On 1/22/18 4:24 PM, Bart Van Assche wrote:
> On Mon, 2018-01-22 at 16:14 -0700, Jens Axboe wrote:
>> On 1/22/18 3:25 PM, Elliott, Robert (Persistent Memory) wrote:
>>> fio engines/sg.c fio_sgio_rw_doio() has that pattern:
>>>
>>> ret = write(f->fd, hdr, sizeof(*hdr));
>>> if (ret < 0)
>>> return ret;
>>> ...
>>> return FIO_Q_QUEUED; [which is 1]
>>>
>>> although there might be special circumstances for the sg interface
>>> making that safe.
>>
>> That's for SG scsi direct IO, I don't think that supports partial
>> IO since it's sending raw SCSI commands.
>>
>> For the regular libaio or sync IO system calls, fio of course checks
>> and handles short IOs correctly. It even logs if it got any.
>
> The entire fio_sgio_rw_doio() function is as follows:
>
> static int fio_sgio_rw_doio(struct fio_file *f, struct io_u *io_u, int do_sync)
> {
> struct sg_io_hdr *hdr = &io_u->hdr;
> int ret;
>
> ret = write(f->fd, hdr, sizeof(*hdr));
> if (ret < 0)
> return ret;
>
> if (do_sync) {
> ret = read(f->fd, hdr, sizeof(*hdr));
> if (ret < 0)
> return ret;
>
> /* record if an io error occurred */
> if (hdr->info & SG_INFO_CHECK)
> io_u->error = EIO;
>
> return FIO_Q_COMPLETED;
> }
>
> return FIO_Q_QUEUED;
> }
>
> I think the 'resid' member of the struct sg_io_hdr that is provided by the
> sg_io kernel driver as a response represents the number of bytes that has not
> been written. So it should be possible to recognize and handle short I/Os in
> that function. From include/scsi/sg.h:
>
> int resid; /* [o] dxfer_len - actual_transferred */
Yeah that's true.
But let's not side track the discussion here, fio+sg isn't relevant to the
topic at hand. Fio and other engines would be (like libaio, or sync and
friends), but those handle short/partial IOs just fine.
That said, if someone wants to submit a patch for the sg engine, I would
of course take it.
--
Jens Axboe
