Re: [RFC v6 PATCH 2/2] mm: mmap: zap pages with read mmap_sem in munmap

[Yang Shi]

On 8/6/18 2:40 AM, Michal Hocko wrote:
> On Fri 03-08-18 14:01:58, Yang Shi wrote:
> > On 8/3/18 2:07 AM, Michal Hocko wrote:
> > > On Fri 27-07-18 02:10:14, Yang Shi wrote:
> > > > 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)
> > > > 	detach vmas
> > > >           deal with special mappings
> > > >           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_LOCKED | VM_HUGETLB | VM_PFNMAP or uprobe, they are
> > > > considered as special mappings. They will be dealt with before zapping
> > > > pages with write mmap_sem held. Basically, just update vm_flags.
> > > Well, I think it would be safer to simply fallback to the current
> > > implementation with these mappings and deal with them on top. This would
> > > make potential issues easier to bisect and partial reverts as well.
> > Do you mean just call do_munmap()? It sounds ok. Although we may waste some
> > cycles to repeat what has done, it sounds not too bad since those special
> > mappings should be not very common.
> VM_HUGETLB is quite spread. Especially for DB workloads.

Wait a minute. In this way, it sounds we go back to my old 
implementation with special handling for those mappings with write 
mmap_sem held, right?

> > > > And, since they are also manipulated by unmap_single_vma() which is
> > > > called by unmap_vma() with read mmap_sem held in this case, to
> > > > prevent from updating vm_flags in read critical section, a new
> > > > parameter, called "skip_flags" is added to unmap_region(), unmap_vmas()
> > > > and unmap_single_vma(). If it is true, then just skip unmap those
> > > > special mappings. Currently, the only place which pass true to this
> > > > parameter is us.
> > > skip parameters are usually ugly and lead to more mess later on. Can we
> > > do without them?
> > We need a way to tell unmap_region() that it is called in a kind of special
> > context which updating vm_flags is not allowed. I didn't think of a better
> > way.
> >
> > We could add a new API to do what unmap_region() does without updating
> > vm_flags, but we would have to duplicate some code.
> I really didn't get to think about a better way myself but I strongly
> suspect we can do without special hacks here. Is updating flags under
> read lock a real problem? Assuming that special mappings are not really
> considered at this stage.

In normal case, I don't think vm_flags can be updated with read 
mmap_sem, but in this patch the vmas have been detached from the rb 
tree, nobody can find them anymore (I'm supposed all vma looking up is 
done by find_vma), so it might be safe to update vm_flags with read 
mmap_sem.

If it is safe, we don't have to have any special handling to those 
special mappings anymore.

> > > > With this 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.
> > > By with this approach you mean detaching right?
> > Yes, the detaching approach.
> Please make it explicit in the changelog.

Sure.

>   
> > > > 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.
> > > I guess you mean to say that lock downgrade approach doesn't lead to
> > > regressions because the overal time mmap_sem is taken is not longer?
> > Yes. And, there is not lock take/retake cost since we don't release it.
> Please also be explicit.

Sure.

>   
> > > > 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 for stability reason.
> > > You have used this argument previously and several people have asked.
> > > I think it is just wrong. Either the concept is safe and all callers can
> > > use it or it is not and then those subtle differences should be called
> > > out. Your previous response was that you simply haven't tested other
> > > paths. Well, that is not an argument, I am afraid. The whole thing
> > > should be done at a proper layer. If there are some difficulties to
> > > achieve that for all callers then OK just be explicit about that. I can
> > > imagine some callers really require the exclusive look when munmap
> > > returns for example.
> > Yes, the statement here sounds ambiguous. There are definitely some
> > difficulties to achieve that in mmap and mremap. Since they acquire write
> > mmap_sem at the very beginning, then do their stuff, which may call
> > do_munmap if overlapped address space has to be changed.
> Do call them out. Maybe even add a comment in the code so that people
> who would like those other paths know what they need to look at.

OK

> > 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.
> >
> > I assumes munmap itself is the main source of the latency issue. mmap/mremap
> > might hit the latency problem if they are trying to map or remap a huge
> > overlapped address space, but it should be rare. So, I leave them untouched.
> That depends on usecases very much. mremap might be called on very large
> areas as well. But let's go in smaller steps and build on top...

Yes, agree. And, I agree to achieve it step by step.

Thanks,
Yang