> On Nov 21, 2020, at 1:28 PM, David Wysochanski <dwysocha@xxxxxxxxxx> wrote: > > On Sat, Nov 21, 2020 at 12:16 PM Chuck Lever <chuck.lever@xxxxxxxxxx> wrote: >> >> >> >>> On Nov 21, 2020, at 12:01 PM, David Wysochanski <dwysocha@xxxxxxxxxx> wrote: >>> >>> On Sat, Nov 21, 2020 at 11:14 AM Chuck Lever <chuck.lever@xxxxxxxxxx> wrote: >>>> >>>> Hi Dave- >>>> >>>>> On Nov 21, 2020, at 8:29 AM, Dave Wysochanski <dwysocha@xxxxxxxxxx> wrote: >>>>> >>>>> These patches update the NFS client to use the new netfs and fscache >>>>> APIs and are at: >>>>> https://github.com/DaveWysochanskiRH/kernel.git >>>>> https://github.com/DaveWysochanskiRH/kernel/commit/94e9633d98a5542ea384b1095290ac6f915fc917 >>>>> https://github.com/DaveWysochanskiRH/kernel/releases/tag/fscache-iter-nfs-20201120 >>>>> >>>>> The patches are based on David Howells fscache-iter tree at >>>>> https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=fscache-iter >>>>> >>>>> The first 6 patches refactor some of the NFS read code to facilitate >>>>> re-use, the next 6 patches do the conversion to the new API, and the >>>>> last patch is a somewhat awkward fix for a problem seen in final >>>>> testing. >>>>> >>>>> Per David Howell's recent post, note that the new fscache API is >>>>> divided into two separate APIs, a 'netfs' API and an 'fscache' API. >>>>> The netfs API was done to help simplify the IO paths of network >>>>> filesystems, and can be called even when fscache is disabled, thus >>>>> simplifing both readpage and readahead implementations. However, >>>>> for now these NFS conversion patches only call the netfs API when >>>>> fscache is enabled, similar to the existing NFS code. >>>>> >>>>> Trond and Anna, I would appreciate your guidance on this patchset. >>>>> At least I would like your feedback regarding the direction >>>>> you would like these patches to go, in particular, the following >>>>> items: >>>>> >>>>> 1. Whether you are ok with using the netfs API unconditionally even >>>>> when fscache is disabled, or prefer this "least invasive to NFS" >>>>> approach. Note the unconditional use of the netfs API is the >>>>> recommended approach per David's post and the AFS and CEPH >>>>> implementations, but I was unsure if you would accept this >>>>> approach or would prefer to minimize changes to NFS. Note if >>>>> we keep the current approach to minimize NFS changes, we will >>>>> have to address some problems with page unlocking such as with >>>>> patch 13 in the series. >>>>> >>>>> 2. Whether to keep the NFS fscache implementation as "read only" >>>>> or if we add write through support. Today we only enable fscache >>>>> when a file is open read-only and disable / invalidate when a file >>>>> is open for write. >>>>> >>>>> Still TODO >>>>> 1. Address known issues (lockdep, page unlocking), depending on >>>>> what is decided as far as implementation direction >>>>> a) nfs_issue_op: takes rcu_read_lock but may calls nfs_page_alloc() >>>>> with GFP_KERNEL which may sleep (dhowells noted this in a review) >>>>> b) nfs_refresh_inode() takes inode->i_lock but may call >>>>> __fscache_invalidate() which may sleep (found with lockdep) >>>>> 2. Fixup NFS fscache stats (NFSIOS_FSCACHE_*) >>>>> * Compare with netfs stats and determine if still needed >>>>> 3. Cleanup dfprintks and/or convert to tracepoints >>>>> 4. Further tests (see "Not tested yet") >>>> >>>> Can you say whether your approach has any performance impact? >>> No I cannot. >>> >>>> In particular, what comparative benchmarks have been run? >>>> >>> No comparisons so far, but note the last bullet - "performance". >>> >>> Are you wondering about performance with/without fscache for this >>> series, or the old vs new fscache, or something else? >> >> I'd like to have some explicit performance-related merge worthiness >> criteria. For example: "No performance regressions, and here's how >> we're going to determine that we're good: fio / iozone / yada with >> NFS/TCP and NFS/RDMA on 100GbE; for very little additional CPU >> cost measured via perf xyzzy. Also some benchmark that measures lock >> contention." >> >> We haven't been especially careful about this in the past when >> reworking the client's primary I/O paths. Nothing unreasonable, but >> it should be stated up front where we want to end up. >> > Makes sense. > >> Another approach might be: we're going to start by making fscache >> opt-in. As confidence increases over time and good performance is >> demonstrated, then we'll unify the fscache and non-cached I/O paths. >> > It sounds like what you want is what I've done in this first implementation. My understanding is that you haven't decided whether to take the opt-in approach or to convert the primary I/O paths right now. > This implementation takes a "least invasive to NFS" approach, staying > with the old fscache on/off logic, even though this was not ideal or what > was recommended as the end game for the netfs API. I'm not taking a position on your two approaches, but I would like that when the time comes to take the approach that involves full integration, we should have an agreed-upon set of performance goals. I don't want that integration to cause our high performance environments (NFS/RDMA, for instance) to lose out from that integration. Btw, it isn't clear yet that we need to use the fscache APIs to introduce proper huge page support. One of our current efforts is to convert xdr_buf from the use of an array of struct page pointers to an array of struct bio_vec pointers. In that case each entry in the array carries the size of the thing that the entry points to, which today is always PAGE_SIZE, but in the future could be much larger. >>>>> Checks run >>>>> 1. checkpatch: PASS* >>>>> * a few warnings, mostly trivial fixups, some unrelated to this set >>>>> 2. kernel builds with each patch: PASS >>>>> * each patch in series built cleanly which ensure bisection >>>>> >>>>> Tests run >>>>> 1. Custom NFS+fscache unit tests for basic operation: PASS* >>>>> * no oops or data corruptions >>>>> * Some op counts are a bit off but these are mostly due >>>>> to statistics not implemented properly (NFSIOS_FSCACHE_*) >>>>> 2. cthon04: PASS (test options "-b -g -s -l", fsc,vers=3,4.0,4.1,4.2,sec=sys) >>>>> * No failures or oopses for any version or test options >>>>> 3. iozone tests (fsc,vers=3,4.0,4.1,4.2,sec=sys): PASS >>>>> * No failures or oopses >>>>> 4. kernel build (fsc,vers=3,4.1,4.2): PASS* >>>>> * all builds finish without errors or data corruption >>>>> * one lockdep "scheduling while atomic" fired with NFS41 and >>>>> was due to one an fscache invalidation code path (known issue 'b' above) >>>>> 5. xfstests/generic (fsc,vers=4.2, nofsc,vers=4.2): PASS* >>>>> * generic/013 (pass but triggers i_lock lockdep warning known issue 'a' above) >>>>> * NOTE: The following tests failed with errors, but they >>>>> also fail on vanilla 5.10-rc4 so are not related to this >>>>> patchset >>>>> * generic/074 (lockep invalid wait context - nfs_free_request()) >>>>> * generic/538 (short read) >>>>> * generic/551 (pread: Unknown error 524, Data verification fail) >>>>> * generic/568 (ERROR: File grew from 4096 B to 8192 B when writing to the fallocated range) >>>>> >>>>> Not tested yet: >>>>> * error injections (for example, connection disruptions, server errors during IO, etc) >>>>> * pNFS >>>>> * many process mixed read/write on same file >>>>> * performance >>>>> Dave Wysochanski (13): >>>>> NFS: Clean up nfs_readpage() and nfs_readpages() >>>>> NFS: In nfs_readpage() only increment NFSIOS_READPAGES when read >>>>> succeeds >>>>> NFS: Refactor nfs_readpage() and nfs_readpage_async() to use >>>>> nfs_readdesc >>>>> NFS: Call readpage_async_filler() from nfs_readpage_async() >>>>> NFS: Add nfs_pageio_complete_read() and remove nfs_readpage_async() >>>>> NFS: Allow internal use of read structs and functions >>>>> NFS: Convert fscache_acquire_cookie and fscache_relinquish_cookie >>>>> NFS: Convert fscache_enable_cookie and fscache_disable_cookie >>>>> NFS: Convert fscache invalidation and update aux_data and i_size >>>>> NFS: Convert to the netfs API and nfs_readpage to use netfs_readpage >>>>> NFS: Convert readpage to readahead and use netfs_readahead for fscache >>>>> NFS: Allow NFS use of new fscache API in build >>>>> NFS: Ensure proper page unlocking when reads fail with retryable >>>>> errors >>>>> >>>>> fs/nfs/Kconfig | 2 +- >>>>> fs/nfs/direct.c | 2 + >>>>> fs/nfs/file.c | 22 ++-- >>>>> fs/nfs/fscache-index.c | 94 -------------- >>>>> fs/nfs/fscache.c | 315 ++++++++++++++++++++------------------------- >>>>> fs/nfs/fscache.h | 132 +++++++------------ >>>>> fs/nfs/inode.c | 4 +- >>>>> fs/nfs/internal.h | 8 ++ >>>>> fs/nfs/nfs4proc.c | 2 +- >>>>> fs/nfs/pagelist.c | 2 + >>>>> fs/nfs/read.c | 248 ++++++++++++++++------------------- >>>>> fs/nfs/write.c | 3 +- >>>>> include/linux/nfs_fs.h | 5 +- >>>>> include/linux/nfs_iostat.h | 2 +- >>>>> include/linux/nfs_page.h | 1 + >>>>> include/linux/nfs_xdr.h | 1 + >>>>> 16 files changed, 339 insertions(+), 504 deletions(-) >>>>> >>>>> -- >>>>> 1.8.3.1 >>>>> >>>> >>>> -- >>>> Chuck Lever >> >> -- >> Chuck Lever -- Chuck Lever
