cifs: Strange O_DIRECT behaviour on non-directio mounts

[Kevin Wolf]

Hi Steve,

I just discussed a problem with a qemu user on IRC, which boiled down to
him trying to open an image file on cifs with O_DIRECT, but not using a
directio mount. I understand that this probably isn't going to work
anytime soon (if at all), but it resulted in a rather unhelpful failure
mode.

What happens is that cifs lets the open() call succeed even with the
unsupported O_DIRECT on that mount, but then fails any I/O on the file
descriptor. I believe this was introduced in commit dca69288 (which I
think is otherwise pretty useful).

With the old behaviour, qemu detected what's going on and suggested to
use a non-O_DIRECT mode to the user, but with the new one, it got rather
unhappy after failing to find a working O_DIRECT alignment and ran into
an assertion failure...

Now I'll certainly fix the latter in qemu, but I also think that the
behaviour of cifs is rather surprising. Any chance that you can make
open() with O_DIRECT fail again on non-directio mounts?

Thanks,
Kevin

[Kevin Wolf]

Post by Kevin Wolf
Hi Steve,
I just discussed a problem with a qemu user on IRC, which boiled down to
him trying to open an image file on cifs with O_DIRECT, but not using a
directio mount. I understand that this probably isn't going to work
anytime soon (if at all), but it resulted in a rather unhelpful failure
mode.
What happens is that cifs lets the open() call succeed even with the
unsupported O_DIRECT on that mount, but then fails any I/O on the file
descriptor. I believe this was introduced in commit dca69288 (which I
think is otherwise pretty useful).
With the old behaviour, qemu detected what's going on and suggested to
use a non-O_DIRECT mode to the user, but with the new one, it got rather
unhappy after failing to find a working O_DIRECT alignment and ran into
an assertion failure...
Now I'll certainly fix the latter in qemu, but I also think that the
behaviour of cifs is rather surprising. Any chance that you can make
open() with O_DIRECT fail again on non-directio mounts?
Thanks,
Kevin

No interest at all, not even to tell me that I'm completely wrong? :-/

Kevin

[Steve French]

What is the server type? (you can view this by seeing the "NOS" field
in /proc/fs/cifs/DebugData or checking to see if "unix" is listed in
the mount when you do "cat /proc/mounts | grep cifs")

When mounted to Samba for example, with cifs unix extensions
negotiated, the cifs client will send open flags (including O_DIRECT
IIRC) to the server on open. When mounted to Windows/NetApp etc. or
when using "nounix" on mount, the flag is not sent on open.

Post by Kevin Wolf
Hi Steve,
I just discussed a problem with a qemu user on IRC, which boiled down to
him trying to open an image file on cifs with O_DIRECT, but not using a
directio mount. I understand that this probably isn't going to work
anytime soon (if at all), but it resulted in a rather unhelpful failure
mode.
What happens is that cifs lets the open() call succeed even with the
unsupported O_DIRECT on that mount, but then fails any I/O on the file
descriptor. I believe this was introduced in commit dca69288 (which I
think is otherwise pretty useful).
With the old behaviour, qemu detected what's going on and suggested to
use a non-O_DIRECT mode to the user, but with the new one, it got rather
unhappy after failing to find a working O_DIRECT alignment and ran into
an assertion failure...
Now I'll certainly fix the latter in qemu, but I also think that the
behaviour of cifs is rather surprising. Any chance that you can make
open() with O_DIRECT fail again on non-directio mounts?
Thanks,
Kevin

[Kevin Wolf]

Post by Steve French
What is the server type? (you can view this by seeing the "NOS" field
in /proc/fs/cifs/DebugData or checking to see if "unix" is listed in
the mount when you do "cat /proc/mounts | grep cifs")
When mounted to Samba for example, with cifs unix extensions
negotiated, the cifs client will send open flags (including O_DIRECT
IIRC) to the server on open. When mounted to Windows/NetApp etc. or
when using "nounix" on mount, the flag is not sent on open.

As I wrote, this was another user whose problem I debugged over IRC, and
he doesn't seem to be available any more, so I can't get any new
information from him. What I do have is the line that 'mount' outputs:

[12:08:21] <veleno> kwolf: this is what is shown by mount: //192.168.128.200/storage-cs on /var/lib/one/datastores/107 type cifs (rw,username=yy,password=xx,iocharset=utf8,file_mode=0777,dir_mode=0777)

The other information that I have is an strace snippet of the qemu
process that tried using a file with O_DIRECT, but no directio mount
option.

open("/var/lib/one//datastores/1/4131/disk.0", O_RDWR|O_DIRECT|O_CLOEXEC) = 6
[...]
pread(6, 0x7f2512bae200, 4096, 0) = -1 EINVAL (Invalid argument)
pread(6, 0x7f2512bae400, 4096, 0) = -1 EINVAL (Invalid argument)
pread(6, 0x7f2512bae800, 4096, 0) = -1 EINVAL (Invalid argument)
pread(6, 0x7f2512baf000, 4096, 0) = -1 EINVAL (Invalid argument)
pread(6, 0x7f2512badda0, 512, 0) = -1 EINVAL (Invalid argument)
pread(6, 0x7f2512badda0, 1024, 0) = -1 EINVAL (Invalid argument)
pread(6, 0x7f2512badda0, 2048, 0) = -1 EINVAL (Invalid argument)
pread(6, 0x7f2512badda0, 4096, 0) = -1 EINVAL (Invalid argument)
write(2, "qemu-system-x86_64: /build/build"..., 129qemu-system-x86_64: /build/buildd/qemu-2.0.0+dfsg/block.c:942: bdrv_open_common: Assertion `bdrv_opt_mem_align(bs) != 0' failed.
) = 129

Note that remounting with the 'cache=none' mount option, the reporter
could successfully use the file even with O_DIRECT. This and the commit
message of dca69288 made me think that this is not related to the
server, but a client side issue.

Kevin

Post by Steve French

Post by Kevin Wolf
Hi Steve,
I just discussed a problem with a qemu user on IRC, which boiled down to
him trying to open an image file on cifs with O_DIRECT, but not using a
directio mount. I understand that this probably isn't going to work
anytime soon (if at all), but it resulted in a rather unhelpful failure
mode.
What happens is that cifs lets the open() call succeed even with the
unsupported O_DIRECT on that mount, but then fails any I/O on the file
descriptor. I believe this was introduced in commit dca69288 (which I
think is otherwise pretty useful).
With the old behaviour, qemu detected what's going on and suggested to
use a non-O_DIRECT mode to the user, but with the new one, it got rather
unhappy after failing to find a working O_DIRECT alignment and ran into
an assertion failure...
Now I'll certainly fix the latter in qemu, but I also think that the
behaviour of cifs is rather surprising. Any chance that you can make
open() with O_DIRECT fail again on non-directio mounts?
Thanks,
Kevin

--
Thanks,
Steve