fs/address_space: add alignment padding for i_map and i_mmap_rwsem to mitigate a...
authorZhu, Lipeng <lipeng.zhu@intel.com>
Sun, 16 Jul 2023 14:56:54 +0000 (22:56 +0800)
committerAndrew Morton <akpm@linux-foundation.org>
Fri, 18 Aug 2023 17:12:38 +0000 (10:12 -0700)
When running UnixBench/Shell Scripts, we observed high false sharing for
accessing i_mmap against i_mmap_rwsem.

UnixBench/Shell Scripts are typical load/execute command test scenarios,
which concurrently launch->execute->exit a lot of shell commands.  A lot
of processes invoke vma_interval_tree_remove which touch "i_mmap", the
call stack:

----vma_interval_tree_remove
    |----unlink_file_vma
    |    free_pgtables
    |    |----exit_mmap
    |    |    mmput
    |    |    |----begin_new_exec
    |    |    |    load_elf_binary
    |    |    |    bprm_execve

Meanwhile, there are a lot of processes touch 'i_mmap_rwsem' to acquire
the semaphore in order to access 'i_mmap'.  In existing 'address_space'
layout, 'i_mmap' and 'i_mmap_rwsem' are in the same cacheline.

The patch places the i_mmap and i_mmap_rwsem in separate cache lines to
avoid this false sharing problem.

With this patch, based on kernel v6.4.0, on Intel Sapphire Rapids
112C/224T platform, the score improves by ~5.3%.  And perf c2c tool shows
the false sharing is resolved as expected, the symbol
vma_interval_tree_remove disappeared in cache line 0 after this change.

Baseline:
=================================================
      Shared Cache Line Distribution Pareto
=================================================
-------------------------------------------------------------
    0    3729     5791        0        0  0xff19b3818445c740
-------------------------------------------------------------
   3.27%    3.02%    0.00%    0.00%   0x18     0       1  0xffffffffa194403b       604       483       389      692       203  [k] vma_interval_tree_insert    [kernel.kallsyms]  vma_interval_tree_insert+75      0  1
   4.13%    3.63%    0.00%    0.00%   0x20     0       1  0xffffffffa19440a2       553       413       415      962       215  [k] vma_interval_tree_remove    [kernel.kallsyms]  vma_interval_tree_remove+18      0  1
   2.04%    1.35%    0.00%    0.00%   0x28     0       1  0xffffffffa219a1d6      1210       855       460     1229       222  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+678    0  1
   0.62%    1.85%    0.00%    0.00%   0x28     0       1  0xffffffffa219a1bf       762       329       577      527       198  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+655    0  1
   0.48%    0.31%    0.00%    0.00%   0x28     0       1  0xffffffffa219a58c      1677      1476       733     1544       224  [k] down_write                  [kernel.kallsyms]  down_write+28                    0  1
   0.05%    0.07%    0.00%    0.00%   0x28     0       1  0xffffffffa219a21d      1040       819       689       33        27  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+749    0  1
   0.00%    0.05%    0.00%    0.00%   0x28     0       1  0xffffffffa17707db         0      1005       786     1373       223  [k] up_write                    [kernel.kallsyms]  up_write+27                      0  1
   0.00%    0.02%    0.00%    0.00%   0x28     0       1  0xffffffffa219a064         0       233       778       32        30  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+308    0  1
  33.82%   34.10%    0.00%    0.00%   0x30     0       1  0xffffffffa1770945       779       495       534     6011       224  [k] rwsem_spin_on_owner         [kernel.kallsyms]  rwsem_spin_on_owner+53           0  1
  17.06%   15.28%    0.00%    0.00%   0x30     0       1  0xffffffffa1770915       593       438       468     2715       224  [k] rwsem_spin_on_owner         [kernel.kallsyms]  rwsem_spin_on_owner+5            0  1
   3.54%    3.52%    0.00%    0.00%   0x30     0       1  0xffffffffa2199f84       881       601       583     1421       223  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+84     0  1

With this change:
-------------------------------------------------------------
   0      556      838        0        0  0xff2780d7965d2780
-------------------------------------------------------------
    0.18%    0.60%    0.00%    0.00%    0x8     0       1  0xffffffffafff27b8       503       453       569       14        13  [k] do_dentry_open              [kernel.kallsyms]  do_dentry_open+456               0  1
    0.54%    0.12%    0.00%    0.00%    0x8     0       1  0xffffffffaffc51ac       510       199       428       15        12  [k] hugepage_vma_check          [kernel.kallsyms]  hugepage_vma_check+252           0  1
    1.80%    2.15%    0.00%    0.00%   0x18     0       1  0xffffffffb079a1d6      1778       799       343      215       136  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+678    0  1
    0.54%    1.31%    0.00%    0.00%   0x18     0       1  0xffffffffb079a1bf       547       296       528       91        71  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+655    0  1
    0.72%    0.72%    0.00%    0.00%   0x18     0       1  0xffffffffb079a58c      1479      1534       676      288       163  [k] down_write                  [kernel.kallsyms]  down_write+28                    0  1
    0.00%    0.12%    0.00%    0.00%   0x18     0       1  0xffffffffafd707db         0      2381       744      282       158  [k] up_write                    [kernel.kallsyms]  up_write+27                      0  1
    0.00%    0.12%    0.00%    0.00%   0x18     0       1  0xffffffffb079a064         0       239       518        6         6  [k] rwsem_down_write_slowpath   [kernel.kallsyms]  rwsem_down_write_slowpath+308    0  1
   46.58%   47.02%    0.00%    0.00%   0x20     0       1  0xffffffffafd70945       704       403       499     1137       219  [k] rwsem_spin_on_owner         [kernel.kallsyms]  rwsem_spin_on_owner+53           0  1
   23.92%   25.78%    0.00%    0.00%   0x20     0       1  0xffffffffafd70915       558       413       500      542       185  [k] rwsem_spin_on_owner         [kernel.kallsyms]  rwsem_spin_on_owner+5            0  1

v1->v2: change padding to exchange fields.

Link: https://lkml.kernel.org/r/20230716145653.20122-1-lipeng.zhu@intel.com
Signed-off-by: Lipeng Zhu <lipeng.zhu@intel.com>
Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Yu Ma <yu.ma@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
include/linux/fs.h

index 6867512907d6e593ec18cbd27dc52471b66418a0..bb00d37a6ca02006a1936c8ec5638667ce6f80bf 100644 (file)
@@ -447,11 +447,11 @@ struct address_space {
        atomic_t                nr_thps;
 #endif
        struct rb_root_cached   i_mmap;
-       struct rw_semaphore     i_mmap_rwsem;
        unsigned long           nrpages;
        pgoff_t                 writeback_index;
        const struct address_space_operations *a_ops;
        unsigned long           flags;
+       struct rw_semaphore     i_mmap_rwsem;
        errseq_t                wb_err;
        spinlock_t              private_lock;
        struct list_head        private_list;