Message ID | 163162767601.438332.9017034724960075707.stgit@warthog.procyon.org.uk (mailing list archive) |
---|---|
Headers | show |
Series | fscache: Replace and remove old I/O API | expand |
On Tue, Sep 14, 2021 at 9:55 AM David Howells <dhowells@redhat.com> wrote: > > > Here's a set of patches that removes the old fscache I/O API by the following > means: > > (1) A simple fallback API is added that can read or write a single page > synchronously. The functions for this have "deprecated" in their names > as they have to be removed at some point. > > (2) An implementation of this is provided in cachefiles. It creates a kiocb > to use DIO to the backing file rather than calling readpage on the > backing filesystem page and then snooping the page wait queue. > > (3) NFS is switched to use the fallback API. > > (4) CIFS is switched to use the fallback API also for the moment. > > (5) 9P is switched to using netfslib. > > (6) The old I/O API is removed from fscache and the page snooping > implementation is removed from cachefiles. > > The reasons for doing this are: > > (A) Using a kiocb to do asynchronous DIO from/to the pages of the backing > file is now a possibility that didn't exist when cachefiles was created. > This is much simpler than the snooping mechanism with a proper callback > path and it also requires fewer copies and less memory. > > (B) We have to stop using bmap() or SEEK_DATA/SEEK_HOLE to work out what > blocks are present in the backing file is dangerous and can lead to data > corruption if the backing filesystem can insert or remove blocks of zeros > arbitrarily in order to optimise its extent list[1]. > > Whilst this patchset doesn't fix that yet, it does simplify the code and > the fix for that can be made in a subsequent patchset. > > (C) In order to fix (B), the cache will need to keep track itself of what > data is present. To make this easier to manage, the intention is to > increase the cache block granularity to, say, 256KiB - importantly, a > size that will span multiple pages - which means the single-page > interface will have to go away. netfslib is designed to deal with > that on behalf of a filesystem, though a filesystem could use raw > cache calls instead and manage things itself. > > These patches can be found also on: > > https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=fscache-iter-3 > > David > > Link: https://lore.kernel.org/r/YO17ZNOcq+9PajfQ@mit.edu [1] > --- > David Howells (8): > fscache: Generalise the ->begin_read_operation method > fscache: Implement an alternate I/O interface to replace the old API > nfs: Move to using the alternate (deprecated) fscache I/O API > 9p: (untested) Convert to using the netfs helper lib to do reads and caching > cifs: (untested) Move to using the alternate (deprecated) fscache I/O API > fscache: Remove the old I/O API > fscache: Remove stats that are no longer used > fscache: Update the documentation to reflect I/O API changes > > > .../filesystems/caching/backend-api.rst | 138 +-- > .../filesystems/caching/netfs-api.rst | 386 +----- > fs/9p/Kconfig | 1 + > fs/9p/cache.c | 137 --- > fs/9p/cache.h | 98 +- > fs/9p/v9fs.h | 9 + > fs/9p/vfs_addr.c | 174 ++- > fs/9p/vfs_file.c | 21 +- > fs/cachefiles/Makefile | 1 - > fs/cachefiles/interface.c | 15 - > fs/cachefiles/internal.h | 38 - > fs/cachefiles/io.c | 28 +- > fs/cachefiles/main.c | 1 - > fs/cachefiles/rdwr.c | 972 --------------- > fs/cifs/file.c | 64 +- > fs/cifs/fscache.c | 105 +- > fs/cifs/fscache.h | 74 +- > fs/fscache/cache.c | 6 - > fs/fscache/cookie.c | 10 - > fs/fscache/internal.h | 58 +- > fs/fscache/io.c | 140 ++- > fs/fscache/object.c | 2 - > fs/fscache/page.c | 1066 ----------------- > fs/fscache/stats.c | 73 +- > fs/nfs/file.c | 14 +- > fs/nfs/fscache-index.c | 26 - > fs/nfs/fscache.c | 161 +-- > fs/nfs/fscache.h | 84 +- > fs/nfs/read.c | 25 +- > fs/nfs/write.c | 7 +- > include/linux/fscache-cache.h | 131 -- > include/linux/fscache.h | 418 ++----- > include/linux/netfs.h | 17 +- > 33 files changed, 508 insertions(+), 3992 deletions(-) > delete mode 100644 fs/cachefiles/rdwr.c > > I tested an earlier version of these with NFS, which identified a couple issues which you fixed. Last I checked my unit tests and xfstests were looking good. I'll do some testing on this latest branch / patches and report back.
On Tue, 2021-09-14 at 14:54 +0100, David Howells wrote: > Here's a set of patches that removes the old fscache I/O API by the following > means: > > (1) A simple fallback API is added that can read or write a single page > synchronously. The functions for this have "deprecated" in their names > as they have to be removed at some point. > > (2) An implementation of this is provided in cachefiles. It creates a kiocb > to use DIO to the backing file rather than calling readpage on the > backing filesystem page and then snooping the page wait queue. > > (3) NFS is switched to use the fallback API. > > (4) CIFS is switched to use the fallback API also for the moment. > > (5) 9P is switched to using netfslib. > > (6) The old I/O API is removed from fscache and the page snooping > implementation is removed from cachefiles. > > The reasons for doing this are: > > (A) Using a kiocb to do asynchronous DIO from/to the pages of the backing > file is now a possibility that didn't exist when cachefiles was created. > This is much simpler than the snooping mechanism with a proper callback > path and it also requires fewer copies and less memory. > > (B) We have to stop using bmap() or SEEK_DATA/SEEK_HOLE to work out what > blocks are present in the backing file is dangerous and can lead to data > corruption if the backing filesystem can insert or remove blocks of zeros > arbitrarily in order to optimise its extent list[1]. > > Whilst this patchset doesn't fix that yet, it does simplify the code and > the fix for that can be made in a subsequent patchset. > > (C) In order to fix (B), the cache will need to keep track itself of what > data is present. To make this easier to manage, the intention is to > increase the cache block granularity to, say, 256KiB - importantly, a > size that will span multiple pages - which means the single-page > interface will have to go away. netfslib is designed to deal with > that on behalf of a filesystem, though a filesystem could use raw > cache calls instead and manage things itself. > > These patches can be found also on: > > https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=fscache-iter-3 > > David > > Link: https://lore.kernel.org/r/YO17ZNOcq+9PajfQ@mit.edu [1] > --- > David Howells (8): > fscache: Generalise the ->begin_read_operation method > fscache: Implement an alternate I/O interface to replace the old API > nfs: Move to using the alternate (deprecated) fscache I/O API > 9p: (untested) Convert to using the netfs helper lib to do reads and caching > cifs: (untested) Move to using the alternate (deprecated) fscache I/O API > fscache: Remove the old I/O API > fscache: Remove stats that are no longer used > fscache: Update the documentation to reflect I/O API changes > > > .../filesystems/caching/backend-api.rst | 138 +-- > .../filesystems/caching/netfs-api.rst | 386 +----- > fs/9p/Kconfig | 1 + > fs/9p/cache.c | 137 --- > fs/9p/cache.h | 98 +- > fs/9p/v9fs.h | 9 + > fs/9p/vfs_addr.c | 174 ++- > fs/9p/vfs_file.c | 21 +- > fs/cachefiles/Makefile | 1 - > fs/cachefiles/interface.c | 15 - > fs/cachefiles/internal.h | 38 - > fs/cachefiles/io.c | 28 +- > fs/cachefiles/main.c | 1 - > fs/cachefiles/rdwr.c | 972 --------------- > fs/cifs/file.c | 64 +- > fs/cifs/fscache.c | 105 +- > fs/cifs/fscache.h | 74 +- > fs/fscache/cache.c | 6 - > fs/fscache/cookie.c | 10 - > fs/fscache/internal.h | 58 +- > fs/fscache/io.c | 140 ++- > fs/fscache/object.c | 2 - > fs/fscache/page.c | 1066 ----------------- > fs/fscache/stats.c | 73 +- > fs/nfs/file.c | 14 +- > fs/nfs/fscache-index.c | 26 - > fs/nfs/fscache.c | 161 +-- > fs/nfs/fscache.h | 84 +- > fs/nfs/read.c | 25 +- > fs/nfs/write.c | 7 +- > include/linux/fscache-cache.h | 131 -- > include/linux/fscache.h | 418 ++----- > include/linux/netfs.h | 17 +- > 33 files changed, 508 insertions(+), 3992 deletions(-) > delete mode 100644 fs/cachefiles/rdwr.c > > Love the diffstat. I looked over the set and didn't see any problems with it. I also ran a quick test with ceph+fscache and didn't run into any issues, but that's not too surprising since this patchset doesn't really change the codepaths that it uses. Reviewed-by: Jeff Layton <jlayton@kernel.org>
On Tue, Sep 14, 2021 at 6:54 AM David Howells <dhowells@redhat.com> wrote: > > (1) A simple fallback API is added that can read or write a single page > synchronously. The functions for this have "deprecated" in their names > as they have to be removed at some point. I'm looking at those patches, and there's no way I'll apply anything that starts out with moving to a "deprecated" interface. Call it "fallback" or "simple" or something that shows the intent, but no, I'm not taking patches that introduce a _new_ interface and call it "deprecated". Linus
On Tue, Sep 14, 2021 at 9:21 AM Linus Torvalds <torvalds@linux-foundation.org> wrote: > > Call it "fallback" or "simple" or something that shows the intent, but > no, I'm not taking patches that introduce a _new_ interface and call > it "deprecated". Put another way: to call something "deprecated", you have to already have the replacement all ready to go. And if you have that, then converting existing code to a deprecated model isn't the way to go. So in neither situation does it make any sense to convert anything to a model that is deprecated. Linus
Linus Torvalds <torvalds@linux-foundation.org> wrote: > > Call it "fallback" or "simple" or something that shows the intent, but > > no, I'm not taking patches that introduce a _new_ interface and call > > it "deprecated". Yeah, I'll change it to "fallback" - I started talking about it like that in the docs anyway. > Put another way: to call something "deprecated", you have to already > have the replacement all ready to go. We're not far off. There's a fair distance (in number of patches) between this patchset and the completion, hence why I marked them as deprecated here, intending to remove them at the end. Between myself, Jeff and Dave we have fscache, cachefiles, afs, ceph and nfs (almost) covered. I have patches for 9p and I've given a partial patch for cifs to Steve and Shyam. David
On Tue, Sep 14, 2021 at 11:30 AM David Wysochanski <dwysocha@redhat.com> wrote: > > On Tue, Sep 14, 2021 at 9:55 AM David Howells <dhowells@redhat.com> wrote: > > > > > > Here's a set of patches that removes the old fscache I/O API by the following > > means: > > > > (1) A simple fallback API is added that can read or write a single page > > synchronously. The functions for this have "deprecated" in their names > > as they have to be removed at some point. > > > > (2) An implementation of this is provided in cachefiles. It creates a kiocb > > to use DIO to the backing file rather than calling readpage on the > > backing filesystem page and then snooping the page wait queue. > > > > (3) NFS is switched to use the fallback API. > > > > (4) CIFS is switched to use the fallback API also for the moment. > > > > (5) 9P is switched to using netfslib. > > > > (6) The old I/O API is removed from fscache and the page snooping > > implementation is removed from cachefiles. > > > > The reasons for doing this are: > > > > (A) Using a kiocb to do asynchronous DIO from/to the pages of the backing > > file is now a possibility that didn't exist when cachefiles was created. > > This is much simpler than the snooping mechanism with a proper callback > > path and it also requires fewer copies and less memory. > > > > (B) We have to stop using bmap() or SEEK_DATA/SEEK_HOLE to work out what > > blocks are present in the backing file is dangerous and can lead to data > > corruption if the backing filesystem can insert or remove blocks of zeros > > arbitrarily in order to optimise its extent list[1]. > > > > Whilst this patchset doesn't fix that yet, it does simplify the code and > > the fix for that can be made in a subsequent patchset. > > > > (C) In order to fix (B), the cache will need to keep track itself of what > > data is present. To make this easier to manage, the intention is to > > increase the cache block granularity to, say, 256KiB - importantly, a > > size that will span multiple pages - which means the single-page > > interface will have to go away. netfslib is designed to deal with > > that on behalf of a filesystem, though a filesystem could use raw > > cache calls instead and manage things itself. > > > > These patches can be found also on: > > > > https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=fscache-iter-3 > > > > David > > > > Link: https://lore.kernel.org/r/YO17ZNOcq+9PajfQ@mit.edu [1] > > --- > > David Howells (8): > > fscache: Generalise the ->begin_read_operation method > > fscache: Implement an alternate I/O interface to replace the old API > > nfs: Move to using the alternate (deprecated) fscache I/O API > > 9p: (untested) Convert to using the netfs helper lib to do reads and caching > > cifs: (untested) Move to using the alternate (deprecated) fscache I/O API > > fscache: Remove the old I/O API > > fscache: Remove stats that are no longer used > > fscache: Update the documentation to reflect I/O API changes > > > > > > .../filesystems/caching/backend-api.rst | 138 +-- > > .../filesystems/caching/netfs-api.rst | 386 +----- > > fs/9p/Kconfig | 1 + > > fs/9p/cache.c | 137 --- > > fs/9p/cache.h | 98 +- > > fs/9p/v9fs.h | 9 + > > fs/9p/vfs_addr.c | 174 ++- > > fs/9p/vfs_file.c | 21 +- > > fs/cachefiles/Makefile | 1 - > > fs/cachefiles/interface.c | 15 - > > fs/cachefiles/internal.h | 38 - > > fs/cachefiles/io.c | 28 +- > > fs/cachefiles/main.c | 1 - > > fs/cachefiles/rdwr.c | 972 --------------- > > fs/cifs/file.c | 64 +- > > fs/cifs/fscache.c | 105 +- > > fs/cifs/fscache.h | 74 +- > > fs/fscache/cache.c | 6 - > > fs/fscache/cookie.c | 10 - > > fs/fscache/internal.h | 58 +- > > fs/fscache/io.c | 140 ++- > > fs/fscache/object.c | 2 - > > fs/fscache/page.c | 1066 ----------------- > > fs/fscache/stats.c | 73 +- > > fs/nfs/file.c | 14 +- > > fs/nfs/fscache-index.c | 26 - > > fs/nfs/fscache.c | 161 +-- > > fs/nfs/fscache.h | 84 +- > > fs/nfs/read.c | 25 +- > > fs/nfs/write.c | 7 +- > > include/linux/fscache-cache.h | 131 -- > > include/linux/fscache.h | 418 ++----- > > include/linux/netfs.h | 17 +- > > 33 files changed, 508 insertions(+), 3992 deletions(-) > > delete mode 100644 fs/cachefiles/rdwr.c > > > > > > I tested an earlier version of these with NFS, which identified a > couple issues which you fixed. Last I checked my unit tests and > xfstests were looking good. I'll do some testing on this latest branch > / patches and report back. For the series, you can add Tested-by: Dave Wysochanski <dwysocha@redhat.com> Testing was limited to NFS enabled code paths. I ran custom unit tests, as well as a series of xfstest generic runs with various NFS versions, both fscache enabled and not enabled, as well as various NFS servers, comparing 5.15.0-rc1 runs vs runs with these patches. I did not see any failures with these new patches that were not already present with 5.15.0-rc1. Here are the list of xfstest generic runs: 1. Hammerspace (pNFS flexfiles) version 4.6.3-166: vers=4.1,fsc 2. Hammerspace (pNFS flexfiles) version 4.6.3-166: vers=4.2 3. Netapp (pNFS filelayout) version 9.5RC1: vers=4.1 4. Netapp (pNFS filelayout) version 9.5RC1: vers=4.1,fsc 5. Red Hat version 8.2 (kernel-4.18.0-193.el8): vers=4.2,fsc 6. Red Hat version 8.2 (kernel-4.18.0-193.el8): vers=4.0,fsc 7. Red Hat version 8.2 (kernel-4.18.0-193.el8): vers=3,fsc