Re: [PATCH RFC] mm: implement granular soft-dirty vma support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 20.12.22 17:26, Muhammad Usama Anjum wrote:
The VM_SOFTDIRTY is used to mark a whole VMA soft-dirty. Sometimes
soft-dirty and non-soft-dirty VMAs are merged making the non-soft-dirty
region soft-dirty. This creates problems as the whole VMA region comes
out to be soft-dirty while in-reality no there is no soft-dirty page.
This can be solved by not merging the VMAs with different soft-dirty
flags. But it has potential to cause regression as the number of VMAs
can increase drastically.

I'm not going to look into the details of this RFC, but (1) it looks over-engineered and (2) is increases the size of each and every VMA in the system.

Let's talk about what happens when we stop merging VMAs where VM_SOFTDIRTY differs:

(a) Not merging VMAs when VM_SOFTDIRTY differs will only affect
    processes with active softdirty tracking (i.e., after
    CLEAR_REFS_SOFT_DIRTY). All other VMAs have VM_SOFTDIRTY set and
    will get merged. Processes without CLEAR_REFS_SOFT_DIRTY behave the
    same.

(b) After CLEAR_REFS_SOFT_DIRTY, old mappings will have VM_SOFTDIRTY set
    but new ones won't. We can't merge them. Consequently, we might not
    merge these VMAs and create more.

(c) The problem about (b) is that it will get worse every time we
    CLEAR_REFS_SOFT_DIRTY, because we're not merging the old ones that
    could get merged.


To tackle (c), we can simply try merging VMAs in clear_refs_write() when clearing VM_SOFTDIRTY. We're already properly holding the mmap lock in write mode, so it's easy to check if we can merge the modified VMA into the previous one or into the next one -- or if we can merge all of them into a single VMA.


For (b), the usual thing we care about is most probably

[!VM_SOFTDIRTY][VM_SOFTDIRTY]

No longer getting merged into a single VMA. This would imply that during (b), we could have doubled the #VMAs.

Of course, there are cases like

[!VM_SOFTDIRTY][VM_SOFTDIRTY][!VM_SOFTDIRTY]

where we could triple them or even a chain like

[!VM_SOFTDIRTY][VM_SOFTDIRTY][!VM_SOFTDIRTY][VM_SOFTDIRTY]...

where the #VMAs could explode. BUT, that would imply that someone has to do fine-grained mmap()'s over an existing mmap()'s (or into holes) and heavily relies on VMA merging to happen. I assume that only with anonymous memory this really works as expected, so I am not sure how likely such a pattern is in applications we care about and if we should really care.

My suggestion would be to

(1) Make the new merging behavior (consider VM_SOFTDIRTY or ignore
    VM_SOFTDIRTY) configurable (per process? for the system) to not
    accidentally degrade existing users of soft-dirty tracking with such
    "special" applications.
(2) Implement conditional VMA merging during CLEAR_REFS_SOFT_DIRTY: if
    we consider VM_SOFTDIRTY when making merging decisions, we want to
    try merging VMAs here after clearing VM_SOFTDIRTY.
(2) For your use case, enable the new behavior and eventually slightly
    bump up the #VMA limit in case you're dealing with "special"
    applications.

(1) would be called something like "precise_softdirty_tracking". Disabled is old handling, enabled is new handling. Precision here means that we're not over-indicating softdirty due to VMA merging.


Anything important I am missing? Are we aware of such applications for your use-case?

--
Thanks,

David / dhildenb




[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [NTFS 3]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [NTFS 3]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux