[PATCH v2 for-rc 0/3] Fixes for 64K page size support

[PATCH v2 for-rc 0/3] Fixes for 64K page size support

[Shiraz Saleem]

This is a three patch series.

The first core hunk corrects the core iterator to use __sg_advance to skip
preceding 4k HCA pages.

The second patch corrects an iWarp issue where the SQ must be PAGE_SIZE
aligned.

The third patch corrects an issue with the RDMA driver use of
ib_umem_find_best_pgsz(). QP and CQ allocations pass PAGE_SIZE as the
only bitmap bit. This is incorrect and should use the precise 4k value.

v1->v2: Add a umem specific block iter next function

Mike Marciniszyn (3):
  RDMA/core: Fix umem iterator when PAGE_SIZE is greater then HCA pgsz
  RDMA/irdma: Ensure iWarp QP queue memory is OS paged aligned
  RDMA/irdma: Fix support for 64k pages

 drivers/infiniband/core/umem.c      | 6 ------
 drivers/infiniband/hw/irdma/verbs.c | 7 ++++++-
 include/rdma/ib_umem.h              | 9 ++++++++-
 include/rdma/ib_verbs.h             | 1 +
 4 files changed, 15 insertions(+), 8 deletions(-)

-- 
1.8.3.1

[PATCH v2 for-rc 1/3] RDMA/core: Fix umem iterator when PAGE_SIZE is greater then HCA pgsz

[Shiraz Saleem]

From: Mike Marciniszyn <mike.marciniszyn@intel.com>

64k pages introduce the situation in this diagram when the HCA
4k page size is being used:

 +-------------------------------------------+ <--- 64k aligned VA
 |                                           |
 |              HCA 4k page                  |
 |                                           |
 +-------------------------------------------+
 |                   o                       |
 |                                           |
 |                   o                       |
 |                                           |
 |                   o                       |
 +-------------------------------------------+
 |                                           |
 |              HCA 4k page                  |
 |                                           |
 +-------------------------------------------+ <--- Live HCA page
 |OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO| <--- offset
 |                                           | <--- VA
 |                MR data                    |
 +-------------------------------------------+
 |                                           |
 |              HCA 4k page                  |
 |                                           |
 +-------------------------------------------+
 |                   o                       |
 |                                           |
 |                   o                       |
 |                                           |
 |                   o                       |
 +-------------------------------------------+
 |                                           |
 |              HCA 4k page                  |
 |                                           |
 +-------------------------------------------+

The VA addresses are coming from rdma-core in this diagram can
be arbitrary, but for 64k pages, the VA may be offset by some
number of HCA 4k pages and followed by some number of HCA 4k
pages.

The current iterator doesn't account for either the preceding
4k pages or the following 4k pages.

Fix the issue by extending the ib_block_iter to contain
the number of DMA pages like comment [1] says and by using
__sg_advance to start the iterator at the first live HCA page.

The changes are contained in a parallel set of iterator start
and next functions that are umem aware and specfic to umem
since there is one user of the rdma_for_each_block() without
umem.

These two fixes prevents the extra pages before and after the
user MR data.

Fix the preceding pages by using the __sq_advance field to start
at the first 4k page containing MR data.

Fix the following pages by saving the number of pgsz blocks in
the iterator state and downcounting on each next.

This fix allows for the elimination of the small page crutch noted
in the Fixes.

Fixes: 10c75ccb54e4 ("RDMA/umem: Prevent small pages from being returned by ib_umem_find_best_pgsz()")
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/rdma/ib_umem.h#n91 [1]
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
---
 drivers/infiniband/core/umem.c | 6 ------
 include/rdma/ib_umem.h         | 9 ++++++++-
 include/rdma/ib_verbs.h        | 1 +
 3 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/drivers/infiniband/core/umem.c b/drivers/infiniband/core/umem.c
index f9ab671c8eda..07c571c7b699 100644
--- a/drivers/infiniband/core/umem.c
+++ b/drivers/infiniband/core/umem.c
@@ -96,12 +96,6 @@ unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
 		return page_size;
 	}
 
-	/* rdma_for_each_block() has a bug if the page size is smaller than the
-	 * page size used to build the umem. For now prevent smaller page sizes
-	 * from being returned.
-	 */
-	pgsz_bitmap &= GENMASK(BITS_PER_LONG - 1, PAGE_SHIFT);
-
 	/* The best result is the smallest page size that results in the minimum
 	 * number of required pages. Compute the largest page size that could
 	 * work based on VA address bits that don't change.
diff --git a/include/rdma/ib_umem.h b/include/rdma/ib_umem.h
index 95896472a82b..726773747764 100644
--- a/include/rdma/ib_umem.h
+++ b/include/rdma/ib_umem.h
@@ -77,6 +77,13 @@ static inline void __rdma_umem_block_iter_start(struct ib_block_iter *biter,
 {
 	__rdma_block_iter_start(biter, umem->sgt_append.sgt.sgl,
 				umem->sgt_append.sgt.nents, pgsz);
+	biter->__sg_advance = ib_umem_offset(umem) & ~(pgsz - 1);
+	biter->__sg_numblocks = ib_umem_num_dma_blocks(umem, pgsz);
+}
+
+static inline bool __rdma_umem_block_iter_next(struct ib_block_iter *biter)
+{
+	     return __rdma_block_iter_next(biter) && biter->__sg_numblocks--;
 }
 
 /**
@@ -92,7 +99,7 @@ static inline void __rdma_umem_block_iter_start(struct ib_block_iter *biter,
  */
 #define rdma_umem_for_each_dma_block(umem, biter, pgsz)                        \
 	for (__rdma_umem_block_iter_start(biter, umem, pgsz);                  \
-	     __rdma_block_iter_next(biter);)
+	     __rdma_umem_block_iter_next(biter);)
 
 #ifdef CONFIG_INFINIBAND_USER_MEM
 
diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h
index fb1a2d6b1969..b7b6b58dd348 100644
--- a/include/rdma/ib_verbs.h
+++ b/include/rdma/ib_verbs.h
@@ -2850,6 +2850,7 @@ struct ib_block_iter {
 	/* internal states */
 	struct scatterlist *__sg;	/* sg holding the current aligned block */
 	dma_addr_t __dma_addr;		/* unaligned DMA address of this block */
+	size_t __sg_numblocks;		/* ib_umem_num_dma_blocks() */
 	unsigned int __sg_nents;	/* number of SG entries */
 	unsigned int __sg_advance;	/* number of bytes to advance in sg in next step */
 	unsigned int __pg_bit;		/* alignment of current block */
-- 
1.8.3.1

Re: [PATCH v2 for-rc 1/3] RDMA/core: Fix umem iterator when PAGE_SIZE is greater then HCA pgsz

[Jason Gunthorpe]

On Wed, Nov 29, 2023 at 02:21:41PM -0600, Shiraz Saleem wrote:
> be arbitrary, but for 64k pages, the VA may be offset by some
> number of HCA 4k pages and followed by some number of HCA 4k
> pages.
> 
> The current iterator doesn't account for either the preceding
> 4k pages or the following 4k pages.
> 
> Fix the issue by extending the ib_block_iter to contain
> the number of DMA pages like comment [1] says and by using
> __sg_advance to start the iterator at the first live HCA page.
> 
> The changes are contained in a parallel set of iterator start
> and next functions that are umem aware and specfic to umem
> since there is one user of the rdma_for_each_block() without
> umem.
> 
> These two fixes prevents the extra pages before and after the
> user MR data.
> 
> Fix the preceding pages by using the __sq_advance field to start
> at the first 4k page containing MR data.
> 
> Fix the following pages by saving the number of pgsz blocks in
> the iterator state and downcounting on each next.
> 
> This fix allows for the elimination of the small page crutch noted
> in the Fixes.
> 
> Fixes: 10c75ccb54e4 ("RDMA/umem: Prevent small pages from being returned by ib_umem_find_best_pgsz()")
> Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/rdma/ib_umem.h#n91 [1]
> Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
> Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
> ---
>  drivers/infiniband/core/umem.c | 6 ------
>  include/rdma/ib_umem.h         | 9 ++++++++-
>  include/rdma/ib_verbs.h        | 1 +
>  3 files changed, 9 insertions(+), 7 deletions(-)

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

I admit this fix looks so easy and simple I wonder if there was some
tricky reason why I didn't do it back then.. Oh well, I've forgotten,
I guess we will find out!

Jason

[PATCH v2 for-rc 2/3] RDMA/irdma: Ensure iWarp QP queue memory is OS paged aligned

[Shiraz Saleem]

From: Mike Marciniszyn <mike.marciniszyn@intel.com>

The SQ is shared for between kernel and used by storing
the kernel page pointer and passing that to a kmap_atomic().

This then requires that the alignment is PAGE_SIZE aligned.

Fix by adding an iWarp specific alignment check.

Fixes: e965ef0e7b2c ("RDMA/irdma: Split QP handler into irdma_reg_user_mr_type_qp")
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
---
 drivers/infiniband/hw/irdma/verbs.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/infiniband/hw/irdma/verbs.c b/drivers/infiniband/hw/irdma/verbs.c
index 5fa88e6cca4e..fb9088392b19 100644
--- a/drivers/infiniband/hw/irdma/verbs.c
+++ b/drivers/infiniband/hw/irdma/verbs.c
@@ -2946,6 +2946,11 @@ static int irdma_reg_user_mr_type_qp(struct irdma_mem_reg_req req,
 	int err;
 	u8 lvl;
 
+	/* iWarp: Catch page not starting on OS page boundary */
+	if (!rdma_protocol_roce(&iwdev->ibdev, 1) &&
+	    ib_umem_offset(iwmr->region))
+		return -EINVAL;
+
 	total = req.sq_pages + req.rq_pages + 1;
 	if (total > iwmr->page_cnt)
 		return -EINVAL;
-- 
1.8.3.1

[PATCH v2 for-rc 3/3] RDMA/irdma: Fix support for 64k pages

[Shiraz Saleem]

From: Mike Marciniszyn <mike.marciniszyn@intel.com>

Virtual QP and CQ require a 4K HW page size but the driver
passes PAGE_SIZE to ib_umem_find_best_pgsz() instead.

Fix this by using the appropriate 4k value in the bitmap passed to
ib_umem_find_best_pgsz().

Fixes: 693a5386eff0 ("RDMA/irdma: Split mr alloc and free into new functions")
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Shiraz Saleem <shiraz.saleem@intel.com>
---
 drivers/infiniband/hw/irdma/verbs.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/infiniband/hw/irdma/verbs.c b/drivers/infiniband/hw/irdma/verbs.c
index fb9088392b19..b5eb8d421988 100644
--- a/drivers/infiniband/hw/irdma/verbs.c
+++ b/drivers/infiniband/hw/irdma/verbs.c
@@ -2914,7 +2914,7 @@ static struct irdma_mr *irdma_alloc_iwmr(struct ib_umem *region,
 	iwmr->type = reg_type;
 
 	pgsz_bitmap = (reg_type == IRDMA_MEMREG_TYPE_MEM) ?
-		iwdev->rf->sc_dev.hw_attrs.page_size_cap : PAGE_SIZE;
+		iwdev->rf->sc_dev.hw_attrs.page_size_cap : SZ_4K;
 
 	iwmr->page_size = ib_umem_find_best_pgsz(region, pgsz_bitmap, virt);
 	if (unlikely(!iwmr->page_size)) {
-- 
1.8.3.1

Re: [PATCH v2 for-rc 0/3] Fixes for 64K page size support

[Jason Gunthorpe]

On Wed, Nov 29, 2023 at 02:21:40PM -0600, Shiraz Saleem wrote:
> This is a three patch series.
> 
> The first core hunk corrects the core iterator to use __sg_advance to skip
> preceding 4k HCA pages.
> 
> The second patch corrects an iWarp issue where the SQ must be PAGE_SIZE
> aligned.
> 
> The third patch corrects an issue with the RDMA driver use of
> ib_umem_find_best_pgsz(). QP and CQ allocations pass PAGE_SIZE as the
> only bitmap bit. This is incorrect and should use the precise 4k value.
> 
> v1->v2: Add a umem specific block iter next function
> 
> Mike Marciniszyn (3):
>   RDMA/core: Fix umem iterator when PAGE_SIZE is greater then HCA pgsz
>   RDMA/irdma: Ensure iWarp QP queue memory is OS paged aligned
>   RDMA/irdma: Fix support for 64k pages

Applied to for-rc, thanks

Jason