RDMA/rxe: Implement memory windows
Subsystems (2)
v8 cover RDMA/rxe: Implement memory windows
This series of patches implement memory windows for the rdma_rxe driver. This is a shorter reimplementation of an earlier patch set. They apply to and depend on the current for-next linux rdma tree.
---
v8:
Dropped wr.mw.flags in the rxe_send_wr struct in rdma_user_rxe.h.
v7:
Fixed a duplicate INIT_RDMA_OBJ_SIZE(ib_mw, ...) in rxe_verbs.c.
v6:
Added rxe_ prefix to subroutine names in lines that changed from Zhu's review of v5.
v5:
Fixed a typo in 10th patch.
v4:
Added a 10th patch to check when MRs have bound MWs and disallow dereg and invalidate operations.
v3:
cleaned up void return and lower case enums from
Zhu's review.
v2:
cleaned up an issue in rdma_user_rxe.h
cleaned up a collision in rxe_resp.c
Bob Pearson (10):
RDMA/rxe: Add bind MW fields to rxe_send_wr RDMA/rxe: Return errors for add index and key RDMA/rxe: Enable MW object pool
RDMA/rxe: Add ib_alloc_mw and ib_dealloc_mw verbs RDMA/rxe: Replace WR_REG_MASK by WR_LOCAL_OP_MASK RDMA/rxe: Move local ops to subroutine RDMA/rxe: Add support for bind MW work requests RDMA/rxe: Implement invalidate MW operations RDMA/rxe: Implement memory access through MWs RDMA/rxe: Disallow MR dereg and invalidate when bound
drivers/infiniband/sw/rxe/Makefile | 1 + drivers/infiniband/sw/rxe/rxe.c | 1 + drivers/infiniband/sw/rxe/rxe_comp.c | 5 +- drivers/infiniband/sw/rxe/rxe_loc.h | 36 +-- drivers/infiniband/sw/rxe/rxe_mr.c | 126 ++++++--- drivers/infiniband/sw/rxe/rxe_mw.c | 343 +++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_opcode.c | 11 +- drivers/infiniband/sw/rxe/rxe_opcode.h | 3 +- drivers/infiniband/sw/rxe/rxe_param.h | 19 +- drivers/infiniband/sw/rxe/rxe_pool.c | 45 ++-- drivers/infiniband/sw/rxe/rxe_pool.h | 8 +- drivers/infiniband/sw/rxe/rxe_req.c | 104 +++++--- drivers/infiniband/sw/rxe/rxe_resp.c | 111 +++++--- drivers/infiniband/sw/rxe/rxe_verbs.c | 15 +- drivers/infiniband/sw/rxe/rxe_verbs.h | 48 +++- include/uapi/rdma/rdma_user_rxe.h | 10 + 16 files changed, 702 insertions(+), 184 deletions(-) create mode 100644 drivers/infiniband/sw/rxe/rxe_mw.c
Signature · 1 line
2.27.0
PATCH v8 1/9 RDMA/rxe: Add bind MW fields to rxe_send_wr
Add fields to struct rxe_send_wr in rdma_user_rxe.h to support bind MW work requests
---
v8:
Dropped the flags field in wr.mw which was no longer needed. The size of the mw struct broke binary compatibility because it extended the size of the wr union beyond 40 bytes.
--- include/uapi/rdma/rdma_user_rxe.h | 10 ++++++++++ 1 file changed, 10 insertions(+)
include/uapi/rdma/rdma_user_rxe.h +10 -0
@@ -99,7 +99,16 @@ struct rxe_send_wr {__u32 remote_qkey;__u16 pkey_index; } ud; +struct { +__aligned_u64 addr; +__aligned_u64 length; +__u32 mr_lkey; +__u32 mw_rkey; +__u32 rkey; +__u32 access; + } mw;/* reg is only used by the kernel and is not part of the uapi */ +#ifdef __KERNEL__ struct { union { struct ib_mr *mr;
@@ -108,6 +117,7 @@ struct rxe_send_wr {__u32 key; __u32 access; } reg; +#endif } wr; };
Signature · 1 line
2.30.2
PATCH v8 2/10 RDMA/rxe: Return errors for add index and key
In reply to Bob PearsonModify rxe_add_index() and rxe_add_key() to return an error if the index or key is aleady present in the pool. Currently they print a warning and silently fail with bad consequences to the caller.
--- drivers/infiniband/sw/rxe/rxe_pool.c | 44 ++++++++++++++++++---------- drivers/infiniband/sw/rxe/rxe_pool.h | 8 ++--- 2 files changed, 32 insertions(+), 20 deletions(-)
drivers/infiniband/sw/rxe/rxe_pool.c +28 -16
@@ -183,7 +183,7 @@ static u32 alloc_index(struct rxe_pool *pool)return index + pool->index.min_index; } -static void insert_index (struct rxe_pool *pool,struct rxe_pool_entry *new) +static int rxe_insert_index (struct rxe_pool *pool,struct rxe_pool_entry *new) {struct rb_node **link = &pool->index.tree.rb_node;struct rb_node *parent =NULL ;
@@ -195,7 +195,7 @@ static void insert_index(struct rxe_pool *pool, struct rxe_pool_entry *new)if (elem->index == new->index) {pr_warn ("element already exists!\n" ); - goto out; +return -EINVAL ; }if (elem->index > new->index)
@@ -206,11 +206,11 @@ static void insert_index(struct rxe_pool *pool, struct rxe_pool_entry *new)rb_link_node (&new->index_node, parent, link);rb_insert_color (&new->index_node, &pool->index.tree); -out: -return ; + +return 0 ; } -static void insert_key (struct rxe_pool *pool,struct rxe_pool_entry *new) +static int rxe_insert_key (struct rxe_pool *pool,struct rxe_pool_entry *new) {struct rb_node **link = &pool->key.tree.rb_node;struct rb_node *parent =NULL ;
@@ -226,7 +226,7 @@ static void insert_key(struct rxe_pool *pool, struct rxe_pool_entry *new)if (cmp ==0 ) {pr_warn ("key already exists!\n" ); - goto out; +return -EINVAL ; }if (cmp >0 )
@@ -237,26 +237,32 @@ static void insert_key(struct rxe_pool *pool, struct rxe_pool_entry *new)rb_link_node (&new->key_node, parent, link);rb_insert_color (&new->key_node, &pool->key.tree); -out: -return ; + +return 0 ; } -void __rxe_add_key_locked (struct rxe_pool_entry *elem,void *key) +int __rxe_add_key_locked (struct rxe_pool_entry *elem,void *key) {struct rxe_pool *pool = elem->pool; +int err;memcpy ((u8 *)elem + pool->key.key_offset, key, pool->key.key_size); -insert_key (pool, elem); + err =rxe_insert_key (pool, elem); + +return err; } -void __rxe_add_key (struct rxe_pool_entry *elem,void *key) +int __rxe_add_key (struct rxe_pool_entry *elem,void *key) {struct rxe_pool *pool = elem->pool;unsigned long flags; +int err;write_lock_irqsave (&pool->pool_lock, flags); -__rxe_add_key_locked (elem, key); + err =__rxe_add_key_locked (elem, key);write_unlock_irqrestore (&pool->pool_lock, flags); + +return err; }void __rxe_drop_key_locked (struct rxe_pool_entry *elem)
@@ -276,22 +282,28 @@ void __rxe_drop_key(struct rxe_pool_entry *elem)write_unlock_irqrestore (&pool->pool_lock, flags); } -void __rxe_add_index_locked (struct rxe_pool_entry *elem) +int __rxe_add_index_locked (struct rxe_pool_entry *elem) {struct rxe_pool *pool = elem->pool; +int err; elem->index =alloc_index (pool); -insert_index (pool, elem); + err =rxe_insert_index (pool, elem); + +return err; } -void __rxe_add_index (struct rxe_pool_entry *elem) +int __rxe_add_index (struct rxe_pool_entry *elem) {struct rxe_pool *pool = elem->pool;unsigned long flags; +int err;write_lock_irqsave (&pool->pool_lock, flags); -__rxe_add_index_locked (elem); + err =__rxe_add_index_locked (elem);write_unlock_irqrestore (&pool->pool_lock, flags); + +return err; }void __rxe_drop_index_locked (struct rxe_pool_entry *elem)
drivers/infiniband/sw/rxe/rxe_pool.h +4 -4
@@ -111,11 +111,11 @@ int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_entry *elem);/* assign an index to an indexed object and insert object into * pool's rb tree holding and not holding the pool_lock */ -void __rxe_add_index_locked (struct rxe_pool_entry *elem); +int __rxe_add_index_locked (struct rxe_pool_entry *elem);#define rxe_add_index_locked (obj) __rxe_add_index_locked(&(obj)->pelem) -void __rxe_add_index (struct rxe_pool_entry *elem); +int __rxe_add_index (struct rxe_pool_entry *elem);#define rxe_add_index (obj) __rxe_add_index(&(obj)->pelem)
@@ -133,11 +133,11 @@ void __rxe_drop_index(struct rxe_pool_entry *elem);/* assign a key to a keyed object and insert object into * pool's rb tree holding and not holding pool_lock */ -void __rxe_add_key_locked (struct rxe_pool_entry *elem,void *key); +int __rxe_add_key_locked (struct rxe_pool_entry *elem,void *key);#define rxe_add_key_locked (obj, key) __rxe_add_key_locked(&(obj)->pelem, key) -void __rxe_add_key (struct rxe_pool_entry *elem,void *key); +int __rxe_add_key (struct rxe_pool_entry *elem,void *key);#define rxe_add_key (obj, key) __rxe_add_key(&(obj)->pelem, key)
Signature · 1 line
2.30.2
PATCH v8 3/10 RDMA/rxe: Enable MW object pool
In reply to Bob PearsonCurrently the rxe driver has a rxe_mw struct object but nothing about memory windows is enabled. This patch turns on memory windows and some minor cleanup.
Set device attribute in rxe.c so max_mw = MAX_MW. Change parameters in rxe_param.h so that MAX_MW is the same as MAX_MR. Reduce the number of MRs and MWs to 4K from 256K. Add device capability bits for 2a and 2b memory windows. Removed RXE_MR_TYPE_MW from the rxe_mr_type enum.
--- drivers/infiniband/sw/rxe/rxe.c | 1 + drivers/infiniband/sw/rxe/rxe_param.h | 19 ++++++++++++------- drivers/infiniband/sw/rxe/rxe_verbs.h | 1 - 3 files changed, 13 insertions(+), 8 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +1 -0
@@ -54,6 +54,7 @@ static void rxe_init_device_param(struct rxe_dev *rxe) rxe->attr.max_cq =RXE_MAX_CQ ; rxe->attr.max_cqe = (1 <<RXE_MAX_LOG_CQE ) -1 ; rxe->attr.max_mr =RXE_MAX_MR ; + rxe->attr.max_mw =RXE_MAX_MW ; rxe->attr.max_pd =RXE_MAX_PD ; rxe->attr.max_qp_rd_atom =RXE_MAX_QP_RD_ATOM ; rxe->attr.max_res_rd_atom =RXE_MAX_RES_RD_ATOM ;
drivers/infiniband/sw/rxe/rxe_param.h +12 -7
@@ -37,7 +37,6 @@ static inline enum ib_mtu eth_mtu_int_to_enum(int mtu)enum rxe_device_param {RXE_MAX_MR_SIZE =-1ull ,RXE_PAGE_SIZE_CAP =0xfffff000 , -RXE_MAX_QP =0x10000 ,RXE_MAX_QP_WR =0x4000 ,RXE_DEVICE_CAP_FLAGS =IB_DEVICE_BAD_PKEY_CNTR |IB_DEVICE_BAD_QKEY_CNTR
@@ -49,7 +48,10 @@ enum rxe_device_param { | IB_DEVICE_RC_RNR_NAK_GEN |IB_DEVICE_SRQ_RESIZE |IB_DEVICE_MEM_MGT_EXTENSIONS - |IB_DEVICE_ALLOW_USER_UNREG , + |IB_DEVICE_ALLOW_USER_UNREG + |IB_DEVICE_MEM_WINDOW + |IB_DEVICE_MEM_WINDOW_TYPE_2A + |IB_DEVICE_MEM_WINDOW_TYPE_2B ,RXE_MAX_SGE =32 ,RXE_MAX_WQE_SIZE =sizeof (struct rxe_send_wqe ) + sizeof(struct ib_sge) * RXE_MAX_SGE,
@@ -58,7 +60,6 @@ enum rxe_device_param {RXE_MAX_SGE_RD =32 ,RXE_MAX_CQ =16384 ,RXE_MAX_LOG_CQE =15 , -RXE_MAX_MR =256 *1024 ,RXE_MAX_PD =0x7ffc ,RXE_MAX_QP_RD_ATOM =128 , RXE_MAX_RES_RD_ATOM = 0x3f000,
@@ -67,7 +68,6 @@ enum rxe_device_param {RXE_MAX_MCAST_QP_ATTACH =56 ,RXE_MAX_TOT_MCAST_QP_ATTACH =0x70000 ,RXE_MAX_AH =100 , -RXE_MAX_SRQ =960 ,RXE_MAX_SRQ_WR =0x4000 ,RXE_MIN_SRQ_WR =1 , RXE_MAX_SRQ_SGE = 27,
@@ -80,16 +80,21 @@ enum rxe_device_param {RXE_NUM_PORT =1 , +RXE_MAX_QP =0x10000 ,RXE_MIN_QP_INDEX =16 ,RXE_MAX_QP_INDEX =0x00020000 , +RXE_MAX_SRQ =0x00001000 ,RXE_MIN_SRQ_INDEX =0x00020001 ,RXE_MAX_SRQ_INDEX =0x00040000 , +RXE_MAX_MR =0x00001000 , +RXE_MAX_MW =0x00001000 ,RXE_MIN_MR_INDEX =0x00000001 , -RXE_MAX_MR_INDEX =0x00040000 , -RXE_MIN_MW_INDEX =0x00040001 , -RXE_MAX_MW_INDEX =0x00060000 , +RXE_MAX_MR_INDEX =0x00010000 , +RXE_MIN_MW_INDEX =0x00010001 , +RXE_MAX_MW_INDEX =0x00020000 , +RXE_MAX_PKT_PER_ACK =64 , RXE_MAX_UNACKED_PSNS = 128,
drivers/infiniband/sw/rxe/rxe_verbs.h +0 -1
@@ -273,7 +273,6 @@ enum rxe_mr_type {RXE_MR_TYPE_NONE ,RXE_MR_TYPE_DMA , RXE_MR_TYPE_MR, - RXE_MR_TYPE_MW, };#define RXE_BUF_PER_MAP (PAGE_SIZE / sizeof(struct rxe_phys_buf))
Signature · 1 line
2.30.2
PATCH v8 4/10 RDMA/rxe: Add ib_alloc_mw and ib_dealloc_mw verbs
In reply to Bob PearsonAdd ib_alloc_mw and ib_dealloc_mw verbs APIs.
Added new file rxe_mw.c focused on MWs. Changed the 8 bit random key generator. Added a cleanup routine for MWs. Added verbs routines to ib_device_ops.
---
v7:
Removed duplicate INIT_RDMA_OBJ_SIZE(ib_mw, ...) in rxe_verbs.c. This was already added in patch 03/10.
--- drivers/infiniband/sw/rxe/Makefile | 1 + drivers/infiniband/sw/rxe/rxe_loc.h | 6 +++ drivers/infiniband/sw/rxe/rxe_mr.c | 20 +++++----- drivers/infiniband/sw/rxe/rxe_mw.c | 53 +++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_pool.c | 1 + drivers/infiniband/sw/rxe/rxe_verbs.c | 2 + drivers/infiniband/sw/rxe/rxe_verbs.h | 2 + 7 files changed, 74 insertions(+), 11 deletions(-) create mode 100644 drivers/infiniband/sw/rxe/rxe_mw.c
drivers/infiniband/sw/rxe/Makefile +1 -0
@@ -15,6 +15,7 @@ rdma_rxe-y := \ rxe_qp.o \ rxe_cq.o \ rxe_mr.o \ + rxe_mw.o \ rxe_opcode.o \ rxe_mmap.o \ rxe_icrc.o \
drivers/infiniband/sw/rxe/rxe_loc.h +6 -0
@@ -76,6 +76,7 @@ enum copy_direction { from_mr_obj, }; +u8 rxe_get_next_key (u32 last_key);void rxe_mr_init_dma (struct rxe_pd *pd,int access,struct rxe_mr *mr); int rxe_mr_init_user(struct rxe_pd *pd, u64 start, u64 length, u64 iova,
@@ -106,6 +107,11 @@ void rxe_mr_cleanup(struct rxe_pool_entry *arg);int advance_dma_data (struct rxe_dma_info *dma,unsigned int length); +/* rxe_mw.c */ +int rxe_alloc_mw (struct ib_mw *ibmw,struct ib_udata *udata); +int rxe_dealloc_mw (struct ib_mw *ibmw); +void rxe_mw_cleanup (struct rxe_pool_entry *arg); +/* rxe_net.c */ void rxe_loopback (struct sk_buff *skb);int rxe_send (struct rxe_pkt_info *pkt,struct sk_buff *skb);
drivers/infiniband/sw/rxe/rxe_mr.c +9 -11
@@ -7,19 +7,17 @@#include "rxe.h" #include "rxe_loc.h" -/* -* lfsr (linear feedback shift register) with period 255 +/* Return a random 8 bit key value that is +* different than the last_key. Set last_key to -1 +* if this is the first key for an MR or MW */ -static u8 rxe_get_key (void ) +u8 rxe_get_next_key (u32 last_key) { -static u32 key =1 ; - - key = key <<1 ; - - key |= (0 != (key &0x100 )) ^ (0 != (key &0x10 )) - ^ (0 != (key &0x80 )) ^ (0 != (key &0x40 )); +u8 key; - key &=0xff ; +do { +get_random_bytes (&key,1 ); + }while (key == last_key);return key; }
@@ -47,7 +45,7 @@ int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length)static void rxe_mr_init (int access,struct rxe_mr *mr) { -u32 lkey = mr->pelem.index <<8 |rxe_get_key (); +u32 lkey = mr->pelem.index <<8 |rxe_get_next_key (-1 );u32 rkey = (access &IB_ACCESS_REMOTE ) ? lkey :0 ; mr->ibmr.lkey = lkey;
drivers/infiniband/sw/rxe/rxe_mw.c new file +53 -0
53 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_pool.c +1 -0
@@ -65,6 +65,7 @@ struct rxe_type_info rxe_type_info[RXE_NUM_TYPES] = { .name = "rxe-mw", .size = sizeof(struct rxe_mw), .elem_offset = offsetof(struct rxe_mw, pelem), + .cleanup = rxe_mw_cleanup, .flags = RXE_POOL_INDEX | RXE_POOL_NO_ALLOC, .max_index = RXE_MAX_MW_INDEX, .min_index = RXE_MIN_MW_INDEX,
drivers/infiniband/sw/rxe/rxe_verbs.c +2 -0
@@ -1060,6 +1060,7 @@ static const struct ib_device_ops rxe_dev_ops = { .alloc_hw_stats = rxe_ib_alloc_hw_stats, .alloc_mr = rxe_alloc_mr, + .alloc_mw = rxe_alloc_mw, .alloc_pd = rxe_alloc_pd, .alloc_ucontext = rxe_alloc_ucontext, .attach_mcast = rxe_attach_mcast,
@@ -1069,6 +1070,7 @@ static const struct ib_device_ops rxe_dev_ops = { .create_srq = rxe_create_srq, .create_user_ah = rxe_create_ah, .dealloc_driver = rxe_dealloc, + .dealloc_mw = rxe_dealloc_mw, .dealloc_pd = rxe_dealloc_pd, .dealloc_ucontext = rxe_dealloc_ucontext, .dereg_mr = rxe_dereg_mr,
drivers/infiniband/sw/rxe/rxe_verbs.h +2 -0
@@ -323,6 +323,8 @@ enum rxe_mw_state {struct rxe_mw {struct ib_mw ibmw;struct rxe_pool_entry pelem; +spinlock_t lock; +enum rxe_mw_state state; };struct rxe_mc_grp {
Signature · 1 line
2.30.2
PATCH v8 5/10 RDMA/rxe: Replace WR_REG_MASK by WR_LOCAL_OP_MASK
In reply to Bob PearsonRxe has two mask bits WR_LOCAL_MASK and WR_REG_MASK with WR_REG_MASK used to indicate any local operation and WR_LOCAL_MASK unused. This patch replaces both of these with one mask bit WR_LOCAL_OP_MASK which is clearer.
--- drivers/infiniband/sw/rxe/rxe_opcode.c | 4 ++-- drivers/infiniband/sw/rxe/rxe_opcode.h | 3 +-- drivers/infiniband/sw/rxe/rxe_req.c | 2 +- drivers/infiniband/sw/rxe/rxe_verbs.c | 2 +- 4 files changed, 5 insertions(+), 6 deletions(-)
drivers/infiniband/sw/rxe/rxe_opcode.c +2 -2
@@ -87,13 +87,13 @@ struct rxe_wr_opcode_info rxe_wr_opcode_info[] = { [IB_WR_LOCAL_INV ] = { .name ="IB_WR_LOCAL_INV" , .mask = { - [IB_QPT_RC ] =WR_REG_MASK , + [IB_QPT_RC ] =WR_LOCAL_OP_MASK , }, }, [IB_WR_REG_MR ] = { .name ="IB_WR_REG_MR" , .mask = { - [IB_QPT_RC ] =WR_REG_MASK , + [IB_QPT_RC ] =WR_LOCAL_OP_MASK , }, }, };
drivers/infiniband/sw/rxe/rxe_opcode.h +1 -2
@@ -19,8 +19,7 @@ enum rxe_wr_mask {WR_SEND_MASK =BIT (2 ),WR_READ_MASK =BIT (3 ),WR_WRITE_MASK =BIT (4 ), -WR_LOCAL_MASK =BIT (5 ), -WR_REG_MASK =BIT (6 ), +WR_LOCAL_OP_MASK =BIT (5 ),WR_READ_OR_WRITE_MASK =WR_READ_MASK |WR_WRITE_MASK , WR_READ_WRITE_OR_SEND_MASK = WR_READ_OR_WRITE_MASK | WR_SEND_MASK,
drivers/infiniband/sw/rxe/rxe_req.c +1 -1
@@ -593,7 +593,7 @@ int rxe_requester(void *arg)if (unlikely (!wqe)) goto exit; -if (wqe->mask &WR_REG_MASK ) { +if (wqe->mask &WR_LOCAL_OP_MASK ) {if (wqe->wr.opcode ==IB_WR_LOCAL_INV ) {struct rxe_dev *rxe =to_rdev (qp->ibqp.device);struct rxe_mr *rmr;
drivers/infiniband/sw/rxe/rxe_verbs.c +1 -1
@@ -577,7 +577,7 @@ static void init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,init_send_wr (qp, &wqe->wr, ibwr);/* local operation */ -if (unlikely (mask &WR_REG_MASK )) { +if (unlikely (mask &WR_LOCAL_OP_MASK )) { wqe->mask = mask; wqe->state = wqe_state_posted;return ;
Signature · 1 line
2.30.2
PATCH v8 6/10 RDMA/rxe: Move local ops to subroutine
In reply to Bob PearsonSimplify rxe_requester() by moving the local operations to a subroutine. Add an error return for illegal send WR opcode. Moved next_index ahead of rxe_run_task which fixed a small bug where work completions were delayed until after the next wqe which was not the intended behavior.
--- drivers/infiniband/sw/rxe/rxe_req.c | 89 +++++++++++++++++------------ 1 file changed, 54 insertions(+), 35 deletions(-)
drivers/infiniband/sw/rxe/rxe_req.c +54 -35
@@ -555,6 +555,56 @@ static void update_state(struct rxe_qp *qp, struct rxe_send_wqe *wqe, jiffies + qp->qp_timeout_jiffies); } +static int rxe_do_local_ops (struct rxe_qp *qp,struct rxe_send_wqe *wqe) +{ +u8 opcode = wqe->wr.opcode; +struct rxe_dev *rxe; +struct rxe_mr *mr; +u32 rkey; + +switch (opcode) { +case IB_WR_LOCAL_INV : + rxe =to_rdev (qp->ibqp.device); + rkey = wqe->wr.ex.invalidate_rkey; + mr =rxe_pool_get_index (&rxe->mr_pool, rkey >>8 ); +if (!mr) { +pr_err ("No MR for rkey %#x\n" , rkey); + wqe->state = wqe_state_error; + wqe->status =IB_WC_LOC_QP_OP_ERR ; +return -EINVAL ; + } + mr->state =RXE_MR_STATE_FREE ; +rxe_drop_ref (mr); +break ; +case IB_WR_REG_MR : + mr =to_rmr (wqe->wr.wr.reg.mr); + +rxe_add_ref (mr); + mr->state =RXE_MR_STATE_VALID ; + mr->access = wqe->wr.wr.reg.access; + mr->ibmr.lkey = wqe->wr.wr.reg.key; + mr->ibmr.rkey = wqe->wr.wr.reg.key; + mr->iova = wqe->wr.wr.reg.mr->iova; +rxe_drop_ref (mr); +break ; +default : +pr_err ("Unexpected send wqe opcode %d\n" , opcode); + wqe->state = wqe_state_error; + wqe->status =IB_WC_LOC_QP_OP_ERR ; +return -EINVAL ; + } + + wqe->state = wqe_state_done; + wqe->status =IB_WC_SUCCESS ; + qp->req.wqe_index =next_index (qp->sq.queue, qp->req.wqe_index); + +if ((wqe->wr.send_flags &IB_SEND_SIGNALED ) || + qp->sq_sig_type ==IB_SIGNAL_ALL_WR ) +rxe_run_task (&qp->comp.task,1 ); + +return 0 ; +} +int rxe_requester (void *arg) {struct rxe_qp *qp = (struct rxe_qp *)arg;
@@ -594,42 +644,11 @@ int rxe_requester(void *arg) goto exit;if (wqe->mask &WR_LOCAL_OP_MASK ) { -if (wqe->wr.opcode ==IB_WR_LOCAL_INV ) { -struct rxe_dev *rxe =to_rdev (qp->ibqp.device); -struct rxe_mr *rmr; - - rmr =rxe_pool_get_index (&rxe->mr_pool, - wqe->wr.ex.invalidate_rkey >>8 ); -if (!rmr) { -pr_err ("No mr for key %#x\n" , - wqe->wr.ex.invalidate_rkey); - wqe->state = wqe_state_error; - wqe->status =IB_WC_MW_BIND_ERR ; - goto exit; - } - rmr->state =RXE_MR_STATE_FREE ; -rxe_drop_ref (rmr); - wqe->state = wqe_state_done; - wqe->status =IB_WC_SUCCESS ; - }else if (wqe->wr.opcode ==IB_WR_REG_MR ) { -struct rxe_mr *rmr =to_rmr (wqe->wr.wr.reg.mr); - - rmr->state =RXE_MR_STATE_VALID ; - rmr->access = wqe->wr.wr.reg.access; - rmr->ibmr.lkey = wqe->wr.wr.reg.key; - rmr->ibmr.rkey = wqe->wr.wr.reg.key; - rmr->iova = wqe->wr.wr.reg.mr->iova; - wqe->state = wqe_state_done; - wqe->status =IB_WC_SUCCESS ; - }else { + ret =rxe_do_local_ops (qp, wqe); +if (unlikely (ret)) goto exit; - } -if ((wqe->wr.send_flags &IB_SEND_SIGNALED ) || - qp->sq_sig_type ==IB_SIGNAL_ALL_WR ) -rxe_run_task (&qp->comp.task,1 ); - qp->req.wqe_index =next_index (qp->sq.queue, - qp->req.wqe_index); - goto next_wqe; +else + goto next_wqe; } if (unlikely(qp_type(qp) == IB_QPT_RC &&
Signature · 1 line
2.30.2
PATCH v8 7/10 RDMA/rxe: Add support for bind MW work requests
In reply to Bob PearsonAdd support for bind MW work requests from user space. Since rdma/core does not support bind mw in ib_send_wr there is no way to support bind mw in kernel space.
Added bind_mw local operation in rxe_req.c. Added bind_mw WR operation in rxe_opcode.c. Added bind_mw WC in rxe_comp.c. Added additional fields to rxe_mw in rxe_verbs.h. Added rxe_do_dealloc_mw() subroutine to cleanup an mw when rxe_dealloc_mw is called. Added code to implement bind_mw operation in rxe_mw.c
--- drivers/infiniband/sw/rxe/rxe_comp.c | 1 + drivers/infiniband/sw/rxe/rxe_loc.h | 1 + drivers/infiniband/sw/rxe/rxe_mw.c | 202 ++++++++++++++++++++++++- drivers/infiniband/sw/rxe/rxe_opcode.c | 7 + drivers/infiniband/sw/rxe/rxe_req.c | 9 ++ drivers/infiniband/sw/rxe/rxe_verbs.h | 15 +- 6 files changed, 230 insertions(+), 5 deletions(-)
drivers/infiniband/sw/rxe/rxe_comp.c +1 -0
@@ -103,6 +103,7 @@ static enum ib_wc_opcode wr_to_wc_opcode(enum ib_wr_opcode opcode)case IB_WR_RDMA_READ_WITH_INV :return IB_WC_RDMA_READ ;case IB_WR_LOCAL_INV :return IB_WC_LOCAL_INV ;case IB_WR_REG_MR :return IB_WC_REG_MR ; +case IB_WR_BIND_MW :return IB_WC_BIND_MW ;default :return 0xff ;
drivers/infiniband/sw/rxe/rxe_loc.h +1 -0
@@ -110,6 +110,7 @@ int advance_dma_data(struct rxe_dma_info *dma, unsigned int length);/* rxe_mw.c */ int rxe_alloc_mw (struct ib_mw *ibmw,struct ib_udata *udata);int rxe_dealloc_mw (struct ib_mw *ibmw); +int rxe_bind_mw (struct rxe_qp *qp,struct rxe_send_wqe *wqe);void rxe_mw_cleanup (struct rxe_pool_entry *arg);/* rxe_net.c */
drivers/infiniband/sw/rxe/rxe_mw.c +201 -1
220 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_opcode.c +7 -0
@@ -96,6 +96,13 @@ struct rxe_wr_opcode_info rxe_wr_opcode_info[] = { [IB_QPT_RC] = WR_LOCAL_OP_MASK, }, }, + [IB_WR_BIND_MW] = { + .name = "IB_WR_BIND_MW", + .mask = { + [IB_QPT_RC] = WR_LOCAL_OP_MASK, + [IB_QPT_UC] = WR_LOCAL_OP_MASK, + }, + }, };struct rxe_opcode_info rxe_opcode[RXE_NUM_OPCODE ] = {
drivers/infiniband/sw/rxe/rxe_req.c +9 -0
@@ -561,6 +561,7 @@ static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe)struct rxe_dev *rxe;struct rxe_mr *mr;u32 rkey; +int ret;switch (opcode) {case IB_WR_LOCAL_INV :
@@ -587,6 +588,14 @@ static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe) mr->iova = wqe->wr.wr.reg.mr->iova;rxe_drop_ref (mr);break ; +case IB_WR_BIND_MW : + ret =rxe_bind_mw (qp, wqe); +if (ret) { + wqe->state = wqe_state_error; + wqe->status =IB_WC_MW_BIND_ERR ; +return -EINVAL ; + } +break ;default :pr_err ("Unexpected send wqe opcode %d\n" , opcode); wqe->state = wqe_state_error;
drivers/infiniband/sw/rxe/rxe_verbs.h +11 -4
@@ -312,6 +312,8 @@ struct rxe_mr {u32 num_map;struct rxe_map **map; + +atomic_t num_mw; };enum rxe_mw_state {
@@ -321,10 +323,15 @@ enum rxe_mw_state { };struct rxe_mw { -struct ib_mw ibmw; -struct rxe_pool_entry pelem; -spinlock_t lock; -enum rxe_mw_state state; +struct ib_mw ibmw; +struct rxe_pool_entry pelem; +spinlock_t lock; +enum rxe_mw_state state; +struct rxe_qp *qp;/* Type 2 only */ +struct rxe_mr *mr; +int access; +u64 addr; +u64 length; };struct rxe_mc_grp {
Signature · 1 line
2.30.2
PATCH v8 8/10 RDMA/rxe: Implement invalidate MW operations
In reply to Bob PearsonImplement invalidate MW and cleaned up invalidate MR operations.
Added code to perform remote invalidate for send with invalidate. Added code to perform local invalidation. Deleted some blank lines in rxe_loc.h.
---
v6:
Added rxe_ to subroutine names in lines that changed.
v3:
Replaced enums in lower case with upper case and moved to rxe_verbs.h which is where enums live.
--- drivers/infiniband/sw/rxe/rxe_comp.c | 4 +- drivers/infiniband/sw/rxe/rxe_loc.h | 29 ++-------- drivers/infiniband/sw/rxe/rxe_mr.c | 81 ++++++++++++++++++--------- drivers/infiniband/sw/rxe/rxe_mw.c | 67 ++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_req.c | 24 ++++---- drivers/infiniband/sw/rxe/rxe_resp.c | 60 ++++++++++++-------- drivers/infiniband/sw/rxe/rxe_verbs.h | 33 ++++++++--- 7 files changed, 204 insertions(+), 94 deletions(-)
drivers/infiniband/sw/rxe/rxe_comp.c +2 -2
@@ -346,7 +346,7 @@ static inline enum comp_state do_read(struct rxe_qp *qp, ret =copy_data (qp->pd,IB_ACCESS_LOCAL_WRITE , &wqe->dma,payload_addr (pkt), -payload_size (pkt), to_mr_obj,NULL ); +payload_size (pkt),RXE_TO_MR_OBJ ,NULL );if (ret)return COMPST_ERROR ;
@@ -366,7 +366,7 @@ static inline enum comp_state do_atomic(struct rxe_qp *qp, ret =copy_data (qp->pd,IB_ACCESS_LOCAL_WRITE , &wqe->dma, &atomic_orig, -sizeof (u64), to_mr_obj,NULL ); +sizeof (u64),RXE_TO_MR_OBJ ,NULL );if (ret)return COMPST_ERROR ;else
drivers/infiniband/sw/rxe/rxe_loc.h +6 -23
@@ -71,46 +71,29 @@ struct rxe_mmap_info *rxe_create_mmap_info(struct rxe_dev *dev, u32 size,int rxe_mmap (struct ib_ucontext *context,struct vm_area_struct *vma);/* rxe_mr.c */ -enum copy_direction { - to_mr_obj, - from_mr_obj, -}; -u8 rxe_get_next_key (u32 last_key);void rxe_mr_init_dma (struct rxe_pd *pd,int access,struct rxe_mr *mr); -int rxe_mr_init_user (struct rxe_pd *pd,u64 start,u64 length,u64 iova,int access,struct rxe_mr *mr); -int rxe_mr_init_fast (struct rxe_pd *pd,int max_pages,struct rxe_mr *mr); -int rxe_mr_copy (struct rxe_mr *mr,u64 iova,void *addr,int length, -enum copy_direction dir,u32 *crcp); - +enum rxe_mr_copy_dir dir,u32 *crcp);int copy_data (struct rxe_pd *pd,int access,struct rxe_dma_info *dma,void *addr,int length, -enum copy_direction dir,u32 *crcp); - +enum rxe_mr_copy_dir dir,u32 *crcp);void *iova_to_vaddr (struct rxe_mr *mr,u64 iova,int length); - -enum lookup_type { - lookup_local, - lookup_remote, -}; -struct rxe_mr *lookup_mr (struct rxe_pd *pd,int access,u32 key, -enum lookup_type type); - +enum rxe_mr_lookup_type type);int mr_check_range (struct rxe_mr *mr,u64 iova,size_t length); - -void rxe_mr_cleanup (struct rxe_pool_entry *arg); -int advance_dma_data (struct rxe_dma_info *dma,unsigned int length); +int rxe_invalidate_mr (struct rxe_qp *qp,u32 rkey); +void rxe_mr_cleanup (struct rxe_pool_entry *arg);/* rxe_mw.c */ int rxe_alloc_mw (struct ib_mw *ibmw,struct ib_udata *udata);int rxe_dealloc_mw (struct ib_mw *ibmw);int rxe_bind_mw (struct rxe_qp *qp,struct rxe_send_wqe *wqe); +int rxe_invalidate_mw (struct rxe_qp *qp,u32 rkey);void rxe_mw_cleanup (struct rxe_pool_entry *arg);/* rxe_net.c */
drivers/infiniband/sw/rxe/rxe_mr.c +55 -26
139 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_mw.c +67 -0
73 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_req.c +13 -11
@@ -464,7 +464,7 @@ static int fill_packet(struct rxe_qp *qp, struct rxe_send_wqe *wqe, }else { err =copy_data (qp->pd,0 , &wqe->dma,payload_addr (pkt), paylen, - from_mr_obj, +RXE_FROM_MR_OBJ , &crc);if (err)return err;
@@ -558,25 +558,25 @@ static void update_state(struct rxe_qp *qp, struct rxe_send_wqe *wqe,static int rxe_do_local_ops (struct rxe_qp *qp,struct rxe_send_wqe *wqe) {u8 opcode = wqe->wr.opcode; -struct rxe_dev *rxe;struct rxe_mr *mr; -u32 rkey;int ret; +u32 rkey;switch (opcode) {case IB_WR_LOCAL_INV : - rxe =to_rdev (qp->ibqp.device); rkey = wqe->wr.ex.invalidate_rkey; - mr =rxe_pool_get_index (&rxe->mr_pool, rkey >>8 ); -if (!mr) { -pr_err ("No MR for rkey %#x\n" , rkey); +if (rkey_is_mw (rkey)) + ret =rxe_invalidate_mw (qp, rkey); +else + ret =rxe_invalidate_mr (qp, rkey); + +if (ret) { wqe->state = wqe_state_error; wqe->status =IB_WC_LOC_QP_OP_ERR ; -return -EINVAL ; +return ret; } - mr->state =RXE_MR_STATE_FREE ; -rxe_drop_ref (mr);break ; +case IB_WR_REG_MR : mr =to_rmr (wqe->wr.wr.reg.mr);
@@ -588,14 +588,16 @@ static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe) mr->iova = wqe->wr.wr.reg.mr->iova;rxe_drop_ref (mr);break ; +case IB_WR_BIND_MW : ret =rxe_bind_mw (qp, wqe);if (ret) { wqe->state = wqe_state_error; wqe->status =IB_WC_MW_BIND_ERR ; -return -EINVAL ; +return ret; }break ; +default :pr_err ("Unexpected send wqe opcode %d\n" , opcode); wqe->state = wqe_state_error;
drivers/infiniband/sw/rxe/rxe_resp.c +36 -24
134 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_verbs.h +25 -8
@@ -275,6 +275,16 @@ enum rxe_mr_type { RXE_MR_TYPE_MR, }; +enum rxe_mr_copy_dir { +RXE_TO_MR_OBJ , +RXE_FROM_MR_OBJ , +}; + +enum rxe_mr_lookup_type { +RXE_LOOKUP_LOCAL , +RXE_LOOKUP_REMOTE , +}; +#define RXE_BUF_PER_MAP (PAGE_SIZE / sizeof(struct rxe_phys_buf))struct rxe_phys_buf {
@@ -286,6 +296,13 @@ struct rxe_map {struct rxe_phys_buf buf[RXE_BUF_PER_MAP ]; }; +static inline int rkey_is_mw (u32 rkey) +{ +u32 index = rkey >>8 ; + +return (index >=RXE_MIN_MW_INDEX ) && (index <=RXE_MAX_MW_INDEX ); +} +struct rxe_mr {struct rxe_pool_entry pelem;struct ib_mr ibmr;
@@ -311,23 +328,23 @@ struct rxe_mr {u32 max_buf;u32 num_map; -struct rxe_map **map; -atomic_t num_mw; + +struct rxe_map **map; };enum rxe_mw_state { -RXE_MW_STATE_INVALID =RXE_MR_STATE_INVALID , -RXE_MW_STATE_FREE =RXE_MR_STATE_FREE , -RXE_MW_STATE_VALID =RXE_MR_STATE_VALID , +RXE_MW_STATE_INVALID =RXE_MR_STATE_INVALID , +RXE_MW_STATE_FREE =RXE_MR_STATE_FREE , +RXE_MW_STATE_VALID =RXE_MR_STATE_VALID , };struct rxe_mw { -struct ib_mw ibmw; -struct rxe_pool_entry pelem; +struct ib_mw ibmw; +struct rxe_pool_entry pelem;spinlock_t lock;enum rxe_mw_state state; -struct rxe_qp *qp;/* Type 2 only */ +struct rxe_qp *qp;/* Type 2 only */ struct rxe_mr *mr;int access;u64 addr;
Signature · 1 line
2.30.2
PATCH v8 9/10 RDMA/rxe: Implement memory access through MWs
In reply to Bob PearsonAdd code to implement memory access through memory windows.
--- drivers/infiniband/sw/rxe/rxe_loc.h | 1 + drivers/infiniband/sw/rxe/rxe_mw.c | 23 +++++++++++ drivers/infiniband/sw/rxe/rxe_resp.c | 55 +++++++++++++++++++-------- drivers/infiniband/sw/rxe/rxe_verbs.h | 11 ++++++ 4 files changed, 75 insertions(+), 15 deletions(-)
drivers/infiniband/sw/rxe/rxe_loc.h +1 -0
@@ -94,6 +94,7 @@ int rxe_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata);int rxe_dealloc_mw (struct ib_mw *ibmw);int rxe_bind_mw (struct rxe_qp *qp,struct rxe_send_wqe *wqe);int rxe_invalidate_mw (struct rxe_qp *qp,u32 rkey); +struct rxe_mw *rxe_lookup_mw (struct rxe_qp *qp,int access,u32 rkey);void rxe_mw_cleanup (struct rxe_pool_entry *arg);/* rxe_net.c */
drivers/infiniband/sw/rxe/rxe_mw.c +23 -0
@@ -312,6 +312,29 @@ int rxe_invalidate_mw(struct rxe_qp *qp, u32 rkey)return ret; } +struct rxe_mw *rxe_lookup_mw (struct rxe_qp *qp,int access,u32 rkey) +{ +struct rxe_dev *rxe =to_rdev (qp->ibqp.device); +struct rxe_pd *pd =to_rpd (qp->ibqp.pd); +struct rxe_mw *mw; +int index = rkey >>8 ; + + mw =rxe_pool_get_index (&rxe->mw_pool, index); +if (!mw) +return NULL ; + +if (unlikely ((rxe_mw_rkey (mw) != rkey) ||rxe_mw_pd (mw) != pd || + (mw->ibmw.type ==IB_MW_TYPE_2 && mw->qp != qp) || + (mw->length ==0 ) || + (access && !(access & mw->access)) || + mw->state !=RXE_MW_STATE_VALID )) { +rxe_drop_ref (mw); +return NULL ; + } + +return mw; +} +void rxe_mw_cleanup (struct rxe_pool_entry *elem) {struct rxe_mw *mw =container_of (elem,typeof (*mw), pelem);
drivers/infiniband/sw/rxe/rxe_resp.c +40 -15
112 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_verbs.h +11 -0
@@ -183,6 +183,7 @@ struct rxe_resp_info {/* RDMA read / atomic only */ u64 va; +u64 offset;struct rxe_mr *mr;u32 resid;u32 rkey;
@@ -480,6 +481,16 @@ static inline u32 mr_rkey(struct rxe_mr *mr)return mr->ibmr.rkey; } +static inline struct rxe_pd *rxe_mw_pd (struct rxe_mw *mw) +{ +return to_rpd (mw->ibmw.pd); +} + +static inline u32 rxe_mw_rkey (struct rxe_mw *mw) +{ +return mw->ibmw.rkey; +} +int rxe_register_device (struct rxe_dev *rxe,const char *ibdev_name);void rxe_mc_cleanup (struct rxe_pool_entry *arg);
Signature · 1 line
2.30.2
PATCH v8 10/10 RDMA/rxe: Disallow MR dereg and invalidate when bound
In reply to Bob PearsonCheck that an MR has no bound MWs before allowing a dereg or invalidate operation.
--- drivers/infiniband/sw/rxe/rxe_loc.h | 1 + drivers/infiniband/sw/rxe/rxe_mr.c | 25 +++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_verbs.c | 11 ----------- 3 files changed, 26 insertions(+), 11 deletions(-)
drivers/infiniband/sw/rxe/rxe_loc.h +1 -0
@@ -87,6 +87,7 @@ struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,int mr_check_range (struct rxe_mr *mr,u64 iova,size_t length);int advance_dma_data (struct rxe_dma_info *dma,unsigned int length);int rxe_invalidate_mr (struct rxe_qp *qp,u32 rkey); +int rxe_dereg_mr (struct ib_mr *ibmr,struct ib_udata *udata);void rxe_mr_cleanup (struct rxe_pool_entry *arg);/* rxe_mw.c */
drivers/infiniband/sw/rxe/rxe_mr.c +25 -0
@@ -546,6 +546,13 @@ int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey) goto err_drop_ref; } +if (atomic_read (&mr->num_mw) >0 ) { +pr_warn ("%s: Attempt to invalidate an MR while bound to MWs\n" , + __func__); + ret = -EINVAL ; + goto err_drop_ref; + } + mr->state =RXE_MR_STATE_FREE ; ret =0 ;
@@ -555,6 +562,24 @@ int rxe_invalidate_mr(struct rxe_qp *qp, u32 rkey)return ret; } +int rxe_dereg_mr (struct ib_mr *ibmr,struct ib_udata *udata) +{ +struct rxe_mr *mr =to_rmr (ibmr); + +if (atomic_read (&mr->num_mw) >0 ) { +pr_warn ("%s: Attempt to deregister an MR while bound to MWs\n" , + __func__); +return -EINVAL ; + } + + mr->state =RXE_MR_STATE_ZOMBIE ; +rxe_drop_ref (mr_pd (mr)); +rxe_drop_index (mr); +rxe_drop_ref (mr); + +return 0 ; +} +void rxe_mr_cleanup (struct rxe_pool_entry *arg) {struct rxe_mr *mr =container_of (arg,typeof (*mr), pelem);
drivers/infiniband/sw/rxe/rxe_verbs.c +0 -11
@@ -913,17 +913,6 @@ static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,return ERR_PTR (err); } -static int rxe_dereg_mr (struct ib_mr *ibmr,struct ib_udata *udata) -{ -struct rxe_mr *mr =to_rmr (ibmr); - - mr->state =RXE_MR_STATE_ZOMBIE ; -rxe_drop_ref (mr_pd (mr)); -rxe_drop_index (mr); -rxe_drop_ref (mr); -return 0 ; -} -static struct ib_mr *rxe_alloc_mr (struct ib_pd *ibpd,enum ib_mr_type mr_type,u32 max_num_sg) {
Signature · 1 line
2.30.2
Jason Gunthorpe
maintainerQuoted 25 lines · Bob Pearson
This series of patches implement memory windows for the
rdma_rxedriver. This is a shorter reimplementation of an earlier patch set. They apply to and depend on the current for-next linux rdma tree.Signed-off-by Bob Pearson <[email protected]>---
v8:
Dropped wr.mw.flags in therxe_send_wrstruct inrdma_user_rxe.h.
v7:
Fixed a duplicateINIT_RDMA_OBJ_SIZE(ib_mw, ...)inrxe_verbs.c.
v6:
Addedrxe_prefix to subroutine names in lines that changed from Zhu's review of v5.
v5:
Fixed a typo in 10th patch.
v4:
Added a 10th patch to check when MRs have bound MWs and disallow dereg and invalidate operations.
v3:
cleaned up void return and lower case enums from
Zhu's review.
v2:
cleaned up an issue inrdma_user_rxe.h
cleaned up a collision inrxe_resp.c
Quoted 11 lines
Bob Pearson (10):
RDMA/rxe: Add bind MW fields torxe_send_wrRDMA/rxe: Return errors for add index and key RDMA/rxe: Enable MW object pool
RDMA/rxe: Addib_alloc_mwandib_dealloc_mwverbs RDMA/rxe: ReplaceWR_REG_MASKbyWR_LOCAL_OP_MASKRDMA/rxe: Move local ops to subroutine RDMA/rxe: Add support for bind MW work requests RDMA/rxe: Implement invalidate MW operations RDMA/rxe: Implement memory access through MWs RDMA/rxe: Disallow MR dereg and invalidate when bound
This is all ready now, right?
Can you put the userspace part on the github please?
Jason
Quoted 37 lines · Jason Gunthorpe
Quoted 25 lines · Bob Pearson
This series of patches implement memory windows for the
rdma_rxedriver. This is a shorter reimplementation of an earlier patch set. They apply to and depend on the current for-next linux rdma tree.Signed-off-by Bob Pearson <[email protected]>---
v8:
Dropped wr.mw.flags in therxe_send_wrstruct inrdma_user_rxe.h.
v7:
Fixed a duplicateINIT_RDMA_OBJ_SIZE(ib_mw, ...)inrxe_verbs.c.
v6:
Addedrxe_prefix to subroutine names in lines that changed from Zhu's review of v5.
v5:
Fixed a typo in 10th patch.
v4:
Added a 10th patch to check when MRs have bound MWs and disallow dereg and invalidate operations.
v3:
cleaned up void return and lower case enums from
Zhu's review.
v2:
cleaned up an issue inrdma_user_rxe.h
cleaned up a collision inrxe_resp.cQuoted 11 lines
Bob Pearson (10):
RDMA/rxe: Add bind MW fields torxe_send_wrRDMA/rxe: Return errors for add index and key RDMA/rxe: Enable MW object pool
RDMA/rxe: Addib_alloc_mwandib_dealloc_mwverbs RDMA/rxe: ReplaceWR_REG_MASKbyWR_LOCAL_OP_MASKRDMA/rxe: Move local ops to subroutine RDMA/rxe: Add support for bind MW work requests RDMA/rxe: Implement invalidate MW operations RDMA/rxe: Implement memory access through MWs RDMA/rxe: Disallow MR dereg and invalidate when bound
Quoted 5 lines
This is all ready now, right?
Can you put the userspace part on the github please?
Jason
I think so. The user space change is in my tree at github but I have to admit I forgot how to submit the pull requests.
I think you sent me instructions a while back but I can't find them.
Bob
Zhu Yanjun
maintainer"
[ 1041.051398] rdma_rxe: loaded [ 1041.054536] infiniband rxe0: set active [ 1041.054540] infiniband rxe0: added enp0s8 [ 1086.287975] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1086.311546] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1086.399826] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1090.232785] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1090.255985] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1090.345427] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1094.024322] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1094.047569] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1094.136103] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1098.989344] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1099.015065] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1099.112970] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1103.040076] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1103.062831] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1103.151157] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1116.121772] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1116.144951] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1116.234607] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1131.655486] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1131.678700] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1131.766776] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1175.517166] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1175.540258] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1175.628214] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1190.760122] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1190.783243] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1190.871167] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elem [ 1255.227916] rdma_rxe: unloaded
"
After I added a rxe device on the netdev, then run rdma-core test tools. Then I remove rxe device, in the end, I unloaded rdma_rxe kernel modules. I found the above logs.
"
[ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elem
"
It seems that some resources leak.
I will make further investigations.
Zhu Yanjun
Quoted 44 lines · Jason Gunthorpe
Quoted 25 lines · Bob Pearson
This series of patches implement memory windows for the
rdma_rxedriver. This is a shorter reimplementation of an earlier patch set. They apply to and depend on the current for-next linux rdma tree.Signed-off-by Bob Pearson <[email protected]>---
v8:
Dropped wr.mw.flags in therxe_send_wrstruct inrdma_user_rxe.h.
v7:
Fixed a duplicateINIT_RDMA_OBJ_SIZE(ib_mw, ...)inrxe_verbs.c.
v6:
Addedrxe_prefix to subroutine names in lines that changed from Zhu's review of v5.
v5:
Fixed a typo in 10th patch.
v4:
Added a 10th patch to check when MRs have bound MWs and disallow dereg and invalidate operations.
v3:
cleaned up void return and lower case enums from
Zhu's review.
v2:
cleaned up an issue inrdma_user_rxe.h
cleaned up a collision inrxe_resp.cQuoted 11 lines
Bob Pearson (10):
RDMA/rxe: Add bind MW fields torxe_send_wrRDMA/rxe: Return errors for add index and key RDMA/rxe: Enable MW object pool
RDMA/rxe: Addib_alloc_mwandib_dealloc_mwverbs RDMA/rxe: ReplaceWR_REG_MASKbyWR_LOCAL_OP_MASKRDMA/rxe: Move local ops to subroutine RDMA/rxe: Add support for bind MW work requests RDMA/rxe: Implement invalidate MW operations RDMA/rxe: Implement memory access through MWs RDMA/rxe: Disallow MR dereg and invalidate when boundThis is all ready now, right?
Can you put the userspace part on the github please?
Jason
There is probably a reference to the PD hanging around. I'll look. -- Bob
-----Original Message-----
From: Zhu Yanjun <[email protected]>
Sent: Friday, June 4, 2021 12:38 AM
To: Jason Gunthorpe <[email protected]>
Cc: Bob Pearson <[email protected]>; [email protected]; RDMA mailing list <[email protected]>
Subject: Re: [PATCH for-next v8 00/10] RDMA/rxe: Implement memory windows
"
[ 1041.051398] rdma_rxe: loaded [ 1041.054536] infiniband rxe0: set active [ 1041.054540] infiniband rxe0: added enp0s8 [ 1086.287975] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1086.311546] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1086.399826] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1090.232785] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1090.255985] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1090.345427] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1094.024322] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1094.047569] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1094.136103] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1098.989344] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1099.015065] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1099.112970] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1103.040076] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1103.062831] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1103.151157] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1116.121772] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1116.144951] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1116.234607] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1131.655486] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1131.678700] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1131.766776] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1175.517166] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1175.540258] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1175.628214] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1190.760122] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1190.783243] rdma_rxe: cqe(1) < current # elements in queue (6) [ 1190.871167] rdma_rxe: cqe(32768) > max_cqe(32767) [ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elem [ 1255.227916] rdma_rxe: unloaded "
After I added a rxe device on the netdev, then run rdma-core test tools. Then I remove rxe device, in the end, I unloaded rdma_rxe kernel modules. I found the above logs.
"
[ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elem "
It seems that some resources leak.
I will make further investigations.
Zhu Yanjun
Quoted 44 lines · Jason Gunthorpe
Quoted 25 lines · Bob Pearson
This series of patches implement memory windows for the
rdma_rxedriver. This is a shorter reimplementation of an earlier patch set. They apply to and depend on the current for-next linux rdma tree.Signed-off-by Bob Pearson <[email protected]>---
v8:
Dropped wr.mw.flags in therxe_send_wrstruct inrdma_user_rxe.h.
v7:
Fixed a duplicateINIT_RDMA_OBJ_SIZE(ib_mw, ...)inrxe_verbs.c.
v6:
Addedrxe_prefix to subroutine names in lines that changed from Zhu's review of v5.
v5:
Fixed a typo in 10th patch.
v4:
Added a 10th patch to check when MRs have bound MWs and disallow dereg and invalidate operations.
v3:
cleaned up void return and lower case enums from
Zhu's review.
v2:
cleaned up an issue inrdma_user_rxe.h
cleaned up a collision inrxe_resp.cQuoted 11 lines
Bob Pearson (10):
RDMA/rxe: Add bind MW fields torxe_send_wrRDMA/rxe: Return errors for add index and key RDMA/rxe: Enable MW object pool
RDMA/rxe: Addib_alloc_mwandib_dealloc_mwverbs RDMA/rxe: ReplaceWR_REG_MASKbyWR_LOCAL_OP_MASKRDMA/rxe: Move local ops to subroutine RDMA/rxe: Add support for bind MW work requests RDMA/rxe: Implement invalidate MW operations RDMA/rxe: Implement memory access through MWs RDMA/rxe: Disallow MR dereg and invalidate when boundThis is all ready now, right?
Can you put the userspace part on the github please?
Jason
Quoted 8 lines · Zhu Yanjun
After I added a rxe device on the netdev, then run rdma-core test tools. Then I remove rxe device, in the end, I unloaded
rdma_rxekernel modules. I found the above logs.
"[ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elem
Quoted 7 lines
"
It seems that some resources leak.
I will make further investigations.
Zhu Yanjun
Zhu,
I suspect this is an older error. I traced all the add and drop ref calls for PDs, then ran the full suite of Python tests and also test_mr which includes the memory window tests by itself and then counted the adds and drops. For test_mr alone I get 85 adds and 85 drops but when I run the whole suite I get 384 adds and 380 drops. Since the memory window code is only exercised in test_mr I think it is OK. Somewhere else there are missing drops. I will try to isolate them.
Bob
Quoted 16 lines · Robert B
Quoted 14 lines · Zhu Yanjun
After I added a rxe device on the netdev, then run rdma-core test tools. Then I remove rxe device, in the end, I unloaded
rdma_rxekernel modules. I found the above logs.
"[ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elem"
It seems that some resources leak.
I will make further investigations.
Quoted 8 lines
Quoted 1 line
Zhu Yanjun
Zhu,
I suspect this is an older error. I traced all the add and drop ref calls for PDs, then ran the full suite of Python tests and also
test_mrwhich includes the memory window tests by itself and then counted the adds and drops. Fortest_mralone I get 85 adds and 85 drops but when I run the whole suite I get 384 adds and 380 drops. Since the memory window code is only exercised intest_mrI think it is OK. Somewhere else there are missing drops. I will try to isolate them.Bob
Zhu,
In rdma_core/tests/test_qpex.py test_qp_ex_rc_atomic_cmp_swp and test_qp_ex_rc_atomic_fetch_add each have two missing drops of PDs. This is either a test bug or a bug in the rxe driver but it has nothing to do with the MW code. We should treat it as a separate error. For some reason these test cases are not cleaning up all resources.
The cleanup code in all these Python tests is very implicit. It just happens by magic so it is hard to figure out where an ibv_destroy_qp or ibv_destroy_cq went missing. It would help if someone who is familiar with these tests could look at it.
Bob
Jason Gunthorpe
maintainerQuoted 41 lines · Robert B
Quoted 29 lines · Robert B
Quoted 9 lines · Zhu Yanjun
After I added a rxe device on the netdev, then run rdma-core test tools. Then I remove rxe device, in the end, I unloaded
rdma_rxekernel modules.
I found the above logs.
"[ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elemQuoted 7 lines
"
It seems that some resources leak.
I will make further investigations.
Zhu Yanjun
Zhu,
I suspect this is an older error. I traced all the add and drop ref calls for PDs, then ran the full suite of Python tests and also
test_mrwhich includes the memory window tests by itself and then counted the adds and drops. Fortest_mralone I get 85 adds and 85 drops but when I run the whole suite I get 384 adds and 380 drops. Since the memory window code is only exercised intest_mrI think it is OK. Somewhere else there are missing drops. I will try to isolate them.Quoted 2 lines
Bob
Zhu,
In
rdma_core/tests/test_qpex.pytest_qp_ex_rc_atomic_cmp_swpandtest_qp_ex_rc_atomic_fetch_addeach have two missing drops of PDs. This is either a test bug or a bug in the rxe driver but it has nothing to do with the MW code. We should treat it as a separate error. For some reason these test cases are not cleaning up all resources.
Quoted 4 lines
The cleanup code in all these Python tests is very implicit. It just happens by magic so it is hard to figure out where an
ibv_destroy_qporibv_destroy_cqwent missing. It would help if someone who is familiar with these tests could look at it.
It is impossible for userspace to leak a kernel resource, when the fd is closed everything is destroyed back to the driver guarenteed by the kernel.
As long as pyverbs has exited pyverbs cannot be the bug
Jason
Quoted 42 lines · Jason Gunthorpe
Quoted 37 lines · Robert B
Quoted 26 lines · Robert B
Quoted 8 lines · Zhu Yanjun
After I added a rxe device on the netdev, then run rdma-core test tools. Then I remove rxe device, in the end, I unloaded
rdma_rxekernel modules.
I found the above logs.
"[ 1249.651921] rdma_rxe: rxe-pd pool destroyed with unfree'd elem [ 1249.651927] rdma_rxe: rxe-qp pool destroyed with unfree'd elem [ 1249.651929] rdma_rxe: rxe-cq pool destroyed with unfree'd elemQuoted 7 lines
"
It seems that some resources leak.
I will make further investigations.
Zhu Yanjun
Zhu,
I suspect this is an older error. I traced all the add and drop ref calls for PDs, then ran the full suite of Python tests and also
test_mrwhich includes the memory window tests by itself and then counted the adds and drops. Fortest_mralone I get 85 adds and 85 drops but when I run the whole suite I get 384 adds and 380 drops. Since the memory window code is only exercised intest_mrI think it is OK. Somewhere else there are missing drops. I will try to isolate them.Quoted 2 lines
Bob
Zhu,
In
rdma_core/tests/test_qpex.pytest_qp_ex_rc_atomic_cmp_swpandtest_qp_ex_rc_atomic_fetch_addeach have two missing drops of PDs. This is either a test bug or a bug in the rxe driver but it has nothing to do with the MW code. We should treat it as a separate error. For some reason these test cases are not cleaning up all resources.Quoted 4 lines
The cleanup code in all these Python tests is very implicit. It just happens by magic so it is hard to figure out where an
ibv_destroy_qporibv_destroy_cqwent missing. It would help if someone who is familiar with these tests could look at it.
Quoted 7 lines
It is impossible for userspace to leak a kernel resource, when the fd is closed everything is destroyed back to the driver guarenteed by the kernel.
As long as pyverbs has exited pyverbs cannot be the bug
Jason
Thanks. That helped. Adding traces for QP references turned up the problem. Someone took a reference on QP in send_atomic_ack() that is never matched by a drop reference. The logic was probably to protect the QP from going away while it held an atomic responder resource. The problem is that since the requester can retry the operation multiple times the responder never knows when to free the resource so it doesn't. It just recycles them FIFO when a new atomic request comes along. They are looked up by PSN. So the only logical solution is to not take the extra reference. If you destroy the QP while a requester is retrying atomic operations it just fails.
I'll submit a patch deleting one line.
Bob