On 08/28/18 17:57, Damien Le Moal wrote:
> Thinking more about this, since zbd_find_zone() only guarantees that the zone
> returned has min_bs data in it, range can in fact still be <= 0 if
> io_u->buflen is larger than min_bs. Is it assumed here that min_bs and
> io_u->buflen are always equal ?
There is no such assumption. io_u->buflen can be larger than min_bs.
E.g. test11 in test-zbd-support triggers that case because it uses the
bsrange option.
> If yes, then range < 0 is not possible, but range == 0 is definitely still
> possible and so the +1 is needed.
Have you considered to leave out the "+ 1" and to handle range == 0
separately?
> Looking at get_next_buflen(), io_u->buflen can be larger than min_bs. So
> either we need to modify zbd_find_zone() to return a zone with at least
> io_u->buflen bytes in it, or we need to handle the range < 0 case...
The range < 0 case is already handled by the following code at the end
of the DDIR_WRITE case:
/*
* Make sure that the buflen is a multiple of the minimal
* block size. Give up if shrinking would make the request too
* small.
*/
new_len = min((unsigned long long)io_u->buflen,
((zb + 1)->start << 9) - io_u->offset);
new_len = new_len / min_bs * min_bs;
if (new_len == io_u->buflen)
goto accept;
if (new_len >= min_bs) {
io_u->buflen = new_len;
dprint(FD_IO, "Changed length from %u into %llu\n",
orig_len, io_u->buflen);
goto accept;
}
log_err("Zone remainder %lld smaller than minimum block size %d\n",
(((zb + 1)->start << 9) - io_u->offset),
min_bs);
goto eof;
Bart.
