Re: Very slow unlockall()

[Vlastimil Babka <vbabka@suse.cz>]

On 2/1/21 8:19 PM, Milan Broz wrote:
> On 01/02/2021 19:55, Vlastimil Babka wrote:
>> On 2/1/21 7:00 PM, Milan Broz wrote:
>>> On 01/02/2021 14:08, Vlastimil Babka wrote:
>>>> On 1/8/21 3:39 PM, Milan Broz wrote:
>>>>> On 08/01/2021 14:41, Michal Hocko wrote:
>>>>>> On Wed 06-01-21 16:20:15, Milan Broz wrote:
>>>>>>> Hi,
>>>>>>>
>>>>>>> we use mlockall(MCL_CURRENT | MCL_FUTURE) / munlockall() in cryptsetup code
>>>>>>> and someone tried to use it with hardened memory allocator library.
>>>>>>>
>>>>>>> Execution time was increased to extreme (minutes) and as we found, the problem
>>>>>>> is in munlockall().
>>>>>>>
>>>>>>> Here is a plain reproducer for the core without any external code - it takes
>>>>>>> unlocking on Fedora rawhide kernel more than 30 seconds!
>>>>>>> I can reproduce it on 5.10 kernels and Linus' git.
>>>>>>>
>>>>>>> The reproducer below tries to mmap large amount memory with PROT_NONE (later never used).
>>>>>>> The real code of course does something more useful but the problem is the same.
>>>>>>>
>>>>>>> #include <stdio.h>
>>>>>>> #include <stdlib.h>
>>>>>>> #include <fcntl.h>
>>>>>>> #include <sys/mman.h>
>>>>>>>
>>>>>>> int main (int argc, char *argv[])
>>>>>>> {
>>>>>>>         void *p  = mmap(NULL, 1UL << 41, PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
>> 
>> So, this is 2TB memory area, but PROT_NONE means it's never actually populated,
>> although mlockall(MCL_CURRENT) should do that. Once you put PROT_READ |
>> PROT_WRITE there, the mlockall() starts taking ages.
>> 
>> So does that reflect your use case? munlockall() with large PROT_NONE areas? If
>> so, munlock_vma_pages_range() is indeed not optimized for that, but I would
>> expect such scenario to be uncommon, so better clarify first.
> 
> It is just a simple reproducer of the underlying problem, as suggested here 
> https://gitlab.com/cryptsetup/cryptsetup/-/issues/617#note_478342301
> 
> We use mlockall() in cryptsetup and with hardened malloc it slows down unlock significantly.
> (For the real case problem please read the whole issue report above.)

OK, finally read through the bug report, and learned two things:

1) the PROT_NONE is indeed intentional part of the reproducer
2) Linux mailing lists still have a bad reputation and people avoid them. That's
sad :( Well, thanks for overcoming that :)

Daniel there says "I think the Linux kernel implementation of mlockall is quite
broken and tries to lock all the reserved PROT_NONE regions in advance which
doesn't make any sense."

From my testing this doesn't seem to be the case, as the mlockall() part is very
fast, so I don't think it faults in and mlocks PROT_NONE areas. It only starts
to be slow when changed to PROT_READ|PROT_WRITE. But the munlockall() part is
slow even with PROT_NONE as we don't skip the PROT_NONE areas there. We probably
can't just skip them, as they might actually contain mlocked pages if those were
faulted first with PROT_READ/PROT_WRITE and only then changed to PROT_NONE.

And the munlock (munlock_vma_pages_range()) is slow, because it uses
follow_page_mask() in a loop incrementing addresses by PAGE_SIZE, so that's
always traversing all levels of page tables from scratch. Funnily enough,
speeding this up was my first linux-mm series years ago. But the speedup only
works if pte's are present, which is not the case for unpopulated PROT_NONE
areas. That use case was unexpected back then. We should probably convert this
code to a proper page table walk. If there are large areas with unpopulated pmd
entries (or even higher levels) we would traverse them very quickly.