficult to implement proper
hierarchy support for blkcg policies.
Always create blkgs recursively and make a child blkg hold a reference
to its parent. blkg->parent is added so that finding the parent is
easy. blkcg_parent() is also added in the process.
This change can be visible to userland. e.g. while issuing IO in a
nested cgroup didn't affect the ancestors at all, now it will
initialize all ancestor blkgs and zero stats for the request_queue
will always appear on them. While this is userland visible, this
shouldn't cause any functional difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
* Rename out_* labels to err_*.
* Do ERR_PTR() conversion once in the error return path.
This patch is cosmetic and to prepare for the hierarchy support.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Reorganize such that
* __blkg_lookup() takes bool param @update_hint to determine whether
to update hint.
* __blkg_lookup_create() no longer performs lookup before trying to
create. Renamed to blkg_create().
* blkg_lookup_create() now performs lookup and then invokes
blkg_create() if lookup fails.
* root_blkg creation in blkcg_activate_policy() updated accordingly.
Note that blkcg_activate_policy() no longer updates lookup hint if
root_blkg already exists.
Except for the last lookup hint bit which is immaterial, this is pure
reorganization and doesn't introduce any visible behavior change.
This is to prepare for proper hierarchy support.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
blkg_alloc() was mistakenly checking blkcg_policy_enabled() twice.
The latter test should have been on whether pol->pd_init_fn() exists.
This doesn't cause actual problems because both blkcg policies
implement pol->pd_init_fn(). Fix it.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
|
|
Currently we attach a character "S" or "A" to the cfqq<pid>, to represent
whether queues is sync or async. Add one more character "N" to represent
whether it is sync-noidle queue or sync queue. So now three different
type of queues will look as follows.
cfq1234S --> sync queus
cfq1234SN --> sync noidle queue
cfq1234A --> Async queue
Previously S/A classification was being printed only if group scheduling
was enabled. This patch also makes sure that this classification is
displayed even if group idling is disabled.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Acked-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
|