Patch series "mm: zap pages with read mmap_sem in munmap for large
mapping", v11.

Background:
Recently, when we ran some vm scalability tests on machines with large memory,
we ran into a couple of mmap_sem scalability issues when unmapping large memory
space, please refer to https://lkml.org/lkml/2017/12/14/733 and
https://lkml.org/lkml/2018/2/20/576.

History:
Then akpm suggested to unmap large mapping section by section and drop mmap_sem
at a time to mitigate it (see https://lkml.org/lkml/2018/3/6/784).

V1 patch series was submitted to the mailing list per Andrew's suggestion
(see https://lkml.org/lkml/2018/3/20/786).  Then I received a lot great
feedback and suggestions.

Then this topic was discussed on LSFMM summit 2018.  In the summit, Michal
Hocko suggested (also in the v1 patches review) to try "two phases"
approach.  Zapping pages with read mmap_sem, then doing via cleanup with
write mmap_sem (for discussion detail, see
https://lwn.net/Articles/753269/)

Approach:
Zapping pages is the most time consuming part, according to the suggestion from
Michal Hocko [1], zapping pages can be done with holding read mmap_sem, like
what MADV_DONTNEED does. Then re-acquire write mmap_sem to cleanup vmas.

But, we can't call MADV_DONTNEED directly, since there are two major drawbacks:
  * The unexpected state from PF if it wins the race in the middle of munmap.
    It may return zero page, instead of the content or SIGSEGV.
  * Can't handle VM_LOCKED | VM_HUGETLB | VM_PFNMAP and uprobe mappings, which
    is a showstopper from akpm

But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
        acquire write mmap_sem
        lookup vmas (find and split vmas)
        deal with special mappings
        detach vmas
        downgrade_write

        zap pages
        free page tables
        release mmap_sem

The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e.  page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.

If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings.  They will be handled by falling back to regular do_munmap()
with exclusive mmap_sem held in this patch since they may update vm flags.

But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case.  So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.

Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.  So, uprobe unmap will not be handled by the regular
path.

With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.

And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.

For the time being, just do this in munmap syscall path.  Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle.  But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization.  So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them.  mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.

This patch (of 3):

When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.

INFO: task ps:14018 blocked for more than 120 seconds.
       Tainted: G            E 4.9.79-009.ali3000.alios7.x86_64 #1
 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
 ps              D    0 14018      1 0x00000004
  ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
  ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
  00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
 Call Trace:
  [<ffffffff817154d0>] ? __schedule+0x250/0x730
  [<ffffffff817159e6>] schedule+0x36/0x80
  [<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
  [<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
  [<ffffffff81717db0>] down_read+0x20/0x40
  [<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
  [<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
  [<ffffffff81241d87>] __vfs_read+0x37/0x150
  [<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
  [<ffffffff81242266>] vfs_read+0x96/0x130
  [<ffffffff812437b5>] SyS_read+0x55/0xc0
  [<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5

It is because munmap holds mmap_sem exclusively from very beginning to all
the way down to the end, and doesn't release it in the middle.  When
unmapping large mapping, it may take long time (take ~18 seconds to unmap
320GB mapping with every single page mapped on an idle machine).

Zapping pages is the most time consuming part, according to the suggestion
from Michal Hocko [1], zapping pages can be done with holding read
mmap_sem, like what MADV_DONTNEED does.  Then re-acquire write mmap_sem to
cleanup vmas.

But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
        acquire write mmap_sem
        lookup vmas (find and split vmas)
        deal with special mappings
        detach vmas
        downgrade_write

        zap pages
        free page tables
        release mmap_sem

The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e.  page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT as
expected.

If the vma has VM_HUGETLB | VM_PFNMAP, they are considered as special
mappings.  They will be handled by without downgrading mmap_sem in this
patch since they may update vm flags.

But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with read
mmap_sem for this specific case.  So, VM_HUGETLB and VM_PFNMAP will be
handled by using the optimized path in the following separate patches for
bisectable sake.

Unmapping uprobe areas may need update mm flags (MMF_RECALC_UPROBES).
However it is fine to have false-positive MMF_RECALC_UPROBES according to
uprobes developer.

With the "detach vmas first" approach we don't have to re-acquire mmap_sem
again to clean up vmas to avoid race window which might get the address
space changed since downgrade_write() doesn't release the lock to lead
regression, which simply downgrades to read lock.

And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.

For the time being, just do this in munmap syscall path.  Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap() might
be called in the middle.  But, the optimized do_munmap would like to be
called without mmap_sem held so that we can do the optimization.  So, if
we want to do the similar optimization for mmap/mremap path, I'm afraid we
would have to redesign them.  mremap might be called on very large area
depending on the usecases, the optimization to it will be considered in
the future.

With the patches, exclusive mmap_sem hold time when munmap a 80GB address
space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to us level
from second.

munmap_test-15002 [008]   594.380138: funcgraph_entry: |
__vm_munmap() {
munmap_test-15002 [008]   594.380146: funcgraph_entry:      !2485684 us
|    unmap_region();
munmap_test-15002 [008]   596.865836: funcgraph_exit:       !2485692 us
|  }

Here the execution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.

[1] https://lwn.net/Articles/753269/

Link: http://lkml.kernel.org/r/1537376621-51150-2-git-send-email-yang.shi@linux.alibaba.comSigned-off-by: Yang Shi &lt;yang.shi@linux.alibaba.com&gt;Suggested-by: Michal Hocko <mhocko@kernel.org>
Suggested-by: Kirill A. Shutemov <kirill@shutemov.name>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Reviewed-by: Matthew Wilcox <willy@infradead.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Laurent Dufour <ldufour@linux.vnet.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>