Message ID | 20241014105514.3206191-1-ryan.roberts@arm.com (mailing list archive) |
---|---|
Headers | show |
Series | Boot-time page size selection for arm64 | expand |
On 10/14/24 03:55, Ryan Roberts wrote: > Hi All, > > Patch bomb incoming... This covers many subsystems, so I've included a core set > of people on the full series and additionally included maintainers on relevant > patches. I haven't included those maintainers on this cover letter since the > numbers were far too big for it to work. But I've included a link to this cover > letter on each patch, so they can hopefully find their way here. For follow up > submissions I'll break it up by subsystem, but for now thought it was important > to show the full picture. > > This RFC series implements support for boot-time page size selection within the > arm64 kernel. arm64 supports 3 base page sizes (4K, 16K, 64K), but to date, page > size has been selected at compile-time, meaning the size is baked into a given > kernel image. As use of larger-than-4K page sizes become more prevalent this > starts to present a problem for distributions. Boot-time page size selection > enables the creation of a single kernel image, which can be told which page size > to use on the kernel command line. > > Why is having an image-per-page size problematic? > ================================================= > > Many traditional distros are now supporting both 4K and 64K. And this means > managing 2 kernel packages, along with drivers for each. For some, it means > multiple installer flavours and multiple ISOs. All of this adds up to a > less-than-ideal level of complexity. Additionally, Android now supports 4K and > 16K kernels. I'm told having to explicitly manage their KABI for each kernel is > painful, and the extra flash space required for both kernel images and the > duplicated modules has been problematic. Boot-time page size selection solves > all of this. > > Additionally, in starting to think about the longer term deployment story for > D128 page tables, which Arm architecture now supports, a lot of the same > problems need to be solved, so this work sets us up nicely for that. > > So what's the down side? > ======================== > > Well nothing's free; Various static allocations in the kernel image must be > sized for the worst case (largest supported page size), so image size is in line > with size of 64K compile-time image. So if you're interested in 4K or 16K, there > is a slight increase to the image size. But I expect that problem goes away if > you're compressing the image - its just some extra zeros. At boot-time, I expect > we could free the unused static storage once we know the page size - although > that would be a follow up enhancement. > > And then there is performance. Since PAGE_SIZE and friends are no longer > compile-time constants, we must look up their values and do arithmetic at > runtime instead of compile-time. My early perf testing suggests this is > inperceptible for real-world workloads, and only has small impact on > microbenchmarks - more on this below. > > Approach > ======== > > The basic idea is to rid the source of any assumptions that PAGE_SIZE and > friends are compile-time constant, but in a way that allows the compiler to > perform the same optimizations as was previously being done if they do turn out > to be compile-time constant. Where constants are required, we use limits; > PAGE_SIZE_MIN and PAGE_SIZE_MAX. See commit log in patch 1 for full description > of all the classes of problems to solve. > > By default PAGE_SIZE_MIN=PAGE_SIZE_MAX=PAGE_SIZE. But an arch may opt-in to > boot-time page size selection by defining PAGE_SIZE_MIN & PAGE_SIZE_MAX. arm64 > does this if the user selects the CONFIG_ARM64_BOOT_TIME_PAGE_SIZE Kconfig, > which is an alternative to selecting a compile-time page size. > > When boot-time page size is active, the arch pgtable geometry macro definitions > resolve to something that can be configured at boot. The arm64 implementation in > this series mainly uses global, __ro_after_init variables. I've tried using > alternatives patching, but that performs worse than loading from memory; I think > due to code size bloat. FWIW, this paragraph was not entirely clear to me until I looked at patch 57 to see that the compile time page size selection had been retained, and could continue to be used as-is. It was somewhat implicit, but not IMHO explicit enough, not a big deal though. Great work, thanks for doing that! This makes me wonder if we could leverage any of that to have a single kernel supporting both LPAE and !LPAE on ARM 32-bit, but that still seems like somewhat more difficult, largely due to the difference in the page table descriptor format (long vs. short).
On 14/10/2024 18:32, Florian Fainelli wrote: > On 10/14/24 03:55, Ryan Roberts wrote: >> Hi All, >> >> Patch bomb incoming... This covers many subsystems, so I've included a core set >> of people on the full series and additionally included maintainers on relevant >> patches. I haven't included those maintainers on this cover letter since the >> numbers were far too big for it to work. But I've included a link to this cover >> letter on each patch, so they can hopefully find their way here. For follow up >> submissions I'll break it up by subsystem, but for now thought it was important >> to show the full picture. >> >> This RFC series implements support for boot-time page size selection within the >> arm64 kernel. arm64 supports 3 base page sizes (4K, 16K, 64K), but to date, page >> size has been selected at compile-time, meaning the size is baked into a given >> kernel image. As use of larger-than-4K page sizes become more prevalent this >> starts to present a problem for distributions. Boot-time page size selection >> enables the creation of a single kernel image, which can be told which page size >> to use on the kernel command line. >> >> Why is having an image-per-page size problematic? >> ================================================= >> >> Many traditional distros are now supporting both 4K and 64K. And this means >> managing 2 kernel packages, along with drivers for each. For some, it means >> multiple installer flavours and multiple ISOs. All of this adds up to a >> less-than-ideal level of complexity. Additionally, Android now supports 4K and >> 16K kernels. I'm told having to explicitly manage their KABI for each kernel is >> painful, and the extra flash space required for both kernel images and the >> duplicated modules has been problematic. Boot-time page size selection solves >> all of this. >> >> Additionally, in starting to think about the longer term deployment story for >> D128 page tables, which Arm architecture now supports, a lot of the same >> problems need to be solved, so this work sets us up nicely for that. >> >> So what's the down side? >> ======================== >> >> Well nothing's free; Various static allocations in the kernel image must be >> sized for the worst case (largest supported page size), so image size is in line >> with size of 64K compile-time image. So if you're interested in 4K or 16K, there >> is a slight increase to the image size. But I expect that problem goes away if >> you're compressing the image - its just some extra zeros. At boot-time, I expect >> we could free the unused static storage once we know the page size - although >> that would be a follow up enhancement. >> >> And then there is performance. Since PAGE_SIZE and friends are no longer >> compile-time constants, we must look up their values and do arithmetic at >> runtime instead of compile-time. My early perf testing suggests this is >> inperceptible for real-world workloads, and only has small impact on >> microbenchmarks - more on this below. >> >> Approach >> ======== >> >> The basic idea is to rid the source of any assumptions that PAGE_SIZE and >> friends are compile-time constant, but in a way that allows the compiler to >> perform the same optimizations as was previously being done if they do turn out >> to be compile-time constant. Where constants are required, we use limits; >> PAGE_SIZE_MIN and PAGE_SIZE_MAX. See commit log in patch 1 for full description >> of all the classes of problems to solve. >> >> By default PAGE_SIZE_MIN=PAGE_SIZE_MAX=PAGE_SIZE. But an arch may opt-in to >> boot-time page size selection by defining PAGE_SIZE_MIN & PAGE_SIZE_MAX. arm64 >> does this if the user selects the CONFIG_ARM64_BOOT_TIME_PAGE_SIZE Kconfig, >> which is an alternative to selecting a compile-time page size. >> >> When boot-time page size is active, the arch pgtable geometry macro definitions >> resolve to something that can be configured at boot. The arm64 implementation in >> this series mainly uses global, __ro_after_init variables. I've tried using >> alternatives patching, but that performs worse than loading from memory; I think >> due to code size bloat. > > FWIW, this paragraph was not entirely clear to me until I looked at patch 57 to > see that the compile time page size selection had been retained, and could > continue to be used as-is. It was somewhat implicit, but not IMHO explicit > enough, not a big deal though. I intended to make that bit clear with the above sentance "arm64 does this if the user selects the CONFIG_ARM64_BOOT_TIME_PAGE_SIZE Kconfig, which is an alternative to selecting a compile-time page size.", but appreciate there is a lot going on here. > > Great work, thanks for doing that! This makes me wonder if we could leverage any > of that to have a single kernel supporting both LPAE and !LPAE on ARM 32-bit, > but that still seems like somewhat more difficult, largely due to the difference > in the page table descriptor format (long vs. short). We will eventually have the exact same problem with FEAT_D128 on arm64. This introduces page tables with 128 bit PTEs. Ideally we would like to support both in a single image, although, we have much more thinking to do on that. But my current view is that this series solves a bunch of problems that makes it easier (PTRS_PER_Pxx and Pxx_SHIFT all become boot-time values, for example, so we can easily represent the different geometries). Yes, we still need to solve the PTE size difference (in our case 64-bit vs 128-bit). I have a couple of proposals for how to do that; the "gold-plated" approach would be to create and use a handle type to represent a PTE/PxD slot in a table. Then increments/decrements would be enforced via explicit helpers that know the size, and direct dereferencing would be impossible. When accessing via helpers we would pass around pte_t/pxd_t values that are the larger size, then narrow then when writing back. Anshuman has a series [1] that starts to move in that direction. If you have any other ideas, it would be good to talk! [1] https://lore.kernel.org/linux-mm/20240917073117.1531207-1-anshuman.khandual@arm.com/ Thanks, Ryan
From: Ryan Roberts <ryan.roberts@arm.com> Sent: Monday, October 14, 2024 3:55 AM > > Hi All, > > Patch bomb incoming... This covers many subsystems, so I've included a core set > of people on the full series and additionally included maintainers on relevant > patches. I haven't included those maintainers on this cover letter since the > numbers were far too big for it to work. But I've included a link to this cover > letter on each patch, so they can hopefully find their way here. For follow up > submissions I'll break it up by subsystem, but for now thought it was important > to show the full picture. > > This RFC series implements support for boot-time page size selection within the > arm64 kernel. arm64 supports 3 base page sizes (4K, 16K, 64K), but to date, page > size has been selected at compile-time, meaning the size is baked into a given > kernel image. As use of larger-than-4K page sizes become more prevalent this > starts to present a problem for distributions. Boot-time page size selection > enables the creation of a single kernel image, which can be told which page size > to use on the kernel command line. > > Why is having an image-per-page size problematic? > ================================================= > > Many traditional distros are now supporting both 4K and 64K. And this means > managing 2 kernel packages, along with drivers for each. For some, it means > multiple installer flavours and multiple ISOs. All of this adds up to a > less-than-ideal level of complexity. Additionally, Android now supports 4K and > 16K kernels. I'm told having to explicitly manage their KABI for each kernel is > painful, and the extra flash space required for both kernel images and the > duplicated modules has been problematic. Boot-time page size selection solves > all of this. > > Additionally, in starting to think about the longer term deployment story for > D128 page tables, which Arm architecture now supports, a lot of the same > problems need to be solved, so this work sets us up nicely for that. > > So what's the down side? > ======================== > > Well nothing's free; Various static allocations in the kernel image must be > sized for the worst case (largest supported page size), so image size is in line > with size of 64K compile-time image. So if you're interested in 4K or 16K, there > is a slight increase to the image size. But I expect that problem goes away if > you're compressing the image - its just some extra zeros. At boot-time, I expect > we could free the unused static storage once we know the page size - although > that would be a follow up enhancement. > > And then there is performance. Since PAGE_SIZE and friends are no longer > compile-time constants, we must look up their values and do arithmetic at > runtime instead of compile-time. My early perf testing suggests this is > inperceptible for real-world workloads, and only has small impact on > microbenchmarks - more on this below. [snip] This is pretty cool. :-) FWIW, I've built a kernel with this patch set, and have it running in a RHEL 8.7 guest on Hyper-V in the Azure public cloud. Ran with 4K, 16K, and 64K page sizes, and the basic smoke tests work. The Hyper-V specific code in the Linux kernel needed a few tweaks to deal with PAGE_SIZE and friends no longer being constant, but it's nothing significant. Getting the kernel built in the first place was a little harder because my .config file is fairly generic with a lot of device drivers and file system code that aren't really needed for Hyper-V guests. I had to weed out the ones that won't build. My RHEL 8.7 install uses LVM, so I hacked the 'dm' code to make it compile and run. As this work moves forward, I can supply the necessary patches for the Hyper-V support. Let me know if you want to include them in the main patch set. I've added a couple of Microsoft's Linux people to this email's addressee list so they are aware of what's going on. Michael Kelley
On 15/10/2024 19:38, Michael Kelley wrote: > From: Ryan Roberts <ryan.roberts@arm.com> Sent: Monday, October 14, 2024 3:55 AM >> >> Hi All, >> >> Patch bomb incoming... This covers many subsystems, so I've included a core set >> of people on the full series and additionally included maintainers on relevant >> patches. I haven't included those maintainers on this cover letter since the >> numbers were far too big for it to work. But I've included a link to this cover >> letter on each patch, so they can hopefully find their way here. For follow up >> submissions I'll break it up by subsystem, but for now thought it was important >> to show the full picture. >> >> This RFC series implements support for boot-time page size selection within the >> arm64 kernel. arm64 supports 3 base page sizes (4K, 16K, 64K), but to date, page >> size has been selected at compile-time, meaning the size is baked into a given >> kernel image. As use of larger-than-4K page sizes become more prevalent this >> starts to present a problem for distributions. Boot-time page size selection >> enables the creation of a single kernel image, which can be told which page size >> to use on the kernel command line. >> >> Why is having an image-per-page size problematic? >> ================================================= >> >> Many traditional distros are now supporting both 4K and 64K. And this means >> managing 2 kernel packages, along with drivers for each. For some, it means >> multiple installer flavours and multiple ISOs. All of this adds up to a >> less-than-ideal level of complexity. Additionally, Android now supports 4K and >> 16K kernels. I'm told having to explicitly manage their KABI for each kernel is >> painful, and the extra flash space required for both kernel images and the >> duplicated modules has been problematic. Boot-time page size selection solves >> all of this. >> >> Additionally, in starting to think about the longer term deployment story for >> D128 page tables, which Arm architecture now supports, a lot of the same >> problems need to be solved, so this work sets us up nicely for that. >> >> So what's the down side? >> ======================== >> >> Well nothing's free; Various static allocations in the kernel image must be >> sized for the worst case (largest supported page size), so image size is in line >> with size of 64K compile-time image. So if you're interested in 4K or 16K, there >> is a slight increase to the image size. But I expect that problem goes away if >> you're compressing the image - its just some extra zeros. At boot-time, I expect >> we could free the unused static storage once we know the page size - although >> that would be a follow up enhancement. >> >> And then there is performance. Since PAGE_SIZE and friends are no longer >> compile-time constants, we must look up their values and do arithmetic at >> runtime instead of compile-time. My early perf testing suggests this is >> inperceptible for real-world workloads, and only has small impact on >> microbenchmarks - more on this below. > > [snip] > > This is pretty cool. :-) FWIW, I've built a kernel with this patch set, and > have it running in a RHEL 8.7 guest on Hyper-V in the Azure public cloud. > Ran with 4K, 16K, and 64K page sizes, and the basic smoke tests work. That's great to hear - thanks for taking the time to test! > > The Hyper-V specific code in the Linux kernel needed a few tweaks to > deal with PAGE_SIZE and friends no longer being constant, but it's nothing > significant. Getting the kernel built in the first place was a little harder > because my .config file is fairly generic with a lot of device drivers and file > system code that aren't really needed for Hyper-V guests. I had to > weed out the ones that won't build. My RHEL 8.7 install uses LVM, so I> hacked the 'dm' code to make it compile and run. Yeah, getting all this sorted is going to be the long tail. I feel I've had enough positive response to this RFC that I should probably just get on and start that work to get a real feel for how much of it there is going to be. > > As this work moves forward, I can supply the necessary patches for > the Hyper-V support. Let me know if you want to include them in the > main patch set. Great! If you are happy to forward them to me, I'll include them in future versions of the series (or more likely, serieses). Thanks, Ryan > > I've added a couple of Microsoft's Linux people to this email's addressee > list so they are aware of what's going on. > > Michael Kelley
> Performance Testing > =================== > > I've run some limited performance benchmarks: > > First, a real-world benchmark that causes a lot of page table manipulation (and > therefore we would expect to see regression here if we are going to see it > anywhere); kernel compilation. It barely registers a change. Values are times, > so smaller is better. All relative to base-4k: > > | | kern | kern | user | user | real | real | > | config | mean | stdev | mean | stdev | mean | stdev | > |-------------|---------|---------|---------|---------|---------|---------| > | base-4k | 0.0% | 1.1% | 0.0% | 0.3% | 0.0% | 0.3% | > | compile-4k | -0.2% | 1.1% | -0.2% | 0.3% | -0.1% | 0.3% | > | boot-4k | 0.1% | 1.0% | -0.3% | 0.2% | -0.2% | 0.2% | > > The Speedometer JavaScript benchmark also shows no change. Values are runs per > min, so bigger is better. All relative to base-4k: > > | config | mean | stdev | > |-------------|---------|---------| > | base-4k | 0.0% | 0.8% | > | compile-4k | 0.4% | 0.8% | > | boot-4k | 0.0% | 0.9% | > > Finally, I've run some microbenchmarks known to stress page table manipulations > (originally from David Hildenbrand). The fork test maps/allocs 1G of anon > memory, then measures the cost of fork(). The munmap test maps/allocs 1G of anon > memory then measures the cost of munmap()ing it. The fork test is known to be > extremely sensitive to any changes that cause instructions to be aligned > differently in cachelines. When using this test for other changes, I've seen > double digit regressions for the slightest thing, so 12% regression on this test > is actually fairly good. This likely represents the extreme worst case for > regressions that will be observed across other microbenchmarks (famous last > words). Values are times, so smaller is better. All relative to base-4k: > ... and here I am, worrying about much smaller degradation in these micro-benchmark ;) You're right, these are pure micro-benchmarks, and while 12% does sound like "much", even stupid compiler code movement can result in such changes in the fork() micro benchmark. So I think this is just fine, and actually "surprisingly" small. And, there is even a way to statically compile a page size and not worry about that at all. As discussed ahead of times, I consider this change very valuable. In RHEL, the biggest issue is actually the test matrix, that cannot really be reduced significantly ... but it will make shipping/packaging easier. CCing Don, who did the separate 64k RHEL flavor kernel.
On 16/10/2024 16:16, David Hildenbrand wrote: >> Performance Testing >> =================== >> >> I've run some limited performance benchmarks: >> >> First, a real-world benchmark that causes a lot of page table manipulation (and >> therefore we would expect to see regression here if we are going to see it >> anywhere); kernel compilation. It barely registers a change. Values are times, >> so smaller is better. All relative to base-4k: >> >> | | kern | kern | user | user | real | real | >> | config | mean | stdev | mean | stdev | mean | stdev | >> |-------------|---------|---------|---------|---------|---------|---------| >> | base-4k | 0.0% | 1.1% | 0.0% | 0.3% | 0.0% | 0.3% | >> | compile-4k | -0.2% | 1.1% | -0.2% | 0.3% | -0.1% | 0.3% | >> | boot-4k | 0.1% | 1.0% | -0.3% | 0.2% | -0.2% | 0.2% | >> >> The Speedometer JavaScript benchmark also shows no change. Values are runs per >> min, so bigger is better. All relative to base-4k: >> >> | config | mean | stdev | >> |-------------|---------|---------| >> | base-4k | 0.0% | 0.8% | >> | compile-4k | 0.4% | 0.8% | >> | boot-4k | 0.0% | 0.9% | >> >> Finally, I've run some microbenchmarks known to stress page table manipulations >> (originally from David Hildenbrand). The fork test maps/allocs 1G of anon >> memory, then measures the cost of fork(). The munmap test maps/allocs 1G of anon >> memory then measures the cost of munmap()ing it. The fork test is known to be >> extremely sensitive to any changes that cause instructions to be aligned >> differently in cachelines. When using this test for other changes, I've seen >> double digit regressions for the slightest thing, so 12% regression on this test >> is actually fairly good. This likely represents the extreme worst case for >> regressions that will be observed across other microbenchmarks (famous last >> words). Values are times, so smaller is better. All relative to base-4k: >> > > ... and here I am, worrying about much smaller degradation in these micro- > benchmark ;) You're right, these are pure micro-benchmarks, and while 12% does > sound like "much", even stupid compiler code movement can result in such changes > in the fork() micro benchmark. > > So I think this is just fine, and actually "surprisingly" small. And, there is > even a way to statically compile a page size and not worry about that at all. > > As discussed ahead of times, I consider this change very valuable. In RHEL, the > biggest issue is actually the test matrix, that cannot really be reduced > significantly ... but it will make shipping/packaging easier. > > CCing Don, who did the separate 64k RHEL flavor kernel. > Thanks, David! I'm planning to investigate and see if I can improve even on that 12%. I have a couple of ideas. But like you say, I don't think this should be a blocker to moving forwards.
On Mon, 14 Oct 2024 11:55:11 +0100 Ryan Roberts <ryan.roberts@arm.com> wrote: >[...] > The series is arranged as follows: > > - patch 1: Add macros required for converting non-arch code to support > boot-time page size selection > - patches 2-36: Remove PAGE_SIZE compile-time constant assumption from all > non-arch code I have just tried to recompile the openSUSE kernel with these patches applied, and I'm running into this: CC arch/arm64/hyperv/hv_core.o In file included from ../arch/arm64/hyperv/hv_core.c:14:0: ../include/linux/hyperv.h:158:5: error: variably modified ‘reserved2’ at file scope u8 reserved2[PAGE_SIZE - 68]; ^~~~~~~~~ It looks like one more place which needs a patch, right? Petr T
On 17/10/2024 13:27, Petr Tesarik wrote: > On Mon, 14 Oct 2024 11:55:11 +0100 > Ryan Roberts <ryan.roberts@arm.com> wrote: > >> [...] >> The series is arranged as follows: >> >> - patch 1: Add macros required for converting non-arch code to support >> boot-time page size selection >> - patches 2-36: Remove PAGE_SIZE compile-time constant assumption from all >> non-arch code > > I have just tried to recompile the openSUSE kernel with these patches > applied, and I'm running into this: > > CC arch/arm64/hyperv/hv_core.o > In file included from ../arch/arm64/hyperv/hv_core.c:14:0: > ../include/linux/hyperv.h:158:5: error: variably modified ‘reserved2’ at file scope > u8 reserved2[PAGE_SIZE - 68]; > ^~~~~~~~~ > > It looks like one more place which needs a patch, right? As mentioned in the cover letter, so far I've only converted enough to get the defconfig *image* building (i.e. no modules). If you are compiling a different config or compiling the modules for defconfig, you will likely run into these types of issues. That said, I do have some patches to fix Hyper-V, which Michael Kelley was kind enough to send me. I understand that Suse might be able to help with wider performance testing - if that's the reason you are trying to compile, you could send me your config and I'll start working on fixing up other drivers? Thanks, Ryan > > Petr T
On 10/14/24 5:55AM, Ryan Roberts wrote: > Hi All, > > Patch bomb incoming... This covers many subsystems, so I've included a core set > of people on the full series and additionally included maintainers on relevant > patches. I haven't included those maintainers on this cover letter since the > numbers were far too big for it to work. But I've included a link to this cover > letter on each patch, so they can hopefully find their way here. For follow up > submissions I'll break it up by subsystem, but for now thought it was important > to show the full picture. > > This RFC series implements support for boot-time page size selection within the > arm64 kernel. arm64 supports 3 base page sizes (4K, 16K, 64K), but to date, page > size has been selected at compile-time, meaning the size is baked into a given > kernel image. As use of larger-than-4K page sizes become more prevalent this > starts to present a problem for distributions. Boot-time page size selection > enables the creation of a single kernel image, which can be told which page size > to use on the kernel command line. This looks really promising. Building and maintaining separate kernels is costly. Being able to build one kernel for three protential page sizes would not only cut down on the overhead of producing kernel packages and images, but also eases benchmarking and testing different page sizes without the need to build and install multiple kernels. I'm also impressed that the patches are less intrusive than I would have expected. I'm looking forward to seeing this project move forward. Thanks, Shaggy > > Why is having an image-per-page size problematic? > ================================================= > > Many traditional distros are now supporting both 4K and 64K. And this means > managing 2 kernel packages, along with drivers for each. For some, it means > multiple installer flavours and multiple ISOs. All of this adds up to a > less-than-ideal level of complexity. Additionally, Android now supports 4K and > 16K kernels. I'm told having to explicitly manage their KABI for each kernel is > painful, and the extra flash space required for both kernel images and the > duplicated modules has been problematic. Boot-time page size selection solves > all of this. > > Additionally, in starting to think about the longer term deployment story for > D128 page tables, which Arm architecture now supports, a lot of the same > problems need to be solved, so this work sets us up nicely for that. > > So what's the down side? > ======================== > > Well nothing's free; Various static allocations in the kernel image must be > sized for the worst case (largest supported page size), so image size is in line > with size of 64K compile-time image. So if you're interested in 4K or 16K, there > is a slight increase to the image size. But I expect that problem goes away if > you're compressing the image - its just some extra zeros. At boot-time, I expect > we could free the unused static storage once we know the page size - although > that would be a follow up enhancement. > > And then there is performance. Since PAGE_SIZE and friends are no longer > compile-time constants, we must look up their values and do arithmetic at > runtime instead of compile-time. My early perf testing suggests this is > inperceptible for real-world workloads, and only has small impact on > microbenchmarks - more on this below. > > Approach > ======== > > The basic idea is to rid the source of any assumptions that PAGE_SIZE and > friends are compile-time constant, but in a way that allows the compiler to > perform the same optimizations as was previously being done if they do turn out > to be compile-time constant. Where constants are required, we use limits; > PAGE_SIZE_MIN and PAGE_SIZE_MAX. See commit log in patch 1 for full description > of all the classes of problems to solve. > > By default PAGE_SIZE_MIN=PAGE_SIZE_MAX=PAGE_SIZE. But an arch may opt-in to > boot-time page size selection by defining PAGE_SIZE_MIN & PAGE_SIZE_MAX. arm64 > does this if the user selects the CONFIG_ARM64_BOOT_TIME_PAGE_SIZE Kconfig, > which is an alternative to selecting a compile-time page size. > > When boot-time page size is active, the arch pgtable geometry macro definitions > resolve to something that can be configured at boot. The arm64 implementation in > this series mainly uses global, __ro_after_init variables. I've tried using > alternatives patching, but that performs worse than loading from memory; I think > due to code size bloat. > > Status > ====== > > When CONFIG_ARM64_BOOT_TIME_PAGE_SIZE is selected, I've only implemented enough > to compile the kernel image itself with defconfig (and a few other bits and > pieces). This is enough to build a kernel that can boot under QEMU or FVP. I'll > happily do the rest of the work to enable all the extra drivers, but wanted to > get feedback on the shape of this effort first. If anyone wants to do any > testing, and has a must-have config, let me know and I'll prioritize enabling it > first. > > The series is arranged as follows: > > - patch 1: Add macros required for converting non-arch code to support > boot-time page size selection > - patches 2-36: Remove PAGE_SIZE compile-time constant assumption from all > non-arch code > - patches 37-38: Some arm64 tidy ups > - patch 39: Add macros required for converting arm64 code to support > boot-time page size selection > - patches 40-56: arm64 changes to support boot-time page size selection > - patch 57: Add arm64 Kconfig option to enable boot-time page size > selection > > Ideally, I'd like to get the basics merged (something like this series), then > incrementally improve it over a handful of kernel releases until we can > demonstrate that we have feature parity with the compile-time build and no > performance blockers. Once at that point, ideally the compile-time build options > would be removed and the code could be cleaned up further. > > One of the bigger peices that I'd propose to add as a follow up, is to make > va-size boot-time selectable too. That will greatly simplify LPA2 fallback > handling. > > Assuming people are ammenable to the rough shape, how would I go about getting > the non-arch changes merged? Since they cover many subsystems, will each piece > need to go independently to each relevant maintainer or could it all be merged > together through the arm64 tree? > > Image Size > ========== > > The below shows the size of a defconfig (+ xfs, squashfs, ftrace, kprobes) > kernel image on disk for base (before any changes applied), compile (with > changes, configured for compile-time page size) and boot (with changes, > configured for boot-time page size). > > You can see the that compile-16k and 64k configs are actually slightly smaller > than the baselines; that's due to optimizing some buffer sizes which didn't need > to depend on page size during the series. The boot-time image is ~1% bigger than > the 64k compile-time image. I believe there is scope to improve this to make it > equal to compile-64k if required: > > | config | size/KB | diff/KB | diff/% | > |-------------|---------|---------|---------| > | base-4k | 54895 | 0 | 0.0% | > | base-16k | 55161 | 266 | 0.5% | > | base-64k | 56775 | 1880 | 3.4% | > | compile-4k | 54895 | 0 | 0.0% | > | compile-16k | 55097 | 202 | 0.4% | > | compile-64k | 56391 | 1496 | 2.7% | > | boot-4K | 57045 | 2150 | 3.9% | > > And below shows the size of the image in memory at run-time, separated for text > and data costs. The boot image has ~1% text cost; most likely due to the fact > that PAGE_SIZE and friends are not compile-time constants so need instructions > to load the values and do arithmetic. I believe we could eventually get the data > cost to match the cost for the compile image for the chosen page size by freeing > the ends of the static buffers not needed for the selected page size: > > | | text | text | text | data | data | data | > | config | size/KB | diff/KB | diff/% | size/KB | diff/KB | diff/% | > |-------------|---------|---------|---------|---------|---------|---------| > | base-4k | 20561 | 0 | 0.0% | 14314 | 0 | 0.0% | > | base-16k | 20439 | -122 | -0.6% | 14625 | 311 | 2.2% | > | base-64k | 20435 | -126 | -0.6% | 15673 | 1359 | 9.5% | > | compile-4k | 20565 | 4 | 0.0% | 14315 | 1 | 0.0% | > | compile-16k | 20443 | -118 | -0.6% | 14517 | 204 | 1.4% | > | compile-64k | 20439 | -122 | -0.6% | 15134 | 820 | 5.7% | > | boot-4K | 20811 | 250 | 1.2% | 15287 | 973 | 6.8% | > > Functional Testing > ================== > > I've build-tested defconfig for all arches supported by tuxmake (which is most) > without issue. > > I've boot-tested arm64 with CONFIG_ARM64_BOOT_TIME_PAGE_SIZE for all page sizes > and a few va-sizes, and additionally have run all the mm-selftests, with no > regressions observed vs the equivalent compile-time page size build (although > the mm-selftests have a few existing failures when run against 16K and 64K > kernels - those should really be investigated and fixed independently). > > Test coverage is lacking for many of the drivers that I've touched, but in many > cases, I'm hoping the changes are simple enough that review might suffice? > > Performance Testing > =================== > > I've run some limited performance benchmarks: > > First, a real-world benchmark that causes a lot of page table manipulation (and > therefore we would expect to see regression here if we are going to see it > anywhere); kernel compilation. It barely registers a change. Values are times, > so smaller is better. All relative to base-4k: > > | | kern | kern | user | user | real | real | > | config | mean | stdev | mean | stdev | mean | stdev | > |-------------|---------|---------|---------|---------|---------|---------| > | base-4k | 0.0% | 1.1% | 0.0% | 0.3% | 0.0% | 0.3% | > | compile-4k | -0.2% | 1.1% | -0.2% | 0.3% | -0.1% | 0.3% | > | boot-4k | 0.1% | 1.0% | -0.3% | 0.2% | -0.2% | 0.2% | > > The Speedometer JavaScript benchmark also shows no change. Values are runs per > min, so bigger is better. All relative to base-4k: > > | config | mean | stdev | > |-------------|---------|---------| > | base-4k | 0.0% | 0.8% | > | compile-4k | 0.4% | 0.8% | > | boot-4k | 0.0% | 0.9% | > > Finally, I've run some microbenchmarks known to stress page table manipulations > (originally from David Hildenbrand). The fork test maps/allocs 1G of anon > memory, then measures the cost of fork(). The munmap test maps/allocs 1G of anon > memory then measures the cost of munmap()ing it. The fork test is known to be > extremely sensitive to any changes that cause instructions to be aligned > differently in cachelines. When using this test for other changes, I've seen > double digit regressions for the slightest thing, so 12% regression on this test > is actually fairly good. This likely represents the extreme worst case for > regressions that will be observed across other microbenchmarks (famous last > words). Values are times, so smaller is better. All relative to base-4k: > > | | fork | fork | munmap | munmap | > | config | mean | stdev | stdev | stdev | > |-------------|---------|---------|---------|---------| > | base-4k | 0.0% | 1.3% | 0.0% | 0.3% | > | compile-4k | 0.1% | 1.3% | -0.9% | 0.1% | > | boot-4k | 12.8% | 1.2% | 3.8% | 1.0% | > > NOTE: The series applies on top of v6.11. > > Thanks, > Ryan > > > Ryan Roberts (57): > mm: Add macros ahead of supporting boot-time page size selection > vmlinux: Align to PAGE_SIZE_MAX > mm/memcontrol: Fix seq_buf size to save memory when PAGE_SIZE is large > mm/page_alloc: Make page_frag_cache boot-time page size compatible > mm: Avoid split pmd ptl if pmd level is run-time folded > mm: Remove PAGE_SIZE compile-time constant assumption > fs: Introduce MAX_BUF_PER_PAGE_SIZE_MAX for array sizing > fs: Remove PAGE_SIZE compile-time constant assumption > fs/nfs: Remove PAGE_SIZE compile-time constant assumption > fs/ext4: Remove PAGE_SIZE compile-time constant assumption > fork: Permit boot-time THREAD_SIZE determination > cgroup: Remove PAGE_SIZE compile-time constant assumption > bpf: Remove PAGE_SIZE compile-time constant assumption > pm/hibernate: Remove PAGE_SIZE compile-time constant assumption > stackdepot: Remove PAGE_SIZE compile-time constant assumption > perf: Remove PAGE_SIZE compile-time constant assumption > kvm: Remove PAGE_SIZE compile-time constant assumption > trace: Remove PAGE_SIZE compile-time constant assumption > crash: Remove PAGE_SIZE compile-time constant assumption > crypto: Remove PAGE_SIZE compile-time constant assumption > sunrpc: Remove PAGE_SIZE compile-time constant assumption > sound: Remove PAGE_SIZE compile-time constant assumption > net: Remove PAGE_SIZE compile-time constant assumption > net: fec: Remove PAGE_SIZE compile-time constant assumption > net: marvell: Remove PAGE_SIZE compile-time constant assumption > net: hns3: Remove PAGE_SIZE compile-time constant assumption > net: e1000: Remove PAGE_SIZE compile-time constant assumption > net: igbvf: Remove PAGE_SIZE compile-time constant assumption > net: igb: Remove PAGE_SIZE compile-time constant assumption > drivers/base: Remove PAGE_SIZE compile-time constant assumption > edac: Remove PAGE_SIZE compile-time constant assumption > optee: Remove PAGE_SIZE compile-time constant assumption > random: Remove PAGE_SIZE compile-time constant assumption > sata_sil24: Remove PAGE_SIZE compile-time constant assumption > virtio: Remove PAGE_SIZE compile-time constant assumption > xen: Remove PAGE_SIZE compile-time constant assumption > arm64: Fix macros to work in C code in addition to the linker script > arm64: Track early pgtable allocation limit > arm64: Introduce macros required for boot-time page selection > arm64: Refactor early pgtable size calculation macros > arm64: Pass desired page size on command line > arm64: Divorce early init from PAGE_SIZE > arm64: Clean up simple cases of CONFIG_ARM64_*K_PAGES > arm64: Align sections to PAGE_SIZE_MAX > arm64: Rework trampoline rodata mapping > arm64: Generalize fixmap for boot-time page size > arm64: Statically allocate and align for worst-case page size > arm64: Convert switch to if for non-const comparison values > arm64: Convert BUILD_BUG_ON to VM_BUG_ON > arm64: Remove PAGE_SZ asm-offset > arm64: Introduce cpu features for page sizes > arm64: Remove PAGE_SIZE from assembly code > arm64: Runtime-fold pmd level > arm64: Support runtime folding in idmap_kpti_install_ng_mappings > arm64: TRAMP_VALIAS is no longer compile-time constant > arm64: Determine THREAD_SIZE at boot-time > arm64: Enable boot-time page size selection > > arch/alpha/include/asm/page.h | 1 + > arch/arc/include/asm/page.h | 1 + > arch/arm/include/asm/page.h | 1 + > arch/arm64/Kconfig | 26 ++- > arch/arm64/include/asm/assembler.h | 78 ++++++- > arch/arm64/include/asm/cpufeature.h | 44 +++- > arch/arm64/include/asm/efi.h | 2 +- > arch/arm64/include/asm/fixmap.h | 28 ++- > arch/arm64/include/asm/kernel-pgtable.h | 150 +++++++++---- > arch/arm64/include/asm/kvm_arm.h | 21 +- > arch/arm64/include/asm/kvm_hyp.h | 11 + > arch/arm64/include/asm/kvm_pgtable.h | 6 +- > arch/arm64/include/asm/memory.h | 62 ++++-- > arch/arm64/include/asm/page-def.h | 3 +- > arch/arm64/include/asm/pgalloc.h | 16 +- > arch/arm64/include/asm/pgtable-geometry.h | 46 ++++ > arch/arm64/include/asm/pgtable-hwdef.h | 28 ++- > arch/arm64/include/asm/pgtable-prot.h | 2 +- > arch/arm64/include/asm/pgtable.h | 133 +++++++++--- > arch/arm64/include/asm/processor.h | 10 +- > arch/arm64/include/asm/sections.h | 1 + > arch/arm64/include/asm/smp.h | 1 + > arch/arm64/include/asm/sparsemem.h | 15 +- > arch/arm64/include/asm/sysreg.h | 54 +++-- > arch/arm64/include/asm/tlb.h | 3 + > arch/arm64/kernel/asm-offsets.c | 4 +- > arch/arm64/kernel/cpufeature.c | 93 ++++++-- > arch/arm64/kernel/efi.c | 2 +- > arch/arm64/kernel/entry.S | 60 +++++- > arch/arm64/kernel/head.S | 46 +++- > arch/arm64/kernel/hibernate-asm.S | 6 +- > arch/arm64/kernel/image-vars.h | 14 ++ > arch/arm64/kernel/image.h | 4 + > arch/arm64/kernel/pi/idreg-override.c | 68 +++++- > arch/arm64/kernel/pi/map_kernel.c | 165 ++++++++++---- > arch/arm64/kernel/pi/map_range.c | 201 ++++++++++++++++-- > arch/arm64/kernel/pi/pi.h | 63 +++++- > arch/arm64/kernel/relocate_kernel.S | 10 +- > arch/arm64/kernel/vdso-wrap.S | 4 +- > arch/arm64/kernel/vdso.c | 7 +- > arch/arm64/kernel/vdso/vdso.lds.S | 4 +- > arch/arm64/kernel/vdso32-wrap.S | 4 +- > arch/arm64/kernel/vdso32/vdso.lds.S | 4 +- > arch/arm64/kernel/vmlinux.lds.S | 48 +++-- > arch/arm64/kvm/arm.c | 10 + > arch/arm64/kvm/hyp/nvhe/Makefile | 1 + > arch/arm64/kvm/hyp/nvhe/host.S | 10 +- > arch/arm64/kvm/hyp/nvhe/hyp.lds.S | 4 +- > arch/arm64/kvm/hyp/nvhe/pgtable-geometry.c | 16 ++ > arch/arm64/kvm/mmu.c | 39 ++-- > arch/arm64/lib/clear_page.S | 7 +- > arch/arm64/lib/copy_page.S | 33 ++- > arch/arm64/lib/mte.S | 27 ++- > arch/arm64/mm/Makefile | 1 + > arch/arm64/mm/fixmap.c | 38 ++-- > arch/arm64/mm/hugetlbpage.c | 40 +--- > arch/arm64/mm/init.c | 26 +-- > arch/arm64/mm/kasan_init.c | 8 +- > arch/arm64/mm/mmu.c | 53 +++-- > arch/arm64/mm/pgd.c | 12 +- > arch/arm64/mm/pgtable-geometry.c | 24 +++ > arch/arm64/mm/proc.S | 128 ++++++++--- > arch/arm64/mm/ptdump.c | 3 +- > arch/arm64/tools/cpucaps | 3 + > arch/csky/include/asm/page.h | 3 + > arch/hexagon/include/asm/page.h | 2 + > arch/loongarch/include/asm/page.h | 2 + > arch/m68k/include/asm/page.h | 1 + > arch/microblaze/include/asm/page.h | 1 + > arch/mips/include/asm/page.h | 1 + > arch/nios2/include/asm/page.h | 2 + > arch/openrisc/include/asm/page.h | 1 + > arch/parisc/include/asm/page.h | 1 + > arch/powerpc/include/asm/page.h | 2 + > arch/riscv/include/asm/page.h | 1 + > arch/s390/include/asm/page.h | 1 + > arch/sh/include/asm/page.h | 1 + > arch/sparc/include/asm/page.h | 3 + > arch/um/include/asm/page.h | 2 + > arch/x86/include/asm/page_types.h | 2 + > arch/xtensa/include/asm/page.h | 1 + > crypto/lskcipher.c | 4 +- > drivers/ata/sata_sil24.c | 46 ++-- > drivers/base/node.c | 6 +- > drivers/base/topology.c | 32 +-- > drivers/block/virtio_blk.c | 2 +- > drivers/char/random.c | 4 +- > drivers/edac/edac_mc.h | 13 +- > drivers/firmware/efi/libstub/arm64.c | 3 +- > drivers/irqchip/irq-gic-v3-its.c | 2 +- > drivers/mtd/mtdswap.c | 4 +- > drivers/net/ethernet/freescale/fec.h | 3 +- > drivers/net/ethernet/freescale/fec_main.c | 5 +- > .../net/ethernet/hisilicon/hns3/hns3_enet.h | 4 +- > drivers/net/ethernet/intel/e1000/e1000_main.c | 6 +- > drivers/net/ethernet/intel/igb/igb.h | 25 +-- > drivers/net/ethernet/intel/igb/igb_main.c | 149 +++++++------ > drivers/net/ethernet/intel/igbvf/netdev.c | 6 +- > drivers/net/ethernet/marvell/mvneta.c | 9 +- > drivers/net/ethernet/marvell/sky2.h | 2 +- > drivers/tee/optee/call.c | 7 +- > drivers/tee/optee/smc_abi.c | 2 +- > drivers/virtio/virtio_balloon.c | 10 +- > drivers/xen/balloon.c | 11 +- > drivers/xen/biomerge.c | 12 +- > drivers/xen/privcmd.c | 2 +- > drivers/xen/xenbus/xenbus_client.c | 5 +- > drivers/xen/xlate_mmu.c | 6 +- > fs/binfmt_elf.c | 11 +- > fs/buffer.c | 2 +- > fs/coredump.c | 8 +- > fs/ext4/ext4.h | 36 ++-- > fs/ext4/move_extent.c | 2 +- > fs/ext4/readpage.c | 2 +- > fs/fat/dir.c | 4 +- > fs/fat/fatent.c | 4 +- > fs/nfs/nfs42proc.c | 2 +- > fs/nfs/nfs42xattr.c | 2 +- > fs/nfs/nfs4proc.c | 2 +- > include/asm-generic/pgtable-geometry.h | 71 +++++++ > include/asm-generic/vmlinux.lds.h | 38 ++-- > include/linux/buffer_head.h | 1 + > include/linux/cpumask.h | 5 + > include/linux/linkage.h | 4 +- > include/linux/mm.h | 17 +- > include/linux/mm_types.h | 15 +- > include/linux/mm_types_task.h | 2 +- > include/linux/mmzone.h | 3 +- > include/linux/netlink.h | 6 +- > include/linux/percpu-defs.h | 4 +- > include/linux/perf_event.h | 2 +- > include/linux/sched.h | 4 +- > include/linux/slab.h | 7 +- > include/linux/stackdepot.h | 6 +- > include/linux/sunrpc/svc.h | 8 +- > include/linux/sunrpc/svc_rdma.h | 4 +- > include/linux/sunrpc/svcsock.h | 2 +- > include/linux/swap.h | 17 +- > include/linux/swapops.h | 6 +- > include/linux/thread_info.h | 10 +- > include/xen/page.h | 2 + > init/main.c | 7 +- > kernel/bpf/core.c | 9 +- > kernel/bpf/ringbuf.c | 54 ++--- > kernel/cgroup/cgroup.c | 8 +- > kernel/crash_core.c | 2 +- > kernel/events/core.c | 2 +- > kernel/fork.c | 71 +++---- > kernel/power/power.h | 2 +- > kernel/power/snapshot.c | 2 +- > kernel/power/swap.c | 129 +++++++++-- > kernel/trace/fgraph.c | 2 +- > kernel/trace/trace.c | 2 +- > lib/stackdepot.c | 6 +- > mm/kasan/report.c | 3 +- > mm/memcontrol.c | 11 +- > mm/memory.c | 4 +- > mm/mmap.c | 2 +- > mm/page-writeback.c | 2 +- > mm/page_alloc.c | 31 +-- > mm/slub.c | 2 +- > mm/sparse.c | 2 +- > mm/swapfile.c | 2 +- > mm/vmalloc.c | 7 +- > net/9p/trans_virtio.c | 4 +- > net/core/hotdata.c | 4 +- > net/core/skbuff.c | 4 +- > net/core/sysctl_net_core.c | 2 +- > net/sunrpc/cache.c | 3 +- > net/unix/af_unix.c | 2 +- > sound/soc/soc-utils.c | 4 +- > virt/kvm/kvm_main.c | 2 +- > 172 files changed, 2185 insertions(+), 951 deletions(-) > create mode 100644 arch/arm64/include/asm/pgtable-geometry.h > create mode 100644 arch/arm64/kvm/hyp/nvhe/pgtable-geometry.c > create mode 100644 arch/arm64/mm/pgtable-geometry.c > create mode 100644 include/asm-generic/pgtable-geometry.h > > -- > 2.43.0 >