On 03/24/2015 06:27 AM, Dave Chinner wrote: > On Thu, Mar 05, 2015 at 09:03:48AM +1100, Dave Chinner wrote: >> On Wed, Mar 04, 2015 at 04:54:50PM +0200, Boaz Harrosh wrote: >>> On 03/04/2015 03:01 PM, Dave Chinner wrote: >>>> On Wed, Mar 04, 2015 at 12:09:40PM +0200, Boaz Harrosh wrote: >>> <> >>>> >>>> So, we definitely need splice to/from DAX enabled inodes to be >>>> rejected. I'll have a look at that... >>>> >>> >>> default_file_splice_read uses kernel_readv which I think might actually >>> work. Do you know what xfstest(s) exercise splice? >> >> We have a rudimentary one only because I discovered a while back >> none existed at all. i.e. splice is effectively untested by >> xfstests. If you want to write some tests to execise it, that'd be >> great.... > > Turns out there's no great need to write splice tests for xfstests - > the current loopback device uses splice, and so all of the tests > that run on loopback are exercising the splice path through the > filesystem. > > I found this out by disabling splice on dax altogether, and then finding out > that lots of tests failed badly, then narrowing it down to: > > $ sudo mount -o dax /dev/ram0 /mnt/test > $ sudo mkfs.xfs -dfile,name=/mnt/test/foo1,size=1g > meta-data=/mnt/test/foo1 isize=512 agcount=4, agsize=65536 blks > = sectsz=512 attr=2, projid32bit=1 > = crc=1 finobt=1 > data = bsize=4096 blocks=262144, imaxpct=25 > = sunit=0 swidth=0 blks > naming =version 2 bsize=4096 ascii-ci=0 ftype=1 > log =internal log bsize=4096 blocks=2560, version=2 > = sectsz=512 sunit=0 blks, lazy-count=1 > realtime =none extsz=4096 blocks=0, rtextents=0 > $ sudo mount -o loop /mnt/test/foo1 /mnt/test/foo > mount: /dev/loop0: can't read superblock > $ > > because the splice read returned EINVAL rather than data. So, yes, > splice canbe made to work with dax if we pass it through the paths > that aren't interacting directly with the page cache. > Cool so current dax code actually does support splice by using default_file_splice_read/write indirectly. therefor I think there is merit in keeping just the one file_operations vector pointing to an internal function and doing an if (IS_DAX()) default_file_splice_read/write() at run time. Because with current code, if CONFIG_FS_DAX is enabled at compile time, then also the regular HD none dax mounts will use the slow default_file_splice_read instead of what ever something better that the FS is doing. Do you think we should do the IS_DAX() switch at generic_file_splice_read and iter_file_splice_write to fix all the Fss in one go or point to an internal FS function and do the switch there? Please advise? I will send up a patch that fixes up ext2 ext4, to see how it looks like. > Cheers, > Dave. > Thanks Dave Boaz _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
