Hi there
The LU-9026 patch will allow building with MOFED 4.0 but there are still known issues attempting to run.
Doug
Do you have any suggestions here
Peter
The "Async QP event type 3" is a IB_EVENT_QP_ACCESS_ERR. This error will stop the connection from continuing and explains all the errors which follow in your logs.
This error happens when a call to ib_create_qp() fails. It could fail if the version of the parameters being passed in is wrong (i.e. the size of the parameter structure is incorrect). This could be related to the other issue I am currently working on that involves MOFED 4. If MOFED 4 has changed the structure we use as a parameter to this call and we have not adapted to that change, we could see an error like this.
Does this error happen on each mount attempt or was this a one off?
This error happen on each mount attempt.(the test lustre filesystem servers OFED is 3.4)
And "Async QP event type 3" also happened when lustre servers mount mgs/mds/....with OFED4.
I ran into this "Async" error at the same time as the issues I talk about in LU-9500. They are related. When I have a fix for LU-9500, this issue will be addressed as well.
I think it's my fault.
Could you try this patch on top of LU-9472 and LU-9500?
diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c
index 047fe3c..ba7829b 100644
--- a/lnet/klnds/o2iblnd/o2iblnd.c
+++ b/lnet/klnds/o2iblnd/o2iblnd.c
@@ -1900,8 +1900,6 @@ again:
return n < 0 ? n : -EINVAL;
}
- mr->iova = iov;
-
wr = &frd->frd_fastreg_wr;
memset(wr, 0, sizeof(*wr));
I'm curious as to why you would not want to set the mr->iova value? Is this an unneeded step?
Hi Doug,
I believe the issue only applies to mlx5 cards using MOFED4.
in MOFED4, mr->iova is set by ib_map_mr_sg()->ib_sg_to_pages()
It doesn't make sense to reset mr->iova after calling ib_map_mr_sg().
That line of code was introduced to address an similar issue, see the comments on
https://review.whamcloud.com/#/c/19168/
I've done some testing using MOFED4 + lustre-release with mlx4 cards forcing fast reg as well, so far I've seen no problems.
I can mount lustre ok ( 2.9.57_69_g0bc1964 + LU-9472 / LU-9500 and this patch (- mr->iova = iov
)
Is it ok to apply only LU-9026 + LU-9472 / LU-9500 and (- mr->iova = iov patchs to 2.9.0?
Sould I apply any other patchs to Lustre2.9.0 for mlx5 cards using MOFED4 ? Thanks for your suggestion
Li: Thank you for the information. I checked and you are right, ib_map_mr_sg() does set mr->iova so that line is not needed (could cause a problem). I will update LU-9500 with a removal of that line.
sebg-crd-pm: Correct, you only need --LU-9026LU-9500 (with the removal of setting mr->iova) + LU-9472.
We have just encountered this issue as well. Is it possible to have LU-9026, LU-9500, and LU-9472 backported to the lustre 2.5 and 2.8 branches?
dinatale2 can you please open a new ticket to track this request?
Hi,
I got create striped directory error in Lustre 2.10 with LU-9500 patch (https://review.whamcloud.com/#/c/28237/) for OFED4.0
[root@hsm client]# lfs mkdir -c 2 dir1
error on LL_IOC_LMV_SETSTRIPE 'dir1' (3): Input/output error
error: mkdir: create stripe dir 'dir1' failed
Should I apply any other patch for this issue? Thanks
This issue is being tracked under LU-9958