[Tarantool-patches] [PATCH luajit 2/2] Linux/ARM64: Make mremap() non-moving due to VA space woes.

[Sergey Ostanevich]

Hi!

Thanks for explanation, LGTM.


> On 2 Aug 2021, at 18:08, Sergey Kaplun <skaplun at tarantool.org> wrote:
> 
> Hi, Sergos!
> 
> Thanks for the review!
> 
> On 01.08.21, Sergey Ostanevich wrote:
>> Hi! Thanks for the patch!
>> 
>> Now direct_resize() will fail if it doesn't fit into the current place
>> with a new size and a direct_alloc() supposedly will be called. This one 
>> doesn't help with 47-bit address AFAIU since it has no extra option -
>> I doubt it exist at all - to ask kernel to fit.
>> 
>> So, how it helps?
> 
> For arm64 LJ_64 mode is enabled.
> It means LJ_ALLOC_MMAP_PROBE is defined and for each huge allocation
> `mmap_probe()` is called. This function tries several allocations
> unless the limit (30) is reached, or any suitable address is found.
> 
>> 
>> 
>> Regards,
>> Sergos
>> 
>> 
>>> On 1 Aug 2021, at 13:36, Igor Munkin <imun at tarantool.org> wrote:
>>> 
>>> Sergey,
>>> 
>>> Thanks for the fixes! LGTM, except the single typo.
>>> 
>>> On 28.07.21, Sergey Kaplun wrote:
>>> 
>>> <snipped>
>>> 
>>>> 
>>>> The new commit message is the following:
>>>> 
>>>> ===================================================================
>>>> Linux/ARM64: Make mremap() non-moving due to VA space woes.
>>>> 
>>>> This reduces overall performance on ARM64, but we have no choice.
>>>> Linux kernel default userspace VA is 48 bit, but we'd need 47 bit.
>>>> mremap() ignores address hints due to a kernel API issue. The mapping
>>>> may move to an undesired address which will cause an assert or crash.
>>>> 
>>>> Reported by Raymond W. Ko.
>>>> 
>>>> (cherry picked from commit 67dbec82f4f05a416a78a560a726553beaa7a223)
>>>> 
>>>> 47-bit VA space is required by LuaJIT for keeping a GC object pointer in
>>>> TValue. In case of huge blobs that are mapped directly, `mremap()` may
>>>> move the chunk out of 47-bit range of VA space on ARM64. `mremap()`
>>>> accepts the fifth argument (new address hint) only with MREMAP_FIXED
>>>> flag. In that case it unmaps any other mapping to specified address.
>>>> 
>>>> To avoid this behaviour this patch restricts `mremap()` to relocate
>>>> the mapping to a new virtual address by set CALL_MREMAP_NOMOVE flag
>>> 
>>> Typo: s/set/setting/.
>>> 
>>>> instead of CALL_MREMAP_MAYMOVE for arm64 architecture.
>>>> 
>>>> Sergey Kaplun:
>>>> * added the description and the test for the problem
>>>> 
>>>> Needed for tarantool/tarantool#6154
>>> 
>>> Minor: Why #5629 is not mentioned?
>>> 
>>>> ===================================================================
>>>> 
>>> 
>>> <snipped>
>>> 
>>>> 
>>>> -- 
>>>> Best regards,
>>>> Sergey Kaplun
>>> 
>>> -- 
>>> Best regards,
>>> IM
>> 
> 
> -- 
> Best regards,
> Sergey Kaplun