Mempool c->fill_iter is used to allocate memory for struct btree_iter in
bch_btree_node_read_done() to iterate all keys of a read-in btree node.
The allocation size is defined in bch_cache_set_alloc() by,
mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
where iter_size is defined by a calculation,
(sb->bucket_size / sb->block_size + 1) * sizeof(struct btree_iter_set)
For 16bit width bucket_size the calculation is OK, but now the bucket
size is extended to 32bit, the bucket size can be 2GB. By the above
calculation, iter_size can be 2048 pages (order 11 is still accepted by
buddy allocator).
But the actual size holds the bkeys in meta data bucket is limited to
meta_bucket_pages() already, which is 16MB. By the above calculation,
if replace sb->bucket_size by meta_bucket_pages() * PAGE_SECTORS, the
result is 16 pages. This is the size large enough for the mempool
allocation to struct btree_iter.
Therefore in worst case every time mempool c->fill_iter allocates, at
most 4080 pages are wasted and won't be used. Therefore this patch uses
meta_bucket_pages() * PAGE_SECTORS to calculate the iter size in
bch_cache_set_alloc(), to avoid extra memory allocation from mempool
c->fill_iter.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The following three sysfs files are created to display according feature
set information of bcache:
/sys/fs/bcache/<cache set UUID>/internal/feature_compat
/sys/fs/bcache/<cache set UUID>/internal/feature_ro_compat
/sys/fs/bcache/<cache set UUID>/internal/feature_incompat
is added by this patch, to display feature sets information of the cache
set.
Now only an incompat feature 'large_bucket' added in bcache, the sysfs
file content is:
[large_bucket]
string large_bucket means the running bcache drive supports incompat
feature 'large_bucket', the wrapping [] means the 'large_bucket' feature
is currently enabled on this cache set.
This patch is ready to display compat and ro_compat features, in future
once bcache code implements such feature sets, the according feature
strings will be displayed in their sysfs files too.
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The large bucket feature is to extend bucket_size from 16bit to 32bit.
When create cache device on zoned device (e.g. zoned NVMe SSD), making
a single bucket cover one or more zones of the zoned device is the
simplest way to support zoned device as cache by bcache.
But current maximum bucket size is 16MB and a typical zone size of zoned
device is 256MB, this is the major motiviation to extend bucket size to
a larger bit width.
This patch is the basic and first change to support large bucket size,
the major changes it makes are,
- Add BCH_FEATURE_INCOMPAT_LARGE_BUCKET for the large bucket feature,
INCOMPAT means it introduces incompatible on-disk format change.
- Add BCH_FEATURE_INCOMPAT_FUNCS(large_bucket, LARGE_BUCKET) routines.
- Adds __le16 bucket_size_hi into struct cache_sb_disk at offset 0x8d0
for the on-disk super block format.
- For the in-memory super block struct cache_sb, member bucket_size is
extended from __u16 to __32.
- Add get_bucket_size() to combine the bucket_size and bucket_size_hi
from struct cache_sb_disk into an unsigned int value.
Since we already have large bucket size helpers meta_bucket_pages(),
meta_bucket_bytes() and alloc_meta_bucket_pages(), they make sure when
bucket size > 8MB, the memory allocation for bcache meta data bucket
won't fail no matter how large the bucket size extended. So these meta
data buckets are handled properly when the bucket size width increase
from 16bit to 32bit, we don't need to worry about them.
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently the bcache internal btree node occupies a whole bucket. When
loading the btree node from cache device into memory, mca_data_alloc()
will call bch_btree_keys_alloc() to allocate memory for the whole bucket
size, ilog2(b->c->btree_pages) is send to bch_btree_keys_alloc() as the
parameter 'page_order'.
c->btree_pages is set as bucket_pages() in bch_cache_set_alloc(), for
bucket size > 8MB, ilog2(b->c->btree_pages) is 12 for 4KB page size. By
default the maximum page order __get_free_pages() accepts is MAX_ORDER
(11), in this condition bch_btree_keys_alloc() will always fail.
Because of other over-page-order allocation failure fails the cache
device registration, such btree node allocation failure wasn't observed
during runtime. After other blocking page allocation failures for bucket
size > 8MB, this btree node allocation issue may trigger potentical risk
e.g. infinite dead-loop to retry btree node allocation after failure.
This patch fixes the potential problem by setting c->btree_pages to
meta_bucket_pages() in bch_cache_set_alloc(). In the condition that
bucket size > 8MB, meta_bucket_pages() will always return a number which
won't exceed the maximum page order of the buddy allocator.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In bch_btree_cache_alloc() when CONFIG_BCACHE_DEBUG is configured,
allocate memory for c->verify_ondisk may fail if the bucket size > 8MB,
which will require __get_free_pages() to allocate continuous pages
with order > 11 (the default MAX_ORDER of Linux buddy allocator). Such
over size allocation will fail, and cause 2 problems,
- When CONFIG_BCACHE_DEBUG is configured, bch_btree_verify() does not
work, because c->verify_ondisk is NULL and bch_btree_verify() returns
immediately.
- bch_btree_cache_alloc() will fail due to c->verify_ondisk allocation
failed, then the whole cache device registration fails. And because of
this failure, the first problem of bch_btree_verify() has no chance to
be triggered.
This patch fixes the above problem by two means,
1) If pages allocation of c->verify_ondisk fails, set it to NULL and
returns bch_btree_cache_alloc() with -ENOMEM.
2) When calling __get_free_pages() to allocate c->verify_ondisk pages,
use ilog2(meta_bucket_pages(&c->sb)) to make sure ilog2() will always
generate a pages order <= MAX_ORDER (or CONFIG_FORCE_MAX_ZONEORDER).
Then the buddy system won't directly reject the allocation request.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Similar to c->uuids, struct cache's prio_buckets and disk_buckets also
have the potential memory allocation failure during cache registration
if the bucket size > 8MB.
ca->prio_buckets can be stored on cache device in multiple buckets, its
in-memory space is allocated by kzalloc() interface but normally
allocated by alloc_pages() because the size > KMALLOC_MAX_CACHE_SIZE.
So allocation of ca->prio_buckets has the MAX_ORDER restriction too. If
the bucket size > 8MB, by default the page allocator will fail because
the page order > 11 (default MAX_ORDER value). ca->prio_buckets should
also use meta_bucket_bytes(), meta_bucket_pages() to decide its memory
size and use alloc_meta_bucket_pages() to allocate pages, to avoid the
allocation failure during cache set registration when bucket size > 8MB.
ca->disk_buckets is a single bucket size memory buffer, it is used to
iterate each bucket of ca->prio_buckets, and compose the bio based on
memory of ca->disk_buckets, then write ca->disk_buckets memory to cache
disk one-by-one for each bucket of ca->prio_buckets. ca->disk_buckets
should have in-memory size exact to the meta_bucket_pages(), this is the
size that ca->prio_buckets will be stored into each on-disk bucket.
This patch fixes the above issues and handle cache's prio_buckets and
disk_buckets properly for bucket size larger than 8MB.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Bcache allocates a whole bucket to store c->uuids on cache device, and
allocates continuous pages to store it in-memory. When the bucket size
exceeds maximum allocable continuous pages, bch_cache_set_alloc() will
fail and cache device registration will fail.
This patch allocates c->uuids by alloc_meta_bucket_pages(), and uses
ilog2(meta_bucket_pages(c)) to indicate order of c->uuids pages when
free it. When writing c->uuids to cache device, its size is decided
by meta_bucket_pages(c) * PAGE_SECTORS. Now c->uuids is properly handled
for bucket size > 8MB.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently the in-memory meta data like c->uuids or c->disk_buckets
are allocated by alloc_bucket_pages(). The macro alloc_bucket_pages()
calls __get_free_pages() to allocated continuous pages with order
indicated by ilog2(bucket_pages(c)),
#define alloc_bucket_pages(gfp, c) \
((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
The maximum order is defined as MAX_ORDER, the default value is 11 (and
can be overwritten by CONFIG_FORCE_MAX_ZONEORDER). In bcache code the
maximum bucket size width is 16bits, this is restricted both by KEY_SIZE
size and bucket_size size from struct cache_sb_disk. The maximum 16bits
width and power-of-2 value is (1<<15) in unit of sector (512byte). It
means the maximum value of bucket size in bytes is (1<<24) bytes a.k.a
4096 pages.
When the bucket size is set to maximum permitted value, ilog2(4096) is
12, which exceeds the default maximum order __get_free_pages() can
accepted, the failed pages allocation will fail cache set registration
procedure and print a kernel oops message for the exceeded pages order.
This patch introduces meta_bucket_pages(), meta_bucket_bytes(), and
alloc_bucket_pages() helper routines. meta_bucket_pages() indicates the
maximum pages can be allocated to meta data bucket, meta_bucket_bytes()
indicates the according maximum bytes, and alloc_bucket_pages() does
the pages allocation for meta bucket. Because meta_bucket_pages()
chooses the smaller value among the bucket size and MAX_ORDER_NR_PAGES,
it still works when MAX_ORDER overwritten by CONFIG_FORCE_MAX_ZONEORDER.
Following patches will use these helper routines to decide maximum pages
can be allocated for different meta data buckets. If the bucket size is
larger than meta_bucket_bytes(), the bcache registration can continue to
success, just the space more than meta_bucket_bytes() inside the bucket
is wasted. Comparing bcache failed for large bucket size, wasting some
space for meta data buckets is acceptable at this moment.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Setting sb->first_bucket and checking sb->keys indeed are only for cache
device, it does not make sense to do them in read_super() for backing
device too.
This patch moves the related code piece into read_super_common()
explicitly for cache device and avoid the confusion.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The new added super block version BCACHE_SB_VERSION_BDEV_WITH_FEATURES
(5) BCACHE_SB_VERSION_CDEV_WITH_FEATURES value (6), is for the feature
set bits.
Devices have super block version equal to the new version will have
three new members for feature set bits in the on-disk super block,
__le64 feature_compat;
__le64 feature_incompat;
__le64 feature_ro_compat;
They are used for further new features which may introduce on-disk
format change, and avoid unncessary super block version increase.
The very basic features handling code skeleton is also initialized in
this patch.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In register_cache_set(), c is pointer to struct cache_set, and ca is
pointer to struct cache, if ca->sb.seq > c->sb.seq, it means this
registering cache has up to date version and other members, the in-
memory version and other members should be updated to the newer value.
But current implementation makes a cache set only has a single cache
device, so the above assumption works well except for a special case.
The execption is when a cache device new created and both ca->sb.seq and
c->sb.seq are 0, because the super block is never flushed out yet. In
the location for the following if() check,
2156 if (ca->sb.seq > c->sb.seq) {
2157 c->sb.version = ca->sb.version;
2158 memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
2159 c->sb.flags = ca->sb.flags;
2160 c->sb.seq = ca->sb.seq;
2161 pr_debug("set version = %llu\n", c->sb.version);
2162 }
c->sb.version is not initialized yet and valued 0. When ca->sb.seq is 0,
the if() check will fail (because both values are 0), and the cache set
version, set_uuid, flags and seq won't be updated.
The above problem is hiden for current code, because the bucket size is
compatible among different super block version. And the next time when
running cache set again, ca->sb.seq will be larger than 0 and cache set
super block version will be updated properly.
But if the large bucket feature is enabled, sb->bucket_size is the low
16bits of the bucket size. For a power of 2 value, when the actual
bucket size exceeds 16bit width, sb->bucket_size will always be 0. Then
read_super_common() will fail because the if() check to
is_power_of_2(sb->bucket_size) is false. This is how the long time
hidden bug is triggered.
This patch modifies the if() check to the following way,
2156 if (ca->sb.seq > c->sb.seq || c->sb.seq == 0) {
Then cache set's version, set_uuid, flags and seq will always be updated
corectly including for a new created cache device.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In bch_cache_set_alloc() there is a big if() checks combined by 11 items
together. When this big if() statement fails, it is difficult to tell
exactly which item fails indeed.
This patch disassembles this big if() checks into 11 single if() checks,
which makes code debug more easier.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The improperly set bucket or block size will trigger error in
read_super_common(). For large bucket size, a more accurate error message
for invalid bucket or block size is necessary.
This patch disassembles the combined if() checks into multiple single
if() check, and provide more accurate error message for each check
failure condition.
Signed-off-by: Coly Li <colyli@suse.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Later patches will introduce feature set bits to on-disk super block and
increase super block version. Current code in read_super() which reads
common part of super block for version BCACHE_SB_VERSION_CDEV and version
BCACHE_SB_VERSION_CDEV_WITH_UUID will be shared with the new version.
Therefore this patch moves the reusable part into read_super_common(),
this preparation patch will make later patches more simplier and only
focus on new feature set bits.
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
offset_to_stripe() returns the stripe number (in type unsigned int) from
an offset (in type uint64_t) by the following calculation,
do_div(offset, d->stripe_size);
For large capacity backing device (e.g. 18TB) with small stripe size
(e.g. 4KB), the result is 4831838208 and exceeds UINT_MAX. The actual
returned value which caller receives is 536870912, due to the overflow.
Indeed in bcache_device_init(), bcache_device->nr_stripes is limited in
range [1, INT_MAX]. Therefore all valid stripe numbers in bcache are
in range [0, bcache_dev->nr_stripes - 1].
This patch adds a upper limition check in offset_to_stripe(): the max
valid stripe number should be less than bcache_device->nr_stripes. If
the calculated stripe number from do_div() is equal to or larger than
bcache_device->nr_stripe, -EINVAL will be returned. (Normally nr_stripes
is less than INT_MAX, exceeding upper limitation doesn't mean overflow,
therefore -EOVERFLOW is not used as error code.)
This patch also changes nr_stripes' type of struct bcache_device from
'unsigned int' to 'int', and return value type of offset_to_stripe()
from 'unsigned int' to 'int', to match their exact data ranges.
All locations where bcache_device->nr_stripes and offset_to_stripe() are
referenced also get updated for the above type change.
Reported-and-tested-by: Ken Raeburn <raeburn@redhat.com>
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Link: https://bugzilla.redhat.com/show_bug.cgi?id=1783075
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For some block devices which large capacity (e.g. 8TB) but small io_opt
size (e.g. 8 sectors), in bcache_device_init() the stripes number calcu-
lated by,
DIV_ROUND_UP_ULL(sectors, d->stripe_size);
might be overflow to the unsigned int bcache_device->nr_stripes.
This patch uses the uint64_t variable to store DIV_ROUND_UP_ULL()
and after the value is checked to be available in unsigned int range,
sets it to bache_device->nr_stripes. Then the overflow is avoided.
Reported-and-tested-by: Ken Raeburn <raeburn@redhat.com>
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Link: https://bugzilla.redhat.com/show_bug.cgi?id=1783075
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Make use of the struct_size() helper instead of an open-coded version
in order to avoid any potential type mistakes.
This code was detected with the help of Coccinelle and, audited and
fixed manually.
Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Make use of the struct_size() helper instead of an open-coded version
in order to avoid any potential type mistakes.
This code was detected with the help of Coccinelle and, audited and
fixed manually.
Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Remove unneeded variable i in bch_dirty_init_thread().
Signed-off-by: Xu Wang <vulab@iscas.ac.cn>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Using for_each_clear_bit() to simplify the code.
Signed-off-by: Xu Wang <vulab@iscas.ac.cn>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are some meta data of bcache are allocated by multiple pages,
and they are used as bio bv_page for I/Os to the cache device. for
example cache_set->uuids, cache->disk_buckets, journal_write->data,
bset_tree->data.
For such meta data memory, all the allocated pages should be treated
as a single memory block. Then the memory management and underlying I/O
code can treat them more clearly.
This patch adds __GFP_COMP flag to all the location allocating >0 order
pages for the above mentioned meta data. Then their pages are treated
as compound pages now.
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit adc0daad36 ("dm: report suspended
device during destroy") broke integrity recalculation.
The problem is dm_suspended() returns true not only during suspend,
but also during resume. So this race condition could occur:
1. dm_integrity_resume calls queue_work(ic->recalc_wq, &ic->recalc_work)
2. integrity_recalc (&ic->recalc_work) preempts the current thread
3. integrity_recalc calls if (unlikely(dm_suspended(ic->ti))) goto unlock_ret;
4. integrity_recalc exits and no recalculating is done.
To fix this race condition, add a function dm_post_suspending that is
only true during the postsuspend phase and use it instead of
dm_suspended().
Signed-off-by: Mikulas Patocka <mpatocka redhat com>
Fixes: adc0daad36 ("dm: report suspended device during destroy")
Cc: stable vger kernel org # v4.18+
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
We get compilation error on 32-bit architectures (e.g. m68k), as:
ERROR: modpost: "__udivdi3" [drivers/md/raid456.ko] undefined!
Since 'sync_blocks' is defined as u64, use do_div() to fix this error.
Fixes: c911c46c01 ("md/raid456: convert macro STRIPE_* to RAID5_STRIPE_*")
Signed-off-by: Yufen Yu <yuyufen@huawei.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
When disk failure happens and the array has a spare drive, resync thread
kicks in and starts to refill the spare. However it may get blocked by
a retry thread that resubmits failed IO to a mirror and itself can get
blocked on a barrier raised by the resync thread.
Acked-by: Nigel Croxon <ncroxon@redhat.com>
Signed-off-by: Vitaly Mayatskikh <vmayatskikh@digitalocean.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Adding a new 'stripe_size' sysfs entry to set and show stripe_size.
stripe_size should not be bigger than PAGE_SIZE, and it requires to
be multiple of 4096. We can adjust stripe_size by writing value into
sysfs entry, likely, set stripe_size as 16KB:
echo 16384 > /sys/block/md1/md/stripe_size
Show current stripe_size value:
cat /sys/block/md1/md/stripe_size
For PAGE_SIZE is equal to 4096, 'stripe_size' can just be read.
Signed-off-by: Yufen Yu <yuyufen@huawei.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
In RAID5, if issued bio size is bigger than stripe_size, it will be
split in the unit of stripe_size and process them one by one. Even
for size less then stripe_size, RAID5 also request data from disk at
least of stripe_size.
Nowdays, stripe_size is equal to the value of PAGE_SIZE. Since filesystem
usually issue bio in the unit of 4KB, there is no problem for PAGE_SIZE
as 4KB. But, for 64KB PAGE_SIZE, bio from filesystem requests 4KB data
while RAID5 issue IO at least stripe_size (64KB) each time. That will
waste resource of disk bandwidth and compute xor.
To avoding the waste, we want to make stripe_size configurable. This
patch just set default stripe_size as 4096. User can also set the value
bigger than 4KB for some special requirements, such as we know the
issued io size is more than 4KB.
To evaluate the new feature, we create raid5 device '/dev/md5' with
4 SSD disk and test it on arm64 machine with 64KB PAGE_SIZE.
1) We format /dev/md5 with mkfs.ext4 and mount ext4 with default
configure on /mnt directory. Then, trying to test it by dbench with
command: dbench -D /mnt -t 1000 10. Result show as:
'stripe_size = 64KB'
Operation Count AvgLat MaxLat
----------------------------------------
NTCreateX 9805011 0.021 64.728
Close 7202525 0.001 0.120
Rename 415213 0.051 44.681
Unlink 1980066 0.079 93.147
Deltree 240 1.793 6.516
Mkdir 120 0.004 0.007
Qpathinfo 8887512 0.007 37.114
Qfileinfo 1557262 0.001 0.030
Qfsinfo 1629582 0.012 0.152
Sfileinfo 798756 0.040 57.641
Find 3436004 0.019 57.782
WriteX 4887239 0.021 57.638
ReadX 15370483 0.005 37.818
LockX 31934 0.003 0.022
UnlockX 31933 0.001 0.021
Flush 687205 13.302 530.088
Throughput 307.799 MB/sec 10 clients 10 procs max_latency=530.091 ms
-------------------------------------------------------
'stripe_size = 4KB'
Operation Count AvgLat MaxLat
----------------------------------------
NTCreateX 11999166 0.021 36.380
Close 8814128 0.001 0.122
Rename 508113 0.051 29.169
Unlink 2423242 0.070 38.141
Deltree 300 1.885 7.155
Mkdir 150 0.004 0.006
Qpathinfo 10875921 0.007 35.485
Qfileinfo 1905837 0.001 0.032
Qfsinfo 1994304 0.012 0.125
Sfileinfo 977450 0.029 26.489
Find 4204952 0.019 9.361
WriteX 5981890 0.019 27.804
ReadX 18809742 0.004 33.491
LockX 39074 0.003 0.025
UnlockX 39074 0.001 0.014
Flush 841022 10.712 458.848
Throughput 376.777 MB/sec 10 clients 10 procs max_latency=458.852 ms
-------------------------------------------------------
It show that setting stripe_size as 4KB has higher thoughput, i.e.
(376.777 vs 307.799) and has smaller latency than that setting as 64KB.
2) We try to evaluate IO throughput for /dev/md5 by fio with config:
[4KB randwrite]
direct=1
numjob=2
iodepth=64
ioengine=libaio
filename=/dev/md5
bs=4KB
rw=randwrite
[64KB write]
direct=1
numjob=2
iodepth=64
ioengine=libaio
filename=/dev/md5
bs=1MB
rw=write
The result as follow:
+ +
| stripe_size(64KB) | stripe_size(4KB)
+----------------------------------------------------+
4KB randwrite | 15MB/s | 100MB/s
+----------------------------------------------------+
1MB write | 1000MB/s | 700MB/s
The result show that when size of io is bigger than 4KB (64KB),
64KB stripe_size has much higher IOPS. But for 4KB randwrite, that
means, size of io issued to device are smaller, 4KB stripe_size
have better performance.
Normally, default value (4096) can get relatively good performance.
But if each issued io is bigger than 4096, setting value more than
4096 may get better performance.
Here, we just set default stripe_size as 4096, and we will try to
support setting different stripe_size by sysfs interface in the
following patch.
Signed-off-by: Yufen Yu <yuyufen@huawei.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
Convert macro STRIPE_SIZE, STRIPE_SECTORS and STRIPE_SHIFT to
RAID5_STRIPE_SIZE(), RAID5_STRIPE_SECTORS() and RAID5_STRIPE_SHIFT().
This patch is prepare for the following adjustable stripe_size.
It will not change any existing functionality.
Signed-off-by: Yufen Yu <yuyufen@huawei.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
This function is just a tiny wrapper around blk_stack_limit and has
two callers. Simplify the stack a bit by open coding it in the two
callers.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Tested-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Lift the code from device mapper into blk_stack_limits to inherity
the stacking limitations. This ensures we do the right thing for
all stacked zoned block devices.
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Tested-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
* for-5.9/drivers: (38 commits)
block: add max_active_zones to blk-sysfs
block: add max_open_zones to blk-sysfs
s390/dasd: Use struct_size() helper
s390/dasd: fix inability to use DASD with DIAG driver
md-cluster: fix wild pointer of unlock_all_bitmaps()
md/raid5-cache: clear MD_SB_CHANGE_PENDING before flushing stripes
md: fix deadlock causing by sysfs_notify
md: improve io stats accounting
md: raid0/linear: fix dereference before null check on pointer mddev
rsxx: switch from 'pci_free_consistent()' to 'dma_free_coherent()'
nvme: remove ns->disk checks
nvme-pci: use standard block status symbolic names
nvme-pci: use the consistent return type of nvme_pci_iod_alloc_size()
nvme-pci: add a blank line after declarations
nvme-pci: fix some comments issues
nvme-pci: remove redundant segment validation
nvme: document quirked Intel models
nvme: expose reconnect_delay and ctrl_loss_tmo via sysfs
nvme: support for zoned namespaces
nvme: support for multiple Command Sets Supported and Effects log pages
...
* for-5.9/block: (124 commits)
blk-cgroup: show global disk stats in root cgroup io.stat
blk-cgroup: make iostat functions visible to stat printing
block: improve discard bio alignment in __blkdev_issue_discard()
block: change REQ_OP_ZONE_RESET and REQ_OP_ZONE_RESET_ALL to be odd numbers
block: defer flush request no matter whether we have elevator
block: make blk_timeout_init() static
block: remove retry loop in ioc_release_fn()
block: remove unnecessary ioc nested locking
block: integrate bd_start_claiming into __blkdev_get
block: use bd_prepare_to_claim directly in the loop driver
block: refactor bd_start_claiming
block: simplify the restart case in __blkdev_get
Revert "blk-rq-qos: remove redundant finish_wait to rq_qos_wait."
block: always remove partitions from blk_drop_partitions()
block: relax jiffies rounding for timeouts
blk-mq: remove redundant validation in __blk_mq_end_request()
blk-mq: Remove unnecessary local variable
writeback: remove bdi->congested_fn
writeback: remove struct bdi_writeback_congested
writeback: remove {set,clear}_wb_congested
...
Enable support for zoned block devices. This is done by:
1) implementing the target report_zones method.
2) adding the DM_TARGET_ZONED_HM flag to the target features.
3) setting DM_CRYPT_NO_WRITE_WORKQUEUE flag to avoid IO
processing via workqueue.
4) Introducing inline write encryption completion to preserve write
ordering.
The last point is implemented by introducing the internal flag
DM_CRYPT_WRITE_INLINE. When set, kcryptd_crypt_write_convert() always
waits inline for the completion of a write request encryption if the
request is not already completed once crypt_convert() returns.
Completion of write request encryption is signaled using the
restart completion by kcryptd_async_done(). This mechanism allows
using ciphers that have an asynchronous implementation, isolating
dm-crypt from any potential request completion reordering for these
ciphers.
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
This is a follow up to [1] that detailed latency problems associated
with dm-crypt's use of workqueues when processing IO.
Current dm-crypt implementation creates a significant IO performance
overhead (at least on small IO block sizes) for both latency and
throughput. We suspect offloading IO request processing into
workqueues and async threads is more harmful these days with the
modern fast storage. I also did some digging into the dm-crypt git
history and much of this async processing is not needed anymore,
because the reasons it was added are mostly gone from the kernel. More
details can be found in [2] (see "Git archeology" section).
This change adds DM_CRYPT_NO_READ_WORKQUEUE and
DM_CRYPT_NO_WRITE_WORKQUEUE flags for read and write BIOs, which
direct dm-crypt to not offload crypto operations into kcryptd
workqueues. In addition, writes are not buffered to be sorted in the
dm-crypt red-black tree, but dispatched immediately. For cases, where
crypto operations cannot happen (hard interrupt context, for example
the read path of some NVME drivers), we offload the work to a tasklet
rather than a workqueue.
These flags only ensure no async BIO processing in the dm-crypt
module. It is worth noting that some Crypto API implementations may
offload encryption into their own workqueues, which are independent of
the dm-crypt and its configuration. However upon enabling these new
flags dm-crypt will instruct Crypto API not to backlog crypto
requests.
To give an idea of the performance gains for certain workloads,
consider the script, and results when tested against various
devices, detailed here:
https://www.redhat.com/archives/dm-devel/2020-July/msg00138.html
[1]: https://www.spinics.net/lists/dm-crypt/msg07516.html
[2]: https://blog.cloudflare.com/speeding-up-linux-disk-encryption/
Signed-off-by: Ignat Korchagin <ignat@cloudflare.com>
Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>
Reviewed-by: Bob Liu <bob.liu@oracle.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Until now, DM bufio's waiting for IO from reclaim context in its
shrinker has caused kswapd to block; which results in systemic IO
stalls and even deadlock, e.g.:
https://www.redhat.com/archives/dm-devel/2020-March/msg00025.html
Here is Dave Chinner's problem description that motivated this fix,
from: https://lore.kernel.org/linux-fsdevel/20190809215733.GZ7777@dread.disaster.area/
"Waiting for IO in kswapd reclaim context is considered harmful -
kswapd context shrinker reclaim should be as non-blocking as possible,
and any back-off to wait for IO to complete should be done by the high
level reclaim core once it's completed an entire reclaim scan cycle of
everything....
What follows from that, and is pertinent in this situation, is that if
you don't block kswapd, then other reclaim contexts are not going to
get stuck waiting for it regardless of the reclaim context they use."
Continued elsewhere:
"The only way to fix this problem once and for all is to stop using
the shrinker as a mechanism to issue and wait on IO. If you need
background writeback of dirty buffers, do it from a WQ_MEM_RECLAIM
workqueue that isn't directly in the memory reclaim path and so can
issue writeback and block safely from a GFP_KERNEL context. Kick the
workqueue from the shrinker context, but get rid of the IO submission
and waiting from the shrinker and all the GFP_NOFS memory reclaim
recursion problems go away."
As such, this commit moves buffer cleanup to a workqueue.
Suggested-by: Dave Chinner <dchinner@redhat.com>
Reported-by: Tahsin Erdogan <tahsin@google.com>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Tested-by: Gabriel Krisman Bertazi <krisman@collabora.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
dm_stop_queue() only uses blk_mq_quiesce_queue() so it doesn't
formally stop the blk-mq queue; therefore there is no point making the
blk_mq_queue_stopped() check -- it will never be stopped.
In addition, even though dm_stop_queue() actually tries to quiesce hw
queues via blk_mq_quiesce_queue(), checking with blk_queue_quiesced()
to avoid unnecessary queue quiesce isn't reliable because: the
QUEUE_FLAG_QUIESCED flag is set before synchronize_rcu() and
dm_stop_queue() may be called when synchronize_rcu() from another
blk_mq_quiesce_queue() is in-progress.
Fixes: 7b17c2f729 ("dm: Fix a race condition related to stopping and starting queues")
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
This interface may help anyone who want to know all badblocks without
querying for each block.
[Bryan: DMEMIT message if no blocks are in the bad block list.]
Signed-off-by: yangerkun <yangerkun@huawei.com>
Signed-off-by: Bryan Gurney <bgurney@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Some messages (queryblock, countbadblocks, removebadblock) are best
reported directly to user directly. Do so with DMEMIT.
[Bryan: maintain __func__ output in DMEMIT messages]
Signed-off-by: yangerkun <yangerkun@huawei.com>
Signed-off-by: Bryan Gurney <bgurney@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Using uninitialized_var() is dangerous as it papers over real bugs[1]
(or can in the future), and suppresses unrelated compiler warnings
(e.g. "unused variable"). If the compiler thinks it is uninitialized,
either simply initialize the variable or make compiler changes.
In preparation for removing[2] the[3] macro[4], remove all remaining
needless uses with the following script:
git grep '\buninitialized_var\b' | cut -d: -f1 | sort -u | \
xargs perl -pi -e \
's/\buninitialized_var\(([^\)]+)\)/\1/g;
s:\s*/\* (GCC be quiet|to make compiler happy) \*/$::g;'
drivers/video/fbdev/riva/riva_hw.c was manually tweaked to avoid
pathological white-space.
No outstanding warnings were found building allmodconfig with GCC 9.3.0
for x86_64, i386, arm64, arm, powerpc, powerpc64le, s390x, mips, sparc64,
alpha, and m68k.
[1] https://lore.kernel.org/lkml/20200603174714.192027-1-glider@google.com/
[2] https://lore.kernel.org/lkml/CA+55aFw+Vbj0i=1TGqCR5vQkCzWJ0QxK6CernOU6eedsudAixw@mail.gmail.com/
[3] https://lore.kernel.org/lkml/CA+55aFwgbgqhbp1fkxvRKEpzyR5J8n1vKT1VZdz9knmPuXhOeg@mail.gmail.com/
[4] https://lore.kernel.org/lkml/CA+55aFz2500WfbKXAx8s67wrm9=yVJu65TpLgN_ybYNv0VEOKA@mail.gmail.com/
Reviewed-by: Leon Romanovsky <leonro@mellanox.com> # drivers/infiniband and mlx4/mlx5
Acked-by: Jason Gunthorpe <jgg@mellanox.com> # IB
Acked-by: Kalle Valo <kvalo@codeaurora.org> # wireless drivers
Reviewed-by: Chao Yu <yuchao0@huawei.com> # erofs
Signed-off-by: Kees Cook <keescook@chromium.org>
We can't guarntee the batched stripe to be set with STRIPE_HANDLE since
there are lots of functions could set the flag, such as sync_request,
ops_complete_* and end_{read,write}_request etc.
Also clear_batch_ready called in handle_stripe ensures the batched list
can't continue to be handled by handle_stripe.
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
To make people understand the function well, let's put the comment to
the right place.
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
We tried to only put the head sh of batch list to handle_list, then the
handle_stripe doesn't handle other members in the batch list. However,
we still got the calltrace in break_stripe_batch_list.
[593764.644269] stripe state: 2003
kernel: [593764.644299] ------------[ cut here ]------------
kernel: [593764.644308] WARNING: CPU: 12 PID: 856 at drivers/md/raid5.c:4625 break_stripe_batch_list+0x203/0x240 [raid456]
[...]
kernel: [593764.644363] Call Trace:
kernel: [593764.644370] handle_stripe+0x907/0x20c0 [raid456]
kernel: [593764.644376] ? __wake_up_common_lock+0x89/0xc0
kernel: [593764.644379] handle_active_stripes.isra.57+0x35f/0x570 [raid456]
kernel: [593764.644382] ? raid5_wakeup_stripe_thread+0x96/0x1f0 [raid456]
kernel: [593764.644385] raid5d+0x480/0x6a0 [raid456]
kernel: [593764.644390] ? md_thread+0x11f/0x160
kernel: [593764.644392] md_thread+0x11f/0x160
kernel: [593764.644394] ? wait_woken+0x80/0x80
kernel: [593764.644396] kthread+0xfc/0x130
kernel: [593764.644398] ? find_pers+0x70/0x70
kernel: [593764.644399] ? kthread_create_on_node+0x70/0x70
kernel: [593764.644401] ret_from_fork+0x1f/0x30
As we can see, the stripe was set with STRIPE_ACTIVE and STRIPE_HANDLE,
and only handle_stripe could set those flags then return. And since the
stipe was already in the batch list, we need to return earlier before
set the two flags.
And after dig a little about git history especially commit 3664847d95
("md/raid5: fix a race condition in stripe batch"), it seems the batched
stipe still could be handled by handle_stipe, then handle_stipe needs to
return earlier if clear_batch_ready to return true.
Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
md_setup_drive knows it works with md devices, so it is rather pointless
to open a file descriptor and issue ioctls. Just call directly into the
relevant low-level md routines after getting a handle to the device using
blkdev_get_by_dev instead.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: NeilBrown <neilb@suse.de>
Acked-by: Song Liu <song@kernel.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Move the loop over the possible arrays into the caller to remove a level
of indentation for the whole function.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: NeilBrown <neilb@suse.de>
Acked-by: Song Liu <song@kernel.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
mdp_major can just move to drivers/md/md.h.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Song Liu <song@kernel.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
devfs is long gone, and autoscan works just fine without this these days.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: NeilBrown <neilb@suse.de>
Acked-by: Song Liu <song@kernel.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Instead of using a spcial RAID_AUTORUN ioctl that only exists for
non-modular builds and is only called from the early init code, just
call the actual function directly.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: NeilBrown <neilb@suse.de>
Acked-by: Song Liu <song@kernel.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Just like the NFS and CIFS root code this better lives with the
driver it is tightly integrated with.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Song Liu <song@kernel.org>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
