This minimal set of patches update the NFS client to use the new
readahead method and convert the fscache read paths to use the new
netfs API, and are at:
https://github.com/DaveWysochanskiRH/kernel/releases/tag/fscache-iter-lib-nfs-20210127
https://github.com/DaveWysochanskiRH/kernel/commit/8693c3602ce02e4038fa732bef02b2641f739e02
The patches are based on David Howells fscache-netfs-lib tree at
https://git.kernel.org/pub/scm/linux/kernel/git/dhowells/linux-fs.git/log/?h=fscache-netfs-lib
The first 6 patches refactor some of the NFS read code to facilitate
re-use, the next 2 patches do the conversion to the new API. Note
that the last patch converts nfs_readpages to nfs_readahead.
Changes since dhowells posting on Jan 25, 2021
- address feedback from Willy (Chuck's feedback in TODO list)
- ensure kernel builds cleanly with each patch
- readahead patch: cleanup handling of 'ret'
Still TODO
1. Fix known bugs (some may disappear with uncondional netfs API)
a) One oops with fscache enabled on parallel read unit test
b) 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)
c) 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")
Tests run
1. Custom NFS+fscache unit tests for basic operation: PASS*
* vers=3,4.0,4.1,4.2,sec=sys,server=localhost (same kernel)
* one fscache enabled parallel read test that oopses due to page lock state
* probably goes away if we unconditionally call netfs API
2. cthon04: PASS
* test options "-b -g -s -l", fsc,vers=3,4.0,4.1,4.2,sec=sys
* No failures, oopses or hangs
3. iozone tests: PASS
* nofsc,fsc,vers=3,4.0,4.1,4.2,sec=sys,server=rhel7,rhel8
* No failures, oopses, or hangs
4. xfstests/generic: PASS*
* no hangs or oopses; a few failures unrelated to these patches
* Ran following configurations
* vers=4.2,fsc,sec=sys,rhel7-server
* vers=4.2,nofsc,sec=sys,rhel8-server (test still running, ok so far)
* vers=4.1,nofsc,sec=sys,netapp-server(pnfs/files)
* vers=4.0,fsc (test still running, ok so far)
Not tested yet:
* error injections (for example, connection disruptions, server errors during IO, etc)
* many process mixed read/write on same file
* performance