On 4/29/2018 8:34 PM, Rohit Zambre wrote:
On Sun, Apr 29, 2018 at 11:12 AM, Yishai Hadas
<yishaih@xxxxxxxxxxxxxxxxxx> wrote:
On 4/23/2018 8:58 PM, Rohit Zambre wrote:
The calculation of need_uuar_lock rightly accounts for at least 1 medium
bfreg.
The code should support even 0 medium bfregs comparing the suggested patch below.
So when the user sets MLX5_TOTAL_UUARS=16 and
MLX5_NUM_LOW_LAT_UUARS=15, the first 15 QPs will map to the 15 low
latency bfregs. To which bfreg will the 16th QP map to?
The 16th QP expects to get bfreg index 0 which is returned from the
kernel once there are *no* low latency high class nor medium bfregs
available. In this case DB is used in rdma-core instead of BF and no
lock is needed. (see usage of bfs[bfi].buf_size)
However, looking in the kernel code, it seems that there is some bug in
'first_med_bfreg()' which had to handle this non-default case when there
are no medium bfregs.
Below patch expects to fix that, can you please give it a try and
confirm the fix ? this should come in parallel with the candidate patch
that I have sent yesterday in rdma-core in need_uuar_lock().
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -480,8 +480,12 @@ static int qp_has_rq(struct ib_qp_init_attr *attr)
return 1;
}
-static int first_med_bfreg(void)
+static int first_med_bfreg(struct mlx5_bfreg_info *bfregi)
{
+ if (bfregi->total_num_bfregs -
+ (bfregi->num_low_latency_bfregs + bfregi->num_dyn_bfregs) == 1)
+ return -ENOMEM;
+
return 1;
}
@@ -537,10 +541,13 @@ static int alloc_high_class_bfreg(struct
mlx5_ib_dev *dev,
static int alloc_med_class_bfreg(struct mlx5_ib_dev *dev,
struct mlx5_bfreg_info *bfregi)
{
- int minidx = first_med_bfreg();
+ int minidx = first_med_bfreg(bfregi);
int i;
- for (i = first_med_bfreg(); i < first_hi_bfreg(dev, bfregi); i++) {
+ if (minidx < 0)
+ return minidx;
+
+ for (i = minidx; i < first_hi_bfreg(dev, bfregi); i++) {
if (bfregi->count[i] < bfregi->count[minidx])
minidx = i;
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
