We have full disk encryption enabled, profiling shows page allocations may
incur a noticeable overhead when writing. The dm-crypt creates an "out"
bio for writing. And fill the "out" bio with the same amount of pages
as "in" bio. But the driver allocates one page at a time in a loop. For
1M bio it means the driver has to call page allocator 256 times. It seems
not that efficient.
Since v5.13 we have page bulk allocator supported, so dm-crypt could use
it to do page allocations more efficiently.
I could just call the page bulk allocator in dm-crypt driver before the
mempool allocator, but it seems ad-hoc and the quick search shows some
others do the similar thing, for example, f2fs compress, block bounce,
g2fs, ufs, etc. So it seems more neat to implement a bulk allocation
API for mempool.
So introduce the mempool page bulk allocator.
The below APIs are introduced:
- mempool_init_pages_bulk()
- mempool_create_pages_bulk()
They initialize the mempool for page bulk allocator. The pool is filled
by alloc_page() in a loop.
- mempool_alloc_pages_bulk_list()
- mempool_alloc_pages_bulk_array()
They do bulk allocation from mempool.
They do the below conceptually:
1. Call bulk page allocator
2. If the allocation is fulfilled then return otherwise try to
allocate the remaining pages from the mempool
3. If it is fulfilled then return otherwise retry from #1 with sleepable
gfp
4. If it is still failed, sleep for a while to wait for the mempool is
refilled, then retry from #1
The populated pages will stay on the list or array until the callers
consume them or free them.
Since mempool allocator is guaranteed to success in the sleepable context,
so the two APIs return true for success or false for fail. It is the
caller's responsibility to handle failure case (partial allocation), just
like the page bulk allocator.
The mempool typically is an object agnostic allocator, but bulk allocation
is only supported by pages, so the mempool bulk allocator is for page
allocation only as well.
With the mempool bulk allocator the IOPS of dm-crypt with 1M I/O would get
improved by approxiamately 6%. The test is done on a VM with 80 vCPU and
64GB memory with an encrypted ram device (the impact from storage hardware
could be minimized so that we could benchmark the dm-crypt layer more
accurately).
Before the patch:
Jobs: 1 (f=1): [w(1)][100.0%][r=0KiB/s,w=402MiB/s][r=0,w=402 IOPS][eta 00m:00s]
crypt: (groupid=0, jobs=1): err= 0: pid=233950: Thu Sep 15 16:23:10 2022
write: IOPS=402, BW=403MiB/s (423MB/s)(23.6GiB/60002msec)
slat (usec): min=2425, max=3819, avg=2480.84, stdev=34.00
clat (usec): min=7, max=165751, avg=156398.72, stdev=4691.03
lat (msec): min=2, max=168, avg=158.88, stdev= 4.69
clat percentiles (msec):
| 1.00th=[ 157], 5.00th=[ 157], 10.00th=[ 157], 20.00th=[ 157],
| 30.00th=[ 157], 40.00th=[ 157], 50.00th=[ 157], 60.00th=[ 157],
| 70.00th=[ 157], 80.00th=[ 157], 90.00th=[ 157], 95.00th=[ 157],
| 99.00th=[ 159], 99.50th=[ 159], 99.90th=[ 165], 99.95th=[ 165],
| 99.99th=[ 167]
bw ( KiB/s): min=405504, max=413696, per=99.71%, avg=411845.53, stdev=1155.04, samples=120
iops : min= 396, max= 404, avg=402.17, stdev= 1.15, samples=120
lat (usec) : 10=0.01%
lat (msec) : 4=0.01%, 10=0.01%, 20=0.02%, 50=0.05%, 100=0.08%
lat (msec) : 250=100.09%
cpu : usr=3.74%, sys=95.66%, ctx=27, majf=0, minf=4
IO depths : 1=0.1%, 2=0.1%, 4=0.1%, 8=0.1%, 16=0.1%, 32=0.1%, >=64=103.1%
submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.1%, >=64=0.0%
issued rwts: total=0,24138,0,0 short=0,0,0,0 dropped=0,0,0,0
latency : target=0, window=0, percentile=100.00%, depth=64
Run status group 0 (all jobs):
WRITE: bw=403MiB/s (423MB/s), 403MiB/s-403MiB/s (423MB/s-423MB/s), io=23.6GiB (25.4GB), run=60002-60002msec
After the patch:
Jobs: 1 (f=1): [w(1)][100.0%][r=0KiB/s,w=430MiB/s][r=0,w=430 IOPS][eta 00m:00s]
crypt: (groupid=0, jobs=1): err= 0: pid=288730: Thu Sep 15 16:25:39 2022
write: IOPS=430, BW=431MiB/s (452MB/s)(25.3GiB/60002msec)
slat (usec): min=2253, max=3213, avg=2319.49, stdev=34.29
clat (usec): min=6, max=149337, avg=146257.68, stdev=4239.52
lat (msec): min=2, max=151, avg=148.58, stdev= 4.24
clat percentiles (msec):
| 1.00th=[ 146], 5.00th=[ 146], 10.00th=[ 146], 20.00th=[ 146],
| 30.00th=[ 146], 40.00th=[ 146], 50.00th=[ 146], 60.00th=[ 146],
| 70.00th=[ 146], 80.00th=[ 146], 90.00th=[ 148], 95.00th=[ 148],
| 99.00th=[ 148], 99.50th=[ 148], 99.90th=[ 150], 99.95th=[ 150],
| 99.99th=[ 150]
bw ( KiB/s): min=438272, max=442368, per=99.73%, avg=440463.57, stdev=1305.60, samples=120
iops : min= 428, max= 432, avg=430.12, stdev= 1.28, samples=120
lat (usec) : 10=0.01%
lat (msec) : 4=0.01%, 10=0.01%, 20=0.02%, 50=0.05%, 100=0.09%
lat (msec) : 250=100.07%
cpu : usr=3.78%, sys=95.37%, ctx=12778, majf=0, minf=4
IO depths : 1=0.1%, 2=0.1%, 4=0.1%, 8=0.1%, 16=0.1%, 32=0.1%, >=64=103.1%
submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.1%, >=64=0.0%
issued rwts: total=0,25814,0,0 short=0,0,0,0 dropped=0,0,0,0
latency : target=0, window=0, percentile=100.00%, depth=64
Run status group 0 (all jobs):
WRITE: bw=431MiB/s (452MB/s), 431MiB/s-431MiB/s (452MB/s-452MB/s), io=25.3GiB (27.1GB), run=60002-60002msec
The function tracing also shows the time consumed by page allocations is
reduced significantly. The test allocated 1M (256 pages) bio in the same
environment.
Before the patch:
It took approximately 600us by excluding the bio_add_page() calls.
2720.630754 | 56) xfs_io-38859 | 2.571 us | mempool_alloc();
2720.630757 | 56) xfs_io-38859 | 0.937 us | bio_add_page();
2720.630758 | 56) xfs_io-38859 | 1.772 us | mempool_alloc();
2720.630760 | 56) xfs_io-38859 | 0.852 us | bio_add_page();
….
2720.631559 | 56) xfs_io-38859 | 2.058 us | mempool_alloc();
2720.631561 | 56) xfs_io-38859 | 0.717 us | bio_add_page();
2720.631562 | 56) xfs_io-38859 | 2.014 us | mempool_alloc();
2720.631564 | 56) xfs_io-38859 | 0.620 us | bio_add_page();
After the patch:
It took approxiamately 30us.
11564.266385 | 22) xfs_io-136183 | + 30.551 us | __alloc_pages_bulk();
Page allocations overhead is around 6% (600us/9853us) in dm-crypt layer shown by
function trace. The data also matches the IOPS data shown by fio.
And the benchmark with 4K size I/O doesn't show measurable regression.
Yang Shi (4):
mm: mempool: extract common initialization code
mm: mempool: introduce page bulk allocator
md: dm-crypt: move crypt_free_buffer_pages ahead
md: dm-crypt: use mempool page bulk allocator
drivers/md/dm-crypt.c | 92 ++++++++++++++++-------------
include/linux/mempool.h | 19 ++++++
mm/mempool.c | 227 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-------
3 files changed, 276 insertions(+), 62 deletions(-)
