On 08/29/18 21:15, Damien Le Moal wrote:
> With zonemode=zbd, for random read operations with read_beyond_wp=0,
> the zbd code will always adjust an I/O offset so that the I/O hits a
> non empty zone. However, the adjustment always sets the I/O offset to
> the start of the zone, resulting in a high device read cache hit rate
> if the device has few zones written.
>
> Improve randomness of read I/Os by adjusting the I/O offset to a random
> value within the range of written data of the chosen zone. Also ensure
> that the modified I/O does not cross over the zone wp position by
> adjusting its size.
>
> Signed-off-by: Damien Le Moal <damien.lemoal@xxxxxxx>
> ---
> zbd.c | 35 +++++++++++++++++++++++++++++++++--
> 1 file changed, 33 insertions(+), 2 deletions(-)
>
> diff --git a/zbd.c b/zbd.c
> index 56197693..19511454 100644
> --- a/zbd.c
> +++ b/zbd.c
> @@ -1122,7 +1122,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
> struct fio_zone_info *zb, *zl;
> uint32_t orig_len = io_u->buflen;
> uint32_t min_bs = td->o.min_bs[io_u->ddir];
> - uint64_t new_len;
> + uint64_t new_len, zofst;
> int64_t range;
>
> if (!f->zbd_info)
> @@ -1168,6 +1168,8 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
> }
> if (zb->cond == BLK_ZONE_COND_OFFLINE ||
> (io_u->offset + io_u->buflen) >> 9 > zb->wp) {
> + struct fio_zone_info *orig_zb = zb;
> +
> pthread_mutex_unlock(&zb->mutex);
> zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
> f->file_offset + f->io_size)];
> @@ -1179,7 +1181,36 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
> io_u->buflen);
> goto eof;
> }
> - io_u->offset = zb->start << 9;
> + /*
> + * zbd_find_zone() returned a zone with a range of at
> + * least min_bs, but range may be less than the I/O
> + * size. Handle this case here and also make sure that
> + * the I/O does not the cross the zone wp.
> + */
> + range = ((zb->wp - zb->start) << 9) / min_bs * min_bs;
> + if ((!td_random(td)) || range <= io_u->buflen) {
> + io_u->offset = zb->start << 9;
> + } else {
> + zofst = ((io_u->offset - (orig_zb->start << 9)) %
> + range) / min_bs * min_bs;
> + if (zofst >= range)
> + io_u->offset =
> + ((zb->wp << 9) / min_bs - 1) *
> + min_bs;
> + else
> + io_u->offset = (zb->start << 9) + zofst;
The "zofst >= range" test doesn't make sense to me. Since "zofst" is the
result of rounding down the result of a "% range" operation it is
guaranteed that zofst < range.
> + }
> + new_len = min((unsigned long long)io_u->buflen,
> + (unsigned long long)(zb->wp << 9) -
> + io_u->offset);
> + new_len = new_len / min_bs * min_bs;
> + if (new_len < io_u->buflen) {
> + io_u->buflen = new_len;
> + dprint(FD_IO, "Changed length from %u into %llu\n",
> + orig_len, io_u->buflen);
> + }
> + assert(zb->start << 9 <= io_u->offset);
> + assert(io_u->offset + io_u->buflen <= zb->wp << 9);
> }
> if ((io_u->offset + io_u->buflen) >> 9 > zb->wp) {
> dprint(FD_ZBD, "%s: %lld + %lld > %" PRIu64 "\n",
>
The approach of this patch does not make sense to me: with this patch
applied, if the read request doesn't fit below the write pointer, a new
zone is selected and if the read request doesn't fit in that zone buflen
is reduced. Wouldn't it be better to shrink the request if zone *zb is
not empty and only to call zbd_find_zone() if zone *zb is empty?
Bart.
