Replace AV by AH in UD sends
Subsystems (2)
v5 cover Replace AV by AH in UD sends
Currently the rdma_rxe driver and its user space provider exchange addressing information for UD sends by having the provider compute an address vector (AV) and send it with each WQE. This is not the way that the RDMA verbs API was intended to operate.
This series of patches modifies the way UD send WQEs work by exchanging an index identifying the AH replacing the 88 byte AV by a 4 byte AH index. In order to not break compatibility with the existing API the rdma_rxe driver will recognise when an older version of the provider is not sending an index (i.e. it is 0) and will use the AV instead.
This series of patches is identical to the previous version but rebased to 5.15.0-rc2+. It applies cleanly to
commit: d30ef6d5c013c19e907f2a3a3d6eee04fcd3de0d (for-next)
---
v5:
Rebase to 5.15.0-rc2+
v4:
Rebase to 5.15.0-rc1+
v3:
Split up commits into smaller steps.
v2:
Rearranged AV in rxe_send_wqe to be in the ud struct but padded to the same offset as the original preserving ABI compatibility.
Bob Pearson (6):
RDMA/rxe: Move AV from rxe_send_wqe to rxe_send_wr RDMA/rxe: Change AH objects to indexed RDMA/rxe: Create AH index and return to user space RDMA/rxe: Replace ah->pd by ah->ibah.pd RDMA/rxe: Lookup kernel AH from ah index in UD WQEs RDMA/rxe: Convert kernel UD post send to use ah_num
drivers/infiniband/sw/rxe/rxe_av.c | 20 +++++++++++++- drivers/infiniband/sw/rxe/rxe_param.h | 4 ++- drivers/infiniband/sw/rxe/rxe_pool.c | 4 ++- drivers/infiniband/sw/rxe/rxe_req.c | 8 +++--- drivers/infiniband/sw/rxe/rxe_verbs.c | 39 ++++++++++++++++++++++----- drivers/infiniband/sw/rxe/rxe_verbs.h | 8 +++++- include/uapi/rdma/rdma_user_rxe.h | 10 ++++++- 7 files changed, 79 insertions(+), 14 deletions(-)
Signature · 22 lines
2.30.2 Bob Pearson (6): RDMA/rxe: Move AV from rxe_send_wqe to rxe_send_wr RDMA/rxe: Change AH objects to indexed RDMA/rxe: Create AH index and return to user space RDMA/rxe: Replace ah->pd by ah->ibah.pd RDMA/rxe: Lookup kernel AH from ah index in UD WQEs RDMA/rxe: Convert kernel UD post send to use ah_num drivers/infiniband/sw/rxe/rxe_av.c | 20 +++++++++++++- drivers/infiniband/sw/rxe/rxe_param.h | 4 ++- drivers/infiniband/sw/rxe/rxe_pool.c | 4 ++- drivers/infiniband/sw/rxe/rxe_req.c | 8 +++--- drivers/infiniband/sw/rxe/rxe_verbs.c | 39 ++++++++++++++++++++++----- drivers/infiniband/sw/rxe/rxe_verbs.h | 8 +++++- include/uapi/rdma/rdma_user_rxe.h | 10 ++++++- 7 files changed, 79 insertions(+), 14 deletions(-) -- 2.30.2
PATCH v5 1/6 RDMA/rxe: Move AV from rxe_send_wqe to rxe_send_wr
Move the struct rxe_av av from struct rxe_send_wqe to struct rxe_send_wr placing it in wr.ud at the same offset as it was previously. This has the effect of increasing the size of struct rxe_send_wr while keeping the size of struct rxe_send_wqe the same. This better reflects the use of this field which is only used for UD sends. This change has no effect on ABI compatibility so the modified rxe driver will operate with older versions of rdma-core.
--- drivers/infiniband/sw/rxe/rxe_av.c | 2 +- drivers/infiniband/sw/rxe/rxe_verbs.c | 3 ++- include/uapi/rdma/rdma_user_rxe.h | 4 +++- 3 files changed, 6 insertions(+), 3 deletions(-)
drivers/infiniband/sw/rxe/rxe_av.c +1 -1
@@ -107,5 +107,5 @@ struct rxe_av *rxe_get_av(struct rxe_pkt_info *pkt)if (qp_type (pkt->qp) ==IB_QPT_RC ||qp_type (pkt->qp) ==IB_QPT_UC )return &pkt->qp->pri_av; -return (pkt->wqe) ? &pkt->wqe->av :NULL ; +return (pkt->wqe) ? &pkt->wqe->wr.wr.ud.av :NULL ; }
drivers/infiniband/sw/rxe/rxe_verbs.c +2 -1
@@ -584,7 +584,8 @@ static void init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,if (qp_type (qp) ==IB_QPT_UD ||qp_type (qp) ==IB_QPT_SMI ||qp_type (qp) ==IB_QPT_GSI ) -memcpy (&wqe->av, &to_rah (ud_wr (ibwr)->ah)->av,sizeof (wqe->av)); +memcpy (&wqe->wr.wr.ud.av, &to_rah (ud_wr (ibwr)->ah)->av, +sizeof (struct rxe_av ));if (unlikely (ibwr->send_flags &IB_SEND_INLINE ))copy_inline_data_to_wqe (wqe, ibwr);
include/uapi/rdma/rdma_user_rxe.h +3 -1
@@ -98,6 +98,9 @@ struct rxe_send_wr {__u32 remote_qpn;__u32 remote_qkey;__u16 pkey_index; +__u16 reserved; +__u32 pad[5 ]; +struct rxe_av av; } ud;struct {__aligned_u64 addr;
@@ -148,7 +151,6 @@ struct rxe_dma_info {struct rxe_send_wqe {struct rxe_send_wr wr; -struct rxe_av av;__u32 status;__u32 state;__aligned_u64 iova;
Signature · 1 line
2.30.2
PATCH v5 2/6 RDMA/rxe: Change AH objects to indexed
In reply to Bob PearsonMake changes to rxe_param.h and rxe_pool.c to allow indexing of AH objects. Valid indices are non-zero so older providers can be detected.
--- drivers/infiniband/sw/rxe/rxe_param.h | 4 +++- drivers/infiniband/sw/rxe/rxe_pool.c | 4 +++- 2 files changed, 6 insertions(+), 2 deletions(-)
drivers/infiniband/sw/rxe/rxe_param.h +3 -1
@@ -67,7 +67,9 @@ enum rxe_device_param {RXE_MAX_MCAST_GRP =8192 ,RXE_MAX_MCAST_QP_ATTACH =56 ,RXE_MAX_TOT_MCAST_QP_ATTACH =0x70000 , -RXE_MAX_AH =100 , +RXE_MAX_AH =16383 , +RXE_MIN_AH_INDEX =1 , +RXE_MAX_AH_INDEX =16383 ,RXE_MAX_SRQ_WR =0x4000 ,RXE_MIN_SRQ_WR =1 , RXE_MAX_SRQ_SGE = 27,
drivers/infiniband/sw/rxe/rxe_pool.c +3 -1
@@ -26,7 +26,9 @@ struct rxe_type_info rxe_type_info[RXE_NUM_TYPES] = { .name = "rxe-ah", .size = sizeof(struct rxe_ah), .elem_offset = offsetof(struct rxe_ah, pelem), - .flags = RXE_POOL_NO_ALLOC, + .flags = RXE_POOL_INDEX | RXE_POOL_NO_ALLOC, + .min_index = RXE_MIN_AH_INDEX, + .max_index = RXE_MAX_AH_INDEX, }, [RXE_TYPE_SRQ] = { .name = "rxe-srq",
Signature · 1 line
2.30.2
PATCH v5 3/6 RDMA/rxe: Create AH index and return to user space
In reply to Bob PearsonMake changes to rdma_user_rxe.h to allow indexing AH objects, passing the index in UD send WRs to the driver and returning the index to the rxe provider.
Modify rxe_create_ah() to add an index to AH when created and if called from a new user provider return it to user space. If called from an old provider mark the AH as not having a useful index. Modify rxe_destroy_ah to drop the index before deleting the object.
--- drivers/infiniband/sw/rxe/rxe_verbs.c | 31 ++++++++++++++++++++++++++- drivers/infiniband/sw/rxe/rxe_verbs.h | 2 ++ include/uapi/rdma/rdma_user_rxe.h | 8 ++++++- 3 files changed, 39 insertions(+), 2 deletions(-)
drivers/infiniband/sw/rxe/rxe_verbs.c +30 -1
@@ -158,9 +158,19 @@ static int rxe_create_ah(struct ib_ah *ibah, struct ib_udata *udata) { -int err;struct rxe_dev *rxe =to_rdev (ibah->device);struct rxe_ah *ah =to_rah (ibah); +struct rxe_create_ah_resp __user *uresp =NULL ; +int err; + +if (udata) { +/* test if new user provider */ +if (udata->outlen >=sizeof (*uresp)) + uresp = udata->outbuf; + ah->is_user = true; + }else { + ah->is_user = false; + } err =rxe_av_chk_attr (rxe, init_attr->ah_attr);if (err)
@@ -170,6 +180,24 @@ static int rxe_create_ah(struct ib_ah *ibah,if (err)return err; +/* create index > 0 */ +rxe_add_index (ah); + ah->ah_num = ah->pelem.index; + +if (uresp) { +/* only if new user provider */ + err =copy_to_user (&uresp->ah_num, &ah->ah_num, +sizeof (uresp->ah_num)); +if (err) { +rxe_drop_index (ah); +rxe_drop_ref (ah); +return -EFAULT ; + } + }else if (ah->is_user) { +/* only if old user provider */ + ah->ah_num =0 ; + } +rxe_init_av (init_attr->ah_attr, &ah->av);return 0 ; }
@@ -202,6 +230,7 @@ static int rxe_destroy_ah(struct ib_ah *ibah, u32 flags) {struct rxe_ah *ah =to_rah (ibah); +rxe_drop_index (ah);rxe_drop_ref (ah);return 0 ; }
drivers/infiniband/sw/rxe/rxe_verbs.h +2 -0
@@ -48,6 +48,8 @@ struct rxe_ah {struct rxe_pool_entry pelem;struct rxe_pd *pd;struct rxe_av av; +bool is_user; +int ah_num; };struct rxe_cqe {
include/uapi/rdma/rdma_user_rxe.h +7 -1
@@ -99,7 +99,8 @@ struct rxe_send_wr {__u32 remote_qkey;__u16 pkey_index;__u16 reserved; -__u32 pad[5 ]; +__u32 ah_num; +__u32 pad[4 ];struct rxe_av av; } ud;struct {
@@ -170,6 +171,11 @@ struct rxe_recv_wqe {struct rxe_dma_info dma; }; +struct rxe_create_ah_resp { +__u32 ah_num; +__u32 reserved; +}; +struct rxe_create_cq_resp {struct mminfo mi; };
Signature · 1 line
2.30.2
PATCH v5 4/6 RDMA/rxe: Replace ah->pd by ah->ibah.pd
In reply to Bob PearsonThe pd field in struct rxe_ah is redundant with the pd field in the rdma-core's ib_ah. Eliminate the pd field in rxe_ah and add an inline to extract the pd from the ibah field.
--- drivers/infiniband/sw/rxe/rxe_verbs.h | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/infiniband/sw/rxe/rxe_verbs.h +5 -1
@@ -46,7 +46,6 @@ struct rxe_pd {struct rxe_ah {struct ib_ah ibah;struct rxe_pool_entry pelem; -struct rxe_pd *pd;struct rxe_av av;bool is_user;int ah_num;
@@ -471,6 +470,11 @@ static inline struct rxe_mw *to_rmw(struct ib_mw *mw) return mw ? container_of(mw, struct rxe_mw, ibmw) : NULL; } +static inline struct rxe_pd *rxe_ah_pd (struct rxe_ah *ah) +{ +return to_rpd (ah->ibah.pd); +} +static inline struct rxe_pd *mr_pd (struct rxe_mr *mr) {return to_rpd (mr->ibmr.pd);
Signature · 1 line
2.30.2
PATCH v5 5/6 RDMA/rxe: Lookup kernel AH from ah index in UD WQEs
In reply to Bob PearsonAdd code to rxe_get_av in rxe_av.c to use the AH index in UD send WQEs to lookup the kernel AH. For old user providers continue to use the AV passed in WQEs. Move setting pkt->rxe to before the call to rxe_get_av() to get access to the AH pool.
--- drivers/infiniband/sw/rxe/rxe_av.c | 20 +++++++++++++++++++- drivers/infiniband/sw/rxe/rxe_req.c | 8 +++++--- 2 files changed, 24 insertions(+), 4 deletions(-)
drivers/infiniband/sw/rxe/rxe_av.c +19 -1
@@ -101,11 +101,29 @@ void rxe_av_fill_ip_info(struct rxe_av *av, struct rdma_ah_attr *attr)struct rxe_av *rxe_get_av (struct rxe_pkt_info *pkt) { +struct rxe_ah *ah; +u32 ah_num; +if (!pkt || !pkt->qp)return NULL ;if (qp_type (pkt->qp) ==IB_QPT_RC ||qp_type (pkt->qp) ==IB_QPT_UC )return &pkt->qp->pri_av; -return (pkt->wqe) ? &pkt->wqe->wr.wr.ud.av :NULL ; +if (!pkt->wqe) +return NULL ; + + ah_num = pkt->wqe->wr.wr.ud.ah_num; +if (ah_num) { +/* only new user provider or kernel client */ + ah =rxe_pool_get_index (&pkt->rxe->ah_pool, ah_num); +if (!ah || ah->ah_num != ah_num ||rxe_ah_pd (ah) != pkt->qp->pd) { +pr_warn ("Unable to find AH matching ah_num\n" ); +return NULL ; + } +return &ah->av; + } + +/* only old user provider for UD sends*/ +return &pkt->wqe->wr.wr.ud.av; }
drivers/infiniband/sw/rxe/rxe_req.c +5 -3
@@ -379,9 +379,8 @@ static struct sk_buff *init_req_packet(struct rxe_qp *qp,/* length from start of bth to end of icrc */ paylen = rxe_opcode[opcode].length + payload + pad +RXE_ICRC_SIZE ; -/* pkt->hdr, rxe, port_num and mask are initialized in ifc -* layer -*/ +/* pkt->hdr, port_num and mask are initialized in ifc layer */ + pkt->rxe = rxe; pkt->opcode = opcode; pkt->qp = qp; pkt->psn = qp->req.psn;
@@ -391,6 +390,9 @@ static struct sk_buff *init_req_packet(struct rxe_qp *qp,/* init skb */ av =rxe_get_av (pkt); +if (!av) +return NULL ; + skb =rxe_init_packet (rxe, av, paylen, pkt);if (unlikely (!skb))return NULL ;
Signature · 1 line
2.30.2
Zhu Yanjun
maintainerRe: [PATCH for-next v5 5/6] RDMA/rxe: Lookup kernel AH from ah index in UD WQEs
Quoted 26 lines · Bob Pearson
Add code to
rxe_get_avinrxe_av.cto use the AH index in UD send WQEs to lookup the kernel AH. For old user providers continue to use the AV passed in WQEs. Move settingpkt->rxeto before the call torxe_get_av()to get access to the AH pool.Signed-off-by Bob Pearson <[email protected]>--- drivers/infiniband/sw/rxe/rxe_av.c | 20 +++++++++++++++++++- drivers/infiniband/sw/rxe/rxe_req.c | 8 +++++--- 2 files changed, 24 insertions(+), 4 deletions(-)drivers/infiniband/sw/rxe/rxe_av.c +3
-0@@ -101,11 +101,29 @@ void rxe_av_fill_ip_info(struct rxe_av *av, struct rdma_ah_attr *attr)struct rxe_av *rxe_get_av (struct rxe_pkt_info *pkt) { +struct rxe_ah *ah; +u32 ah_num; +if (!pkt || !pkt->qp)return NULL ;
drivers/infiniband/sw/rxe/rxe_av.c void rxe_av_fill_ip_info(struct rxe_av *av, struct rdma_ah_attr *attr) · line 118
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_av.c +6
-1if (qp_type (pkt->qp) ==IB_QPT_RC ||qp_type (pkt->qp) ==IB_QPT_UC )return &pkt->qp->pri_av; -return (pkt->wqe) ? &pkt->wqe->wr.wr.ud.av :NULL ; +if (!pkt->wqe) +return NULL ; + + ah_num = pkt->wqe->wr.wr.ud.ah_num; +if (ah_num) { +/* only new user provider or kernel client */
struct rxe_ah *ah;
ah is only used in this snippet. Is it better to move to here? It is only a trivial problem.
Zhu Yanjun
Quoted 40 lines
drivers/infiniband/sw/rxe/rxe_av.c +10
-0+ ah =rxe_pool_get_index (&pkt->rxe->ah_pool, ah_num); +if (!ah || ah->ah_num != ah_num ||rxe_ah_pd (ah) != pkt->qp->pd) { +pr_warn ("Unable to find AH matching ah_num\n" ); +return NULL ; + } +return &ah->av; + } + +/* only old user provider for UD sends*/ +return &pkt->wqe->wr.wr.ud.av; }drivers/infiniband/sw/rxe/rxe_req.c +5
-4@@ -379,9 +379,8 @@ static struct sk_buff *init_req_packet(struct rxe_qp *qp,/* length from start of bth to end of icrc */ paylen = rxe_opcode[opcode].length + payload + pad +RXE_ICRC_SIZE ; -/* pkt->hdr, rxe, port_num and mask are initialized in ifc -* layer -*/ +/* pkt->hdr, port_num and mask are initialized in ifc layer */ + pkt->rxe = rxe; pkt->opcode = opcode; pkt->qp = qp; pkt->psn = qp->req.psn;@@ -391,6 +390,9 @@ static struct sk_buff *init_req_packet(struct rxe_qp *qp,/* init skb */ av =rxe_get_av (pkt); +if (!av) +return NULL ; + skb =rxe_init_packet (rxe, av, paylen, pkt);if (unlikely (!skb))return NULL ; --2.30.2
RE: [PATCH for-next v5 5/6] RDMA/rxe: Lookup kernel AH from ah index in UD WQEs
Zhu,
It's a matter of preference. I find that for me always putting all the local variables at the top of a subroutine saves time and reduces bugs. I know where to look. They're always there. And there are no tricky scope issues to think about. If you can't see them because they are off the screen the subroutine is probably too big.
BTW do you have a new email address? I just saw one go by.
Bob
-----Original Message-----
From: Zhu Yanjun <[email protected]>
Sent: Wednesday, October 6, 2021 6:56 AM
To: Bob Pearson <[email protected]>
Cc: Jason Gunthorpe <[email protected]>; RDMA mailing list <[email protected]>
Subject: Re: [PATCH for-next v5 5/6] RDMA/rxe: Lookup kernel AH from ah index in UD WQEs
Quoted 27 lines · Bob Pearson
Add code to
rxe_get_avinrxe_av.cto use the AH index in UD send WQEs to lookup the kernel AH. For old user providers continue to use the AV passed in WQEs. Move settingpkt->rxeto before the call torxe_get_av()to get access to the AH pool.Signed-off-by Bob Pearson <[email protected]>--- drivers/infiniband/sw/rxe/rxe_av.c | 20 +++++++++++++++++++- drivers/infiniband/sw/rxe/rxe_req.c | 8 +++++--- 2 files changed, 24 insertions(+), 4 deletions(-)drivers/infiniband/sw/rxe/rxe_av.c +3
-0@@ -101,11 +101,29 @@ void rxe_av_fill_ip_info(struct rxe_av *av, struct rdma_ah_attr *attr)struct rxe_av *rxe_get_av (struct rxe_pkt_info *pkt) { +struct rxe_ah *ah; +u32 ah_num; +if (!pkt || !pkt->qp)return NULL ;
drivers/infiniband/sw/rxe/rxe_av.c void rxe_av_fill_ip_info(struct rxe_av *av, · line 118
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_av.c +6
-1if (qp_type (pkt->qp) ==IB_QPT_RC ||qp_type (pkt->qp) ==IB_QPT_UC )return &pkt->qp->pri_av; -return (pkt->wqe) ? &pkt->wqe->wr.wr.ud.av :NULL ; +if (!pkt->wqe) +return NULL ; + + ah_num = pkt->wqe->wr.wr.ud.ah_num; +if (ah_num) { +/* only new user provider or kernel client */
struct rxe_ah *ah;
ah is only used in this snippet. Is it better to move to here? It is only a trivial problem.
Zhu Yanjun
Quoted 43 lines
drivers/infiniband/sw/rxe/rxe_av.c +10
-0+ ah =rxe_pool_get_index (&pkt->rxe->ah_pool, ah_num); +if (!ah || ah->ah_num != ah_num ||rxe_ah_pd (ah) != pkt->qp->pd) { +pr_warn ("Unable to find AH matching ah_num\n" ); +return NULL ; + } +return &ah->av; + } + +/* only old user provider for UD sends*/ +return &pkt->wqe->wr.wr.ud.av; }drivers/infiniband/sw/rxe/rxe_req.c +2
-3@@ -379,9 +379,8 @@ static struct sk_buff *init_req_packet(struct rxe_qp *qp,/* length from start of bth to end of icrc */ paylen = rxe_opcode[opcode].length + payload + pad +RXE_ICRC_SIZE ; -/* pkt->hdr, rxe, port_num and mask are initialized in ifc -* layer -*/ +/* pkt->hdr, port_num and mask are initialized in ifc layer */ + pkt->rxe = rxe; pkt->opcode = opcode; pkt->qp = qp;pkt->psn = qp->req.psn;drivers/infiniband/sw/rxe/rxe_req.c +3
-0@@ -391,6 +390,9 @@ static struct sk_buff *init_req_packet(structrxe_qp *qp,/* init skb */ av =rxe_get_av (pkt); +if (!av) +return NULL ; + skb =rxe_init_packet (rxe, av, paylen, pkt);if (unlikely (!skb))return NULL;--
2.30.2
Zhu Yanjun
maintainerRe: [PATCH for-next v5 5/6] RDMA/rxe: Lookup kernel AH from ah index in UD WQEs
Quoted 5 lines · Robert B
Zhu,
It's a matter of preference. I find that for me always putting all the local variables at the top of a subroutine saves time and reduces bugs. I know where to look. They're always there. And there are no tricky scope issues to think about. If you can't see them because they are off the screen the subroutine is probably too big.
Yeah. It is a matter of preference. I like to put all the variables near where they are used.
Do not worry. I am fine with your preference.
Zhu Yanjun
Quoted 98 lines
BTW do you have a new email address? I just saw one go by.
Bob
-----Original Message-----
From: Zhu Yanjun <[email protected]>
Sent: Wednesday, October 6, 2021 6:56 AM
To: Bob Pearson <[email protected]>
Cc: Jason Gunthorpe <[email protected]>; RDMA mailing list <[email protected]>
Subject: Re: [PATCH for-next v5 5/6] RDMA/rxe: Lookup kernel AH from ah index in UD WQEsQuoted 27 lines · Bob Pearson
Add code to
rxe_get_avinrxe_av.cto use the AH index in UD send WQEs to lookup the kernel AH. For old user providers continue to use the AV passed in WQEs. Move settingpkt->rxeto before the call torxe_get_av()to get access to the AH pool.Signed-off-by Bob Pearson <[email protected]>--- drivers/infiniband/sw/rxe/rxe_av.c | 20 +++++++++++++++++++- drivers/infiniband/sw/rxe/rxe_req.c | 8 +++++--- 2 files changed, 24 insertions(+), 4 deletions(-)drivers/infiniband/sw/rxe/rxe_av.c +3
-0@@ -101,11 +101,29 @@ void rxe_av_fill_ip_info(struct rxe_av *av, struct rdma_ah_attr *attr)struct rxe_av *rxe_get_av (struct rxe_pkt_info *pkt) { +struct rxe_ah *ah; +u32 ah_num; +if (!pkt || !pkt->qp)return NULL ;drivers/infiniband/sw/rxe/rxe_av.c void rxe_av_fill_ip_info(struct rxe_av *av, · line 118
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_av.c +6
-1if (qp_type (pkt->qp) ==IB_QPT_RC ||qp_type (pkt->qp) ==IB_QPT_UC )return &pkt->qp->pri_av; -return (pkt->wqe) ? &pkt->wqe->wr.wr.ud.av :NULL ; +if (!pkt->wqe) +return NULL ; + + ah_num = pkt->wqe->wr.wr.ud.ah_num; +if (ah_num) { +/* only new user provider or kernel client */ struct
rxe_ah*ah;
ah is only used in this snippet. Is it better to move to here? It is only a trivial problem.Zhu Yanjun
Quoted 43 lines
drivers/infiniband/sw/rxe/rxe_av.c +10
-0+ ah =rxe_pool_get_index (&pkt->rxe->ah_pool, ah_num); +if (!ah || ah->ah_num != ah_num ||rxe_ah_pd (ah) != pkt->qp->pd) { +pr_warn ("Unable to find AH matching ah_num\n" ); +return NULL ; + } +return &ah->av; + } + +/* only old user provider for UD sends*/ +return &pkt->wqe->wr.wr.ud.av; }drivers/infiniband/sw/rxe/rxe_req.c +2
-3@@ -379,9 +379,8 @@ static struct sk_buff *init_req_packet(struct rxe_qp *qp,/* length from start of bth to end of icrc */ paylen = rxe_opcode[opcode].length + payload + pad +RXE_ICRC_SIZE ; -/* pkt->hdr, rxe, port_num and mask are initialized in ifc -* layer -*/ +/* pkt->hdr, port_num and mask are initialized in ifc layer */ + pkt->rxe = rxe; pkt->opcode = opcode; pkt->qp = qp;pkt->psn = qp->req.psn;drivers/infiniband/sw/rxe/rxe_req.c +3
-0@@ -391,6 +390,9 @@ static struct sk_buff *init_req_packet(structrxe_qp *qp,/* init skb */ av =rxe_get_av (pkt); +if (!av) +return NULL ; + skb =rxe_init_packet (rxe, av, paylen, pkt);if (unlikely (!skb))return NULL;--
2.30.2
PATCH v5 6/6 RDMA/rxe: Convert kernel UD post send to use ah_num
In reply to Bob PearsonModify ib_post_send for kernel UD sends to put the AH index into the WQE instead of the address vector.
--- drivers/infiniband/sw/rxe/rxe_verbs.c | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-)
drivers/infiniband/sw/rxe/rxe_verbs.c +3 -6
@@ -537,8 +537,11 @@ static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr,if (qp_type (qp) ==IB_QPT_UD ||qp_type (qp) ==IB_QPT_SMI ||qp_type (qp) ==IB_QPT_GSI ) { +struct ib_ah *ibah =ud_wr (ibwr)->ah; + wr->wr.ud.remote_qpn =ud_wr (ibwr)->remote_qpn; wr->wr.ud.remote_qkey =ud_wr (ibwr)->remote_qkey; + wr->wr.ud.ah_num =to_rah (ibah)->ah_num;if (qp_type (qp) ==IB_QPT_GSI ) wr->wr.ud.pkey_index =ud_wr (ibwr)->pkey_index;if (wr->opcode ==IB_WR_SEND_WITH_IMM )
@@ -610,12 +613,6 @@ static void init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,return ; } -if (qp_type (qp) ==IB_QPT_UD || -qp_type (qp) ==IB_QPT_SMI || -qp_type (qp) ==IB_QPT_GSI ) -memcpy (&wqe->wr.wr.ud.av, &to_rah (ud_wr (ibwr)->ah)->av, -sizeof (struct rxe_av )); -if (unlikely (ibwr->send_flags &IB_SEND_INLINE ))copy_inline_data_to_wqe (wqe, ibwr);else
Signature · 1 line
2.30.2
Jason Gunthorpe
maintainerQuoted 11 lines · Bob Pearson
Currently the
rdma_rxedriver and its user space provider exchange addressing information for UD sends by having the provider compute an address vector (AV) and send it with each WQE. This is not the way that the RDMA verbs API was intended to operate.This series of patches modifies the way UD send WQEs work by exchanging an index identifying the AH replacing the 88 byte AV by a 4 byte AH index. In order to not break compatibility with the existing API the
rdma_rxedriver will recognise when an older version of the provider is not sending an index (i.e. it is 0) and will use the AV instead.
Quoted 7 lines
This series of patches is identical to the previous version but rebased to 5.15.0-rc2+. It applies cleanly to
commit: d30ef6d5c013c19e907f2a3a3d6eee04fcd3de0d (for-next)v5:
Rebase to 5.15.0-rc2+
This is not the right base, I said you needed something path Rao's patch like current rdma for-next since it gets conflicts:
Applying: RDMA/rxe: Move AV from rxe_send_wqe to rxe_send_wr
Applying: RDMA/rxe: Change AH objects to indexed
Using index info to reconstruct a base tree...
M drivers/infiniband/sw/rxe/rxe_param.h
Falling back to patching base and 3-way merge...
Auto-merging drivers/infiniband/sw/rxe/rxe_param.h
CONFLICT (content): Merge conflict in drivers/infiniband/sw/rxe/rxe_param.h
error: Failed to merge in the changes.
Patch failed at 0002 RDMA/rxe: Change AH objects to indexed
hint: Use 'git am --show-current-patch=diff' to see the failed patch When you have resolved this problem, run "git am --continue". If you prefer to skip this patch, run "git am --skip" instead. To restore the original branch and stop patching, run "git am --abort".
Try again
Jason
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 9 lines · Jason Gunthorpe
Quoted 2 lines · Bob Pearson
On looking, Rao's patch is not in for-next. Last one was January. Which branch are you looking at?
Oh, it is still in the wip branch, try now
Jason
I see the issue. Rao is asking for 2^20 objects max by default which will require 128KiB of memory in the index reservation bit mask for each of them. There are 4 indexed objects QP by qpn, SRQ by srqn, MR by rkey and MW by rkey. That's 512KiB of memory which seems excessive to me for many use cases where the number of objects is fairly small.
The bit mask is used to allocate and free the indices and there is also a red black tree that is used to look up objects by their index (or key if they use keys instead.)
If there is a usual way to address these kinds of issues in Linux maybe we should consider that. If not there are a couple of approaches we could take. First would be to get rid of the index bit mask and just hand out randomly selected indices in (a bigger range) and detect collisions when we insert the object into the red black tree and retry. This is basically what happens with 'keys' for example mgids for multicast group elements. Alternatively we could leave the max big but limit the allocated indices to a smaller amount until the total number of allocated indices reached some threshold and then extend the bit mask table. Then only the use cases that really needed the big index range would pay the price for the memory.
Random indices would slightly reduce some of the security issues that have been pointed out about the InfiniBand transport.
I am looking for suggestions on how to go forward here.
Bob
Jason Gunthorpe
maintainerQuoted 17 lines · Bob Pearson
Quoted 9 lines · Jason Gunthorpe
Quoted 2 lines · Bob Pearson
On looking, Rao's patch is not in for-next. Last one was January. Which branch are you looking at?
Oh, it is still in the wip branch, try now
Jason
I see the issue. Rao is asking for 2^20 objects max by default which will require 128KiB of memory in the index reservation bit mask for each of them. There are 4 indexed objects QP by qpn, SRQ by srqn, MR by rkey and MW by rkey. That's 512KiB of memory which seems excessive to me for many use cases where the number of objects is fairly small.
Quoted 5 lines
The bit mask is used to allocate and free the indices and there is also a red black tree that is used to look up objects by their index (or key if they use keys instead.)
If there is a usual way to address these kinds of issues in Linux maybe we should consider that.
Use an allocating xarray
But for these AV patches just fix the merge conflict to something sane and go ahead
Jason
Quoted 19 lines · Jason Gunthorpe
Quoted 18 lines · Bob Pearson
Quoted 8 lines · Jason Gunthorpe
Quoted 2 lines · Bob Pearson
On looking, Rao's patch is not in for-next. Last one was January. Which branch are you looking at?
Oh, it is still in the wip branch, try now
Jason
I see the issue. Rao is asking for 2^20 objects max by default which will require 128KiB of memory in the index reservation bit mask for each of them. There are 4 indexed objects QP by qpn, SRQ by srqn, MR by rkey and MW by rkey. That's 512KiB of memory which seems excessive to me for many use cases where the number of objects is fairly small.
The bit mask is used to allocate and free the indices and there is also a red black tree that is used to look up objects by their index (or key if they use keys instead.)
Quoted 8 lines
Quoted 2 lines
If there is a usual way to address these kinds of issues in Linux maybe we should consider that.
Use an allocating xarray
But for these AV patches just fix the merge conflict to something sane and go ahead
Jason
I did not want to increase the values too high but we discussed it so I did. Let me know if I need to modify the patch and reduce the values.
Shoaib
Quoted 25 lines · Shoaib Rao
Quoted 23 lines · Jason Gunthorpe
Quoted 15 lines · Bob Pearson
Quoted 8 lines · Jason Gunthorpe
Quoted 2 lines · Bob Pearson
On looking, Rao's patch is not in for-next. Last one was January. Which branch are you looking at?
Oh, it is still in the wip branch, try now
Jason
I see the issue. Rao is asking for 2^20 objects max by default which will require 128KiB of memory in the index reservation bit mask for each of them. There are 4 indexed objects QP by qpn, SRQ by srqn, MR by rkey and MW by rkey. That's 512KiB of memory which seems excessive to me for many use cases where the number of objects is fairly small.
Quoted 5 lines
The bit mask is used to allocate and free the indices and there is also a red black tree that is used to look up objects by their index (or key if they use keys instead.)
If there is a usual way to address these kinds of issues in Linux maybe we should consider that.
Use an allocating xarray
Quoted 9 lines
Quoted 4 lines
But for these AV patches just fix the merge conflict to something sane and go ahead
Jason
I did not want to increase the values too high but we discussed it so I did. Let me know if I need to modify the patch and reduce the values.
Shoaib
If we convert the rxe_pools to use xarrays as Jason suggests it looks like this issue goes away. I'm looking at that.
Bob
Quoted 29 lines · Bob Pearson
Quoted 28 lines · Shoaib Rao
Quoted 27 lines · Jason Gunthorpe
Quoted 15 lines · Bob Pearson
Quoted 8 lines · Jason Gunthorpe
Quoted 2 lines · Bob Pearson
On looking, Rao's patch is not in for-next. Last one was January. Which branch are you looking at?
Oh, it is still in the wip branch, try now
Jason
I see the issue. Rao is asking for 2^20 objects max by default which will require 128KiB of memory in the index reservation bit mask for each of them. There are 4 indexed objects QP by qpn, SRQ by srqn, MR by rkey and MW by rkey. That's 512KiB of memory which seems excessive to me for many use cases where the number of objects is fairly small.
Quoted 5 lines
The bit mask is used to allocate and free the indices and there is also a red black tree that is used to look up objects by their index (or key if they use keys instead.)
If there is a usual way to address these kinds of issues in Linux maybe we should consider that.
Use an allocating xarray
But for these AV patches just fix the merge conflict to something sane and go ahead
Jason
Quoted 8 lines
Quoted 4 lines
I did not want to increase the values too high but we discussed it so I did. Let me know if I need to modify the patch and reduce the values.
Shoaib
If we convert the
rxe_poolsto use xarrays as Jason suggests it looks like this issue goes away. I'm looking at that.Bob
Thanks Bob. Let me know if there is anything that I can help out with.
Shoaib