RFR: 8256155: 2M large pages for code when LargePageSizeInBytes is set to 1G for heap [v2]

Thomas Stuefe stuefe at openjdk.java.net
Thu Nov 19 13:11:10 UTC 2020


On Thu, 19 Nov 2020 08:19:59 GMT, Stefan Johansson <sjohanss at openjdk.org> wrote:

>> Hi Stefan,
>> 
>> Thanks so much for your review. 
>> 
>>> Hi and welcome :)
>>> 
>>> I haven't started reviewing the code in detail but a first quick glance raised a couple of questions/comments:
>>> 
>>> * Why do we have a special case for `exec` when selecting a large page size?
>> 
>> To my knowledge 2M is the smallest large pages size supported by Linux at the moment. Hardcoding 2M pages was an attempt to simplify the reservation of code memory using LargePages. In most cases currently code memory is reserved in default page size of the system when using 1G LargePages because it does not require 1G or larger reservations. In modern Linux variants default page size seems to be 4k on x86_64. In other architectures it could be up to 64k. The purpose of the patch is to enable the use of smaller LargePages for reservations less than 1G when LargePages are enabled and 1G is set as LargePageSizeInBytes, so as not to fall back to 4k-64k pages for these reservations.
>> 
>> Perhaps I should just select the page size <= bytes requested and remove 'exec' special case.
>> 
>>> * If we need the special case, `exec_large_page_size()` should not be hard code to return 2m but rather use `os::default_large_page_size()`.
>> 
>> os::default_large_page_size() will not necessarily be small enough for code memory reservations if the os::default_large_page_size() = 1G, in those cases we would get 4k on most linux x86_64 variants. My attempt is to ensure the smallest large_page_size availabe is used for code memory reservations. Perhaps my 2M hardcoding was a mistake and I should discover this size and select it based on the bytes being reserved.
>
>> Hi Stefan,
>> 
>> Thanks so much for your review.
>> 
>> > Hi and welcome :)
>> > I haven't started reviewing the code in detail but a first quick glance raised a couple of questions/comments:
>> > 
>> > * Why do we have a special case for `exec` when selecting a large page size?
>> 
>> To my knowledge 2M is the smallest large pages size supported by Linux at the moment. Hardcoding 2M pages was an attempt to simplify the reservation of code memory using LargePages. In most cases currently code memory is reserved in default page size of the system when using 1G LargePages because it does not require 1G or larger reservations. In modern Linux variants default page size seems to be 4k on x86_64. In other architectures it could be up to 64k. The purpose of the patch is to enable the use of smaller LargePages for reservations less than 1G when LargePages are enabled and 1G is set as LargePageSizeInBytes, so as not to fall back to 4k-64k pages for these reservations.
>> 
>> Perhaps I should just select the page size <= bytes requested and remove 'exec' special case.
>> 
> Yes, I see no reason to keep that special case and we want to keep this code as general as possible. Looking at the code in `os::Linux::find_default_large_page_size()` it looks like S390 supports 1M large pages, so we cannot assume 2M. I suggest using a technique similar to the one used in `os::Linux::find_large_page_size` to find supported page sizes. If you scan `/sys/kernel/mm/hugepages` and populate  `_page_sizes` using the information found we know we only get supported sizes. 
> 
>> > * If we need the special case, `exec_large_page_size()` should not be hard code to return 2m but rather use `os::default_large_page_size()`.
>> 
>> os::default_large_page_size() will not necessarily be small enough for code memory reservations if the os::default_large_page_size() = 1G, in those cases we would get 4k on most linux x86_64 variants. My attempt is to ensure the smallest large_page_size availabe is used for code memory reservations. Perhaps my 2M hardcoding was a mistake and I should discover this size and select it based on the bytes being reserved.
> 
> You are correct that the default size might indeed be 1G, so using something like I suggest above to figure  out the available page sizes and then using an appropriate one given the size of the mapping sounds good.
> 
> Please also avoid force pushing changes to open PRs since it makes it harder to follow what changes between updates. It is fine for a PR to contain multiple commits and if you need to update with things from the main branch you should merge rather than rebase. 
> 
> Cheers,
> Stefan

Hi Markus,

thanks, and a belated welcome!

Some initial background:

We at SAP are maintainers for a number of ports, among others AIX and linux ppc/s390 as well as some propietary ones (e.g. HPUX or ia64). So I wear my platform glasses when looking at this code.

IMHO the virtual memory layer in hotspot - os::reserve_memory() and all its friends - could do with a revamp. At least a consistent API documentation :-/. Supposed to be an API-independent abstraction, its facade breaks in many places. See e.g. JDK-8255978, JDK-8253649 (all windows), AIX sysV shmem handling, @AntonKozlov's valiant attempt to add MAP_JIT code heap reservation on MacOS (https://github.com/openjdk/jdk/pull/294), or the relative difficulty with which support for JEP 316 (from Intel) had been added.

Hence my initial caution. Every new feature increases complexity for us maintainers. Especially if it continues the bad tradition of not documenting or commenting anything. Since I do not know whether Intel sticks around to maintain this contribution (bit of a mixed track record there, see e.g. JDK-8256181), we must plan on maintenance falling to us.

That said, now that I understand better what you want to do, your plan certainly makes sense and is useful. 

One of the more pressing concerns I have is that the changes to reserve_memory() would somehow be observable from the outside and/or leak back into the os layer when calling os::commit_memory/uncommit_memory/release_memory. This is the case with @AntonKozlov's MAP_JIT change: it requires a matching commit call in os::commit_memory() to be made for executable memory allocated with os::reserve_memory(), and therefore exposed one weakness of the os::reserve_memory() API, that its very difficult to pass along meta information about memory mappings.

I think this is not the case here, but I'm not sure and we should be sure.

**More remarks inline.**

> Hi Thomas,
> 
> Thanks so much for your review. Please bear with me as this if my first patch to JDK community. But I have pushed patches to other open source communities (OpenDaylight, ONAP, OpenStack) and worked as a committer in some.
> 
> **Responses below inline:**
> 
> > Hi,
> > this seems like a improvement for a very specific scenario (2M instead of 1G pages on x86(?) Linux(?)). At the moment this feels more like an early prototype. The lack of comments/documentation is not helping.
> > Both JBS and PR are a bit taciturn. It would help if you could elaborate a bit. E.g. is this just for Linux? for x86 only? since the ticket talks about 4K pages, which are not universal across all architectures.
> 
> I appreciate the feedback. Perhaps the lack of detail in the pull request/JDK issue is a function of my zoomed focus on the specific purpose and lack of understanding about how much detail is normally included. The purpose of the patch/issue is to enable code hugepage memory reservations on Linux when the JDK is configured with 1G hugepages (LargePages in JDK parlance).

Please beef up the JBS issue a bit. If you do not have access to it, you can send the text to me I will update it. Or even easier, just update the PR description and we copy the text to the JBS.

JBS tickets are supposed to keep information about what we did and why for a long time. When formulating the text, just imagine the reader to be someone in the future with general knowledge in your field but without particular knowledge about this very case. I know this is a vague description though; for an example, see e.g. https://bugs.openjdk.java.net/browse/JDK-8255978.

> 
> To my knowledge, in most cases currently code memory is reserved in default page size of the system when using 1G LargePages because it does not require 1G or larger reservations. In modern Linux variants default page size seems to be 4k on x86_64. In other architectures it could be up to 64k. The purpose of the patch is to enable the use of smaller LargePages for reservations less than 1G when LargePages are enabled and 1G is set as LargePageSizeInBytes, so as not to fall back to 4k-64k pages for these reservations.

Right, and as Stefan suggested, this should be kept more "fluid" and not be hard coded to 2M, nor to just one additional large page. Maybe the system has four page sizes (our propietary HPUX has that, not that it matters here).

> 
> > I can glean some of what you want to do from the patch itself, but the spec is vague so there is no way to verify if the patch matches the spec.
> > What does page size have to do with exec permission? This should not be tied to exec. The whole patch should not contain the word "exec" :)
> 
> I'd appreciate any advice on writing a less vague spec. I have used exec as a stand-in for code memory reservations in my descriptions, mostly due to the fact that a 'bool exec' is used in functions that reserve HugePages and this later is translated into 'PROT_EXEC' when mmap is called, "exec" is passed in but not used in SHM. These are the particular memory reservations we wanted the patch to affect when using 1G LargePages. However I will remove those references if unwarranted.

<snip>

> 
> > What memory regions are supposed to be affected by this? JBS ticket talks about "code, card table and other".
> 
> Code is the target memory region. However there are some other instances where large_page reservation is happening due to the addition of 2M pages as an option. Some calls fail and error when adding 2M pages to _page_sizes array in the company of 1G pages. See line https://github.com/openjdk/jdk/pull/1153/files/daba99ac5f46dadb263caafa7ff87566d6d7dc58#diff-aeec57d804d56002f26a85359fc4ac8b48cfc249d57c656a30a63fc6bf3457adR3790
> This is where 2m pages are added.
> 
> However at https://github.com/openjdk/jdk/pull/1153/files/daba99ac5f46dadb263caafa7ff87566d6d7dc58#diff-aeec57d804d56002f26a85359fc4ac8b48cfc249d57c656a30a63fc6bf3457adR4077
> we get failures after the addition of 2M sizes to _page_sizes array, due to some smaller reservations that happen regardless of the LargePageSizeInBytes=1G
> 
> So in
> https://github.com/openjdk/jdk/pull/1153/files/daba99ac5f46dadb263caafa7ff87566d6d7dc58#diff-aeec57d804d56002f26a85359fc4ac8b48cfc249d57c656a30a63fc6bf3457adR4229
> we make sure that we select the largest page size that works with the bytes requested for reservation. Perhaps we shouldn't have exec as a special case, as the large page returned will be the same based on the size requested.


We need to decide on whether we want to do this for the code heap only or for every reservation done with reserve_memory_special (I really dislike that name btw). In your proposal you "piggyback" on the exec property as a standin for "code heap", which is not clean and also not necessarily true. So:

a) If we only want to do this for the code heap, we could think about creating an own API for allocating the code heap. E.g. os::reserve_code_space() and os::release_code_space(). This is one of the ideas @AntonKozlov came up with to circumvent the need for a fully fledged revamp of these APIs while still being able to move his PR forward.

b) If we want to do this for all callers of reserve_memory_special(), we should also remove any mention of "exec" and just implement that.

I currently favour (b) but would like to know opinions of others.

> 
> > One problem I see is that the notion of "we have a small standard page and a single large pinned page size" is - I believe - baked in into a few places. Are there any places where an implicit assumption of the page size or their "pinned-ness" could break things now (see also below remark about UseSHM)? For instance, are these pages pinned on all our platforms, and if no, could code be affected which commits/uncommits and assumes a certain page size?
> > What tests have you run? On what platforms? Also platforms with different page sizes? How well did you test UseSHM?
> 
> My architecture knowledge outside of x86_64 is limited. I've been looking into this and thinking about it and I will have some more comments in the next day or so. For UseSHM I ran some negative tests but I will do some more rigorous testing and report back.

Okay. We do not expect every contributor to have exotic test machines, but this means we will have to do that testing. We need to know to plan in these efforts.

> 
> > The latter is interesting because arguably there is the bigger behavioral change. TLBFS path was using a mixture of large and small pages anyway, so adding another page size into the mix is not a big stretch. But for SHM, things would change: where before reserve_memory_special would return NULL and we'd invoke fallback reservation, now we return a region consisting of 2M pinned pages.
> > For SHM, I think you need to make sure that alignment matches SHMLBA?
> 
> Looking into this.
> 
> > It was not clear from the patch or the JBS item whether you propose to change the semantics of LargePageSizeInBytes. E.g. what happens if the value specified explicitely is smaller than your exec_page size? Your patch seems to give preference to exec_page_size. But this would be a behavioral change, and may need a CSR.
> 
> I'm open to removing exec references and just enabling multiple page sizes which would allow for 2M pages to be used by code memory reservations.
> 
> > Finally, comments would be nice. Clear API specs. Extending regression tests for reserve_memory_special would be good too, to test the new behavior (for gtests examples, see test/hotspot/gtest/runtime in the source folder).
> 
> Thanks. I will push an updated patch w/ comments. Will attempt clear API spec and regression tests for reserve_memory_special but will need some guidance on those.

When I write API specs I basically mean "new code should comment better". That can be as simple as a one liner above your os::Linux::select_large_page_size() function. 

About regression tests, we have a google-test suite (see test/hotspot/gtest) which would be the appropiate point to put in tests. 

> 
> > The linux-2m-page-specific code in the platform-generic G1 test seems wrong.
> 
> Any advice here. My change specifically changes the behavior of the pages returned in the test for linux platforms but should not have effects on other platforms. I don't know how this would generally happen for JDK tests in this case. It seems to me that the JDK will act differently on different platforms. How is this normally handled?

I defer to the G1 folks for that.

> 
> > Cheers, Thomas
> 
> Thanks again for the review.

Sure. Thanks for the much more clear information.

Cheers, Thomas

-------------

PR: https://git.openjdk.java.net/jdk/pull/1153


More information about the hotspot-gc-dev mailing list