Linux 4.14 compat: vfs_read/write

[rfehren]

Commit torvalds/linux@bd8df82b
- Unexport vfs_read and vfs_write

Patch received basic run-time testing applied to the spl-0.7-release branch and built against 4.14-rc3.

[behlendorf]

Thanks for noticing this and providing a patch. I've added some review comments inline. When you're refreshing this can you please also do the following.

• Rebase on SPL master.
• Open an additional dummy PR against ZFS with the line Requires-spl: refs/pull/659/head in the commit message. This way the SPL commit will be run through our automated testing against a variety of kernels. PR Test SPL PR - style fixes zfs#6726 is a good example of this.
• Add your signed-off-by to the commit git commit -s.

[behlendorf]

It looks like kernel_write() was exported in v3.8 torvalds/linux@7bb307e and kernel_read() in 3.9 torvalds/linux@3dc20cb. This is arguably when we should have moved over to these new interfaces.

What we need to do to handle this is to add a configure check for both kernel_read and kernel_write and a compatibility wrapper for each of these called spl_kernel_read() and spl_kernel_write() for consistency. The wrappers should prefer the kernel_* interfaces if they're determined to be available. And since these functions are only called from spl-vnode.c they can be added there and declared static.

There are a couple things to be careful of when implementing the wrappers.

• The arguments for the functions do not appear in the same order.
• vfs_read/write functions don't handle the set_fs() bit but the kernel_read/write functions do. This logic will need to be pulled in to the wrapper for only the vfs_read/write case and out of vn_rdwr.

[rfehren]

Unfortunately, I won't have the time to work further on this issue. Concerning your comments: I would actually leave it the way it is. All previous ZFS versions compiled against kernels < 4.14 have been using vfs_read/write and were tested against this. Why risk catching a regression with such kernels due to a possibly buggy early implemenation of kernel_read/write? I don't see a benefit, only potential problems when using kernel_read/write with kernels < 4.14.
I signed off the commit rebased and pushed.

[bmaxa]

I have problem with this patch, cannot mount files on older kernels. Drives mount ok.
I also have this errors in dmesg on kernel 4.14 when importing pools,
noop error is because I use blk-mq and there noop is called none.
143.026192] attempt to access beyond end of device
[ 143.026194] unknown-block(8,0): rw=0, want=468862135, limit=468862128
[ 143.026195] Buffer I/O error on dev sda1, logical block 58607758, async page read
[ 144.193110] attempt to access beyond end of device
[ 144.193115] unknown-block(8,16): rw=0, want=3907029175, limit=3907029168
[ 144.193119] attempt to access beyond end of device
[ 144.193121] unknown-block(8,16): rw=0, want=3907029175, limit=3907029168
[ 144.193123] Buffer I/O error on dev sdb1, logical block 488378638, async page read
[ 145.009731] ZFS: Unable to set "noop" scheduler for /dev/disk/by-id/ata-ST2000VN000-1H3164_W1H2DSKA-part1 (sdc): 256
[ 145.256794] ZFS: Unable to set "noop" scheduler for /dev/disk/by-id/ata-ST2000VN000-1H3164_W1H2DSKA-part1 (sdc): 256

[behlendorf]

Replaced by #667.