If a block size percentage is ispecified as 0 in bssplit, the block size
defined is not ignored by the loop in get_next_buflen(). In particular,
if the first (smallest) block size specified has a 0 percentage, the
loop is existed and that block size used as the next IO size, resulting
in a behavior equivalent to specifying 100%. E.g. using
--bssplit=64k/0,1024k/100 results in 100% of issued IOs to be 64KB
instead of 1MB.
Fix this by ignoring bssplit entries that have a 0 percentage. This is
safe as the initialization of the bssplit array ensure that the sum of
all percentages is always 100, guaranteeing that a block size will be
chosen for the next IO size.
Signed-off-by: Damien Le Moal <damien.lemoal@xxxxxxx>
---
io_u.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/io_u.c b/io_u.c
index 910b7deb..80df2854 100644
--- a/io_u.c
+++ b/io_u.c
@@ -557,10 +557,10 @@ static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *i
for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
struct bssplit *bsp = &td->o.bssplit[ddir][i];
+ if (!bsp->perc)
+ continue;
buflen = bsp->bs;
perc += bsp->perc;
- if (!perc)
- break;
if ((r / perc <= frand_max / 100ULL) &&
io_u_fits(td, io_u, buflen))
break;
--
2.21.0
