| Message ID | 20250318094932.2643614-3-matsuda-daisuke@fujitsu.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | RDMA/rxe: RDMA FLUSH and ATOMIC WRITE with ODP | expand |
On Tue, Mar 18, 2025 at 06:49:32PM +0900, Daisuke Matsuda wrote: > Add rxe_odp_do_atomic_write() so that ODP specific steps are applied to > ATOMIC WRITE requests. > > Signed-off-by: Daisuke Matsuda <matsuda-daisuke@fujitsu.com> > --- > drivers/infiniband/sw/rxe/rxe.c | 1 + > drivers/infiniband/sw/rxe/rxe_loc.h | 5 +++ > drivers/infiniband/sw/rxe/rxe_mr.c | 12 ------- > drivers/infiniband/sw/rxe/rxe_odp.c | 53 ++++++++++++++++++++++++++++ > drivers/infiniband/sw/rxe/rxe_resp.c | 11 +++++- > include/rdma/ib_verbs.h | 1 + > 6 files changed, 70 insertions(+), 13 deletions(-) > > diff --git a/drivers/infiniband/sw/rxe/rxe.c b/drivers/infiniband/sw/rxe/rxe.c > index df66f8f9efa1..21ce2d876b42 100644 > --- a/drivers/infiniband/sw/rxe/rxe.c > +++ b/drivers/infiniband/sw/rxe/rxe.c > @@ -110,6 +110,7 @@ static void rxe_init_device_param(struct rxe_dev *rxe) > rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC; > 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; > } > } <...> > +static inline int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > +{ > + return RESPST_ERR_UNSUPPORTED_OPCODE; > +} > #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ You are returning "enum resp_states", while function expects to return "int". You should return -EOPNOTSUPP. > > #endif /* RXE_LOC_H */ > diff --git a/drivers/infiniband/sw/rxe/rxe_mr.c b/drivers/infiniband/sw/rxe/rxe_mr.c > index 93e4b5acd3ac..d40fbe10633f 100644 > --- a/drivers/infiniband/sw/rxe/rxe_mr.c > +++ b/drivers/infiniband/sw/rxe/rxe_mr.c > @@ -547,16 +547,6 @@ int rxe_mr_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > struct page *page; > u64 *va; > > - /* ODP is not supported right now. WIP. */ > - if (mr->umem->is_odp) > - return RESPST_ERR_UNSUPPORTED_OPCODE; > - > - /* See IBA oA19-28 */ > - if (unlikely(mr->state != RXE_MR_STATE_VALID)) { > - rxe_dbg_mr(mr, "mr not in valid state\n"); > - return RESPST_ERR_RKEY_VIOLATION; > - } > - > if (mr->ibmr.type == IB_MR_TYPE_DMA) { > page_offset = iova & (PAGE_SIZE - 1); > page = ib_virt_dma_to_page(iova); > @@ -584,10 +574,8 @@ int rxe_mr_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > } > > va = kmap_local_page(page); > - > /* Do atomic write after all prior operations have completed */ > smp_store_release(&va[page_offset >> 3], value); > - > kunmap_local(va); > > return 0; > diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c > index 9a9aae967486..f3443c604a7f 100644 > --- a/drivers/infiniband/sw/rxe/rxe_odp.c > +++ b/drivers/infiniband/sw/rxe/rxe_odp.c > @@ -378,3 +378,56 @@ int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, > > return 0; > } > + > +#if defined CONFIG_64BIT > +/* only implemented or called for 64 bit architectures */ > +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > +{ > + struct ib_umem_odp *umem_odp = to_ib_umem_odp(mr->umem); > + unsigned int page_offset; > + unsigned long index; > + struct page *page; > + int err; > + u64 *va; > + > + /* See IBA oA19-28 */ > + err = mr_check_range(mr, iova, sizeof(value)); > + if (unlikely(err)) { > + rxe_dbg_mr(mr, "iova out of range\n"); > + return RESPST_ERR_RKEY_VIOLATION; Please don't redefine returned errors. > + } <...> > +#else > +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > +{ > + return RESPST_ERR_UNSUPPORTED_OPCODE; > +} > +#endif You already have empty declaration in rxe_loc.h, use it. Thanks
On Tue, Mar 18, 2025 7:10 PM Leon Romanovsky wrote: > On Tue, Mar 18, 2025 at 06:49:32PM +0900, Daisuke Matsuda wrote: > > <...> > > > +static inline int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > > +{ > > + return RESPST_ERR_UNSUPPORTED_OPCODE; > > +} > > #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ > > You are returning "enum resp_states", while function expects to return "int". You should return -EOPNOTSUPP. Other than my patches, there are some functions that do the same thing. I would like to post a patch to make them consistent, but I think we need reach an agreement on the design of rxe responder before taking up. Please see my opinion below. > > > > > #endif /* RXE_LOC_H */ <...> > > diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c > > index 9a9aae967486..f3443c604a7f 100644 > > --- a/drivers/infiniband/sw/rxe/rxe_odp.c > > +++ b/drivers/infiniband/sw/rxe/rxe_odp.c > > @@ -378,3 +378,56 @@ int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, > > > > return 0; > > } > > + > > +#if defined CONFIG_64BIT > > +/* only implemented or called for 64 bit architectures */ > > +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > > +{ > > + struct ib_umem_odp *umem_odp = to_ib_umem_odp(mr->umem); > > + unsigned int page_offset; > > + unsigned long index; > > + struct page *page; > > + int err; > > + u64 *va; > > + > > + /* See IBA oA19-28 */ > > + err = mr_check_range(mr, iova, sizeof(value)); > > + if (unlikely(err)) { > > + rxe_dbg_mr(mr, "iova out of range\n"); > > + return RESPST_ERR_RKEY_VIOLATION; > > Please don't redefine returned errors. As a general principle, I think your comment is totally correct. The problem is that rxe_receiver(), the responder of rxe, is originally designed as a state machine, and the returned values of "enum resp_states" are used to specify the next state. One thing to note is that rxe_receiver() run solely in workqueue, so the errors generated in the bottom half context are never returned to userspace. In that regard, I think redefining the error codes with different enum values can be justified. The responder using the state machine is easy to understand and maintain. TBH, I am not inclined to change the design, but would like to know the opinions of other people. > > > + } > > <...> > > > +#else > > +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > > +{ > > + return RESPST_ERR_UNSUPPORTED_OPCODE; > > +} > > +#endif > > You already have empty declaration in rxe_loc.h, use it. That's right. I will change it. Thanks, Daisuke > > Thanks
On Wed, Mar 19, 2025 at 02:58:51AM +0000, Daisuke Matsuda (Fujitsu) wrote: > On Tue, Mar 18, 2025 7:10 PM Leon Romanovsky wrote: > > On Tue, Mar 18, 2025 at 06:49:32PM +0900, Daisuke Matsuda wrote: > > > > <...> > > > > > +static inline int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > > > +{ > > > + return RESPST_ERR_UNSUPPORTED_OPCODE; > > > +} > > > #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ > > > > You are returning "enum resp_states", while function expects to return "int". You should return -EOPNOTSUPP. > > Other than my patches, there are some functions that do the same thing. Yes, but you are adding new code and in the new code you should try to have correlated function declaration and returned values. > I would like to post a patch to make them consistent, but I think we need > reach an agreement on the design of rxe responder before taking up. > Please see my opinion below. > > > > > > > > > #endif /* RXE_LOC_H */ > > <...> > > > > diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c > > > index 9a9aae967486..f3443c604a7f 100644 > > > --- a/drivers/infiniband/sw/rxe/rxe_odp.c > > > +++ b/drivers/infiniband/sw/rxe/rxe_odp.c > > > @@ -378,3 +378,56 @@ int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, > > > > > > return 0; > > > } > > > + > > > +#if defined CONFIG_64BIT > > > +/* only implemented or called for 64 bit architectures */ > > > +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > > > +{ > > > + struct ib_umem_odp *umem_odp = to_ib_umem_odp(mr->umem); > > > + unsigned int page_offset; > > > + unsigned long index; > > > + struct page *page; > > > + int err; > > > + u64 *va; > > > + > > > + /* See IBA oA19-28 */ > > > + err = mr_check_range(mr, iova, sizeof(value)); > > > + if (unlikely(err)) { > > > + rxe_dbg_mr(mr, "iova out of range\n"); > > > + return RESPST_ERR_RKEY_VIOLATION; > > > > Please don't redefine returned errors. > > As a general principle, I think your comment is totally correct. > The problem is that rxe_receiver(), the responder of rxe, is originally designed > as a state machine, and the returned values of "enum resp_states" are used > to specify the next state. > > One thing to note is that rxe_receiver() run solely in workqueue, so the errors > generated in the bottom half context are never returned to userspace. In that regard, > I think redefining the error codes with different enum values can be justified. In places where rxe_odp_do_atomic_write() respond is important, you can write something like: err = rxe_odp_do_atomic_write(...) if (err == -EPERM) state = RESPST_ERR_RKEY_VIOLATION ... or declare rxe_odp_do_atomic_write() to return enum resp_state. Thanks
Hi Leon, Thank you for taking a look. I've submitted v3 patches to address your comment. I will also work on rechecking the inconsistency in whole rxe driver after the patches are merged and for-next is rebased. Thanks, Daisuke On Wed, Mar 19, 2025 5:58 PM Leon Romanovsky wrote: > On Wed, Mar 19, 2025 at 02:58:51AM +0000, Daisuke Matsuda (Fujitsu) wrote: > > On Tue, Mar 18, 2025 7:10 PM Leon Romanovsky wrote: > > > On Tue, Mar 18, 2025 at 06:49:32PM +0900, Daisuke Matsuda wrote: > > > > > > <...> > > > > > > > +static inline int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > > > > +{ > > > > + return RESPST_ERR_UNSUPPORTED_OPCODE; > > > > +} > > > > #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ > > > > > > You are returning "enum resp_states", while function expects to return "int". You should return -EOPNOTSUPP. > > > > Other than my patches, there are some functions that do the same thing. > > Yes, but you are adding new code and in the new code you should try to > have correlated function declaration and returned values. > > > I would like to post a patch to make them consistent, but I think we need > > reach an agreement on the design of rxe responder before taking up. > > Please see my opinion below. > > > > > > > > > > > > > #endif /* RXE_LOC_H */ > > > > <...> > > > > > > diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c > > > > index 9a9aae967486..f3443c604a7f 100644 > > > > --- a/drivers/infiniband/sw/rxe/rxe_odp.c > > > > +++ b/drivers/infiniband/sw/rxe/rxe_odp.c > > > > @@ -378,3 +378,56 @@ int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, > > > > > > > > return 0; > > > > } > > > > + > > > > +#if defined CONFIG_64BIT > > > > +/* only implemented or called for 64 bit architectures */ > > > > +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) > > > > +{ > > > > + struct ib_umem_odp *umem_odp = to_ib_umem_odp(mr->umem); > > > > + unsigned int page_offset; > > > > + unsigned long index; > > > > + struct page *page; > > > > + int err; > > > > + u64 *va; > > > > + > > > > + /* See IBA oA19-28 */ > > > > + err = mr_check_range(mr, iova, sizeof(value)); > > > > + if (unlikely(err)) { > > > > + rxe_dbg_mr(mr, "iova out of range\n"); > > > > + return RESPST_ERR_RKEY_VIOLATION; > > > > > > Please don't redefine returned errors. > > > > As a general principle, I think your comment is totally correct. > > The problem is that rxe_receiver(), the responder of rxe, is originally designed > > as a state machine, and the returned values of "enum resp_states" are used > > to specify the next state. > > > > One thing to note is that rxe_receiver() run solely in workqueue, so the errors > > generated in the bottom half context are never returned to userspace. In that regard, > > I think redefining the error codes with different enum values can be justified. > > In places where rxe_odp_do_atomic_write() respond is important, you can > write something like: > err = rxe_odp_do_atomic_write(...) > if (err == -EPERM) > state = RESPST_ERR_RKEY_VIOLATION > ... > > or declare rxe_odp_do_atomic_write() to return enum resp_state. > > Thanks
diff --git a/drivers/infiniband/sw/rxe/rxe.c b/drivers/infiniband/sw/rxe/rxe.c index df66f8f9efa1..21ce2d876b42 100644 --- a/drivers/infiniband/sw/rxe/rxe.c +++ b/drivers/infiniband/sw/rxe/rxe.c @@ -110,6 +110,7 @@ static void rxe_init_device_param(struct rxe_dev *rxe) rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC; 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; } } diff --git a/drivers/infiniband/sw/rxe/rxe_loc.h b/drivers/infiniband/sw/rxe/rxe_loc.h index 0012bebe96ef..8b1517c0894c 100644 --- a/drivers/infiniband/sw/rxe/rxe_loc.h +++ b/drivers/infiniband/sw/rxe/rxe_loc.h @@ -196,6 +196,7 @@ int rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode, u64 compare, u64 swap_add, u64 *orig_val); int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, unsigned int length); +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value); #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int rxe_odp_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova, @@ -219,6 +220,10 @@ static inline int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, { return -EOPNOTSUPP; } +static inline int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) +{ + return RESPST_ERR_UNSUPPORTED_OPCODE; +} #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ #endif /* RXE_LOC_H */ diff --git a/drivers/infiniband/sw/rxe/rxe_mr.c b/drivers/infiniband/sw/rxe/rxe_mr.c index 93e4b5acd3ac..d40fbe10633f 100644 --- a/drivers/infiniband/sw/rxe/rxe_mr.c +++ b/drivers/infiniband/sw/rxe/rxe_mr.c @@ -547,16 +547,6 @@ int rxe_mr_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) struct page *page; u64 *va; - /* ODP is not supported right now. WIP. */ - if (mr->umem->is_odp) - return RESPST_ERR_UNSUPPORTED_OPCODE; - - /* See IBA oA19-28 */ - if (unlikely(mr->state != RXE_MR_STATE_VALID)) { - rxe_dbg_mr(mr, "mr not in valid state\n"); - return RESPST_ERR_RKEY_VIOLATION; - } - if (mr->ibmr.type == IB_MR_TYPE_DMA) { page_offset = iova & (PAGE_SIZE - 1); page = ib_virt_dma_to_page(iova); @@ -584,10 +574,8 @@ int rxe_mr_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) } va = kmap_local_page(page); - /* Do atomic write after all prior operations have completed */ smp_store_release(&va[page_offset >> 3], value); - kunmap_local(va); return 0; diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c index 9a9aae967486..f3443c604a7f 100644 --- a/drivers/infiniband/sw/rxe/rxe_odp.c +++ b/drivers/infiniband/sw/rxe/rxe_odp.c @@ -378,3 +378,56 @@ int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, return 0; } + +#if defined CONFIG_64BIT +/* only implemented or called for 64 bit architectures */ +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) +{ + struct ib_umem_odp *umem_odp = to_ib_umem_odp(mr->umem); + unsigned int page_offset; + unsigned long index; + struct page *page; + int err; + u64 *va; + + /* See IBA oA19-28 */ + err = mr_check_range(mr, iova, sizeof(value)); + if (unlikely(err)) { + rxe_dbg_mr(mr, "iova out of range\n"); + return RESPST_ERR_RKEY_VIOLATION; + } + + err = rxe_odp_map_range_and_lock(mr, iova, sizeof(value), + RXE_PAGEFAULT_DEFAULT); + if (err) + return RESPST_ERR_RKEY_VIOLATION; + + page_offset = rxe_odp_iova_to_page_offset(umem_odp, iova); + index = rxe_odp_iova_to_index(umem_odp, iova); + page = hmm_pfn_to_page(umem_odp->pfn_list[index]); + if (!page) { + mutex_unlock(&umem_odp->umem_mutex); + return RESPST_ERR_RKEY_VIOLATION; + } + /* See IBA A19.4.2 */ + if (unlikely(page_offset & 0x7)) { + mutex_unlock(&umem_odp->umem_mutex); + rxe_dbg_mr(mr, "misaligned address\n"); + return RESPST_ERR_MISALIGNED_ATOMIC; + } + + va = kmap_local_page(page); + /* Do atomic write after all prior operations have completed */ + smp_store_release(&va[page_offset >> 3], value); + kunmap_local(va); + + mutex_unlock(&umem_odp->umem_mutex); + + return 0; +} +#else +int rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value) +{ + return RESPST_ERR_UNSUPPORTED_OPCODE; +} +#endif diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c b/drivers/infiniband/sw/rxe/rxe_resp.c index 304e3de740ad..fd7bac5bce18 100644 --- a/drivers/infiniband/sw/rxe/rxe_resp.c +++ b/drivers/infiniband/sw/rxe/rxe_resp.c @@ -749,7 +749,16 @@ static enum resp_states atomic_write_reply(struct rxe_qp *qp, value = *(u64 *)payload_addr(pkt); iova = qp->resp.va + qp->resp.offset; - err = rxe_mr_do_atomic_write(mr, iova, value); + /* See IBA oA19-28 */ + if (unlikely(mr->state != RXE_MR_STATE_VALID)) { + rxe_dbg_mr(mr, "mr not in valid state\n"); + return RESPST_ERR_RKEY_VIOLATION; + } + + if (mr->umem->is_odp) + err = rxe_odp_do_atomic_write(mr, iova, value); + else + err = rxe_mr_do_atomic_write(mr, iova, value); if (err) return err; diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index da07d3e2db1d..bfa1bff3c720 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -326,6 +326,7 @@ enum ib_odp_transport_cap_bits { IB_ODP_SUPPORT_ATOMIC = 1 << 4, IB_ODP_SUPPORT_SRQ_RECV = 1 << 5, IB_ODP_SUPPORT_FLUSH = 1 << 6, + IB_ODP_SUPPORT_ATOMIC_WRITE = 1 << 7, }; struct ib_odp_caps {
Add rxe_odp_do_atomic_write() so that ODP specific steps are applied to ATOMIC WRITE requests. Signed-off-by: Daisuke Matsuda <matsuda-daisuke@fujitsu.com> --- drivers/infiniband/sw/rxe/rxe.c | 1 + drivers/infiniband/sw/rxe/rxe_loc.h | 5 +++ drivers/infiniband/sw/rxe/rxe_mr.c | 12 ------- drivers/infiniband/sw/rxe/rxe_odp.c | 53 ++++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_resp.c | 11 +++++- include/rdma/ib_verbs.h | 1 + 6 files changed, 70 insertions(+), 13 deletions(-)