RDMA/irdma: Add support for dmabuf pin memory regions
Subsystems (2)
PATCH v3 1/1 RDMA/irdma: Add support for dmabuf pin memory regions
From: Zhu Yanjun <[email protected]>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does not support on-demand-paging(ODP). So it uses habanalabs as the dmabuf exporter, and irdma as the importer to allow for peer2peer access through libibverbs.
In this commit, the function ib_umem_dmabuf_get_pinned() is used. This function is introduced in EFA dmabuf[1] which allows the driver to get a dmabuf umem which is pinned and does not require move_notify callback implementation. The returned umem is pinned and DMA mapped like standard cpu umems, and is released through ib_umem_release().
[1]https://lore.kernel.org/lkml/[email protected]/t/
---
V2->V3: Remove unnecessary variable initialization;
Use error handler;
V1->V2: Thanks Shiraz Saleem, he gave me a lot of good suggestions.
This commit is based on the shared functions from refactored
irdma_reg_user_mr.
---
drivers/infiniband/hw/irdma/verbs.c | 45 +++++++++++++++++++++++++++++
1 file changed, 45 insertions(+)drivers/infiniband/hw/irdma/verbs.c +45 -0
@@ -2977,6 +2977,50 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,return ERR_PTR (err); } +static struct ib_mr *irdma_reg_user_mr_dmabuf (struct ib_pd *pd,u64 start, +u64 len,u64 virt, +int fd,int access, +struct ib_udata *udata) +{ +struct irdma_device *iwdev =to_iwdev (pd->device); +struct ib_umem_dmabuf *umem_dmabuf; +struct irdma_mr *iwmr; +int err; + +if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size) +return ERR_PTR (-EINVAL ); + +if (udata->inlen <IRDMA_MEM_REG_MIN_REQ_LEN ) +return ERR_PTR (-EINVAL ); + + umem_dmabuf =ib_umem_dmabuf_get_pinned (pd->device, start, len, fd, access); +if (IS_ERR (umem_dmabuf)) { + err =PTR_ERR (umem_dmabuf); +ibdev_dbg (&iwdev->ibdev,"Failed to get dmabuf umem[%d]\n" , err); +return ERR_PTR (err); + } + + iwmr =irdma_alloc_iwmr (&umem_dmabuf->umem, pd, virt,IRDMA_MEMREG_TYPE_MEM ); +if (IS_ERR (iwmr)) { + err =PTR_ERR (iwmr); + goto err_release; + } + + err =irdma_reg_user_mr_type_mem (iwmr, access); +if (err) + goto err_iwmr; + +return &iwmr->ibmr; + +err_iwmr: +irdma_free_iwmr (iwmr); + +err_release: +ib_umem_release (&umem_dmabuf->umem); + +return ERR_PTR (err); +} +/** * irdma_reg_phys_mr - register kernel physical memory * @pd: ibpd pointer
@@ -4483,6 +4527,7 @@ static const struct ib_device_ops irdma_dev_ops = { .query_port = irdma_query_port, .query_qp = irdma_query_qp, .reg_user_mr = irdma_reg_user_mr, + .reg_user_mr_dmabuf = irdma_reg_user_mr_dmabuf, .req_notify_cq = irdma_req_notify_cq, .resize_cq = irdma_resize_cq, INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
Signature · 1 line
2.27.0
Quoted 15 lines · Zhu Yanjun
From: Zhu Yanjun <[email protected]>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does not support on-demand-paging(ODP). So it uses habanalabs as the dmabuf exporter, and irdma as the importer to allow for peer2peer access through libibverbs.
In this commit, the function
ib_umem_dmabuf_get_pinned()is used. This function is introduced in EFA dmabuf[1] which allows the driver to get a dmabuf umem which is pinned and does not requiremove_notifycallback implementation. The returned umem is pinned and DMA mapped like standard cpu umems, and is released throughib_umem_release().
Quoted 7 lines
Signed-off-by Zhu Yanjun <[email protected]>---
V2->V3: Remove unnecessary variable initialization; Use error handler;
V1->V2: Thanks Shiraz Saleem, he gave me a lot of good suggestions.
This commit is based on the shared functions from refactoredirdma_reg_user_mr.
Shiraz Saleem is on vacation and returning to office on Feb.13, 2023. We can wait for him.
Zhu Yanjun
Quoted 67 lines
--- drivers/infiniband/hw/irdma/verbs.c | 45 +++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+)drivers/infiniband/hw/irdma/verbs.c +45
-0@@ -2977,6 +2977,50 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,return ERR_PTR (err); } +static struct ib_mr *irdma_reg_user_mr_dmabuf (struct ib_pd *pd,u64 start, +u64 len,u64 virt, +int fd,int access, +struct ib_udata *udata) +{ +struct irdma_device *iwdev =to_iwdev (pd->device); +struct ib_umem_dmabuf *umem_dmabuf; +struct irdma_mr *iwmr; +int err; + +if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size) +return ERR_PTR (-EINVAL ); + +if (udata->inlen <IRDMA_MEM_REG_MIN_REQ_LEN ) +return ERR_PTR (-EINVAL ); + + umem_dmabuf =ib_umem_dmabuf_get_pinned (pd->device, start, len, fd, access); +if (IS_ERR (umem_dmabuf)) { + err =PTR_ERR (umem_dmabuf); +ibdev_dbg (&iwdev->ibdev,"Failed to get dmabuf umem[%d]\n" , err); +return ERR_PTR (err); + } + + iwmr =irdma_alloc_iwmr (&umem_dmabuf->umem, pd, virt,IRDMA_MEMREG_TYPE_MEM ); +if (IS_ERR (iwmr)) { + err =PTR_ERR (iwmr); + goto err_release; + } + + err =irdma_reg_user_mr_type_mem (iwmr, access); +if (err) + goto err_iwmr; + +return &iwmr->ibmr; + +err_iwmr: +irdma_free_iwmr (iwmr); + +err_release: +ib_umem_release (&umem_dmabuf->umem); + +return ERR_PTR (err); +} +/** * irdma_reg_phys_mr - register kernel physical memory * @pd: ibpd pointer @@ -4483,6 +4527,7 @@ static const struct ib_device_ops irdma_dev_ops = { .query_port = irdma_query_port, .query_qp = irdma_query_qp, .reg_user_mr = irdma_reg_user_mr, + .reg_user_mr_dmabuf = irdma_reg_user_mr_dmabuf, .req_notify_cq = irdma_req_notify_cq, .resize_cq = irdma_resize_cq, INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd),
Quoted 42 lines
Subject: [PATCHv3 for-next 1/1] RDMA/irdma: Add support for dmabuf pin memory regions
From: Zhu Yanjun <[email protected]>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does not support on-demand-paging(ODP). So it uses habanalabs as the dmabuf exporter, and irdma as the importer to allow for peer2peer access through libibverbs.
In this commit, the function
ib_umem_dmabuf_get_pinned()is used. This function is introduced in EFA dmabuf[1] which allows the driver to get a dmabuf umem which is pinned and does not requiremove_notifycallback implementation. The returned umem is pinned and DMA mapped like standard cpu umems, and is released throughib_umem_release().[1]https://lore.kernel.org/lkml/[email protected]/t/
Signed-off-by Zhu Yanjun <[email protected]>---
V2->V3: Remove unnecessary variable initialization; Use error handler; V1->V2: Thanks Shiraz Saleem, he gave me a lot of good suggestions. This commit is based on the shared functions from refactored irdma_reg_user_mr. --- drivers/infiniband/hw/irdma/verbs.c | 45 +++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+)drivers/infiniband/hw/irdma/verbs.c +5
-0@@ -2977,6 +2977,50 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len, return ERR_PTR(err); } +static struct ib_mr *irdma_reg_user_mr_dmabuf (struct ib_pd *pd,u64 start, +u64 len,u64 virt, +int fd,int access, +struct ib_udata *udata) +{
drivers/infiniband/hw/irdma/verbs.c static struct ib_mr *irdma_reg_user_mr(struct ib_pd · line 2995
Quoted 10 lines
drivers/infiniband/hw/irdma/verbs.c +10
-0+struct irdma_device *iwdev =to_iwdev (pd->device); +struct ib_umem_dmabuf *umem_dmabuf; +struct irdma_mr *iwmr; +int err; + +if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size) +return ERR_PTR (-EINVAL ); + +if (udata->inlen <IRDMA_MEM_REG_MIN_REQ_LEN ) +return ERR_PTR (-EINVAL );
Do we need this? we don't copy anything from udata. There is no info passed via ABI struct irdma_mem_reg_req.
Quoted 36 lines
drivers/infiniband/hw/irdma/verbs.c +9
-0+ + umem_dmabuf =ib_umem_dmabuf_get_pinned (pd->device, start, len, fd, access); +if (IS_ERR (umem_dmabuf)) { + err =PTR_ERR (umem_dmabuf); +ibdev_dbg (&iwdev->ibdev,"Failed to get dmabuf umem[%d]\n" , err); +return ERR_PTR (err); + } + + iwmr = irdma_alloc_iwmr(&umem_dmabuf->umem, pd, virt,
IRDMA_MEMREG_TYPE_MEM);drivers/infiniband/hw/irdma/verbs.c +20
-0+if (IS_ERR (iwmr)) { + err =PTR_ERR (iwmr); + goto err_release; + } + + err =irdma_reg_user_mr_type_mem (iwmr, access); +if (err) + goto err_iwmr; + +return &iwmr->ibmr; + +err_iwmr: +irdma_free_iwmr (iwmr); + +err_release: +ib_umem_release (&umem_dmabuf->umem); + +return ERR_PTR (err); +} +/** * irdma_reg_phys_mr - register kernel physical memory * @pd: ibpd pointer
drivers/infiniband/hw/irdma/verbs.c static const struct ib_device_ops irdma_dev_ops = {
Quoted 8 lines
drivers/infiniband/hw/irdma/verbs.c +1
-1@@ -4483,6 +4527,7 @@ static const struct ib_device_ops irdma_dev_ops = { .query_port = irdma_query_port, .query_qp = irdma_query_qp, .reg_user_mr = irdma_reg_user_mr, + .reg_user_mr_dmabuf = irdma_reg_user_mr_dmabuf, .req_notify_cq = irdma_req_notify_cq, .resize_cq = irdma_resize_cq, INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd), --
Quoted 37 lines · Shiraz
Quoted 37 lines
Subject: [PATCHv3 for-next 1/1] RDMA/irdma: Add support for dmabuf pin memory regions
From: Zhu Yanjun <[email protected]>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does not support on-demand-paging(ODP). So it uses habanalabs as the dmabuf exporter, and irdma as the importer to allow for peer2peer access through libibverbs.
In this commit, the function
ib_umem_dmabuf_get_pinned()is used. This function is introduced in EFA dmabuf[1] which allows the driver to get a dmabuf umem which is pinned and does not requiremove_notifycallback implementation. The returned umem is pinned and DMA mapped like standard cpu umems, and is released throughib_umem_release().[1]https://lore.kernel.org/lkml/[email protected]/t/
Signed-off-by Zhu Yanjun <[email protected]>---
V2->V3: Remove unnecessary variable initialization; Use error handler; V1->V2: Thanks Shiraz Saleem, he gave me a lot of good suggestions. This commit is based on the shared functions from refactored irdma_reg_user_mr. --- drivers/infiniband/hw/irdma/verbs.c | 45 +++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+)drivers/infiniband/hw/irdma/verbs.c +0
-0@@ -2977,6 +2977,50 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len, return ERR_PTR(err); }
Quoted 16 lines
Quoted 5 lines
drivers/infiniband/hw/irdma/verbs.c +5
-0+static struct ib_mr *irdma_reg_user_mr_dmabuf (struct ib_pd *pd,u64 start, +u64 len,u64 virt, +int fd,int access, +struct ib_udata *udata) +{drivers/infiniband/hw/irdma/verbs.c static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len, · line 2994
Quoted 10 lines
drivers/infiniband/hw/irdma/verbs.c +10
-0+struct irdma_device *iwdev =to_iwdev (pd->device); +struct ib_umem_dmabuf *umem_dmabuf; +struct irdma_mr *iwmr; +int err; + +if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size) +return ERR_PTR (-EINVAL ); + +if (udata->inlen <IRDMA_MEM_REG_MIN_REQ_LEN ) +return ERR_PTR (-EINVAL );Do we need this? we don't copy anything from udata. There is no info passed via ABI struct
irdma_mem_reg_req.
Got it. I will remove the inlen test and send a new commit out.
Quoted 38 lines
Quoted 38 lines
drivers/infiniband/hw/irdma/verbs.c +9
-0+ + umem_dmabuf =ib_umem_dmabuf_get_pinned (pd->device, start, len, fd, access); +if (IS_ERR (umem_dmabuf)) { + err =PTR_ERR (umem_dmabuf); +ibdev_dbg (&iwdev->ibdev,"Failed to get dmabuf umem[%d]\n" , err); +return ERR_PTR (err); + } + + iwmr = irdma_alloc_iwmr(&umem_dmabuf->umem, pd, virt,
IRDMA_MEMREG_TYPE_MEM);drivers/infiniband/hw/irdma/verbs.c +20
-0+if (IS_ERR (iwmr)) { + err =PTR_ERR (iwmr); + goto err_release; + } + + err =irdma_reg_user_mr_type_mem (iwmr, access); +if (err) + goto err_iwmr; + +return &iwmr->ibmr; + +err_iwmr: +irdma_free_iwmr (iwmr); + +err_release: +ib_umem_release (&umem_dmabuf->umem); + +return ERR_PTR (err); +} +/** * irdma_reg_phys_mr - register kernel physical memory * @pd: ibpd pointer @@ -4483,6 +4527,7 @@ static const struct ib_device_ops irdma_dev_ops = { .query_port = irdma_query_port,
Quoted 8 lines
drivers/infiniband/hw/irdma/verbs.c static const struct ib_device_ops irdma_dev_ops = {
Quoted 7 lines
drivers/infiniband/hw/irdma/verbs.c +1
-1.query_qp = irdma_query_qp, .reg_user_mr = irdma_reg_user_mr, + .reg_user_mr_dmabuf = irdma_reg_user_mr_dmabuf, .req_notify_cq = irdma_req_notify_cq, .resize_cq = irdma_resize_cq, INIT_RDMA_OBJ_SIZE(ib_pd, irdma_pd, ibpd), --Reviewed-by Shiraz Saleem <[email protected]>
Thanks.
Zhu Yanjun
Quoted 1 line
Jason Gunthorpe
maintainerQuoted 40 lines · Zhu Yanjun
From: Zhu Yanjun <[email protected]>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does not support on-demand-paging(ODP). So it uses habanalabs as the dmabuf exporter, and irdma as the importer to allow for peer2peer access through libibverbs.
In this commit, the function
ib_umem_dmabuf_get_pinned()is used. This function is introduced in EFA dmabuf[1] which allows the driver to get a dmabuf umem which is pinned and does not requiremove_notifycallback implementation. The returned umem is pinned and DMA mapped like standard cpu umems, and is released throughib_umem_release().[1]https://lore.kernel.org/lkml/[email protected]/t/
Signed-off-by Zhu Yanjun <[email protected]>Reviewed-by Shiraz Saleem <[email protected]>---
V2->V3: Remove unnecessary variable initialization; Use error handler; V1->V2: Thanks Shiraz Saleem, he gave me a lot of good suggestions. This commit is based on the shared functions from refactored irdma_reg_user_mr. --- drivers/infiniband/hw/irdma/verbs.c | 45 +++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+)drivers/infiniband/hw/irdma/verbs.c +5
-0@@ -2977,6 +2977,50 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,return ERR_PTR (err); } +static struct ib_mr *irdma_reg_user_mr_dmabuf (struct ib_pd *pd,u64 start, +u64 len,u64 virt, +int fd,int access, +struct ib_udata *udata) +{
drivers/infiniband/hw/irdma/verbs.c static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len, · line 2994
Quoted 10 lines
drivers/infiniband/hw/irdma/verbs.c +10
-0+struct irdma_device *iwdev =to_iwdev (pd->device); +struct ib_umem_dmabuf *umem_dmabuf; +struct irdma_mr *iwmr; +int err; + +if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size) +return ERR_PTR (-EINVAL ); + +if (udata->inlen <IRDMA_MEM_REG_MIN_REQ_LEN ) +return ERR_PTR (-EINVAL );
Shiraz is correct, I'm wondering how this even works. This is a new style uAPI without UVERBS_ATTR_UHW so inlen should always be 0.
How did you manage to test this??
Jason
Quoted 45 lines · Jason Gunthorpe
Quoted 44 lines · Zhu Yanjun
From: Zhu Yanjun <[email protected]>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does not support on-demand-paging(ODP). So it uses habanalabs as the dmabuf exporter, and irdma as the importer to allow for peer2peer access through libibverbs.
In this commit, the function
ib_umem_dmabuf_get_pinned()is used. This function is introduced in EFA dmabuf[1] which allows the driver to get a dmabuf umem which is pinned and does not requiremove_notifycallback implementation. The returned umem is pinned and DMA mapped like standard cpu umems, and is released throughib_umem_release().[1]https://lore.kernel.org/lkml/[email protected]/t/
Signed-off-by Zhu Yanjun <[email protected]>Reviewed-by Shiraz Saleem <[email protected]>---
V2->V3: Remove unnecessary variable initialization; Use error handler; V1->V2: Thanks Shiraz Saleem, he gave me a lot of good suggestions. This commit is based on the shared functions from refactored irdma_reg_user_mr. --- drivers/infiniband/hw/irdma/verbs.c | 45 +++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+)drivers/infiniband/hw/irdma/verbs.c +9
-0@@ -2977,6 +2977,50 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,return ERR_PTR (err); } +static struct ib_mr *irdma_reg_user_mr_dmabuf (struct ib_pd *pd,u64 start, +u64 len,u64 virt, +int fd,int access, +struct ib_udata *udata) +{ +struct irdma_device *iwdev =to_iwdev (pd->device); +struct ib_umem_dmabuf *umem_dmabuf; +struct irdma_mr *iwmr; +int err;
Quoted 8 lines
drivers/infiniband/hw/irdma/verbs.c static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len, · line 2994
Quoted 6 lines
drivers/infiniband/hw/irdma/verbs.c +6
-0+ +if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size) +return ERR_PTR (-EINVAL ); + +if (udata->inlen <IRDMA_MEM_REG_MIN_REQ_LEN ) +return ERR_PTR (-EINVAL );Shiraz is correct, I'm wondering how this even works. This is a new style uAPI without
UVERBS_ATTR_UHWso inlen should always be 0.
Got it. Thanks Shiraz and Jason.
I will remove the test of inlen in the latest commit.
Best Regards,
Zhu Yanjun
Quoted 4 lines
How did you manage to test this??
Jason
Jason Gunthorpe
maintainerQuoted 55 lines · Zhu Yanjun
Quoted 44 lines · Jason Gunthorpe
Quoted 43 lines · Zhu Yanjun
From: Zhu Yanjun <[email protected]>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does not support on-demand-paging(ODP). So it uses habanalabs as the dmabuf exporter, and irdma as the importer to allow for peer2peer access through libibverbs.
In this commit, the function
ib_umem_dmabuf_get_pinned()is used. This function is introduced in EFA dmabuf[1] which allows the driver to get a dmabuf umem which is pinned and does not requiremove_notifycallback implementation. The returned umem is pinned and DMA mapped like standard cpu umems, and is released throughib_umem_release().[1]https://lore.kernel.org/lkml/[email protected]/t/
Signed-off-by Zhu Yanjun <[email protected]>Reviewed-by Shiraz Saleem <[email protected]>---
V2->V3: Remove unnecessary variable initialization; Use error handler; V1->V2: Thanks Shiraz Saleem, he gave me a lot of good suggestions. This commit is based on the shared functions from refactored irdma_reg_user_mr. --- drivers/infiniband/hw/irdma/verbs.c | 45 +++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+)drivers/infiniband/hw/irdma/verbs.c +9
-0@@ -2977,6 +2977,50 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,return ERR_PTR (err); } +static struct ib_mr *irdma_reg_user_mr_dmabuf (struct ib_pd *pd,u64 start, +u64 len,u64 virt, +int fd,int access, +struct ib_udata *udata) +{ +struct irdma_device *iwdev =to_iwdev (pd->device); +struct ib_umem_dmabuf *umem_dmabuf; +struct irdma_mr *iwmr; +int err;Quoted 8 lines
drivers/infiniband/hw/irdma/verbs.c static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len, · line 2994
Quoted 6 lines
drivers/infiniband/hw/irdma/verbs.c +6
-0+ +if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size) +return ERR_PTR (-EINVAL ); + +if (udata->inlen <IRDMA_MEM_REG_MIN_REQ_LEN ) +return ERR_PTR (-EINVAL );Shiraz is correct, I'm wondering how this even works. This is a new style uAPI without
UVERBS_ATTR_UHWso inlen should always be 0.
Quoted 3 lines
Got it. Thanks Shiraz and Jason.
I will remove the test of inlen in the latest commit.
Please answer how did you manage to test this?
Jason