Re: [PATCH v4 0/9] Update to zstd-1.4.6
On 2 Oct 2020, at 2:54, Christoph Hellwig wrote: On Wed, Sep 30, 2020 at 08:05:45PM +, Nick Terrell wrote: On Sep 29, 2020, at 11:53 PM, Christoph Hellwig wrote: As you keep resend this I keep retelling you that should not do it. Please provide a proper Linux API, and switch to that. Versioned APIs have absolutely no business in the Linux kernel. The API is not versioned. We provide a stable ABI for a large section of our API, and the parts that aren???t ABI stable don???t change in semantics, and undergo long deprecation periods before being removed. The change of callers is a one-time change to transition from the existing API in the kernel, which was never upstream's API, to upstream's API. Again, please transition it to a sane kernel API. We don't have an "upstream" in this case. The upstream is the zstd project where all this code originates, and where the active development takes place. As Eric Biggers pointed out, it also receives a lot of Q/A separate from the kernel. I think we gain a great deal by leveraging the testing and documentation of the zstd project in the kernel interfaces we use. We lose some consistency with the kernel coding style, but we gain the ability to search for docs, issues, and fixes directly against the zstd project and git repo. -chris
Re: [GIT PULL][PATCH v4 0/9] Update to zstd-1.4.6
On Thu, Oct 01, 2020 at 06:35:34PM +, Nick Terrell wrote: > I???m open to suggestions on how to get a zstd update done better. I don???t > know of any way to break this patch up into smaller patches that all compile. > The code is all generated directly from upstream and modified to work in the > kernel by automated scripts. Documentation/process/submitting-patches.rst: "Separate your changes - Separate each **logical change** into a separate patch. For example, if your changes include both bug fixes and performance enhancements for a single driver, separate those changes into two or more patches. If your changes include an API update, and a new driver which uses that new API, separate those into two patches." It's not that hard, is it? Please do your very basic homework instead of pretending to be a special snowflake and then come back.
Re: [PATCH v4 0/9] Update to zstd-1.4.6
On Wed, Sep 30, 2020 at 08:05:45PM +, Nick Terrell wrote: > > > > On Sep 29, 2020, at 11:53 PM, Christoph Hellwig wrote: > > > > As you keep resend this I keep retelling you that should not do it. > > Please provide a proper Linux API, and switch to that. Versioned APIs > > have absolutely no business in the Linux kernel. > > The API is not versioned. We provide a stable ABI for a large section of our > API, > and the parts that aren???t ABI stable don???t change in semantics, and > undergo long > deprecation periods before being removed. > > The change of callers is a one-time change to transition from the existing API > in the kernel, which was never upstream's API, to upstream's API. Again, please transition it to a sane kernel API. We don't have an "upstream" in this case.
Re: [GIT PULL][PATCH v4 0/9] Update to zstd-1.4.6
> On Oct 1, 2020, at 3:18 AM, David Sterba wrote: > > On Wed, Sep 30, 2020 at 08:49:49PM +, Nick Terrell wrote: >>> On Sep 29, 2020, at 11:53 PM, Nick Terrell wrote: >>> >>> From: Nick Terrell >> >> It has been brought to my attention that patch 3 hasn’t made it to patchwork, >> likely because it is too large. I’ll include a pull request in the next >> cover letter, >> together with the patches (if needed). > > The patch 3/9 saved to a file is 1.6M, over 35000 lines, the diffstat > says: > > 66 files changed, 24268 insertions(+), 12889 deletions(-) > > Seriously, this is wrong in so many ways. There's the rationale for > one-time change etc, but the actual result is beyond what I would accept > and would not encourage anyone to merge as-is. I’m open to suggestions on how to get a zstd update done better. I don’t know of any way to break this patch up into smaller patches that all compile. The code is all generated directly from upstream and modified to work in the kernel by automated scripts. I think the benefits of updating zstd are pretty clear: bug fixes, 3 years of testing, features, debuggability, support from zstd upstream, and significant performance improvements. So I hope we can come up with a way forward to get this merged. This large of a patch is a one-time change. But, the zstd updates in general will be large, containing 100s of commits worth of changes (as opposed to ~3500 and a structure change in this diff). E.g. the upstream diff between two upstream versions range from 50KB - 500KB. Zstd is an actively maintained project, so there is going to be churn when consuming it. But it also means that we’re actively supporting the project if any problems occur. My view is that kernel developers don’t need to review upstreams zstd’s code. We should focus on the diff from upstream, and ensuring that everything works in the kernel environment. The imported code from upstream zstd is ~30K LOC, which is too large for anyone to reasonably review. As mentioned in the patch, this commit shows the diff from upstream zstd, which is much more manageable: https://github.com/terrelln/linux/commit/467c9ea1df1100db48c020c3c8b282a2a30f5116 I’ve generated it by importing upstream zstd as-is into the kernel file structure. Then running the automation to generate the kernel patch from upstream and importing it into the kernel on top of the upstream patch. Best, Nick
Re: [GIT PULL][PATCH v4 0/9] Update to zstd-1.4.6
On Wed, Sep 30, 2020 at 08:49:49PM +, Nick Terrell wrote: > > On Sep 29, 2020, at 11:53 PM, Nick Terrell wrote: > > > > From: Nick Terrell > > It has been brought to my attention that patch 3 hasn’t made it to patchwork, > likely because it is too large. I’ll include a pull request in the next cover > letter, > together with the patches (if needed). The patch 3/9 saved to a file is 1.6M, over 35000 lines, the diffstat says: 66 files changed, 24268 insertions(+), 12889 deletions(-) Seriously, this is wrong in so many ways. There's the rationale for one-time change etc, but the actual result is beyond what I would accept and would not encourage anyone to merge as-is.
Re: [GIT PULL][PATCH v4 0/9] Update to zstd-1.4.6
> On Sep 29, 2020, at 11:53 PM, Nick Terrell wrote: > > From: Nick Terrell It has been brought to my attention that patch 3 hasn’t made it to patchwork, likely because it is too large. I’ll include a pull request in the next cover letter, together with the patches (if needed). Please pull from g...@github.com:terrelln/linux.git tags/v4-zstd-1.4.6 to get these changes. > This patchset upgrades the zstd library to the latest upstream release. The > current zstd version in the kernel is a modified version of upstream > zstd-1.3.1. > At the time it was integrated, zstd wasn't ready to be used in the kernel > as-is. > But, it is now possible to use upstream zstd directly in the kernel. > > I have not yet release zstd-1.4.6 upstream. I want the zstd version in the > kernel > to match up with a known upstream release, so we know exactly what code is > running. Whenever this patchset is ready for merge, I will cut a release at > the > upstream commit that gets merged. This should not be necessary for future > releases. > > The kernel zstd library is automatically generated from upstream zstd. A > script > makes the necessary changes and imports it into the kernel. The changes are: > > 1. Replace all libc dependencies with kernel replacements and rewrite > includes. > 2. Remove unncessary portability macros like: #if defined(_MSC_VER). > 3. Use the kernel xxhash instead of bundling it. > > This automation gets tested every commit by upstream's continuous integration. > When we cut a new zstd release, we will submit a patch to the kernel to update > the zstd version in the kernel. > > I've updated zstd to upstream with one big patch because every commit must > build, > so that precludes partial updates. Since the commit is 100% generated, I hope > the > review burden is lightened. I considered replaying upstream commits, but that > is > not possible because there have been ~3500 upstream commits since the last > zstd > import, and the commits don't all build individually. The bulk update > preserves > bisectablity because bugs can be bisected to the zstd version update. At that > point the update can be reverted, and we can work with upstream to find and > fix > the bug. After this big switch in how the kernel consumes zstd, future patches > will be smaller, because they will only have one upstream release worth of > changes each. > > This patchset changes the zstd API from a custom kernel API to the upstream > API. > I considered wrapping the upstream API with a wrapper that is closer to the > kernel style guide. Following advise from https://lkml.org/lkml/2020/9/17/814 > I've chosen to use the upstream API directly, to minimize opportunities to > introduce bugs, and because using the upstream API directly makes debugging > and > communication with upstream easier. > > This patchset comes in 3 parts: > 1. The first 2 patches prepare for the zstd upgrade. The first patch adds a > compatibility wrapper so zstd can be upgraded without modifying any callers. > The second patch adds an indirection for the lib/decompress_unzstd.c > including > of all decompression source files. > 2. Import zstd-1.4.6. This patch is completely generated from upstream using > automated tooling. > 3. Update all callers to the zstd-1.4.6 API then delete the compatibility > wrapper. > > I tested every caller of zstd on x86_64. I tested both after the 1.4.6 upgrade > using the compatibility wrapper, and after the final patch in this series. > > I tested kernel and initramfs decompression in i386 and arm. > > I ran benchmarks to compare the current zstd in the kernel with zstd-1.4.6. > I benchmarked on x86_64 using QEMU with KVM enabled on an Intel i9-9900k. > I found: > * BtrFS zstd compression at levels 1 and 3 is 5% faster > * BtrFS zstd decompression+read is 15% faster > * SquashFS zstd decompression+read is 15% faster > * F2FS zstd compression+write at level 3 is 8% faster > * F2FS zstd decompression+read is 20% faster > * ZRAM decompression+read is 30% faster > * Kernel zstd decompression is 35% faster > * Initramfs zstd decompression+build is 5% faster > > The latest zstd also offers bug fixes and a 1 KB reduction in stack uage > during > compression. For example the recent problem with large kernel decompression > has > been fixed upstream for over 2 years https://lkml.org/lkml/2020/9/29/27. > > Please let me know if there is anything that I can do to ease the way for > these > patches. I think it is important because it gets large performance > improvements, > contains bug fixes, and is switching to a more maintainable model of consuming > upstream zstd directly, making it easy to keep up to date. > > Best, > Nick Terrell > > v1 -> v2: > * Successfully tested F2FS with help from Chao Yu to fix my test. > * (1/9) Fix ZSTD_initCStream() wrapper to handle pledged_src_size=0 means > unknown. > This fixes F2FS with the zstd-1.4.6 compatibility wrapper, exposed by the > test. > > v2
Re: [PATCH v4 0/9] Update to zstd-1.4.6
> On Sep 29, 2020, at 11:53 PM, Christoph Hellwig wrote: > > As you keep resend this I keep retelling you that should not do it. > Please provide a proper Linux API, and switch to that. Versioned APIs > have absolutely no business in the Linux kernel. The API is not versioned. We provide a stable ABI for a large section of our API, and the parts that aren’t ABI stable don’t change in semantics, and undergo long deprecation periods before being removed. The change of callers is a one-time change to transition from the existing API in the kernel, which was never upstream's API, to upstream's API. -Nick > On Tue, Sep 29, 2020 at 11:53:09PM -0700, Nick Terrell wrote: >> From: Nick Terrell >> >> This patchset upgrades the zstd library to the latest upstream release. The >> current zstd version in the kernel is a modified version of upstream >> zstd-1.3.1. >> At the time it was integrated, zstd wasn't ready to be used in the kernel >> as-is. >> But, it is now possible to use upstream zstd directly in the kernel. >> >> I have not yet release zstd-1.4.6 upstream. I want the zstd version in the >> kernel >> to match up with a known upstream release, so we know exactly what code is >> running. Whenever this patchset is ready for merge, I will cut a release at >> the >> upstream commit that gets merged. This should not be necessary for future >> releases. >> >> The kernel zstd library is automatically generated from upstream zstd. A >> script >> makes the necessary changes and imports it into the kernel. The changes are: >> >> 1. Replace all libc dependencies with kernel replacements and rewrite >> includes. >> 2. Remove unncessary portability macros like: #if defined(_MSC_VER). >> 3. Use the kernel xxhash instead of bundling it. >> >> This automation gets tested every commit by upstream's continuous >> integration. >> When we cut a new zstd release, we will submit a patch to the kernel to >> update >> the zstd version in the kernel. >> >> I've updated zstd to upstream with one big patch because every commit must >> build, >> so that precludes partial updates. Since the commit is 100% generated, I >> hope the >> review burden is lightened. I considered replaying upstream commits, but >> that is >> not possible because there have been ~3500 upstream commits since the last >> zstd >> import, and the commits don't all build individually. The bulk update >> preserves >> bisectablity because bugs can be bisected to the zstd version update. At that >> point the update can be reverted, and we can work with upstream to find and >> fix >> the bug. After this big switch in how the kernel consumes zstd, future >> patches >> will be smaller, because they will only have one upstream release worth of >> changes each. >> >> This patchset changes the zstd API from a custom kernel API to the upstream >> API. >> I considered wrapping the upstream API with a wrapper that is closer to the >> kernel style guide. Following advise from https://lkml.org/lkml/2020/9/17/814 >> I've chosen to use the upstream API directly, to minimize opportunities to >> introduce bugs, and because using the upstream API directly makes debugging >> and >> communication with upstream easier. >> >> This patchset comes in 3 parts: >> 1. The first 2 patches prepare for the zstd upgrade. The first patch adds a >> compatibility wrapper so zstd can be upgraded without modifying any >> callers. >> The second patch adds an indirection for the lib/decompress_unzstd.c >> including >> of all decompression source files. >> 2. Import zstd-1.4.6. This patch is completely generated from upstream using >> automated tooling. >> 3. Update all callers to the zstd-1.4.6 API then delete the compatibility >> wrapper. >> >> I tested every caller of zstd on x86_64. I tested both after the 1.4.6 >> upgrade >> using the compatibility wrapper, and after the final patch in this series. >> >> I tested kernel and initramfs decompression in i386 and arm. >> >> I ran benchmarks to compare the current zstd in the kernel with zstd-1.4.6. >> I benchmarked on x86_64 using QEMU with KVM enabled on an Intel i9-9900k. >> I found: >> * BtrFS zstd compression at levels 1 and 3 is 5% faster >> * BtrFS zstd decompression+read is 15% faster >> * SquashFS zstd decompression+read is 15% faster >> * F2FS zstd compression+write at level 3 is 8% faster >> * F2FS zstd decompression+read is 20% faster >> * ZRAM decompression+read is 30% faster >> * Kernel zstd decompression is 35% faster >> * Initramfs zstd decompression+build is 5% faster >> >> The latest zstd also offers bug fixes and a 1 KB reduction in stack uage >> during >> compression. For example the recent problem with large kernel decompression >> has >> been fixed upstream for over 2 years https://lkml.org/lkml/2020/9/29/27. >> >> Please let me know if there is anything that I can do to ease the way for >> these >> patches. I think it is important because it gets large performance >>
Re: [PATCH v4 0/9] Update to zstd-1.4.6
As you keep resend this I keep retelling you that should not do it. Please provide a proper Linux API, and switch to that. Versioned APIs have absolutely no business in the Linux kernel. On Tue, Sep 29, 2020 at 11:53:09PM -0700, Nick Terrell wrote: > From: Nick Terrell > > This patchset upgrades the zstd library to the latest upstream release. The > current zstd version in the kernel is a modified version of upstream > zstd-1.3.1. > At the time it was integrated, zstd wasn't ready to be used in the kernel > as-is. > But, it is now possible to use upstream zstd directly in the kernel. > > I have not yet release zstd-1.4.6 upstream. I want the zstd version in the > kernel > to match up with a known upstream release, so we know exactly what code is > running. Whenever this patchset is ready for merge, I will cut a release at > the > upstream commit that gets merged. This should not be necessary for future > releases. > > The kernel zstd library is automatically generated from upstream zstd. A > script > makes the necessary changes and imports it into the kernel. The changes are: > > 1. Replace all libc dependencies with kernel replacements and rewrite > includes. > 2. Remove unncessary portability macros like: #if defined(_MSC_VER). > 3. Use the kernel xxhash instead of bundling it. > > This automation gets tested every commit by upstream's continuous integration. > When we cut a new zstd release, we will submit a patch to the kernel to update > the zstd version in the kernel. > > I've updated zstd to upstream with one big patch because every commit must > build, > so that precludes partial updates. Since the commit is 100% generated, I hope > the > review burden is lightened. I considered replaying upstream commits, but that > is > not possible because there have been ~3500 upstream commits since the last > zstd > import, and the commits don't all build individually. The bulk update > preserves > bisectablity because bugs can be bisected to the zstd version update. At that > point the update can be reverted, and we can work with upstream to find and > fix > the bug. After this big switch in how the kernel consumes zstd, future patches > will be smaller, because they will only have one upstream release worth of > changes each. > > This patchset changes the zstd API from a custom kernel API to the upstream > API. > I considered wrapping the upstream API with a wrapper that is closer to the > kernel style guide. Following advise from https://lkml.org/lkml/2020/9/17/814 > I've chosen to use the upstream API directly, to minimize opportunities to > introduce bugs, and because using the upstream API directly makes debugging > and > communication with upstream easier. > > This patchset comes in 3 parts: > 1. The first 2 patches prepare for the zstd upgrade. The first patch adds a >compatibility wrapper so zstd can be upgraded without modifying any > callers. >The second patch adds an indirection for the lib/decompress_unzstd.c > including >of all decompression source files. > 2. Import zstd-1.4.6. This patch is completely generated from upstream using >automated tooling. > 3. Update all callers to the zstd-1.4.6 API then delete the compatibility >wrapper. > > I tested every caller of zstd on x86_64. I tested both after the 1.4.6 upgrade > using the compatibility wrapper, and after the final patch in this series. > > I tested kernel and initramfs decompression in i386 and arm. > > I ran benchmarks to compare the current zstd in the kernel with zstd-1.4.6. > I benchmarked on x86_64 using QEMU with KVM enabled on an Intel i9-9900k. > I found: > * BtrFS zstd compression at levels 1 and 3 is 5% faster > * BtrFS zstd decompression+read is 15% faster > * SquashFS zstd decompression+read is 15% faster > * F2FS zstd compression+write at level 3 is 8% faster > * F2FS zstd decompression+read is 20% faster > * ZRAM decompression+read is 30% faster > * Kernel zstd decompression is 35% faster > * Initramfs zstd decompression+build is 5% faster > > The latest zstd also offers bug fixes and a 1 KB reduction in stack uage > during > compression. For example the recent problem with large kernel decompression > has > been fixed upstream for over 2 years https://lkml.org/lkml/2020/9/29/27. > > Please let me know if there is anything that I can do to ease the way for > these > patches. I think it is important because it gets large performance > improvements, > contains bug fixes, and is switching to a more maintainable model of consuming > upstream zstd directly, making it easy to keep up to date. > > Best, > Nick Terrell > > v1 -> v2: > * Successfully tested F2FS with help from Chao Yu to fix my test. > * (1/9) Fix ZSTD_initCStream() wrapper to handle pledged_src_size=0 means > unknown. > This fixes F2FS with the zstd-1.4.6 compatibility wrapper, exposed by the > test. > > v2 -> v3: > * (3/9) Silence warnings by Kernel Test Robot: >
[PATCH v4 0/9] Update to zstd-1.4.6
From: Nick Terrell This patchset upgrades the zstd library to the latest upstream release. The current zstd version in the kernel is a modified version of upstream zstd-1.3.1. At the time it was integrated, zstd wasn't ready to be used in the kernel as-is. But, it is now possible to use upstream zstd directly in the kernel. I have not yet release zstd-1.4.6 upstream. I want the zstd version in the kernel to match up with a known upstream release, so we know exactly what code is running. Whenever this patchset is ready for merge, I will cut a release at the upstream commit that gets merged. This should not be necessary for future releases. The kernel zstd library is automatically generated from upstream zstd. A script makes the necessary changes and imports it into the kernel. The changes are: 1. Replace all libc dependencies with kernel replacements and rewrite includes. 2. Remove unncessary portability macros like: #if defined(_MSC_VER). 3. Use the kernel xxhash instead of bundling it. This automation gets tested every commit by upstream's continuous integration. When we cut a new zstd release, we will submit a patch to the kernel to update the zstd version in the kernel. I've updated zstd to upstream with one big patch because every commit must build, so that precludes partial updates. Since the commit is 100% generated, I hope the review burden is lightened. I considered replaying upstream commits, but that is not possible because there have been ~3500 upstream commits since the last zstd import, and the commits don't all build individually. The bulk update preserves bisectablity because bugs can be bisected to the zstd version update. At that point the update can be reverted, and we can work with upstream to find and fix the bug. After this big switch in how the kernel consumes zstd, future patches will be smaller, because they will only have one upstream release worth of changes each. This patchset changes the zstd API from a custom kernel API to the upstream API. I considered wrapping the upstream API with a wrapper that is closer to the kernel style guide. Following advise from https://lkml.org/lkml/2020/9/17/814 I've chosen to use the upstream API directly, to minimize opportunities to introduce bugs, and because using the upstream API directly makes debugging and communication with upstream easier. This patchset comes in 3 parts: 1. The first 2 patches prepare for the zstd upgrade. The first patch adds a compatibility wrapper so zstd can be upgraded without modifying any callers. The second patch adds an indirection for the lib/decompress_unzstd.c including of all decompression source files. 2. Import zstd-1.4.6. This patch is completely generated from upstream using automated tooling. 3. Update all callers to the zstd-1.4.6 API then delete the compatibility wrapper. I tested every caller of zstd on x86_64. I tested both after the 1.4.6 upgrade using the compatibility wrapper, and after the final patch in this series. I tested kernel and initramfs decompression in i386 and arm. I ran benchmarks to compare the current zstd in the kernel with zstd-1.4.6. I benchmarked on x86_64 using QEMU with KVM enabled on an Intel i9-9900k. I found: * BtrFS zstd compression at levels 1 and 3 is 5% faster * BtrFS zstd decompression+read is 15% faster * SquashFS zstd decompression+read is 15% faster * F2FS zstd compression+write at level 3 is 8% faster * F2FS zstd decompression+read is 20% faster * ZRAM decompression+read is 30% faster * Kernel zstd decompression is 35% faster * Initramfs zstd decompression+build is 5% faster The latest zstd also offers bug fixes and a 1 KB reduction in stack uage during compression. For example the recent problem with large kernel decompression has been fixed upstream for over 2 years https://lkml.org/lkml/2020/9/29/27. Please let me know if there is anything that I can do to ease the way for these patches. I think it is important because it gets large performance improvements, contains bug fixes, and is switching to a more maintainable model of consuming upstream zstd directly, making it easy to keep up to date. Best, Nick Terrell v1 -> v2: * Successfully tested F2FS with help from Chao Yu to fix my test. * (1/9) Fix ZSTD_initCStream() wrapper to handle pledged_src_size=0 means unknown. This fixes F2FS with the zstd-1.4.6 compatibility wrapper, exposed by the test. v2 -> v3: * (3/9) Silence warnings by Kernel Test Robot: https://github.com/facebook/zstd/pull/2324 Stack size warnings remain, but these aren't new, and the functions it warns on are either unused or not in the maximum stack path. This patchset reduces zstd compression stack usage by 1 KB overall. I've gotten the low hanging fruit, and more stack reduction would require significant changes that have the potential to introduce new bugs. However, I do hope to continue to reduce zstd stack usage in future versions. v3 -> v4: * (3/9) Fix errors and warnings
