Re: [PATCH] RDMA/mlx5: return the EFAULT per ibv_advise_mr(3)

On 05/08/2021 14:43, Leon Romanovsky wrote:

On Wed, Aug 04, 2021 at 03:50:22PM -0300, Jason Gunthorpe wrote:

quoted

On Wed, Aug 04, 2021 at 08:35:30AM +0300, Leon Romanovsky wrote:

quoted

On Tue, Aug 03, 2021 at 03:13:41PM -0300, Jason Gunthorpe wrote:

quoted

On Tue, Aug 03, 2021 at 08:56:54PM +0300, Leon Romanovsky wrote:

quoted

On Tue, Aug 03, 2021 at 01:25:07PM -0300, Jason Gunthorpe wrote:

quoted

On Sun, Aug 01, 2021 at 05:20:50PM +0800, Li Zhijian wrote:

quoted

ibv_advise_mr(3) says:
EFAULT In one of the following: o When the range requested is out of the  MR  bounds,
        or  when  parts of it are not part of the process address space. o One of the
        lkeys provided in the scatter gather list is invalid or with wrong write access

Actually get_prefetchable_mr() will return NULL if it see above conditions

No, get_prefetchable_mr() returns NULL if the mkey is invalid

And what is this?
   1701 static struct mlx5_ib_mr *
   1702 get_prefetchable_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
   1703                     u32 lkey)

...

   1721         /* prefetch with write-access must be supported by the MR */
   1722         if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
   1723             !mr->umem->writable) {
   1724                 mr = NULL;
   1725                 goto end;
   1726         }

I would say that is an invalid mkey

I see it is as wrong write access.

It just means the man page is wrong

ok, it can be a solution too.

It sounds good.  I will try to update the manpage ibv_advise_mr(3) instead.


Thanks
Zhijian