RDMA/mlx5: Lower setting the umem's PAS for SRQ

Some of the SRQ types are created using a WQ, and the WQ requires a different parameter set to mlx5_umem_find_best_quantized_pgoff() as it has a 5 bit page_offset. Add the umem to the mlx5_srq_attr and defer computing the PAS data until the code has figured out what kind of mailbox to use. Compute the PAS directly from the umem for each of the four unique mailbox types. This also avoids allocating memory to store the user PAS, instead it is written directly to the mailbox as in most other cases.

Jason Gunthorpe committed 6 years ago

Commit 8a7904a1 parent 878f7b33 files · +79 −29Browse files

Unified · Split

3 files

drivers/infiniband/hw/mlx5/srq.c +2 −25
drivers/infiniband/hw/mlx5/srq.h +1 −0
drivers/infiniband/hw/mlx5/srq_cmd.c +76 −4

drivers/infiniband/hw/mlx5/srq.c+2−25

@@ -51,8 +51,6 @@ static int create_srq_user(struct ib_pd pd, struct mlx5_ib_srq srq, · 50 unmodified lines
51	51	udata, struct mlx5_ib_ucontext, ibucontext);
52	52	size_t ucmdlen;
53	53	int err;
54		− unsigned int page_offset_quantized;
55		− unsigned int page_size;
56	54	u32 uidx = MLX5_IB_DEFAULT_UIDX;
57	55
58	56	ucmdlen = min(udata->inlen, sizeof(ucmd));
@@ -84,32 +82,14 @@ static int create_srq_user(struct ib_pd pd, struct mlx5_ib_srq srq, · 25 unmodified lines
84	82	err = PTR_ERR(srq->umem);
85	83	return err;
86	84	}
87		−
88		− page_size = mlx5_umem_find_best_quantized_pgoff(
89		− srq->umem, srqc, log_page_size, MLX5_ADAPTER_PAGE_SHIFT,
90		− page_offset, 64, &page_offset_quantized);
91		− if (!page_size) {
92		− mlx5_ib_warn(dev, "bad offset\n");
93		− goto err_umem;
94		− }
95		−
96		− in->pas = kvcalloc(ib_umem_num_dma_blocks(srq->umem, page_size),
97		− sizeof(*in->pas), GFP_KERNEL);
98		− if (!in->pas) {
99		− err = -ENOMEM;
100		− goto err_umem;
101		− }
102		−
103		− mlx5_ib_populate_pas(srq->umem, page_size, in->pas, 0);
	85	+ in->umem = srq->umem;
104	86
105	87	err = mlx5_ib_db_map_user(ucontext, udata, ucmd.db_addr, &srq->db);
106	88	if (err) {
107	89	mlx5_ib_dbg(dev, "map doorbell failed\n");
108		− goto err_in;
	90	+ goto err_umem;
109	91	}
110	92
111		− in->log_page_size = order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT;
112		− in->page_offset = page_offset_quantized;
113	93	in->uid = (in->type != IB_SRQT_XRC) ? to_mpd(pd)->uid : 0;
114	94	if (MLX5_CAP_GEN(dev->mdev, cqe_version) == MLX5_CQE_VERSION_V1 &&
115	95	in->type != IB_SRQT_BASIC)
@@ -117,9 +97,6 @@ static int create_srq_user(struct ib_pd pd, struct mlx5_ib_srq srq, · 1 unmodified line
117	97
118	98	return 0;
119	99
120		−err_in:
121		− kvfree(in->pas);
122		−
123	100	err_umem:
124	101	ib_umem_release(srq->umem);
125	102

drivers/infiniband/hw/mlx5/srq.h+1−0

@@ -28,6 +28,7 @@ struct mlx5_srq_attr { · 27 unmodified lines
28	28	u32 user_index;
29	29	u64 db_record;
30	30	__be64 *pas;
	31	+ struct ib_umem *umem;
31	32	u32 tm_log_list_size;
32	33	u32 tm_next_tag;
33	34	u32 tm_hw_phase_cnt;

drivers/infiniband/hw/mlx5/srq_cmd.c+76−4

@@ -92,6 +92,25 @@ struct mlx5_core_srq mlx5_cmd_get_srq(struct mlx5_ib_dev dev, u32 srqn) · 91 unmodified lines
92	92	return srq;
93	93	}
94	94
	95	+static int __set_srq_page_size(struct mlx5_srq_attr *in,
	96	+ unsigned long page_size)
	97	+{
	98	+ if (!page_size)
	99	+ return -EINVAL;
	100	+ in->log_page_size = order_base_2(page_size) - MLX5_ADAPTER_PAGE_SHIFT;
	101	+
	102	+ if (WARN_ON(get_pas_size(in) !=
	103	+ ib_umem_num_dma_blocks(in->umem, page_size) * sizeof(u64)))
	104	+ return -EINVAL;
	105	+ return 0;
	106	+}
	107	+
	108	+#define set_srq_page_size(in, typ, log_pgsz_fld) \
	109	+ __set_srq_page_size(in, mlx5_umem_find_best_quantized_pgoff( \
	110	+ (in)->umem, typ, log_pgsz_fld, \
	111	+ MLX5_ADAPTER_PAGE_SHIFT, page_offset, \
	112	+ 64, &(in)->page_offset))
	113	+
95	114	static int create_srq_cmd(struct mlx5_ib_dev dev, struct mlx5_core_srq srq,
96	115	struct mlx5_srq_attr *in)
97	116	{
@@ -103,6 +122,12 @@ static int create_srq_cmd(struct mlx5_ib_dev dev, struct mlx5_core_srq srq, · 5 unmodified lines
103	122	int inlen;
104	123	int err;
105	124
	125	+ if (in->umem) {
	126	+ err = set_srq_page_size(in, srqc, log_page_size);
	127	+ if (err)
	128	+ return err;
	129	+ }
	130	+
106	131	pas_size = get_pas_size(in);
107	132	inlen = MLX5_ST_SZ_BYTES(create_srq_in) + pas_size;
108	133	create_in = kvzalloc(inlen, GFP_KERNEL);
@@ -114,7 +139,13 @@ static int create_srq_cmd(struct mlx5_ib_dev dev, struct mlx5_core_srq srq, · 5 unmodified lines
114	139	pas = MLX5_ADDR_OF(create_srq_in, create_in, pas);
115	140
116	141	set_srqc(srqc, in);
117		− memcpy(pas, in->pas, pas_size);
	142	+ if (in->umem)
	143	+ mlx5_ib_populate_pas(
	144	+ in->umem,
	145	+ 1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
	146	+ pas, 0);
	147	+ else
	148	+ memcpy(pas, in->pas, pas_size);
118	149
119	150	MLX5_SET(create_srq_in, create_in, opcode,
120	151	MLX5_CMD_OP_CREATE_SRQ);
@@ -194,6 +225,12 @@ static int create_xrc_srq_cmd(struct mlx5_ib_dev *dev, · 73 unmodified lines
194	225	int inlen;
195	226	int err;
196	227
	228	+ if (in->umem) {
	229	+ err = set_srq_page_size(in, xrc_srqc, log_page_size);
	230	+ if (err)
	231	+ return err;
	232	+ }
	233	+
197	234	pas_size = get_pas_size(in);
198	235	inlen = MLX5_ST_SZ_BYTES(create_xrc_srq_in) + pas_size;
199	236	create_in = kvzalloc(inlen, GFP_KERNEL);
@@ -207,7 +244,13 @@ static int create_xrc_srq_cmd(struct mlx5_ib_dev *dev, · 7 unmodified lines
207	244
208	245	set_srqc(xrc_srqc, in);
209	246	MLX5_SET(xrc_srqc, xrc_srqc, user_index, in->user_index);
210		− memcpy(pas, in->pas, pas_size);
	247	+ if (in->umem)
	248	+ mlx5_ib_populate_pas(
	249	+ in->umem,
	250	+ 1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
	251	+ pas, 0);
	252	+ else
	253	+ memcpy(pas, in->pas, pas_size);
211	254	MLX5_SET(create_xrc_srq_in, create_in, opcode,
212	255	MLX5_CMD_OP_CREATE_XRC_SRQ);
213	256
@@ -289,11 +332,18 @@ static int create_rmp_cmd(struct mlx5_ib_dev dev, struct mlx5_core_srq srq, · 75 unmodified lines
289	332	void *create_in = NULL;
290	333	void *rmpc;
291	334	void *wq;
	335	+ void *pas;
292	336	int pas_size;
293	337	int outlen;
294	338	int inlen;
295	339	int err;
296	340
	341	+ if (in->umem) {
	342	+ err = set_srq_page_size(in, wq, log_wq_pg_sz);
	343	+ if (err)
	344	+ return err;
	345	+ }
	346	+
297	347	pas_size = get_pas_size(in);
298	348	inlen = MLX5_ST_SZ_BYTES(create_rmp_in) + pas_size;
299	349	outlen = MLX5_ST_SZ_BYTES(create_rmp_out);
@@ -309,8 +359,16 @@ static int create_rmp_cmd(struct mlx5_ib_dev dev, struct mlx5_core_srq srq, · 9 unmodified lines
309	359
310	360	MLX5_SET(rmpc, rmpc, state, MLX5_RMPC_STATE_RDY);
311	361	MLX5_SET(create_rmp_in, create_in, uid, in->uid);
	362	+ pas = MLX5_ADDR_OF(rmpc, rmpc, wq.pas);
	363	+
312	364	set_wq(wq, in);
313		− memcpy(MLX5_ADDR_OF(rmpc, rmpc, wq.pas), in->pas, pas_size);
	365	+ if (in->umem)
	366	+ mlx5_ib_populate_pas(
	367	+ in->umem,
	368	+ 1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
	369	+ pas, 0);
	370	+ else
	371	+ memcpy(pas, in->pas, pas_size);
314	372
315	373	MLX5_SET(create_rmp_in, create_in, opcode, MLX5_CMD_OP_CREATE_RMP);
316	374	err = mlx5_cmd_exec(dev->mdev, create_in, inlen, create_out, outlen);
@@ -421,10 +479,17 @@ static int create_xrq_cmd(struct mlx5_ib_dev dev, struct mlx5_core_srq srq, · 104 unmodified lines
421	479	void *create_in;
422	480	void *xrqc;
423	481	void *wq;
	482	+ void *pas;
424	483	int pas_size;
425	484	int inlen;
426	485	int err;
427	486
	487	+ if (in->umem) {
	488	+ err = set_srq_page_size(in, wq, log_wq_pg_sz);
	489	+ if (err)
	490	+ return err;
	491	+ }
	492	+
428	493	pas_size = get_pas_size(in);
429	494	inlen = MLX5_ST_SZ_BYTES(create_xrq_in) + pas_size;
430	495	create_in = kvzalloc(inlen, GFP_KERNEL);
@@ -433,9 +498,16 @@ static int create_xrq_cmd(struct mlx5_ib_dev dev, struct mlx5_core_srq srq, · 2 unmodified lines
433	498
434	499	xrqc = MLX5_ADDR_OF(create_xrq_in, create_in, xrq_context);
435	500	wq = MLX5_ADDR_OF(xrqc, xrqc, wq);
	501	+ pas = MLX5_ADDR_OF(xrqc, xrqc, wq.pas);
436	502
437	503	set_wq(wq, in);
438		− memcpy(MLX5_ADDR_OF(xrqc, xrqc, wq.pas), in->pas, pas_size);
	504	+ if (in->umem)
	505	+ mlx5_ib_populate_pas(
	506	+ in->umem,
	507	+ 1UL << (in->log_page_size + MLX5_ADAPTER_PAGE_SHIFT),
	508	+ pas, 0);
	509	+ else
	510	+ memcpy(pas, in->pas, pas_size);
439	511
440	512	if (in->type == IB_SRQT_TM) {
441	513	MLX5_SET(xrqc, xrqc, topology, MLX5_XRQC_TOPOLOGY_TAG_MATCHING);