Hello all, I am doing some 520byte per sector testing and utilizing
the sg engine. so far it is working ok (doesn't seem to be able to
have iodepth > 16 but that is workable with jobs...)
I am running into an issue with 32K block sizes though.
If I want to run 4K transfers I will transfer 520*8 = 4,160bytes
16K = 520*32 = 16,640 bytes
32K = 520*64 = 33,280 bytes
Unfortunately 33,280 fails with "read/write not sector aligned"
Checking the code engines/sg.c
static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
{
struct sg_io_hdr *hdr = &io_u->hdr;
struct sg_options *o = td->eo;
struct sgio_data *sd = td->io_ops_data;
unsigned long long nr_blocks, lba;
int offset;
if (io_u->xfer_buflen & (sd->bs - 1)) {
log_err("read/write not sector aligned\n");
return EINVAL;
}
Issue is with how it is calculating whether the transfer size is
aligned or not. It all works for 512/4096 aligned IO yet it breaks
down when doing 520 bytes aligned IO
Chart below shows different block sizes
519: 10 0000 0111 (sd->bs - 1)
1040: 100 0001 0000 & (sd->bs - 1) = 1
4160: 1 0000 0100 0000 & (sd->bs - 1) = 0 OK
8320: 10 0000 1000 0000 & (sd->bs - 1) = 0 OK
16640: 100 0001 0000 0000 & (sd->bs - 1) = 0 OK
33280: 1000 0010 0000 0000 & (sd->bs - 1) = 1
It actually doesn't work for anything below 4160 and block sizes
between 33280 and 66,560
It can be fixed by changing the following
if (io_u->xfer_buflen & (sd->bs - 1)) {
to
if (io_u->xfer_buflen % sd->bs != 0 ) {
I am sure there is a better way but it works for me. Is there anyway
to get this change in (or something similar)?
Thank you,
Kurt
