RDMA/rxe: Prefetching pages with explicit ODP
Subsystems (2)
v2 cover RDMA/rxe: Prefetching pages with explicit ODP
There is ibv_advise_mr(3) that can be used by applications to optimize memory access. This series enables the feature on rxe driver, which has already been available in mlx5.
There is a tiny change on the rdma-core util. cf. https://github.com/linux-rdma/rdma-core/pull/1605
Daisuke Matsuda (2):
RDMA/rxe: Implement synchronous prefetch for ODP MRs RDMA/rxe: Enable asynchronous prefetch for ODP MRs
drivers/infiniband/sw/rxe/rxe.c | 7 ++ drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++ drivers/infiniband/sw/rxe/rxe_odp.c | 165 ++++++++++++++++++++++++++++ 3 files changed, 182 insertions(+)
Signature · 1 line
2.43.0
PATCH v2 1/2 RDMA/rxe: Implement synchronous prefetch for ODP MRs
Minimal implementation of ibv_advise_mr(3) requires synchronous calls being successful with the IBV_ADVISE_MR_FLAG_FLUSH flag. Asynchronous requests, which are best-effort, will be added subsequently.
--- drivers/infiniband/sw/rxe/rxe.c | 7 +++ drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++++ drivers/infiniband/sw/rxe/rxe_odp.c | 86 +++++++++++++++++++++++++++++ 3 files changed, 103 insertions(+)
drivers/infiniband/sw/rxe/rxe.c +7 -0
@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev)mutex_destroy (&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; +/* initialize rxe device parameters */ static void rxe_init_device_param (struct rxe_dev *rxe,struct net_device *ndev) {
@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_SRQ_RECV ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_FLUSH ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_ATOMIC_WRITE ; + +/* set handler for ODP prefetching API - ibv_advise_mr(3) */ +ib_set_device_ops (&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }
drivers/infiniband/sw/rxe/rxe_loc.h +10 -0
@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode,int rxe_odp_flush_pmem_iova (struct rxe_mr *mr,u64 iova,unsigned int length);enum resp_states rxe_odp_do_atomic_write (struct rxe_mr *mr,u64 iova,u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs);#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int rxe_odp_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, {return RESPST_ERR_UNSUPPORTED_OPCODE ; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +return -EOPNOTSUPP ; +} +#endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */
drivers/infiniband/sw/rxe/rxe_odp.c +86 -0
89 lines not on this page · raw message
Signature · 1 line
2.43.0
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 73 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>--- drivers/infiniband/sw/rxe/rxe.c | 7 +++ drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++++ drivers/infiniband/sw/rxe/rxe_odp.c | 86 +++++++++++++++++++++++++++++ 3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev)mutex_destroy (&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; +/* initialize rxe device parameters */ static void rxe_init_device_param (struct rxe_dev *rxe,struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_SRQ_RECV ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_FLUSH ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_ATOMIC_WRITE ; + +/* set handler for ODP prefetching API - ibv_advise_mr(3) */ +ib_set_device_ops (&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +10
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode,int rxe_odp_flush_pmem_iova (struct rxe_mr *mr,u64 iova,unsigned int length);enum resp_states rxe_odp_do_atomic_write (struct rxe_mr *mr,u64 iova,u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs);#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int rxe_odp_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, {return RESPST_ERR_UNSUPPORTED_OPCODE ; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +return -EOPNOTSUPP ; +} +#endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */ drivers/infiniband/sw/rxe/rxe_odp.c +1
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)return RESPST_NONE ; } +
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 437
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 pf_flags,struct ib_sge *sg_list, +u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); +unsigned int i; +int ret =0 ; + +for (i =0 ; i < num_sge; ++i) {
i is unsigned int, num_sge is u32. Perhaps they all use u32 type? It is a minor problem.
Other than that, I am fine with this commit.
I have made tests with rdma-core. Both the synchronous and asynchrounos modes can work well.
Zhu Yanjun
Quoted 75 lines
drivers/infiniband/sw/rxe/rxe_odp.c +75
-0+struct rxe_mr *mr; +struct ib_umem_odp *umem_odp; + + mr =lookup_mr (pd,IB_ACCESS_LOCAL_WRITE , + sg_list[i].lkey,RXE_LOOKUP_LOCAL ); + +if (IS_ERR (mr)) { +rxe_dbg_pd (pd,"mr with lkey %x not found\n" , sg_list[i].lkey); +return PTR_ERR (mr); + } + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { +rxe_dbg_mr (mr,"missing write permission\n" ); +rxe_put (mr); +return -EPERM ; + } + + ret =rxe_odp_do_pagefault_and_lock (mr, sg_list[i].addr, + sg_list[i].length, pf_flags); +if (ret <0 ) { +if (sg_list[i].length ==0 ) +continue ; + +rxe_dbg_mr (mr,"failed to prefetch the mr\n" ); +rxe_put (mr); +return ret; + } + + umem_odp =to_ib_umem_odp (mr->umem); +mutex_unlock (&umem_odp->umem_mutex); + +rxe_put (mr); + } + +return 0 ; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge) +{ +u32 pf_flags =RXE_PAGEFAULT_DEFAULT ; + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH ) + pf_flags |=RXE_PAGEFAULT_RDONLY ; + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) + pf_flags |=RXE_PAGEFAULT_SNAPSHOT ; + +/* Synchronous call */ +if (flags &IB_UVERBS_ADVISE_MR_FLAG_FLUSH ) +return rxe_ib_prefetch_sg_list (ibpd, advice, pf_flags, sg_list, + num_sge); + +/* Asynchronous call is "best-effort" */ + +return 0 ; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags, +struct ib_sge *sg_list, +u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +if (advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) +return -EOPNOTSUPP ; + +return rxe_ib_advise_mr_prefetch (ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 75 lines · Zhu Yanjun
Quoted 74 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>---
drivers/infiniband/sw/rxe/rxe.c | 7 +++drivers/infiniband/sw/rxe/rxe_loc.h| 10 ++++drivers/infiniband/sw/rxe/rxe_odp.c| 86 +++++++++++++++++++++++++++++3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev) mutex_destroy(&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; + /* initialize rxe device parameters */ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE; + + /* set handler for ODP prefetching API - ibv_advise_mr(3) */ + ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +3
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode, int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length); enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */static inline int
rxe_odp_mr_init_user(structrxe_dev*rxe, u64 start, u64 length, u64 iova,drivers/infiniband/sw/rxe/rxe_loc.h +7
-0@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, { return RESPST_ERR_UNSUPPORTED_OPCODE; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + return -EOPNOTSUPP; +} + #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */drivers/infiniband/sw/rxe/rxe_odp.c +6
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct ib_sge *sg_list, + u32 num_sge) +{
Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 437
Quoted 5 lines
drivers/infiniband/sw/rxe/rxe_odp.c +5
-0+ struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + unsigned int i; + int ret = 0; + + for (i = 0; i < num_sge; ++i) {i is unsigned int,
num_sgeis u32. Perhaps they all use u32 type? It is a minor problem.
Since we have dropped support for 32-bit archtectures, these types should practically be identical.
Quoted 3 lines
Other than that, I am fine with this commit.
I have made tests with rdma-core. Both the synchronous and asynchrounos modes can work well.
Thank you for the review and testing.
Daisuke
Quoted 81 lines
Reviewed-by Zhu Yanjun <[email protected]>Zhu Yanjun
Quoted 75 lines
drivers/infiniband/sw/rxe/rxe_odp.c +75
-0+ struct rxe_mr *mr; + struct ib_umem_odp *umem_odp; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + + if (IS_ERR(mr)) { + rxe_dbg_pd(pd, "mr with lkey %x not found\n", sg_list[i].lkey); + return PTR_ERR(mr); + } + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { + rxe_dbg_mr(mr, "missing write permission\n"); + rxe_put(mr); + return -EPERM; + } + + ret = rxe_odp_do_pagefault_and_lock(mr, sg_list[i].addr, + sg_list[i].length, pf_flags); + if (ret < 0) { + if (sg_list[i].length == 0) + continue; + + rxe_dbg_mr(mr, "failed to prefetch the mr\n"); + rxe_put(mr); + return ret; + } + + umem_odp = to_ib_umem_odp(mr->umem); + mutex_unlock(&umem_odp->umem_mutex); + + rxe_put(mr); + } + + return 0; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge) +{ + u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) + pf_flags |= RXE_PAGEFAULT_RDONLY; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + pf_flags |= RXE_PAGEFAULT_SNAPSHOT; + + /* Synchronous call */ + if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH) + return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list, + num_sge); + + /* Asynchronous call is "best-effort" */ + + return 0; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, + struct ib_sge *sg_list, + u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + return -EOPNOTSUPP; + + return rxe_ib_advise_mr_prefetch(ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
In reply to Daisuke MatsudaQuoted 129 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>--- drivers/infiniband/sw/rxe/rxe.c | 7 +++ drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++++ drivers/infiniband/sw/rxe/rxe_odp.c | 86 +++++++++++++++++++++++++++++ 3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev)mutex_destroy (&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; +/* initialize rxe device parameters */ static void rxe_init_device_param (struct rxe_dev *rxe,struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_SRQ_RECV ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_FLUSH ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_ATOMIC_WRITE ; + +/* set handler for ODP prefetching API - ibv_advise_mr(3) */ +ib_set_device_ops (&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +10
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode,int rxe_odp_flush_pmem_iova (struct rxe_mr *mr,u64 iova,unsigned int length);enum resp_states rxe_odp_do_atomic_write (struct rxe_mr *mr,u64 iova,u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs);#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int rxe_odp_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, {return RESPST_ERR_UNSUPPORTED_OPCODE ; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +return -EOPNOTSUPP ; +} +#endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */ drivers/infiniband/sw/rxe/rxe_odp.c +57
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)return RESPST_NONE ; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 pf_flags,struct ib_sge *sg_list, +u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); +unsigned int i; +int ret =0 ; + +for (i =0 ; i < num_sge; ++i) { +struct rxe_mr *mr; +struct ib_umem_odp *umem_odp; + + mr =lookup_mr (pd,IB_ACCESS_LOCAL_WRITE , + sg_list[i].lkey,RXE_LOOKUP_LOCAL ); + +if (IS_ERR (mr)) { +rxe_dbg_pd (pd,"mr with lkey %x not found\n" , sg_list[i].lkey); +return PTR_ERR (mr); + } + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { +rxe_dbg_mr (mr,"missing write permission\n" ); +rxe_put (mr); +return -EPERM ; + } + + ret =rxe_odp_do_pagefault_and_lock (mr, sg_list[i].addr, + sg_list[i].length, pf_flags); +if (ret <0 ) { +if (sg_list[i].length ==0 ) +continue ; + +rxe_dbg_mr (mr,"failed to prefetch the mr\n" ); +rxe_put (mr); +return ret; + } + + umem_odp =to_ib_umem_odp (mr->umem); +mutex_unlock (&umem_odp->umem_mutex); + +rxe_put (mr); + } + +return 0 ; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge) +{ +u32 pf_flags =RXE_PAGEFAULT_DEFAULT ; + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH ) + pf_flags |=RXE_PAGEFAULT_RDONLY ;
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 493
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+ +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) + pf_flags |=RXE_PAGEFAULT_SNAPSHOT ; + +/* Synchronous call */ +if (flags &IB_UVERBS_ADVISE_MR_FLAG_FLUSH ) +return rxe_ib_prefetch_sg_list (ibpd, advice, pf_flags, sg_list, + num_sge); + +/* Asynchronous call is "best-effort" */
Asynchronous call is not implemented now, why does this comment appear?
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ +return 0 ; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags, +struct ib_sge *sg_list, +u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +if (advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) +return -EOPNOTSUPP ; + +return rxe_ib_advise_mr_prefetch (ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 130 lines · Zhu Yanjun
Quoted 129 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>---
drivers/infiniband/sw/rxe/rxe.c | 7 +++drivers/infiniband/sw/rxe/rxe_loc.h| 10 ++++drivers/infiniband/sw/rxe/rxe_odp.c| 86 +++++++++++++++++++++++++++++3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev) mutex_destroy(&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; + /* initialize rxe device parameters */ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE; + + /* set handler for ODP prefetching API - ibv_advise_mr(3) */ + ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +3
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode, int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length); enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */static inline int
rxe_odp_mr_init_user(structrxe_dev*rxe, u64 start, u64 length, u64 iova,drivers/infiniband/sw/rxe/rxe_loc.h +7
-0@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, { return RESPST_ERR_UNSUPPORTED_OPCODE; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + return -EOPNOTSUPP; +} + #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */drivers/infiniband/sw/rxe/rxe_odp.c +61
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct ib_sge *sg_list, + u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + unsigned int i; + int ret = 0; + + for (i = 0; i < num_sge; ++i) { + struct rxe_mr *mr; + struct ib_umem_odp *umem_odp; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + + if (IS_ERR(mr)) { + rxe_dbg_pd(pd, "mr with lkey %x not found\n", sg_list[i].lkey); + return PTR_ERR(mr); + } + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { + rxe_dbg_mr(mr, "missing write permission\n"); + rxe_put(mr); + return -EPERM; + } + + ret = rxe_odp_do_pagefault_and_lock(mr, sg_list[i].addr, + sg_list[i].length, pf_flags); + if (ret < 0) { + if (sg_list[i].length == 0) + continue; + + rxe_dbg_mr(mr, "failed to prefetch the mr\n"); + rxe_put(mr); + return ret; + } + + umem_odp = to_ib_umem_odp(mr->umem); + mutex_unlock(&umem_odp->umem_mutex); + + rxe_put(mr); + } + + return 0; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge) +{ + u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) + pf_flags |= RXE_PAGEFAULT_RDONLY; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + pf_flags |= RXE_PAGEFAULT_SNAPSHOT; +
Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 493
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_odp.c +6
-0+ /* Synchronous call */ + if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH) + return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list, + num_sge); + + /* Asynchronous call is "best-effort" */Asynchronous call is not implemented now, why does this comment appear?
Even without the 2nd patch, async calls are reported as successful.
The comment is inserted to show the reason, which is based on the
description from 'man 3 ibv_advise_mr' as follows:
===
An application may pre-fetch any address range within an ODP MR when using the IBV_ADVISE_MR_ADVICE_PREFETCH or IBV_ADVISE_MR_ADVICE_PREFETCH_WRITE advice. Semantically, this operation is best-effort. That means the kernel does not guarantee that underlying pages are updated in the HCA or the pre-fetched pages would remain resident.
===
Thanks,
Daisuke
Quoted 23 lines
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ + return 0; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, + struct ib_sge *sg_list, + u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + return -EOPNOTSUPP; + + return rxe_ib_advise_mr_prefetch(ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 149 lines · Daisuke Matsuda
Quoted 139 lines · Zhu Yanjun
Quoted 138 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>---
drivers/infiniband/sw/rxe/rxe.c | 7 +++
drivers/infiniband/sw/rxe/rxe_loc.h| 10 ++++
drivers/infiniband/sw/rxe/rxe_odp.c| 86
+++++++++++++++++++++++++++++3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev) mutex_destroy(&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; + /* initialize rxe device parameters */ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE; + + /* set handler for ODP prefetching API - ibv_advise_mr(3) */ + ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +3
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode, int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length); enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */static inline int
rxe_odp_mr_init_user(structrxe_dev*rxe, u64 start, u64 length, u64 iova,drivers/infiniband/sw/rxe/rxe_loc.h +7
-0@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, { return RESPST_ERR_UNSUPPORTED_OPCODE; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + return -EOPNOTSUPP; +} + #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */#endif /*
RXE_LOC_H*/drivers/infiniband/sw/rxe/rxe_odp.c +19
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct ib_sge *sg_list, + u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + unsigned int i; + int ret = 0; + + for (i = 0; i < num_sge; ++i) { + struct rxe_mr *mr; + struct ib_umem_odp *umem_odp; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + + if (IS_ERR(mr)) { + rxe_dbg_pd(pd, "mr with lkey %x not found\n",sg_list[i].lkey);
+ returnPTR_ERR(mr);
+ }drivers/infiniband/sw/rxe/rxe_odp.c +41
-0+ + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { + rxe_dbg_mr(mr, "missing write permission\n"); + rxe_put(mr); + return -EPERM; + } + + ret = rxe_odp_do_pagefault_and_lock(mr, sg_list[i].addr, + sg_list[i].length, pf_flags); + if (ret < 0) { + if (sg_list[i].length == 0) + continue; + + rxe_dbg_mr(mr, "failed to prefetch the mr\n"); + rxe_put(mr); + return ret; + } + + umem_odp = to_ib_umem_odp(mr->umem); + mutex_unlock(&umem_odp->umem_mutex); + + rxe_put(mr); + } + + return 0; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge) +{ + u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) + pf_flags |= RXE_PAGEFAULT_RDONLY; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + pf_flags |= RXE_PAGEFAULT_SNAPSHOT; + + /* Synchronous call */Quoted 8 lines
Quoted 6 lines
+ if (flags &
IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
+ return rxe_ib_prefetch_sg_list(ibpd, advice,pf_flags,sg_list,
+num_sge);
+
+ /* Asynchronous call is "best-effort" */Asynchronous call is not implemented now, why does this comment appear?
Quoted 1 line
Even without the 2nd patch, async calls are reported as successful.
Async call is not implemented. How to call "async calls are reported as successful"?
Zhu Yanjun
Quoted 33 lines
The comment is inserted to show the reason, which is based on the
description from 'man 3ibv_advise_mr' as follows:
===
An application may pre-fetch any address range within an ODP MR when using theIBV_ADVISE_MR_ADVICE_PREFETCHorIBV_ADVISE_MR_ADVICE_PREFETCH_WRITEadvice. Semantically, this operation is best-effort. That means the kernel does not guarantee that underlying pages are updated in the HCA or the pre-fetched pages would remain resident.
===Thanks,
DaisukeQuoted 23 lines
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ + return 0; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, + struct ib_sge *sg_list, + u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + return -EOPNOTSUPP; + + return rxe_ib_advise_mr_prefetch(ibpd, advice, flags, + sg_list, num_sge); +}
Signature · 2 lines
Best Regards, Yanjun.Zhu
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 142 lines · Zhu Yanjun
Quoted 140 lines · Daisuke Matsuda
Quoted 138 lines · Zhu Yanjun
Quoted 125 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>---
drivers/infiniband/sw/rxe/rxe.c | 7 +++drivers/infiniband/sw/rxe/rxe_loc.h| 10 ++++drivers/infiniband/sw/rxe/rxe_odp.c| 86 +++++++++++++++++++++++++++++3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev) mutex_destroy(&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; + /* initialize rxe device parameters */ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE; + + /* set handler for ODP prefetching API - ibv_advise_mr(3) */ + ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +3
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode, int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length); enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */static inline int
rxe_odp_mr_init_user(structrxe_dev*rxe, u64 start, u64 length, u64 iova,drivers/infiniband/sw/rxe/rxe_loc.h +7
-0@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, { return RESPST_ERR_UNSUPPORTED_OPCODE; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + return -EOPNOTSUPP; +} + #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */drivers/infiniband/sw/rxe/rxe_odp.c +57
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct ib_sge *sg_list, + u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + unsigned int i; + int ret = 0; + + for (i = 0; i < num_sge; ++i) { + struct rxe_mr *mr; + struct ib_umem_odp *umem_odp; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + + if (IS_ERR(mr)) { + rxe_dbg_pd(pd, "mr with lkey %x not found\n", sg_list[i].lkey); + return PTR_ERR(mr); + } + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { + rxe_dbg_mr(mr, "missing write permission\n"); + rxe_put(mr); + return -EPERM; + } + + ret = rxe_odp_do_pagefault_and_lock(mr, sg_list[i].addr, + sg_list[i].length, pf_flags); + if (ret < 0) { + if (sg_list[i].length == 0) + continue; + + rxe_dbg_mr(mr, "failed to prefetch the mr\n"); + rxe_put(mr); + return ret; + } + + umem_odp = to_ib_umem_odp(mr->umem); + mutex_unlock(&umem_odp->umem_mutex); + + rxe_put(mr); + } + + return 0; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge) +{ + u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) + pf_flags |= RXE_PAGEFAULT_RDONLY;drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 492
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+ + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + pf_flags |= RXE_PAGEFAULT_SNAPSHOT; + + /* Synchronous call */ + if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH) + return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list, + num_sge); + + /* Asynchronous call is "best-effort" */Asynchronous call is not implemented now, why does this comment appear?
Quoted 4 lines
Quoted 1 line
Even without the 2nd patch, async calls are reported as successful.
Async call is not implemented. How to call "async calls are reported as successful"?
Please see the manual.
cf. https://manpages.debian.org/testing/libibverbs-dev/ibv_advise_mr.3.en.html
If IBV_ADVISE_MR_FLAG_FLUSH is not given to 'flags' parameter, then this function 'rxe_ib_advise_mr_prefetch()' simply returns 0. Consequently, ibv_advise_mr(3) and underlying ioctl(2) get no error. This behaviour is allowd in the spec as I quoted in the last reply.
It might be nice to return -EOPNOTSUPP just below the comment instead, but not doing so is acceptable according to the spec. Additionally, such change will be overwritten in the next patch after all.
Thanks,
Daisuke
Quoted 37 lines
Zhu Yanjun
Quoted 33 lines
The comment is inserted to show the reason, which is based on the
description from 'man 3ibv_advise_mr' as follows:
===
An application may pre-fetch any address range within an ODP MR when using theIBV_ADVISE_MR_ADVICE_PREFETCHorIBV_ADVISE_MR_ADVICE_PREFETCH_WRITEadvice. Semantically, this operation is best-effort. That means the kernel does not guarantee that underlying pages are updated in the HCA or the pre-fetched pages would remain resident.
===Thanks,
DaisukeQuoted 23 lines
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ + return 0; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, + struct ib_sge *sg_list, + u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + return -EOPNOTSUPP; + + return rxe_ib_advise_mr_prefetch(ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 149 lines · Daisuke Matsuda
Quoted 142 lines · Zhu Yanjun
Quoted 140 lines · Daisuke Matsuda
Quoted 138 lines · Zhu Yanjun
Quoted 125 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>--- drivers/infiniband/sw/rxe/rxe.c | 7 +++ drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++++ drivers/infiniband/sw/rxe/rxe_odp.c | 86 +++++++++++++++++++++++++++++ 3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev)mutex_destroy (&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; +/* initialize rxe device parameters */ static void rxe_init_device_param (struct rxe_dev *rxe,struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_SRQ_RECV ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_FLUSH ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_ATOMIC_WRITE ; + +/* set handler for ODP prefetching API - ibv_advise_mr(3) */ +ib_set_device_ops (&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +10
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode,int rxe_odp_flush_pmem_iova (struct rxe_mr *mr,u64 iova,unsigned int length);enum resp_states rxe_odp_do_atomic_write (struct rxe_mr *mr,u64 iova,u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs);#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int rxe_odp_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, {return RESPST_ERR_UNSUPPORTED_OPCODE ; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +return -EOPNOTSUPP ; +} +#endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */ drivers/infiniband/sw/rxe/rxe_odp.c +57
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)return RESPST_NONE ; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 pf_flags,struct ib_sge *sg_list, +u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); +unsigned int i; +int ret =0 ; + +for (i =0 ; i < num_sge; ++i) { +struct rxe_mr *mr; +struct ib_umem_odp *umem_odp; + + mr =lookup_mr (pd,IB_ACCESS_LOCAL_WRITE , + sg_list[i].lkey,RXE_LOOKUP_LOCAL ); + +if (IS_ERR (mr)) { +rxe_dbg_pd (pd,"mr with lkey %x not found\n" , sg_list[i].lkey); +return PTR_ERR (mr); + } + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { +rxe_dbg_mr (mr,"missing write permission\n" ); +rxe_put (mr); +return -EPERM ; + } + + ret =rxe_odp_do_pagefault_and_lock (mr, sg_list[i].addr, + sg_list[i].length, pf_flags); +if (ret <0 ) { +if (sg_list[i].length ==0 ) +continue ; + +rxe_dbg_mr (mr,"failed to prefetch the mr\n" ); +rxe_put (mr); +return ret; + } + + umem_odp =to_ib_umem_odp (mr->umem); +mutex_unlock (&umem_odp->umem_mutex); + +rxe_put (mr); + } + +return 0 ; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge) +{ +u32 pf_flags =RXE_PAGEFAULT_DEFAULT ; + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH ) + pf_flags |=RXE_PAGEFAULT_RDONLY ;drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 492
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+ +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) + pf_flags |=RXE_PAGEFAULT_SNAPSHOT ; + +/* Synchronous call */ +if (flags &IB_UVERBS_ADVISE_MR_FLAG_FLUSH ) +return rxe_ib_prefetch_sg_list (ibpd, advice, pf_flags, sg_list, + num_sge); + +/* Asynchronous call is "best-effort" */ Asynchronous call is not implemented now, why does this comment appear?
Quoted 4 lines
Quoted 1 line
Even without the 2nd patch, async calls are reported as successful.
Async call is not implemented. How to call "async calls are reported as successful"?
Quoted 7 lines
Please see the manual.
cf. https://manpages.debian.org/testing/libibverbs-dev/ibv_advise_mr.3.en.htmlIf
IBV_ADVISE_MR_FLAG_FLUSHis not given to 'flags' parameter, then this function 'rxe_ib_advise_mr_prefetch()' simply returns 0. Consequently,ibv_advise_mr(3)and underlying ioctl(2) get no error. This behaviour is allowd in the spec as I quoted in the last reply.
The functionality wasn't implemented, you added the comments first. You're still weaseling when people point this out.
Quoted 4 lines
It might be nice to return
-EOPNOTSUPPjust below the comment instead, but not doing so is acceptable according to the spec. Additionally, such change will be overwritten in the next patch after all.
Move comments to the next patch. In this patch, return -EOPNOTSUPP.
--G--
Quoted 43 lines
Thanks,
DaisukeQuoted 37 lines
Zhu Yanjun
Quoted 33 lines
The comment is inserted to show the reason, which is based on the
description from 'man 3ibv_advise_mr' as follows:
===
An application may pre-fetch any address range within an ODP MR when using theIBV_ADVISE_MR_ADVICE_PREFETCHorIBV_ADVISE_MR_ADVICE_PREFETCH_WRITEadvice. Semantically, this operation is best-effort. That means the kernel does not guarantee that underlying pages are updated in the HCA or the pre-fetched pages would remain resident.
===Thanks,
DaisukeQuoted 23 lines
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ +return 0 ; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags, +struct ib_sge *sg_list, +u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +if (advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) +return -EOPNOTSUPP ; + +return rxe_ib_advise_mr_prefetch (ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 162 lines · Greg Sword
Quoted 149 lines · Daisuke Matsuda
Quoted 142 lines · Zhu Yanjun
Quoted 140 lines · Daisuke Matsuda
Quoted 138 lines · Zhu Yanjun
Quoted 125 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>--- drivers/infiniband/sw/rxe/rxe.c | 7 +++ drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++++ drivers/infiniband/sw/rxe/rxe_odp.c | 86 +++++++++++++++++++++++++++++ 3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev)mutex_destroy (&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; +/* initialize rxe device parameters */ static void rxe_init_device_param (struct rxe_dev *rxe,struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_SRQ_RECV ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_FLUSH ; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |=IB_ODP_SUPPORT_ATOMIC_WRITE ; + +/* set handler for ODP prefetching API - ibv_advise_mr(3) */ +ib_set_device_ops (&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +10
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode,int rxe_odp_flush_pmem_iova (struct rxe_mr *mr,u64 iova,unsigned int length);enum resp_states rxe_odp_do_atomic_write (struct rxe_mr *mr,u64 iova,u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs);#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int rxe_odp_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, {return RESPST_ERR_UNSUPPORTED_OPCODE ; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +return -EOPNOTSUPP ; +} +#endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */ drivers/infiniband/sw/rxe/rxe_odp.c +57
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)return RESPST_NONE ; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 pf_flags,struct ib_sge *sg_list, +u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); +unsigned int i; +int ret =0 ; + +for (i =0 ; i < num_sge; ++i) { +struct rxe_mr *mr; +struct ib_umem_odp *umem_odp; + + mr =lookup_mr (pd,IB_ACCESS_LOCAL_WRITE , + sg_list[i].lkey,RXE_LOOKUP_LOCAL ); + +if (IS_ERR (mr)) { +rxe_dbg_pd (pd,"mr with lkey %x not found\n" , sg_list[i].lkey); +return PTR_ERR (mr); + } + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { +rxe_dbg_mr (mr,"missing write permission\n" ); +rxe_put (mr); +return -EPERM ; + } + + ret =rxe_odp_do_pagefault_and_lock (mr, sg_list[i].addr, + sg_list[i].length, pf_flags); +if (ret <0 ) { +if (sg_list[i].length ==0 ) +continue ; + +rxe_dbg_mr (mr,"failed to prefetch the mr\n" ); +rxe_put (mr); +return ret; + } + + umem_odp =to_ib_umem_odp (mr->umem); +mutex_unlock (&umem_odp->umem_mutex); + +rxe_put (mr); + } + +return 0 ; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags,struct ib_sge *sg_list,u32 num_sge) +{ +u32 pf_flags =RXE_PAGEFAULT_DEFAULT ; + +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH ) + pf_flags |=RXE_PAGEFAULT_RDONLY ;drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 492
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+ +if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) + pf_flags |=RXE_PAGEFAULT_SNAPSHOT ; + +/* Synchronous call */ +if (flags &IB_UVERBS_ADVISE_MR_FLAG_FLUSH ) +return rxe_ib_prefetch_sg_list (ibpd, advice, pf_flags, sg_list, + num_sge); + +/* Asynchronous call is "best-effort" */ Asynchronous call is not implemented now, why does this comment appear?
Quoted 4 lines
Quoted 1 line
Even without the 2nd patch, async calls are reported as successful.
Async call is not implemented. How to call "async calls are reported as successful"?
Quoted 7 lines
Please see the manual.
cf. https://manpages.debian.org/testing/libibverbs-dev/ibv_advise_mr.3.en.htmlIf
IBV_ADVISE_MR_FLAG_FLUSHis not given to 'flags' parameter, then this function 'rxe_ib_advise_mr_prefetch()' simply returns 0. Consequently,ibv_advise_mr(3)and underlying ioctl(2) get no error. This behaviour is allowd in the spec as I quoted in the last reply.The functionality wasn't implemented, you added the comments first. You're still weaseling when people point this out.
Quoted 1 line
Quoted 5 lines
Quoted 3 lines
It might be nice to return
-EOPNOTSUPPjust below the comment instead, but not doing so is acceptable according to the spec. Additionally, such change will be overwritten in the next patch after all.Move comments to the next patch. In this patch, return
-EOPNOTSUPP.
Any opinion from Zhu?
I may post a new revision to change the intermediate code, but the final result after applying the patchset will be the same. You are the maintainer of rxe, so I will follow that.
Thanks,
Daisuke
Quoted 46 lines
--G--
Quoted 43 lines
Thanks,
DaisukeQuoted 37 lines
Zhu Yanjun
Quoted 33 lines
The comment is inserted to show the reason, which is based on the
description from 'man 3ibv_advise_mr' as follows:
===
An application may pre-fetch any address range within an ODP MR when using theIBV_ADVISE_MR_ADVICE_PREFETCHorIBV_ADVISE_MR_ADVICE_PREFETCH_WRITEadvice. Semantically, this operation is best-effort. That means the kernel does not guarantee that underlying pages are updated in the HCA or the pre-fetched pages would remain resident.
===Thanks,
DaisukeQuoted 23 lines
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ +return 0 ; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 flags, +struct ib_sge *sg_list, +u32 num_sge, +struct uverbs_attr_bundle *attrs) +{ +if (advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice !=IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT ) +return -EOPNOTSUPP ; + +return rxe_ib_advise_mr_prefetch (ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 179 lines · Daisuke Matsuda
Quoted 177 lines · Greg Sword
Quoted 160 lines · Daisuke Matsuda
Quoted 153 lines · Zhu Yanjun
Quoted 151 lines · Daisuke Matsuda
Quoted 149 lines · Zhu Yanjun
Quoted 142 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>---
drivers/infiniband/sw/rxe/rxe.c | 7 +++drivers/infiniband/sw/rxe/rxe_loc.h| 10 ++++drivers/infiniband/sw/rxe/rxe_odp.c| 86 +++++++++++++++++++++
++++++++3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +6
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev) mutex_destroy(&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; + /* initialize rxe device parameters */ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE; + + /* set handler for ODP prefetching API -
ibv_advise_mr(3)*/
+ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops); }
}drivers/infiniband/sw/rxe/rxe_loc.h +3
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode, int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length); enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */static inline int
rxe_odp_mr_init_user(structrxe_dev*rxe, u64 start, u64 length, u64 iova,drivers/infiniband/sw/rxe/rxe_loc.h +7
-0@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, { return RESPST_ERR_UNSUPPORTED_OPCODE; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + return -EOPNOTSUPP; +} + #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */#endif /*
RXE_LOC_H*/drivers/infiniband/sw/rxe/rxe_odp.c +19
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct ib_sge *sg_list, + u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + unsigned int i; + int ret = 0; + + for (i = 0; i < num_sge; ++i) { + struct rxe_mr *mr; + struct ib_umem_odp *umem_odp; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + + if (IS_ERR(mr)) { + rxe_dbg_pd(pd, "mr with lkey %x not found\n",sg_list[i].lkey);
+ returnPTR_ERR(mr);
+ }drivers/infiniband/sw/rxe/rxe_odp.c +31
-0+ + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { + rxe_dbg_mr(mr, "missing write permission\n"); + rxe_put(mr); + return -EPERM; + } + + ret = rxe_odp_do_pagefault_and_lock(mr, sg_list[i].addr, + sg_list[i].length, pf_flags); + if (ret < 0) { + if (sg_list[i].length == 0) + continue; + + rxe_dbg_mr(mr, "failed to prefetch the mr\n"); + rxe_put(mr); + return ret; + } + + umem_odp = to_ib_umem_odp(mr->umem); + mutex_unlock(&umem_odp->umem_mutex); + + rxe_put(mr); + } + + return 0; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32
num_sge)drivers/infiniband/sw/rxe/rxe_odp.c +12
-0+{ + u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) + pf_flags |= RXE_PAGEFAULT_RDONLY; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + pf_flags |= RXE_PAGEFAULT_SNAPSHOT; + + /* Synchronous call */ + if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH) + return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags,Quoted 4 lines
sg_list,
+num_sge);
+
+ /* Asynchronous call is "best-effort" */Asynchronous call is not implemented now, why does this comment appear?
Quoted 4 lines
Quoted 1 line
Even without the 2nd patch, async calls are reported as successful.
Async call is not implemented. How to call "async calls are reported as successful"?
Quoted 7 lines
Please see the manual.
cf. https://manpages.debian.org/testing/libibverbs-dev/ ibv_advise_mr.3.en.htmlIf
IBV_ADVISE_MR_FLAG_FLUSHis not given to 'flags' parameter, then this function 'rxe_ib_advise_mr_prefetch()' simply returns 0. Consequently,ibv_advise_mr(3)and underlying ioctl(2) get no error. This behaviour is allowd in the spec as I quoted in the last reply.The functionality wasn't implemented, you added the comments first. You're still weaseling when people point this out.
Quoted 4 lines
It might be nice to return
-EOPNOTSUPPjust below the comment instead, but not doing so is acceptable according to the spec. Additionally, such change will be overwritten in the next patch after all.
Quoted 6 lines
Quoted 1 line
Move comments to the next patch. In this patch, return
-EOPNOTSUPP.Any opinion from Zhu?
I may post a new revision to change the intermediate code, but the final result after applying the patchset will be the same. You are the maintainer of rxe, so I will follow that.
Thanks a lot. I am fine with your commit.
Please "Move comments to the next patch. In this patch, return -EOPNOTSUPP".
I think that it is a good idea. The final result should be the same. Please send out the latest commit following the suggestions.
Thanks a lot for your contributions and efforts.
Best Regards,
Zhu Yanjun
Quoted 52 lines
Thanks,
DaisukeQuoted 46 lines
--G--
Quoted 43 lines
Thanks,
DaisukeQuoted 37 lines
Zhu Yanjun
Quoted 33 lines
The comment is inserted to show the reason, which is based on the
description from 'man 3ibv_advise_mr' as follows:
===
An application may pre-fetch any address range within an ODP MR when using theIBV_ADVISE_MR_ADVICE_PREFETCHorIBV_ADVISE_MR_ADVICE_PREFETCH_WRITEadvice. Semantically, this operation is best-effort. That means the kernel does not guarantee that underlying pages are updated in the HCA or the pre-fetched pages would remain resident.
===Thanks,
DaisukeQuoted 23 lines
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ + return 0; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, + struct ib_sge *sg_list, + u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + return -EOPNOTSUPP; + + return rxe_ib_advise_mr_prefetch(ibpd, advice, flags, + sg_list, num_sge); +}
Re: [PATCH for-next v2 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs
Quoted 176 lines · Zhu Yanjun
Quoted 168 lines · Daisuke Matsuda
Quoted 166 lines · Greg Sword
Quoted 149 lines · Daisuke Matsuda
Quoted 142 lines · Zhu Yanjun
Quoted 140 lines · Daisuke Matsuda
Quoted 138 lines · Zhu Yanjun
Quoted 125 lines · Daisuke Matsuda
Minimal implementation of
ibv_advise_mr(3)requires synchronous calls being successful with theIBV_ADVISE_MR_FLAG_FLUSHflag. Asynchronous requests, which are best-effort, will be added subsequently.Signed-off-by Daisuke Matsuda <[email protected]>---
drivers/infiniband/sw/rxe/rxe.c | 7 +++drivers/infiniband/sw/rxe/rxe_loc.h| 10 ++++drivers/infiniband/sw/rxe/rxe_odp.c| 86 +++++++++++++++++++++ ++++++++3 files changed, 103 insertions(+)drivers/infiniband/sw/rxe/rxe.c +7
-0@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev) mutex_destroy(&rxe->usdev_lock); } +static const struct ib_device_ops rxe_ib_dev_odp_ops = { + .advise_mr = rxe_ib_advise_mr, +}; + /* initialize rxe device parameters */ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) {@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH; rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE; + + /* set handler for ODP prefetching API - ibv_advise_mr(3) */ + ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops); } }drivers/infiniband/sw/rxe/rxe_loc.h +3
-0@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode, int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length); enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value); +int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */static inline int
rxe_odp_mr_init_user(structrxe_dev*rxe, u64 start, u64 length, u64 iova,drivers/infiniband/sw/rxe/rxe_loc.h +7
-0@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, { return RESPST_ERR_UNSUPPORTED_OPCODE; } +static inline int rxe_ib_advise_mr (struct ib_pd *pd,enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + return -EOPNOTSUPP; +} + #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */drivers/infiniband/sw/rxe/rxe_odp.c +57
-0@@ -424,3 +424,89 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } + +static int rxe_ib_prefetch_sg_list (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct ib_sge *sg_list, + u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + unsigned int i; + int ret = 0; + + for (i = 0; i < num_sge; ++i) { + struct rxe_mr *mr; + struct ib_umem_odp *umem_odp; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + + if (IS_ERR(mr)) { + rxe_dbg_pd(pd, "mr with lkey %x not found\n", sg_list[i].lkey); + return PTR_ERR(mr); + } + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + !mr->umem->writable) { + rxe_dbg_mr(mr, "missing write permission\n"); + rxe_put(mr); + return -EPERM; + } + + ret = rxe_odp_do_pagefault_and_lock(mr, sg_list[i].addr, + sg_list[i].length, pf_flags); + if (ret < 0) { + if (sg_list[i].length == 0) + continue; + + rxe_dbg_mr(mr, "failed to prefetch the mr\n"); + rxe_put(mr); + return ret; + } + + umem_odp = to_ib_umem_odp(mr->umem); + mutex_unlock(&umem_odp->umem_mutex); + + rxe_put(mr); + } + + return 0; +} + +static int rxe_ib_advise_mr_prefetch (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, struct ib_sge *sg_list, u32 num_sge) +{ + u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) + pf_flags |= RXE_PAGEFAULT_RDONLY;drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 492
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+ + if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + pf_flags |= RXE_PAGEFAULT_SNAPSHOT; + + /* Synchronous call */ + if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH) + return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list, + num_sge); + + /* Asynchronous call is "best-effort" */Asynchronous call is not implemented now, why does this comment appear?
Quoted 4 lines
Quoted 1 line
Even without the 2nd patch, async calls are reported as successful.
Async call is not implemented. How to call "async calls are reported as successful"?
Quoted 7 lines
Please see the manual.
cf. https://manpages.debian.org/testing/libibverbs-dev/ ibv_advise_mr.3.en.htmlIf
IBV_ADVISE_MR_FLAG_FLUSHis not given to 'flags' parameter, then this function 'rxe_ib_advise_mr_prefetch()' simply returns 0. Consequently,ibv_advise_mr(3)and underlying ioctl(2) get no error. This behaviour is allowd in the spec as I quoted in the last reply.The functionality wasn't implemented, you added the comments first. You're still weaseling when people point this out.
Quoted 4 lines
It might be nice to return
-EOPNOTSUPPjust below the comment instead, but not doing so is acceptable according to the spec. Additionally, such change will be overwritten in the next patch after all.Quoted 6 lines
Quoted 1 line
Move comments to the next patch. In this patch, return
-EOPNOTSUPP.Any opinion from Zhu?
I may post a new revision to change the intermediate code, but the final result after applying the patchset will be the same. You are the maintainer of rxe, so I will follow that.
Quoted 5 lines
Thanks a lot. I am fine with your commit.
Please "Move comments to the next patch. In this patch, return-EOPNOTSUPP".
I think that it is a good idea. The final result should be the same. Please send out the latest commit following the suggestions.Thanks a lot for your contributions and efforts.
Thank you for the reply and support.
I've posted a new version with the following changes:
- Added return -EOPNOTSUPP; (1st patch)
- Changed type definition from 'unsigned int' to 'u32' as you suggested (1st patch)
- Deleted redundant if-continue statements in rxe_ib_prefetch_sg_list() (1st patch)
Thanks,
Daisuke
Quoted 57 lines
Best Regards,
Zhu YanjunQuoted 52 lines
Thanks,
DaisukeQuoted 46 lines
--G--
Quoted 43 lines
Thanks,
DaisukeQuoted 37 lines
Zhu Yanjun
Quoted 33 lines
The comment is inserted to show the reason, which is based on the
description from 'man 3ibv_advise_mr' as follows:
===
An application may pre-fetch any address range within an ODP MR when using theIBV_ADVISE_MR_ADVICE_PREFETCHorIBV_ADVISE_MR_ADVICE_PREFETCH_WRITEadvice. Semantically, this operation is best-effort. That means the kernel does not guarantee that underlying pages are updated in the HCA or the pre-fetched pages would remain resident.
===Thanks,
DaisukeQuoted 23 lines
Zhu Yanjun
Quoted 19 lines
drivers/infiniband/sw/rxe/rxe_odp.c +19
-0+ + return 0; +} + +int rxe_ib_advise_mr (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 flags, + struct ib_sge *sg_list, + u32 num_sge, + struct uverbs_attr_bundle *attrs) +{ + if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE && + advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT) + return -EOPNOTSUPP; + + return rxe_ib_advise_mr_prefetch(ibpd, advice, flags, + sg_list, num_sge); +}
PATCH v2 2/2 RDMA/rxe: Enable asynchronous prefetch for ODP MRs
In reply to Daisuke MatsudaCalling ibv_advise_mr(3) with flags other than IBV_ADVISE_MR_FLAG_FLUSH invokes asynchronous requests. It is best-effort, and thus can safely be deferred to the system-wide workqueue.
--- drivers/infiniband/sw/rxe/rxe_odp.c | 81 ++++++++++++++++++++++++++++- 1 file changed, 80 insertions(+), 1 deletion(-)
drivers/infiniband/sw/rxe/rxe_odp.c +80 -1
@@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)return RESPST_NONE ; } +struct prefetch_mr_work { +struct work_struct work; +u32 pf_flags; +u32 num_sge; +struct { +u64 io_virt; +struct rxe_mr *mr; +size_t length; + } frags[]; +}; + +static void rxe_ib_prefetch_mr_work (struct work_struct *w) +{ +struct prefetch_mr_work *work = + container_of(w, struct prefetch_mr_work, work); +int ret; +u32 i; + +/* We rely on IB/core that work is executed if we have num_sge != 0 only. */ +WARN_ON (!work->num_sge); +for (i =0 ; i < work->num_sge; ++i) { +struct ib_umem_odp *umem_odp; + + ret =rxe_odp_do_pagefault_and_lock (work->frags[i].mr, work->frags[i].io_virt, + work->frags[i].length, work->pf_flags); +if (ret <0 ) { +rxe_dbg_mr (work->frags[i].mr,"failed to prefetch the mr\n" ); +continue ; + } + + umem_odp =to_ib_umem_odp (work->frags[i].mr->umem); +mutex_unlock (&umem_odp->umem_mutex); + } + +kvfree (work); +} + +static int rxe_init_prefetch_work (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 pf_flags,struct prefetch_mr_work *work, +struct ib_sge *sg_list,u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); +u32 i; + +INIT_WORK (&work->work, rxe_ib_prefetch_mr_work); + work->pf_flags = pf_flags; + +for (i =0 ; i < num_sge; ++i) { +struct rxe_mr *mr; + + mr =lookup_mr (pd,IB_ACCESS_LOCAL_WRITE , + sg_list[i].lkey,RXE_LOOKUP_LOCAL ); +if (IS_ERR (mr)) { + work->num_sge = i; +return PTR_ERR (mr); + } + work->frags[i].io_virt = sg_list[i].addr; + work->frags[i].length = sg_list[i].length; + work->frags[i].mr = mr; + +rxe_put (mr); + } + work->num_sge = num_sge; +return 0 ; +} + static int rxe_ib_prefetch_sg_list(struct ib_pd *ibpd, enum ib_uverbs_advise_mr_advice advice, u32 pf_flags, struct ib_sge *sg_list,
@@ -478,6 +545,8 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd, u32 flags, struct ib_sge *sg_list, u32 num_sge) { u32 pf_flags = RXE_PAGEFAULT_DEFAULT; +struct prefetch_mr_work *work; +int rc;if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH ) pf_flags |=RXE_PAGEFAULT_RDONLY ;
@@ -490,7 +559,17 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd,return rxe_ib_prefetch_sg_list (ibpd, advice, pf_flags, sg_list, num_sge); -/* Asynchronous call is "best-effort" */ +/* Asynchronous call is "best-effort" and allowed to fail */ + work =kvzalloc (struct_size (work, frags, num_sge),GFP_KERNEL ); +if (!work) +return -ENOMEM ; + + rc =rxe_init_prefetch_work (ibpd, advice, pf_flags, work, sg_list, num_sge); +if (rc) { +kvfree (work); +return rc; + } +queue_work (system_unbound_wq, &work->work);return 0 ; }
Signature · 1 line
2.43.0
Re: [PATCH for-next v2 2/2] RDMA/rxe: Enable asynchronous prefetch for ODP MRs
Quoted 5 lines · Daisuke Matsuda
Calling
ibv_advise_mr(3)with flags other thanIBV_ADVISE_MR_FLAG_FLUSHinvokes asynchronous requests. It is best-effort, and thus can safely be deferred to the system-wide workqueue.Signed-off-by Daisuke Matsuda <[email protected]>
I have made tests with rdma-core after applying this patch series. It seems that it can work well. I read through this commit. Other than the following minor problems, I am fine with this commit.
Quoted 12 lines
--- drivers/infiniband/sw/rxe/rxe_odp.c | 81 ++++++++++++++++++++++++++++- 1 file changed, 80 insertions(+), 1 deletion(-)drivers/infiniband/sw/rxe/rxe_odp.c +0
-0@@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)return RESPST_NONE ; }
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 437
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+struct prefetch_mr_work { +struct work_struct work; +u32 pf_flags; +u32 num_sge; +struct { +u64 io_virt; +struct rxe_mr *mr; +size_t length; + } frags[]; +};
The struct prefetch_mr_work should be moved into header file? IMO, it is better to move this struct to rxe_loc.h?
Quoted 12 lines
drivers/infiniband/sw/rxe/rxe_odp.c +12
-0+ +static void rxe_ib_prefetch_mr_work (struct work_struct *w) +{ +struct prefetch_mr_work *work = + container_of(w, struct prefetch_mr_work, work); +int ret; +u32 i; + +/* We rely on IB/core that work is executed if we have num_sge != 0 only. */ +WARN_ON (!work->num_sge); +for (i =0 ; i < work->num_sge; ++i) { +struct ib_umem_odp *umem_odp;
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 459
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_odp.c +10
-0+ + ret =rxe_odp_do_pagefault_and_lock (work->frags[i].mr, work->frags[i].io_virt, + work->frags[i].length, work->pf_flags); +if (ret <0 ) { +rxe_dbg_mr (work->frags[i].mr,"failed to prefetch the mr\n" ); +continue ; + } + + umem_odp =to_ib_umem_odp (work->frags[i].mr->umem); +mutex_unlock (&umem_odp->umem_mutex);
Obviously this function is dependent on the mutex lock umem_mutex. So in the beginning of this function, it is better to add lockdep_assert_held(&umem_odp->umem_mutex)?
Zhu Yanjun
Quoted 66 lines
drivers/infiniband/sw/rxe/rxe_odp.c +48
-1+ } + +kvfree (work); +} + +static int rxe_init_prefetch_work (struct ib_pd *ibpd, +enum ib_uverbs_advise_mr_advice advice, +u32 pf_flags,struct prefetch_mr_work *work, +struct ib_sge *sg_list,u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); +u32 i; + +INIT_WORK (&work->work, rxe_ib_prefetch_mr_work); + work->pf_flags = pf_flags; + +for (i =0 ; i < num_sge; ++i) { +struct rxe_mr *mr; + + mr =lookup_mr (pd,IB_ACCESS_LOCAL_WRITE , + sg_list[i].lkey,RXE_LOOKUP_LOCAL ); +if (IS_ERR (mr)) { + work->num_sge = i; +return PTR_ERR (mr); + } + work->frags[i].io_virt = sg_list[i].addr; + work->frags[i].length = sg_list[i].length; + work->frags[i].mr = mr; + +rxe_put (mr); + } + work->num_sge = num_sge; +return 0 ; +} + static int rxe_ib_prefetch_sg_list(struct ib_pd *ibpd, enum ib_uverbs_advise_mr_advice advice, u32 pf_flags, struct ib_sge *sg_list,@@ -478,6 +545,8 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd, u32 flags, struct ib_sge *sg_list, u32 num_sge) { u32 pf_flags = RXE_PAGEFAULT_DEFAULT; +struct prefetch_mr_work *work; +int rc;if (advice ==IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH ) pf_flags |=RXE_PAGEFAULT_RDONLY ;@@ -490,7 +559,17 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd,return rxe_ib_prefetch_sg_list (ibpd, advice, pf_flags, sg_list, num_sge); -/* Asynchronous call is "best-effort" */ +/* Asynchronous call is "best-effort" and allowed to fail */ + work =kvzalloc (struct_size (work, frags, num_sge),GFP_KERNEL ); +if (!work) +return -ENOMEM ; + + rc =rxe_init_prefetch_work (ibpd, advice, pf_flags, work, sg_list, num_sge); +if (rc) { +kvfree (work); +return rc; + } +queue_work (system_unbound_wq, &work->work);return 0 ; }
Re: [PATCH for-next v2 2/2] RDMA/rxe: Enable asynchronous prefetch for ODP MRs
Quoted 27 lines · Zhu Yanjun
Quoted 5 lines · Daisuke Matsuda
Calling
ibv_advise_mr(3)with flags other thanIBV_ADVISE_MR_FLAG_FLUSHinvokes asynchronous requests. It is best-effort, and thus can safely be deferred to the system-wide workqueue.Signed-off-by Daisuke Matsuda <[email protected]>I have made tests with rdma-core after applying this patch series. It seems that it can work well. I read through this commit. Other than the following minor problems, I am fine with this commit.
Reviewed-by Zhu Yanjun <[email protected]>Quoted 15 lines
---
drivers/infiniband/sw/rxe/rxe_odp.c| 81 ++++++++++++++++++++++++++++-1 file changed, 80 insertions(+), 1 deletion(-)drivers/infiniband/sw/rxe/rxe_odp.c +4
-0@@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } +struct prefetch_mr_work { + struct work_struct work; + u32 pf_flags; + u32 num_sge;
Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 437
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_odp.c +6
-0+ struct { + u64 io_virt; + struct rxe_mr *mr; + size_t length; + } frags[]; +};The struct
prefetch_mr_workshould be moved into header file? IMO, it is better to move this struct torxe_loc.h?
This struct is not likely to be used in other files.
I think leaving it here would be easier for other developers to understand because relevant codes are gathered. If there is any specific reason to move, I will do so.
Quoted 17 lines
Quoted 17 lines
drivers/infiniband/sw/rxe/rxe_odp.c +16
-0+ +static void rxe_ib_prefetch_mr_work (struct work_struct *w) +{ + struct prefetch_mr_work *work = + container_of(w, struct prefetch_mr_work, work); + int ret; + u32 i; + + /* We rely on IB/core that work is executed if we have num_sge != 0 only. */ + WARN_ON(!work->num_sge); + for (i = 0; i < work->num_sge; ++i) { + struct ib_umem_odp *umem_odp; + + ret = rxe_odp_do_pagefault_and_lock(work->frags[i].mr, work->frags[i].io_virt, + work->frags[i].length, work->pf_flags); + if (ret < 0) {
Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 459
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_odp.c +6
-0+ rxe_dbg_mr(work->frags[i].mr, "failed to prefetch the mr\n"); + continue; + } + + umem_odp = to_ib_umem_odp(work->frags[i].mr->umem); + mutex_unlock(&umem_odp->umem_mutex);Obviously this function is dependent on the mutex lock
umem_mutex. So in the beginning of this function, it is better to addlockdep_assert_held(&umem_odp->umem_mutex)?
The mutex **must not** be locked at the beginning, so perhaps we can add lockdep_assert_not_held() instead at the beginning, but this one is not used frequently, and we can do without that.
umem_mutex is locked in rxe_odp_do_pagefault_and_lock() in the for loop.
The function calls ib_umem_odp_map_dma_and_lock(), which locks the mutex only when pagefault is successful. If ib_umem_odp_map_dma_and_lock() fails, an error is returned and then the mutex is not locked.
Daisuke
Quoted 67 lines
Zhu Yanjun
Quoted 63 lines
drivers/infiniband/sw/rxe/rxe_odp.c +48
-1+ } + + kvfree(work); +} + +static int rxe_init_prefetch_work (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct prefetch_mr_work *work, + struct ib_sge *sg_list, u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + u32 i; + + INIT_WORK(&work->work, rxe_ib_prefetch_mr_work); + work->pf_flags = pf_flags; + + for (i = 0; i < num_sge; ++i) { + struct rxe_mr *mr; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + if (IS_ERR(mr)) { + work->num_sge = i; + return PTR_ERR(mr); + } + work->frags[i].io_virt = sg_list[i].addr; + work->frags[i].length = sg_list[i].length; + work->frags[i].mr = mr; + + rxe_put(mr); + } + work->num_sge = num_sge; + return 0; +} + static int rxe_ib_prefetch_sg_list(struct ib_pd *ibpd, enum ib_uverbs_advise_mr_advice advice, u32 pf_flags, struct ib_sge *sg_list,@@ -478,6 +545,8 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd, u32 flags, struct ib_sge *sg_list, u32 num_sge) { u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + struct prefetch_mr_work *work; + int rc; if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) pf_flags |= RXE_PAGEFAULT_RDONLY;@@ -490,7 +559,17 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd, return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list, num_sge); - /* Asynchronous call is "best-effort" */ + /* Asynchronous call is "best-effort" and allowed to fail */ + work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL); + if (!work) + return -ENOMEM; + + rc = rxe_init_prefetch_work(ibpd, advice, pf_flags, work, sg_list, num_sge); + if (rc) { + kvfree(work); + return rc; + } + queue_work(system_unbound_wq, &work->work); return 0; }
Re: [PATCH for-next v2 2/2] RDMA/rxe: Enable asynchronous prefetch for ODP MRs
Quoted 73 lines · Daisuke Matsuda
Quoted 28 lines · Zhu Yanjun
Quoted 5 lines · Daisuke Matsuda
Calling
ibv_advise_mr(3)with flags other thanIBV_ADVISE_MR_FLAG_FLUSHinvokes asynchronous requests. It is best-effort, and thus can safely be deferred to the system-wide workqueue.Signed-off-by Daisuke Matsuda <[email protected]>I have made tests with rdma-core after applying this patch series. It seems that it can work well. I read through this commit. Other than the following minor problems, I am fine with this commit.
Reviewed-by Zhu Yanjun <[email protected]>Quoted 16 lines
---
drivers/infiniband/sw/rxe/rxe_odp.c| 81 ++++++++++++++++++++++++++++-1 file changed, 80 insertions(+), 1 deletion(-)drivers/infiniband/sw/rxe/rxe_odp.c +4
-0@@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } +struct prefetch_mr_work { + struct work_struct work; + u32 pf_flags; + u32 num_sge;Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 437
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_odp.c +6
-0+ struct { + u64 io_virt; + struct rxe_mr *mr; + size_t length; + } frags[]; +};The struct
prefetch_mr_workshould be moved into header file? IMO, it is better to move this struct torxe_loc.h?This struct is not likely to be used in other files.
I think leaving it here would be easier for other developers to understand because relevant codes are gathered. If there is any specific reason to move, I will do so.Quoted 20 lines
Quoted 20 lines
drivers/infiniband/sw/rxe/rxe_odp.c +9
-0+ +static void rxe_ib_prefetch_mr_work (struct work_struct *w) +{ + struct prefetch_mr_work *work = + container_of(w, struct prefetch_mr_work, work); + int ret; + u32 i; + + /* We rely on IB/core that work is executed if we have num_sge!= 0 only. */
+WARN_ON(!work->num_sge);
+ for (i = 0; i <work->num_sge; ++i) {
+ structib_umem_odp*umem_odp;
+
+ ret = rxe_odp_do_pagefault_and_lock(work->frags[i].mr, work->frags[i].io_virt,
+ work->frags[i].length,work->pf_flags); + if (ret < 0) {
+ rxe_dbg_mr(work->frags[i].mr, "failed to prefetch theQuoted 8 lines
Quoted 6 lines
mr\n");
+ continue;
+ }
+
+umem_odp=to_ib_umem_odp(work->frags[i].mr->umem); +mutex_unlock(&umem_odp->umem_mutex);Obviously this function is dependent on the mutex lock
umem_mutex. So in the beginning of this function, it is better to addlockdep_assert_held(&umem_odp->umem_mutex)?
Quoted 6 lines
The mutex **must not** be locked at the beginning, so perhaps we can add
lockdep_assert_not_held()instead at the beginning, but this one is not used frequently, and we can do without that.
umem_mutexis locked inrxe_odp_do_pagefault_and_lock()in the for loop.
The function callsib_umem_odp_map_dma_and_lock(), which locks the mutex only when pagefault is successful.
Sure. umem_mutex is locked in other functions. It is not necessary to use lockdep_assert_held at the beginning of this function.
Zhu Yanjun
Quoted 75 lines
If
ib_umem_odp_map_dma_and_lock()fails, an error is returned and then the mutex is not locked.Daisuke
Quoted 70 lines
Zhu Yanjun
Quoted 66 lines
+ }
drivers/infiniband/sw/rxe/rxe_odp.c +34
-0+ + kvfree(work); +} + +static int rxe_init_prefetch_work (struct ib_pd *ibpd, + enum ib_uverbs_advise_mr_advice advice, + u32 pf_flags, struct prefetch_mr_work *work, + struct ib_sge *sg_list, u32 num_sge) +{ + struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd); + u32 i; + + INIT_WORK(&work->work, rxe_ib_prefetch_mr_work); + work->pf_flags = pf_flags; + + for (i = 0; i < num_sge; ++i) { + struct rxe_mr *mr; + + mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE, + sg_list[i].lkey, RXE_LOOKUP_LOCAL); + if (IS_ERR(mr)) { + work->num_sge = i; + return PTR_ERR(mr); + } + work->frags[i].io_virt = sg_list[i].addr; + work->frags[i].length = sg_list[i].length; + work->frags[i].mr = mr; + + rxe_put(mr); + } + work->num_sge = num_sge; + return 0; +} +static int rxe_ib_prefetch_sg_list(struct
ib_pd*ibpd, enumib_uverbs_advise_mr_adviceadvice, u32pf_flags, structib_sge*sg_list,drivers/infiniband/sw/rxe/rxe_odp.c +13
-1@@ -478,6 +545,8 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd, u32 flags, struct ib_sge *sg_list, u32 num_sge) { u32 pf_flags = RXE_PAGEFAULT_DEFAULT; + struct prefetch_mr_work *work; + int rc; if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH) pf_flags |= RXE_PAGEFAULT_RDONLY;@@ -490,7 +559,17 @@ static int rxe_ib_advise_mr_prefetch(structib_pd *ibpd, return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list, num_sge); - /* Asynchronous call is "best-effort" */ + /* Asynchronous call is "best-effort" and allowed to fail */ + work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL); + if (!work) + return -ENOMEM; + + rc = rxe_init_prefetch_work(ibpd, advice, pf_flags, work, sg_list, num_sge); + if (rc) { + kvfree(work); + return rc; + } + queue_work(system_unbound_wq, &work->work); return 0; }
Re: [PATCH for-next v2 2/2] RDMA/rxe: Enable asynchronous prefetch for ODP MRs
In reply to Daisuke MatsudaQuoted 40 lines · Daisuke Matsuda
Quoted 28 lines · Zhu Yanjun
Quoted 5 lines · Daisuke Matsuda
Calling
ibv_advise_mr(3)with flags other thanIBV_ADVISE_MR_FLAG_FLUSHinvokes asynchronous requests. It is best-effort, and thus can safely be deferred to the system-wide workqueue.Signed-off-by Daisuke Matsuda <[email protected]>I have made tests with rdma-core after applying this patch series. It seems that it can work well. I read through this commit. Other than the following minor problems, I am fine with this commit.
Reviewed-by Zhu Yanjun <[email protected]>Quoted 16 lines
---
drivers/infiniband/sw/rxe/rxe_odp.c| 81 ++++++++++++++++++++++++++++-1 file changed, 80 insertions(+), 1 deletion(-)drivers/infiniband/sw/rxe/rxe_odp.c +4
-0@@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) return RESPST_NONE; } +struct prefetch_mr_work { + struct work_struct work; + u32 pf_flags; + u32 num_sge;Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_odp.c enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) · line 437
Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_odp.c +6
-0+ struct { + u64 io_virt; + struct rxe_mr *mr; + size_t length; + } frags[]; +};The struct
prefetch_mr_workshould be moved into header file? IMO, it is better to move this struct torxe_loc.h?
Quoted 1 line
This struct is not likely to be used in other files.
Normally a struct should be in a header file. Not mandatory.
Zhu Yanjun
Quoted 2 lines
I think leaving it here would be easier for other developers to understand because relevant codes are gathered. If there is any specific reason to move, I will do so.
Quoted 6 lines · Daisuke Matsuda
There is
ibv_advise_mr(3)that can be used by applications to optimize memory access. This series enables the feature on rxe driver, which has already been available in mlx5.There is a tiny change on the rdma-core util. cf. https://github.com/linux-rdma/rdma-core/pull/1605
Hi, Daisuke
Thanks a lot for your efforts to this patch series. It is very nice. With this patch series, we can make prefetch for ODP MRs of RXE.
I read through this patch series. And it seems fine with me.^_^
IIRC, you have added ODP testcases in rdma-core. To verify this prefetch work well for ODP MRs, can you add this synchronous/asynchronous prefetch to the ODP testcases in rdma-core?
Thus, we can verify this patch series. And in the future, we can use these testcases to confirm this prefetch feature work well.
To now, it seems that no tool can verify this prefetch feature.
Thanks a lot.
Zhu Yanjun
Quoted 10 lines
Daisuke Matsuda (2):
RDMA/rxe: Implement synchronous prefetch for ODP MRs RDMA/rxe: Enable asynchronous prefetch for ODP MRsdrivers/infiniband/sw/rxe/rxe.c | 7 ++ drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++ drivers/infiniband/sw/rxe/rxe_odp.c | 165 ++++++++++++++++++++++++++++ 3 files changed, 182 insertions(+)
Quoted 12 lines · Zhu Yanjun
Quoted 6 lines · Daisuke Matsuda
There is
ibv_advise_mr(3)that can be used by applications to optimize memory access. This series enables the feature on rxe driver, which has already been available in mlx5.There is a tiny change on the rdma-core util. cf. https://github.com/linux-rdma/rdma-core/pull/1605
Hi, Daisuke
Thanks a lot for your efforts to this patch series. It is very nice. With this patch series, we can make prefetch for ODP MRs of RXE.
Quoted 7 lines
I read through this patch series. And it seems fine with me.^_^
IIRC, you have added ODP testcases in rdma-core. To verify this prefetch work well for ODP MRs, can you add this synchronous/asynchronous prefetch to the ODP testcases in rdma-core?
Thus, we can verify this patch series. And in the future, we can use these testcases to confirm this prefetch feature work well.
To now, it seems that no tool can verify this prefetch feature.
Hi, Thank you for taking a look!
There are already relevant testcases implemented in ./tests/test_odp.py:
- test_odp_sync_prefetch_rc_traffic
- test_odp_async_prefetch_rc_traffic
- test_odp_prefetch_sync_no_page_fault_rc_traffic
- test_odp_prefetch_async_no_page_fault_rc_traffic
On my node (x86-64/linux-6.15.0-rc1+), this series can pass all these tests.
Other than that, this feature is required by librpma as far as I know. Though it is not maintained anymore, perhaps this could also be used for testing. cf. https://github.com/pmem/rpma
Thanks,
Daisuke
Quoted 15 lines
Thanks a lot.
Zhu YanjunQuoted 10 lines
Daisuke Matsuda (2):
RDMA/rxe: Implement synchronous prefetch for ODP MRs RDMA/rxe: Enable asynchronous prefetch for ODP MRsdrivers/infiniband/sw/rxe/rxe.c | 7 ++
drivers/infiniband/sw/rxe/rxe_loc.h| 10 ++drivers/infiniband/sw/rxe/rxe_odp.c| 165 ++++++++++++++++++++++++++++3 files changed, 182 insertions(+)