Re: [PATCH] mlx4: vmalloc for mlx4_ib_wq.wrid and mlx4_ib_srq.wrid

[Wengang Wang]

Hi Or,

在 2015年09月24日 13:33, Or Gerlitz 写道:

On 9/24/2015 5:10 AM, Wengang Wang wrote:
Use __vmalloc to allocate memory for mlx4_ib_wq.wrid and 
mlx4_ib_srq.wrid.


Several hits that the kmalloc for wrid failed with the following like
call back stack:


Using vmalloc and friends should be done with care, specifically, we'd 
like
to go there only when needed (namely when kmalloc fails), else we can 
get into

another set of troubles.

Please use the practice introduced in commit 89dd86d "mlx4_core: Allow 
large mlx4_buddy bitmaps"
to go the vmalloc way only when needed, note you can just call kvfree 
later, no need to branch
when freeing things (as was fixed later in commit 914efb0 "mlx4: don't 
duplicate kvfree()")


And have the patch title to be

IB/mlx4: Use vmalloc  for WR buffers when needed

I don't think the OOM oops in the change log helps, skip it.



OK, will look at commit 89dd86d and try again.

thanks for so quick response.

thanks
wengang

Or.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html


--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH] mlx4: vmalloc for mlx4_ib_wq.wrid and mlx4_ib_srq.wrid

[Wengang Wang]

Use __vmalloc to allocate memory for mlx4_ib_wq.wrid and mlx4_ib_srq.wrid.

Several hits that the kmalloc for wrid failed with the following like
call back stack:

kworker/u:4: page allocation failure: order:4, mode:0x2000d0
Pid: 16388, comm: kworker/u:4 Not tainted
Call Trace:
 [] warn_alloc_failed+0xf3/0x160
 [] ? __alloc_pages_direct_compact+0x1fa/0x200
 [] __alloc_pages_slowpath+0x4a6/0x7b0
 [] __alloc_pages_nodemask+0x2fb/0x320
 [] kmem_getpages+0x67/0x1c0
 [] fallback_alloc+0x187/0x250
 [] cache_alloc_node+0x9a/0x150
 [] __kmalloc+0x18b/0x340
 [] ? create_qp_common+0x431/0x8e0 [mlx4_ib]
 [] create_qp_common+0x431/0x8e0 [mlx4_ib]
 [] ? kzalloc.clone.1+0xe/0x10 [mlx4_ib]
 [] mlx4_ib_create_qp+0x207/0x310 [mlx4_ib]
 [] ib_create_qp+0x41/0x1c0 [ib_core]
 [] ipoib_cm_create_tx_qp+0xc8/0x130 [ib_ipoib]
 [] ? __vmalloc_node+0x35/0x40
 [] ipoib_cm_tx_init+0x65/0x380 [ib_ipoib]
 [] ? sched_clock_cpu+0xcd/0x110
 [] ? xen_mc_flush+0xb0/0x1b0
 [] ipoib_cm_tx_start+0x230/0x3d0 [ib_ipoib]
 [] process_one_work+0x180/0x420
 [] worker_thread+0x12e/0x390
 [] ? manage_workers+0x180/0x180
 [] kthread+0xce/0xe0
 [] ? xen_end_context_switch+0x1e/0x30
 [] ? kthread_freezable_should_stop+0x70/0x70
 [] ret_from_fork+0x7c/0xb0
 [] ? kthread_freezable_should_stop+0x70/0x70

It needs 16 contigous pages and failed. At the time there actually is
100MB+ free memory:

Node 0 Normal: 10268*4kB (UM) 7443*8kB (UEM) 1647*16kB (UM) 35*32kB (UR)
1*64kB (R) 4*128kB (R) 1*256kB (R) 0*512kB 1*1024kB (R) 0*2048kB 0*4096kB =
129944kB

I also hit same errors order 3.

Signed-off-by: Wengang Wang 
---
 drivers/infiniband/hw/mlx4/qp.c  | 15 +--
 drivers/infiniband/hw/mlx4/srq.c |  6 --
 2 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/drivers/infiniband/hw/mlx4/qp.c b/drivers/infiniband/hw/mlx4/qp.c
index 4ad9be3..754ceb9 100644
--- a/drivers/infiniband/hw/mlx4/qp.c
+++ b/drivers/infiniband/hw/mlx4/qp.c
@@ -34,6 +34,7 @@
 #include 
 #include 
 #include 
+#include 
 
 #include 
 #include 
@@ -786,8 +787,10 @@ static int create_qp_common(struct mlx4_ib_dev *dev, 
struct ib_pd *pd,
if (err)
goto err_mtt;
 
-   qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), gfp);
-   qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), gfp);
+   qp->sq.wrid  = __vmalloc(qp->sq.wqe_cnt * sizeof(u64), gfp,
+PAGE_KERNEL);
+   qp->rq.wrid  = __vmalloc(qp->rq.wqe_cnt * sizeof(u64), gfp,
+PAGE_KERNEL);
if (!qp->sq.wrid || !qp->rq.wrid) {
err = -ENOMEM;
goto err_wrid;
@@ -874,8 +877,8 @@ err_wrid:
if (qp_has_rq(init_attr))

mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), >db);
} else {
-   kfree(qp->sq.wrid);
-   kfree(qp->rq.wrid);
+   vfree(qp->sq.wrid);
+   vfree(qp->rq.wrid);
}
 
 err_mtt:
@@ -1050,8 +1053,8 @@ static void destroy_qp_common(struct mlx4_ib_dev *dev, 
struct mlx4_ib_qp *qp,
  >db);
ib_umem_release(qp->umem);
} else {
-   kfree(qp->sq.wrid);
-   kfree(qp->rq.wrid);
+   vfree(qp->sq.wrid);
+   vfree(qp->rq.wrid);
if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
free_proxy_bufs(>ib_dev, qp);
diff --git a/drivers/infiniband/hw/mlx4/srq.c b/drivers/infiniband/hw/mlx4/srq.c
index dce5dfe..6d21bb2 100644
--- a/drivers/infiniband/hw/mlx4/srq.c
+++ b/drivers/infiniband/hw/mlx4/srq.c
@@ -34,6 +34,7 @@
 #include 
 #include 
 #include 
+#include 
 
 #include "mlx4_ib.h"
 #include "user.h"
@@ -170,7 +171,8 @@ struct ib_srq *mlx4_ib_create_srq(struct ib_pd *pd,
if (err)
goto err_mtt;
 
-   srq->wrid = kmalloc(srq->msrq.max * sizeof (u64), GFP_KERNEL);
+   srq->wrid = __vmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL,
+ PAGE_KERNEL);
if (!srq->wrid) {
err = -ENOMEM;
goto err_mtt;
@@ -204,7 +206,7 @@ err_wrid:
if (pd->uobject)
mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), 
>db);
else
-   kfree(srq->wrid);
+   vfree(srq->wrid);
 
 err_mtt:
mlx4_mtt_cleanup(dev->dev, >mtt);
-- 
2.1.0

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH] mlx4: vmalloc for mlx4_ib_wq.wrid and mlx4_ib_srq.wrid

[Or Gerlitz]

On 9/24/2015 5:10 AM, Wengang Wang wrote:

Use __vmalloc to allocate memory for mlx4_ib_wq.wrid and mlx4_ib_srq.wrid.

Several hits that the kmalloc for wrid failed with the following like
call back stack:


Using vmalloc and friends should be done with care, specifically, we'd like
to go there only when needed (namely when kmalloc fails), else we can 
get into

another set of troubles.

Please use the practice introduced in commit 89dd86d "mlx4_core: Allow 
large mlx4_buddy bitmaps"
to go the vmalloc way only when needed, note you can just call kvfree 
later, no need to branch
when freeing things (as was fixed later in commit 914efb0 "mlx4: don't 
duplicate kvfree()")


And have the patch title to be

IB/mlx4: Use vmalloc  for WR buffers when needed

I don't think the OOM oops in the change log helps, skip it.

Or.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html