On 29/10/2022 01:53, Jason Gunthorpe wrote:
On Tue, Sep 27, 2022 at 01:53:31PM +0800, Li Zhijian wrote:
@@ -122,6 +129,7 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
int num_buf;
void *vaddr;
int err;
+ bool is_pmem = false;
int i;
umem = ib_umem_get(&rxe->ib_dev, start, length, access);
@@ -149,6 +157,7 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
num_buf = 0;
map = mr->map;
if (length > 0) {
+ is_pmem = true;
buf = map[0]->buf;
for_each_sgtable_page (&umem->sgt_append.sgt, &sg_iter, 0) {
@@ -166,6 +175,10 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
goto err_cleanup_map;
}
+ /* True only if the *whole* MR is pmem */
+ if (is_pmem)
+ is_pmem = vaddr_in_pmem(vaddr);
+
I'm not so keen on this use of resources, but this should be written more
like
phys = page_to_phys(sg_page_iter_page(&sg_iter))
region_intersects(phys + sg_iter->offset, sg_iter->length,.. )
And you understand this will make memory registration of every RXE
user a bit slower?
Good catch, i missed it before.
I tested it qemu guest in which pmem is backing to a normal file in host.
In this case, this testing take ~+9% overhead(1.2S -> 1.3S) for 1G size mr. most the time was taken by gup.
the real pmem environment will be tested later.
To minimize side effect, i updated the code to do pmem mr checking on if the require_pmem is true.
region_intersects(phys + sg_iter->offset, sg_iter->length,.. )
I haven't fully apply this suggestion since i think my assumption that a page can only associate to a unique/same
memory zone is true. So i only check 1 byte of each page.
index 5d014cef916e..e4e7c180fa0d 100644
--- a/drivers/infiniband/sw/rxe/rxe_mr.c
+++ b/drivers/infiniband/sw/rxe/rxe_mr.c
@@ -112,6 +112,13 @@ void rxe_mr_init_dma(int access, struct rxe_mr *mr)
mr->ibmr.type = IB_MR_TYPE_DMA;
}
+static bool paddr_in_pmem(unsigned long paddr)
+{
+ return REGION_INTERSECTS ==
+ region_intersects(paddr, 1, IORESOURCE_MEM,
+ IORES_DESC_PERSISTENT_MEMORY);
+}
+
int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
int access, struct rxe_mr *mr)
{
@@ -122,6 +129,7 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
int num_buf;
void *vaddr;
int err;
+ bool require_pmem = access & IB_ACCESS_FLUSH_PERSISTENT;
umem = ib_umem_get(&rxe->ib_dev, start, length, access);
if (IS_ERR(umem)) {
@@ -149,6 +157,7 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
num_buf = 0;
map = mr->map;
if (length > 0) {
+ struct page *pg;
buf = map[0]->buf;
for_each_sgtable_page (&umem->sgt_append.sgt, &sg_iter, 0) {
@@ -158,13 +167,20 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
num_buf = 0;
}
- vaddr = page_address(sg_page_iter_page(&sg_iter));
+ pg = sg_page_iter_page(&sg_iter);
+ vaddr = page_address(pg);
if (!vaddr) {
pr_warn("%s: Unable to get virtual address\n",
__func__);
err = -ENOMEM;
goto err_release_umem;
}
+
+ if (require_pmem && !paddr_in_pmem(page_to_phys(pg))) {
+ err = -EINVAL;
+ goto err_release_umem;
+ }
+
buf->addr = (uintptr_t)vaddr;
num_buf++;
buf++;
And actual pmem will be painfully slow.
It seems like we are doing something wrong here..
Do you think we don't need this patch ?
@@ -174,6 +187,12 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
}
}
+ if (!is_pmem && access & IB_ACCESS_FLUSH_PERSISTENT) {
+ pr_warn("Cannot register IB_ACCESS_FLUSH_PERSISTENT for non-pmem memory\n");
+ err = -EINVAL;
+ goto err_release_umem;
+ }
Do not pr_warn on syscall paths
Got it, will remove it.
Thanks
Zhijian
Jason
