Re: [PATCH] bsg : Add support for io vectors in bsg

From: Pete Wyckoff
Date: Mon Jan 14 2008 - 11:19:01 EST


dougg@xxxxxxxxxx wrote on Fri, 11 Jan 2008 19:16 -0500:
> James Bottomley wrote:
>> On Thu, 2008-01-10 at 16:46 -0500, Pete Wyckoff wrote:
>>> James.Bottomley@xxxxxxxxxxxxxxxxxxxxx wrote on Thu, 10 Jan 2008 14:55 -0600:
>>>> On Thu, 2008-01-10 at 15:43 -0500, Pete Wyckoff wrote:
>>>>> fujita.tomonori@xxxxxxxxxxxxx wrote on Wed, 09 Jan 2008 09:11 +0900:
>>>>>> On Tue, 8 Jan 2008 17:09:18 -0500
>>>>>> Pete Wyckoff <pw@xxxxxxx> wrote:
>>>>>>> I took another look at the compat approach, to see if it is feasible
>>>>>>> to keep the compat handling somewhere else, without the use of #ifdef
>>>>>>> CONFIG_COMPAT and size-comparison code inside bsg.c. I don't see how.
>>>>>>> The use of iovec is within a write operation on a char device. It's
>>>>>>> not amenable to a compat_sys_ or a .compat_ioctl approach.
>>>>>>>
>>>>>>> I'm partial to #1 because the use of architecture-independent fields
>>>>>>> matches the rest of struct sg_io_v4. But if you don't want to have
>>>>>>> another iovec type in the kernel, could we do #2 but just return
>>>>>>> -EINVAL if the need for compat is detected? I.e. change
>>>>>>> dout_iovec_count to dout_iovec_length and do the math?
>>>>>> If you are ok with removing the write/read interface and just have
>>>>>> ioctl, we could can handle comapt stuff like others do. But I think
>>>>>> that you (OSD people) really want to keep the write/read
>>>>>> interface. Sorry, I think that there is no workaround to support iovec
>>>>>> in bsg.
>>>>> I don't care about read/write in particular. But we do need some
>>>>> way to launch asynchronous SCSI commands, and currently read/write
>>>>> are the only way to do that in bsg. The reason is to keep multiple
>>>>> spindles busy at the same time.
>>>> Won't multi-threading the ioctl calls achieve the same effect? Or do
>>>> you trip over BKL there?
>>> There's no BKL on (new) ioctls anymore, at least. A thread per
>>> device would be feasible perhaps. But if you want any sort of
>>> pipelining out of the device, esp. in the remote iSCSI case, you
>>> need to have a good number of commands outstanding to each device.
>>> So a thread per command per device. Typical iSCSI queue depth of
>>> 128 times 16 devices for a small setup is a lot of threads.
>>
>> I was actually thinking of a thread per outstanding command.
>>
>>> The pthread/pipe latency overhead is not insignificant for fast
>>> storage networks too.
>>>
>>>>> How about these new ioctls instead of read/write:
>>>>>
>>>>> SG_IO_SUBMIT - start a new blk_execute_rq_nowait()
>>>>> SG_IO_TEST - complete and return a previous req
>>>>> SG_IO_WAIT - wait for a req to finish, interruptibly
>>>>>
>>>>> Then old write users will instead do ioctl SUBMIT. Read users will
>>>>> do TEST for non-blocking fd, or WAIT for blocking. And SG_IO could
>>>>> be implemented as SUBMIT + WAIT.
>>>>>
>>>>> Then we can do compat_ioctl and convert up iovecs out-of-line before
>>>>> calling the normal functions.
>>>>>
>>>>> Let me know if you want a patch for this.
>>>> Really, the thought of re-inventing yet another async I/O interface
>>>> isn't very appealing.
>>> I'm fine with read/write, except Tomo is against handling iovecs
>>> because of the compat complexity with struct iovec being different
>>> on 32- vs 64-bit. There is a standard way to do "compat" ioctl that
>>> hides this handling in a different file (not bsg.c), which is the
>>> only reason I'm even considering these ioctls. I don't care about
>>> compat setups per se.
>>>
>>> Is there another async I/O mechanism? Userspace builds the CDBs,
>>> just needs some way to drop them in SCSI ML. BSG is almost perfect
>>> for this, but doesn't do iovec, leading to lots of memcpy.
>>
>> No, it's just that async interfaces in Linux have a long and fairly
>> unhappy history.
>
> The sg driver's async interface has been pretty stable for
> a long time. The sync SG_IO ioctl is built on top of the
> async interface. That makes the async interface extremely
> well tested.
>
> The write()/read() async interface in sg does have one
> problem: when a command is dispatched via a write()
> it would be very useful to get back a tag but that
> violates write()'s second argument: 'const void * buf'.
> That tag could be useful both for identification of the
> response and by task management functions.
>
> I was hoping that the 'flags' field in sgv4 could be used
> to implement the variants:
> SG_IO_SUBMIT - start a new blk_execute_rq_nowait()
> SG_IO_TEST - complete and return a previous req
> SG_IO_WAIT - wait for a req to finish, interruptibly
>
> that way the existing SG_IO ioctl is sufficient.
>
> And if Tomo doesn't want to do it in the bsg driver,
> then it could be done it the sg driver.

The sg driver already has async via read/write, and it works fine.
Perhaps someone wants the ioctl versions too, but that's not my main
goal here.

I think that sg doesn't bother with compat iovec handling. It just
uses SZ_SG_IOVEC (defined as sizeof(sg_iovec_t)) and doesn't check
if the userspace iovec happens to be smaller. Sg does have a compat
ioctl function; it just doesn't support SG_IO. So, on 64-bit
kernel, read/write from 32-bit userspace with iovec will get
undefined results due to the mis-interpretation of the iovec fields,
while ioctl from 32-bit will fail with ENOIOCTLCMD (EINVAL to
userspace). Doesn't bother me at all, just an observation. I'd
love to be able to take a similar approach with bsg: only support
iovec in 32-32 or 64-64 environments where kernel iovec == user
iovec.

I'm not going to patch up sg to add the SG_IO async ioctls. My
greedy need requires bidirectional transfers and big CDBs, both of
which could be hacked into sg_io_hdr_t and sg, but it sure wouldn't
be pretty. These were some of the reasons you proposed sgv4 as I
recall.

-- Pete
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/