FWIW I don't have ODP hardware either. So I can't test this either.
On Sat, Mar 02, 2019 at 12:24:35PM -0800, john.hubb...@gmail.com wrote:
> From: John Hubbard <jhubb...@nvidia.com>
>
> 1. Bug fix: the error handling release pages starting
> at the first page that experienced an error.
>
> 2. Refinement: release_pages() is better than put_page()
> in a loop.
>
> 3. Dead code removal: the check for (user_virt & ~page_mask)
> is checking for a condition that can never happen,
> because earlier:
>
> user_virt = user_virt & page_mask;
>
> ...so, remove that entire phrase.
>
> 4. Minor: As long as I'm here, shorten up a couple of long lines
> in the same function, without harming the ability to
> grep for the printed error message.
>
> Cc: Ira Weiny <ira.we...@intel.com>
> Cc: Jason Gunthorpe <j...@ziepe.ca>
> Cc: Andrew Morton <a...@linux-foundation.org>
> Cc: Doug Ledford <dledf...@redhat.com>
> Cc: linux-r...@vger.kernel.org
> Cc: linux...@kvack.org
> Signed-off-by: John Hubbard <jhubb...@nvidia.com>
> ---
>
> v2: Fixes a kbuild test robot reported build failure, by directly
> including pagemap.h
>
> drivers/infiniband/core/umem_odp.c | 25 ++++++++++---------------
> 1 file changed, 10 insertions(+), 15 deletions(-)
>
> diff --git a/drivers/infiniband/core/umem_odp.c
> b/drivers/infiniband/core/umem_odp.c
> index acb882f279cb..83872c1f3f2c 100644
> --- a/drivers/infiniband/core/umem_odp.c
> +++ b/drivers/infiniband/core/umem_odp.c
> @@ -40,6 +40,7 @@
> #include <linux/vmalloc.h>
> #include <linux/hugetlb.h>
> #include <linux/interval_tree_generic.h>
> +#include <linux/pagemap.h>
>
> #include <rdma/ib_verbs.h>
> #include <rdma/ib_umem.h>
> @@ -648,25 +649,17 @@ int ib_umem_odp_map_dma_pages(struct ib_umem_odp
> *umem_odp, u64 user_virt,
>
> if (npages < 0) {
> if (npages != -EAGAIN)
> - pr_warn("fail to get %zu user pages with error
> %d\n", gup_num_pages, npages);
> + pr_warn("fail to get %zu user pages with error
> %d\n",
> + gup_num_pages, npages);
> else
> - pr_debug("fail to get %zu user pages with error
> %d\n", gup_num_pages, npages);
> + pr_debug("fail to get %zu user pages with error
> %d\n",
> + gup_num_pages, npages);
> break;
> }
>
> bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
> mutex_lock(&umem_odp->umem_mutex);
> for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) {
> - if (user_virt & ~page_mask) {
> - p += PAGE_SIZE;
> - if (page_to_phys(local_page_list[j]) != p) {
> - ret = -EFAULT;
> - break;
> - }
> - put_page(local_page_list[j]);
> - continue;
> - }
> -
I think this is trying to account for compound pages. (ie page_mask could
represent more than PAGE_SIZE which is what user_virt is being incrimented by.)
But putting the page in that case seems to be the wrong thing to do?
Yes this was added by Artemy[1] now cc'ed.
> ret = ib_umem_odp_map_dma_single_page(
> umem_odp, k, local_page_list[j],
> access_mask, current_seq);
> @@ -684,9 +677,11 @@ int ib_umem_odp_map_dma_pages(struct ib_umem_odp
> *umem_odp, u64 user_virt,
> mutex_unlock(&umem_odp->umem_mutex);
>
> if (ret < 0) {
> - /* Release left over pages when handling errors. */
> - for (++j; j < npages; ++j)
> - put_page(local_page_list[j]);
> + /*
> + * Release pages, starting at the the first page
> + * that experienced an error.
> + */
> + release_pages(&local_page_list[j], npages - j);
My concern here is that release_pages handle compound pages, perhaps
differently from the above code so calling it may may not work? But I've not
really spent a lot of time on it...
:-/
Ira
[1]
commit 403cd12e2cf759ead5cbdcb62bf9872b9618d400
Author: Artemy Kovalyov <artem...@mellanox.com>
Date: Wed Apr 5 09:23:55 2017 +0300
IB/umem: Add contiguous ODP support
Currenlty ODP supports only regular MMU pages.
Add ODP support for regions consisting of physically contiguous chunks
of arbitrary order (huge pages for instance) to improve performance.
Signed-off-by: Artemy Kovalyov <artem...@mellanox.com>
Signed-off-by: Leon Romanovsky <l...@kernel.org>
Signed-off-by: Doug Ledford <dledf...@redhat.com>
