Message-ID: <340ff8ef-9ff5-7175-c234-4132bbdfc5f7@cybernetics.com>
Date:   Tue, 7 Jun 2022 14:38:34 -0400
From:   Tony Battersby <tonyb@...ernetics.com>
To:     linux-mm@...ck.org, linux-kernel@...r.kernel.org
Cc:     iommu@...ts.linux-foundation.org, kernel-team@...com,
        Matthew Wilcox <willy@...radead.org>,
        Keith Busch <kbusch@...nel.org>,
        Andy Shevchenko <andy.shevchenko@...il.com>,
        Robin Murphy <robin.murphy@....com>,
        Tony Lindgren <tony@...mide.com>
Subject: [PATCH v6 00/11] mpt3sas and dmapool scalability

This patch series improves dmapool scalability by replacing linear scans
with red-black trees.

Note that Keith Busch is also working on improving dmapool scalability,
so for now I would recommend not merging my scalability patches until
Keith's approach can be evaluated.  In the meantime, my patches can
serve as a benchmark comparison.  I also have a number of cleanup
patches in my series that could be useful on their own.

Changes since v5:
1. inline pool_free_page() into dma_pool_destroy() to avoid adding
   unused code
2. convert scnprintf() to sysfs_emit()
3. avoid adding a hole in struct dma_pool
4. fix big O usage in description

References:

v5
https://lore.kernel.org/linux-mm/9b08ab7c-b80b-527d-9adf-7716b0868fbc@cybernetics.com/

Keith Busch's dmapool performance enhancements
https://lore.kernel.org/linux-mm/20220428202714.17630-1-kbusch@kernel.org/

Below is my original description of the motivation for these patches.

drivers/scsi/mpt3sas is running into a scalability problem with the
kernel's DMA pool implementation.  With a LSI/Broadcom SAS 9300-8i
12Gb/s HBA and max_sgl_entries=256, during modprobe, mpt3sas does the
equivalent of:

chain_dma_pool = dma_pool_create(size = 128);
for (i = 0; i < 373959; i++)
    {
    dma_addr[i] = dma_pool_alloc(chain_dma_pool);
    }

And at rmmod, system shutdown, or system reboot, mpt3sas does the
equivalent of:

for (i = 0; i < 373959; i++)
    {
    dma_pool_free(chain_dma_pool, dma_addr[i]);
    }
dma_pool_destroy(chain_dma_pool);

With this usage, both dma_pool_alloc() and dma_pool_free() exhibit
O(n) complexity, although dma_pool_free() is much worse due to
implementation details.  On my system, the dma_pool_free() loop above
takes about 9 seconds to run.  Note that the problem was even worse
before commit 74522a92bbf0 ("scsi: mpt3sas: Optimize I/O memory
consumption in driver."), where the dma_pool_free() loop could take ~30
seconds.

mpt3sas also has some other DMA pools, but chain_dma_pool is the only
one with so many allocations:

cat /sys/devices/pci0000:80/0000:80:07.0/0000:85:00.0/pools
(manually cleaned up column alignment)
poolinfo - 0.1
reply_post_free_array pool  1      21     192     1
reply_free pool             1      1      41728   1
reply pool                  1      1      1335296 1
sense pool                  1      1      970272  1
chain pool                  373959 386048 128     12064
reply_post_free pool        12     12     166528  12

The patches in this series improve the scalability of the DMA pool
implementation, which significantly reduces the running time of the
DMA alloc/free loops.  With the patches applied, "modprobe mpt3sas",
"rmmod mpt3sas", and system shutdown/reboot with mpt3sas loaded are
significantly faster.  Here are some benchmarks (of DMA alloc/free
only, not the entire modprobe/rmmod):

dma_pool_create() + dma_pool_alloc() loop, size = 128, count = 373959
  original:        350 ms ( 1x)
  dmapool patches:  18 ms (19x)

dma_pool_free() loop + dma_pool_destroy(), size = 128, count = 373959
  original:        8901 ms (   1x)
  dmapool patches:   19 ms ( 477x)

Powered by blists - more mailing lists