From: Jason Gunthorpe <jgg@mellanox.com>
Date: Mon, 19 Aug 2019 14:17:01 +0300
Subject: RDMA/odp: Make it clearer when a umem is an implicit ODP umem
Patch-mainline: v5.4-rc1
Git-commit: fd7dbf035edcfb035977423e2a5102832c1427f4
References: jsc#SLE-8449
Implicit ODP umems are special, they don't have any page lists, they don't
exist in the interval tree and they are never DMA mapped.
Instead of trying to guess this based on a zero length use an explicit
flag.
Further, do not allow non-implicit umems to be 0 size.
Link: https://lore.kernel.org/r/20190819111710.18440-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Acked-by: Thomas Bogendoerfer <tbogendoerfer@suse.de>
---
drivers/infiniband/core/umem_odp.c | 54 ++++++++++++++++++++-----------------
drivers/infiniband/hw/mlx5/mr.c | 2 -
drivers/infiniband/hw/mlx5/odp.c | 2 -
include/rdma/ib_umem_odp.h | 8 +++++
4 files changed, 40 insertions(+), 26 deletions(-)
--- a/drivers/infiniband/core/umem_odp.c
+++ b/drivers/infiniband/core/umem_odp.c
@@ -176,18 +176,15 @@ static void add_umem_to_per_mm(struct ib
struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
down_write(&per_mm->umem_rwsem);
- if (likely(ib_umem_start(umem_odp) != ib_umem_end(umem_odp))) {
- /*
- * Note that the representation of the intervals in the
- * interval tree considers the ending point as contained in
- * the interval, while the function ib_umem_end returns the
- * first address which is not contained in the umem.
- */
- umem_odp->interval_tree.start = ib_umem_start(umem_odp);
- umem_odp->interval_tree.last = ib_umem_end(umem_odp) - 1;
- interval_tree_insert(&umem_odp->interval_tree,
- &per_mm->umem_tree);
- }
+ /*
+ * Note that the representation of the intervals in the interval tree
+ * considers the ending point as contained in the interval, while the
+ * function ib_umem_end returns the first address which is not
+ * contained in the umem.
+ */
+ umem_odp->interval_tree.start = ib_umem_start(umem_odp);
+ umem_odp->interval_tree.last = ib_umem_end(umem_odp) - 1;
+ interval_tree_insert(&umem_odp->interval_tree, &per_mm->umem_tree);
up_write(&per_mm->umem_rwsem);
}
@@ -196,11 +193,8 @@ static void remove_umem_from_per_mm(stru
struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm;
down_write(&per_mm->umem_rwsem);
- if (likely(ib_umem_start(umem_odp) != ib_umem_end(umem_odp)))
- interval_tree_remove(&umem_odp->interval_tree,
- &per_mm->umem_tree);
+ interval_tree_remove(&umem_odp->interval_tree, &per_mm->umem_tree);
complete_all(&umem_odp->notifier_completion);
-
up_write(&per_mm->umem_rwsem);
}
@@ -320,6 +314,9 @@ struct ib_umem_odp *ib_alloc_odp_umem(st
int pages = size >> PAGE_SHIFT;
int ret;
+ if (!size)
+ return ERR_PTR(-EINVAL);
+
odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
if (!odp_data)
return ERR_PTR(-ENOMEM);
@@ -381,6 +378,9 @@ int ib_umem_odp_get(struct ib_umem_odp *
struct mm_struct *mm = umem->owning_mm;
int ret_val;
+ if (umem_odp->umem.address == 0 && umem_odp->umem.length == 0)
+ umem_odp->is_implicit_odp = 1;
+
umem_odp->page_shift = PAGE_SHIFT;
if (access & IB_ACCESS_HUGETLB) {
struct vm_area_struct *vma;
@@ -401,7 +401,10 @@ int ib_umem_odp_get(struct ib_umem_odp *
init_completion(&umem_odp->notifier_completion);
- if (ib_umem_odp_num_pages(umem_odp)) {
+ if (!umem_odp->is_implicit_odp) {
+ if (!ib_umem_odp_num_pages(umem_odp))
+ return -EINVAL;
+
umem_odp->page_list =
vzalloc(array_size(sizeof(*umem_odp->page_list),
ib_umem_odp_num_pages(umem_odp)));
@@ -420,7 +423,9 @@ int ib_umem_odp_get(struct ib_umem_odp *
ret_val = get_per_mm(umem_odp);
if (ret_val)
goto out_dma_list;
- add_umem_to_per_mm(umem_odp);
+
+ if (!umem_odp->is_implicit_odp)
+ add_umem_to_per_mm(umem_odp);
return 0;
@@ -439,13 +444,14 @@ void ib_umem_odp_release(struct ib_umem_
* It is the driver's responsibility to ensure, before calling us,
* that the hardware will not attempt to access the MR any more.
*/
- ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
- ib_umem_end(umem_odp));
-
- remove_umem_from_per_mm(umem_odp);
+ if (!umem_odp->is_implicit_odp) {
+ ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
+ ib_umem_end(umem_odp));
+ remove_umem_from_per_mm(umem_odp);
+ vfree(umem_odp->dma_list);
+ vfree(umem_odp->page_list);
+ }
put_per_mm(umem_odp);
- vfree(umem_odp->dma_list);
- vfree(umem_odp->page_list);
}
/*
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -1599,7 +1599,7 @@ static void dereg_mr(struct mlx5_ib_dev
/* Wait for all running page-fault handlers to finish. */
synchronize_srcu(&dev->mr_srcu);
/* Destroy all page mappings */
- if (umem_odp->page_list)
+ if (!umem_odp->is_implicit_odp)
mlx5_ib_invalidate_range(umem_odp,
ib_umem_start(umem_odp),
ib_umem_end(umem_odp));
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -586,7 +586,7 @@ static int pagefault_mr(struct mlx5_ib_d
struct ib_umem_odp *odp;
size_t size;
- if (!odp_mr->page_list) {
+ if (odp_mr->is_implicit_odp) {
odp = implicit_mr_get_data(mr, io_virt, bcnt);
if (IS_ERR(odp))
--- a/include/rdma/ib_umem_odp.h
+++ b/include/rdma/ib_umem_odp.h
@@ -69,6 +69,14 @@ struct ib_umem_odp {
/* Tree tracking */
struct interval_tree_node interval_tree;
+ /*
+ * An implicit odp umem cannot be DMA mapped, has 0 length, and serves
+ * only as an anchor for the driver to hold onto the per_mm. FIXME:
+ * This should be removed and drivers should work with the per_mm
+ * directly.
+ */
+ bool is_implicit_odp;
+
struct completion notifier_completion;
int dying;
unsigned int page_shift;