On Wed, May 22, 2019 at 01:48:52PM -0400, Jerome Glisse wrote:
> From 0b429b2ffbec348e283693cb97d7ffce760d89da Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?J=C3=A9r=C3=B4me=20Glisse?= <jgli...@redhat.com>
> Date: Sat, 8 Dec 2018 15:47:55 -0500
> Subject: [PATCH] RDMA/odp: convert to use HMM for ODP v5
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Convert ODP to use HMM so that we can build on common infrastructure
> for different class of devices that want to mirror a process address
> space into a device. There is no functional changes.
>
> Changes since v4:
> - Rebase on top of rdma-next
> Changes since v3:
> - Rebase on top of 5.2-rc1
> Changes since v2:
> - Update to match changes to HMM API
> Changes since v1:
> - improved comments
> - simplified page alignment computation
>
> Signed-off-by: Jérôme Glisse <jgli...@redhat.com>
> Cc: Jason Gunthorpe <j...@mellanox.com>
> Cc: Leon Romanovsky <leo...@mellanox.com>
> Cc: Doug Ledford <dledf...@redhat.com>
> Cc: Artemy Kovalyov <artem...@mellanox.com>
> Cc: Moni Shoua <mo...@mellanox.com>
> Cc: Mike Marciniszyn <mike.marcinis...@intel.com>
> Cc: Kaike Wan <kaike....@intel.com>
> Cc: Dennis Dalessandro <dennis.dalessan...@intel.com>
> drivers/infiniband/core/umem_odp.c | 506 +++++++++--------------------
> drivers/infiniband/hw/mlx5/mem.c | 20 +-
> drivers/infiniband/hw/mlx5/mr.c | 2 +-
> drivers/infiniband/hw/mlx5/odp.c | 104 +++---
> include/rdma/ib_umem_odp.h | 47 +--
> 5 files changed, 233 insertions(+), 446 deletions(-)
The kconfig stuff is missing, and it doesn't compile in various cases
I tried.
The kconfig stuff for hmm is also really obnoxious, you can't just
enabe HMM_MIRROR, you have to track down all the little required
elements to get it to turn on..
Once I did get it to compile, I also get warnings:
mm/hmm.c: In function ‘hmm_vma_walk_pud’:
mm/hmm.c:782:28: warning: unused variable ‘pfn’ [-Wunused-variable]
unsigned long i, npages, pfn;
^~~
mm/hmm.c: In function ‘hmm_range_snapshot’:
mm/hmm.c:1027:19: warning: unused variable ‘h’ [-Wunused-variable]
struct hstate *h = hstate_vma(vma);
Because this kernel doesn't have CONFIG_HUGETLB_PAGE
Please fold this into your patch if it has to be resent.. I think it
fixes the compilation problems.
diff --git a/drivers/infiniband/Kconfig b/drivers/infiniband/Kconfig
index cbfbea49f126cd..e3eefd0917985a 100644
--- a/drivers/infiniband/Kconfig
+++ b/drivers/infiniband/Kconfig
@@ -63,7 +63,7 @@ config INFINIBAND_USER_MEM
config INFINIBAND_ON_DEMAND_PAGING
bool "InfiniBand on-demand paging support"
depends on INFINIBAND_USER_MEM
- select MMU_NOTIFIER
+ depends on HMM_MIRROR
default y
---help---
On demand paging support for the InfiniBand subsystem.
diff --git a/include/rdma/ib_umem_odp.h b/include/rdma/ib_umem_odp.h
index e1476e9ebb7906..f760103c07349a 100644
--- a/include/rdma/ib_umem_odp.h
+++ b/include/rdma/ib_umem_odp.h
@@ -115,6 +115,16 @@ static inline size_t ib_umem_odp_num_pages(struct
ib_umem_odp *umem_odp)
#define ODP_DMA_ADDR_MASK (~(ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT))
+#define ODP_READ_BIT (1<<0ULL)
+#define ODP_WRITE_BIT (1<<1ULL)
+/*
+ * The device bit is not use by ODP but is there to full-fill HMM API which
+ * also support device with device memory (like GPU). So from ODP/RDMA POV
+ * this can be ignored.
+ */
+#define ODP_DEVICE_BIT (1<<2ULL)
+#define ODP_FLAGS_BITS 3
+
#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
struct ib_ucontext_per_mm {
@@ -138,16 +148,6 @@ struct ib_umem_odp *ib_alloc_odp_umem(struct ib_umem_odp
*root_umem,
unsigned long addr, size_t size);
void ib_umem_odp_release(struct ib_umem_odp *umem_odp);
-#define ODP_READ_BIT (1<<0ULL)
-#define ODP_WRITE_BIT (1<<1ULL)
-/*
- * The device bit is not use by ODP but is there to full-fill HMM API which
- * also support device with device memory (like GPU). So from ODP/RDMA POV
- * this can be ignored.
- */
-#define ODP_DEVICE_BIT (1<<2ULL)
-#define ODP_FLAGS_BITS 3
-
long ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp,
struct hmm_range *range);
