xref: /openbmc/linux/drivers/infiniband/core/rw.c (revision 5a7a9e03)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2016 HGST, a Western Digital Company.
4  */
5 #include <linux/moduleparam.h>
6 #include <linux/slab.h>
7 #include <linux/pci-p2pdma.h>
8 #include <rdma/mr_pool.h>
9 #include <rdma/rw.h>
10 
11 enum {
12 	RDMA_RW_SINGLE_WR,
13 	RDMA_RW_MULTI_WR,
14 	RDMA_RW_MR,
15 	RDMA_RW_SIG_MR,
16 };
17 
18 static bool rdma_rw_force_mr;
19 module_param_named(force_mr, rdma_rw_force_mr, bool, 0);
20 MODULE_PARM_DESC(force_mr, "Force usage of MRs for RDMA READ/WRITE operations");
21 
22 /*
23  * Report whether memory registration should be used. Memory registration must
24  * be used for iWarp devices because of iWARP-specific limitations. Memory
25  * registration is also enabled if registering memory might yield better
26  * performance than using multiple SGE entries, see rdma_rw_io_needs_mr()
27  */
28 static inline bool rdma_rw_can_use_mr(struct ib_device *dev, u8 port_num)
29 {
30 	if (rdma_protocol_iwarp(dev, port_num))
31 		return true;
32 	if (dev->attrs.max_sgl_rd)
33 		return true;
34 	if (unlikely(rdma_rw_force_mr))
35 		return true;
36 	return false;
37 }
38 
39 /*
40  * Check if the device will use memory registration for this RW operation.
41  * For RDMA READs we must use MRs on iWarp and can optionally use them as an
42  * optimization otherwise.  Additionally we have a debug option to force usage
43  * of MRs to help testing this code path.
44  */
45 static inline bool rdma_rw_io_needs_mr(struct ib_device *dev, u8 port_num,
46 		enum dma_data_direction dir, int dma_nents)
47 {
48 	if (dir == DMA_FROM_DEVICE) {
49 		if (rdma_protocol_iwarp(dev, port_num))
50 			return true;
51 		if (dev->attrs.max_sgl_rd && dma_nents > dev->attrs.max_sgl_rd)
52 			return true;
53 	}
54 	if (unlikely(rdma_rw_force_mr))
55 		return true;
56 	return false;
57 }
58 
59 static inline u32 rdma_rw_fr_page_list_len(struct ib_device *dev,
60 					   bool pi_support)
61 {
62 	u32 max_pages;
63 
64 	if (pi_support)
65 		max_pages = dev->attrs.max_pi_fast_reg_page_list_len;
66 	else
67 		max_pages = dev->attrs.max_fast_reg_page_list_len;
68 
69 	/* arbitrary limit to avoid allocating gigantic resources */
70 	return min_t(u32, max_pages, 256);
71 }
72 
73 static inline int rdma_rw_inv_key(struct rdma_rw_reg_ctx *reg)
74 {
75 	int count = 0;
76 
77 	if (reg->mr->need_inval) {
78 		reg->inv_wr.opcode = IB_WR_LOCAL_INV;
79 		reg->inv_wr.ex.invalidate_rkey = reg->mr->lkey;
80 		reg->inv_wr.next = &reg->reg_wr.wr;
81 		count++;
82 	} else {
83 		reg->inv_wr.next = NULL;
84 	}
85 
86 	return count;
87 }
88 
89 /* Caller must have zero-initialized *reg. */
90 static int rdma_rw_init_one_mr(struct ib_qp *qp, u8 port_num,
91 		struct rdma_rw_reg_ctx *reg, struct scatterlist *sg,
92 		u32 sg_cnt, u32 offset)
93 {
94 	u32 pages_per_mr = rdma_rw_fr_page_list_len(qp->pd->device,
95 						    qp->integrity_en);
96 	u32 nents = min(sg_cnt, pages_per_mr);
97 	int count = 0, ret;
98 
99 	reg->mr = ib_mr_pool_get(qp, &qp->rdma_mrs);
100 	if (!reg->mr)
101 		return -EAGAIN;
102 
103 	count += rdma_rw_inv_key(reg);
104 
105 	ret = ib_map_mr_sg(reg->mr, sg, nents, &offset, PAGE_SIZE);
106 	if (ret < 0 || ret < nents) {
107 		ib_mr_pool_put(qp, &qp->rdma_mrs, reg->mr);
108 		return -EINVAL;
109 	}
110 
111 	reg->reg_wr.wr.opcode = IB_WR_REG_MR;
112 	reg->reg_wr.mr = reg->mr;
113 	reg->reg_wr.access = IB_ACCESS_LOCAL_WRITE;
114 	if (rdma_protocol_iwarp(qp->device, port_num))
115 		reg->reg_wr.access |= IB_ACCESS_REMOTE_WRITE;
116 	count++;
117 
118 	reg->sge.addr = reg->mr->iova;
119 	reg->sge.length = reg->mr->length;
120 	return count;
121 }
122 
123 static int rdma_rw_init_mr_wrs(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
124 		u8 port_num, struct scatterlist *sg, u32 sg_cnt, u32 offset,
125 		u64 remote_addr, u32 rkey, enum dma_data_direction dir)
126 {
127 	struct rdma_rw_reg_ctx *prev = NULL;
128 	u32 pages_per_mr = rdma_rw_fr_page_list_len(qp->pd->device,
129 						    qp->integrity_en);
130 	int i, j, ret = 0, count = 0;
131 
132 	ctx->nr_ops = DIV_ROUND_UP(sg_cnt, pages_per_mr);
133 	ctx->reg = kcalloc(ctx->nr_ops, sizeof(*ctx->reg), GFP_KERNEL);
134 	if (!ctx->reg) {
135 		ret = -ENOMEM;
136 		goto out;
137 	}
138 
139 	for (i = 0; i < ctx->nr_ops; i++) {
140 		struct rdma_rw_reg_ctx *reg = &ctx->reg[i];
141 		u32 nents = min(sg_cnt, pages_per_mr);
142 
143 		ret = rdma_rw_init_one_mr(qp, port_num, reg, sg, sg_cnt,
144 				offset);
145 		if (ret < 0)
146 			goto out_free;
147 		count += ret;
148 
149 		if (prev) {
150 			if (reg->mr->need_inval)
151 				prev->wr.wr.next = &reg->inv_wr;
152 			else
153 				prev->wr.wr.next = &reg->reg_wr.wr;
154 		}
155 
156 		reg->reg_wr.wr.next = &reg->wr.wr;
157 
158 		reg->wr.wr.sg_list = &reg->sge;
159 		reg->wr.wr.num_sge = 1;
160 		reg->wr.remote_addr = remote_addr;
161 		reg->wr.rkey = rkey;
162 		if (dir == DMA_TO_DEVICE) {
163 			reg->wr.wr.opcode = IB_WR_RDMA_WRITE;
164 		} else if (!rdma_cap_read_inv(qp->device, port_num)) {
165 			reg->wr.wr.opcode = IB_WR_RDMA_READ;
166 		} else {
167 			reg->wr.wr.opcode = IB_WR_RDMA_READ_WITH_INV;
168 			reg->wr.wr.ex.invalidate_rkey = reg->mr->lkey;
169 		}
170 		count++;
171 
172 		remote_addr += reg->sge.length;
173 		sg_cnt -= nents;
174 		for (j = 0; j < nents; j++)
175 			sg = sg_next(sg);
176 		prev = reg;
177 		offset = 0;
178 	}
179 
180 	if (prev)
181 		prev->wr.wr.next = NULL;
182 
183 	ctx->type = RDMA_RW_MR;
184 	return count;
185 
186 out_free:
187 	while (--i >= 0)
188 		ib_mr_pool_put(qp, &qp->rdma_mrs, ctx->reg[i].mr);
189 	kfree(ctx->reg);
190 out:
191 	return ret;
192 }
193 
194 static int rdma_rw_init_map_wrs(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
195 		struct scatterlist *sg, u32 sg_cnt, u32 offset,
196 		u64 remote_addr, u32 rkey, enum dma_data_direction dir)
197 {
198 	u32 max_sge = dir == DMA_TO_DEVICE ? qp->max_write_sge :
199 		      qp->max_read_sge;
200 	struct ib_sge *sge;
201 	u32 total_len = 0, i, j;
202 
203 	ctx->nr_ops = DIV_ROUND_UP(sg_cnt, max_sge);
204 
205 	ctx->map.sges = sge = kcalloc(sg_cnt, sizeof(*sge), GFP_KERNEL);
206 	if (!ctx->map.sges)
207 		goto out;
208 
209 	ctx->map.wrs = kcalloc(ctx->nr_ops, sizeof(*ctx->map.wrs), GFP_KERNEL);
210 	if (!ctx->map.wrs)
211 		goto out_free_sges;
212 
213 	for (i = 0; i < ctx->nr_ops; i++) {
214 		struct ib_rdma_wr *rdma_wr = &ctx->map.wrs[i];
215 		u32 nr_sge = min(sg_cnt, max_sge);
216 
217 		if (dir == DMA_TO_DEVICE)
218 			rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
219 		else
220 			rdma_wr->wr.opcode = IB_WR_RDMA_READ;
221 		rdma_wr->remote_addr = remote_addr + total_len;
222 		rdma_wr->rkey = rkey;
223 		rdma_wr->wr.num_sge = nr_sge;
224 		rdma_wr->wr.sg_list = sge;
225 
226 		for (j = 0; j < nr_sge; j++, sg = sg_next(sg)) {
227 			sge->addr = sg_dma_address(sg) + offset;
228 			sge->length = sg_dma_len(sg) - offset;
229 			sge->lkey = qp->pd->local_dma_lkey;
230 
231 			total_len += sge->length;
232 			sge++;
233 			sg_cnt--;
234 			offset = 0;
235 		}
236 
237 		rdma_wr->wr.next = i + 1 < ctx->nr_ops ?
238 			&ctx->map.wrs[i + 1].wr : NULL;
239 	}
240 
241 	ctx->type = RDMA_RW_MULTI_WR;
242 	return ctx->nr_ops;
243 
244 out_free_sges:
245 	kfree(ctx->map.sges);
246 out:
247 	return -ENOMEM;
248 }
249 
250 static int rdma_rw_init_single_wr(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
251 		struct scatterlist *sg, u32 offset, u64 remote_addr, u32 rkey,
252 		enum dma_data_direction dir)
253 {
254 	struct ib_rdma_wr *rdma_wr = &ctx->single.wr;
255 
256 	ctx->nr_ops = 1;
257 
258 	ctx->single.sge.lkey = qp->pd->local_dma_lkey;
259 	ctx->single.sge.addr = sg_dma_address(sg) + offset;
260 	ctx->single.sge.length = sg_dma_len(sg) - offset;
261 
262 	memset(rdma_wr, 0, sizeof(*rdma_wr));
263 	if (dir == DMA_TO_DEVICE)
264 		rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
265 	else
266 		rdma_wr->wr.opcode = IB_WR_RDMA_READ;
267 	rdma_wr->wr.sg_list = &ctx->single.sge;
268 	rdma_wr->wr.num_sge = 1;
269 	rdma_wr->remote_addr = remote_addr;
270 	rdma_wr->rkey = rkey;
271 
272 	ctx->type = RDMA_RW_SINGLE_WR;
273 	return 1;
274 }
275 
276 static void rdma_rw_unmap_sg(struct ib_device *dev, struct scatterlist *sg,
277 			     u32 sg_cnt, enum dma_data_direction dir)
278 {
279 	if (is_pci_p2pdma_page(sg_page(sg)))
280 		pci_p2pdma_unmap_sg(dev->dma_device, sg, sg_cnt, dir);
281 	else
282 		ib_dma_unmap_sg(dev, sg, sg_cnt, dir);
283 }
284 
285 static int rdma_rw_map_sg(struct ib_device *dev, struct scatterlist *sg,
286 			  u32 sg_cnt, enum dma_data_direction dir)
287 {
288 	if (is_pci_p2pdma_page(sg_page(sg))) {
289 		if (WARN_ON_ONCE(ib_uses_virt_dma(dev)))
290 			return 0;
291 		return pci_p2pdma_map_sg(dev->dma_device, sg, sg_cnt, dir);
292 	}
293 	return ib_dma_map_sg(dev, sg, sg_cnt, dir);
294 }
295 
296 /**
297  * rdma_rw_ctx_init - initialize a RDMA READ/WRITE context
298  * @ctx:	context to initialize
299  * @qp:		queue pair to operate on
300  * @port_num:	port num to which the connection is bound
301  * @sg:		scatterlist to READ/WRITE from/to
302  * @sg_cnt:	number of entries in @sg
303  * @sg_offset:	current byte offset into @sg
304  * @remote_addr:remote address to read/write (relative to @rkey)
305  * @rkey:	remote key to operate on
306  * @dir:	%DMA_TO_DEVICE for RDMA WRITE, %DMA_FROM_DEVICE for RDMA READ
307  *
308  * Returns the number of WQEs that will be needed on the workqueue if
309  * successful, or a negative error code.
310  */
311 int rdma_rw_ctx_init(struct rdma_rw_ctx *ctx, struct ib_qp *qp, u8 port_num,
312 		struct scatterlist *sg, u32 sg_cnt, u32 sg_offset,
313 		u64 remote_addr, u32 rkey, enum dma_data_direction dir)
314 {
315 	struct ib_device *dev = qp->pd->device;
316 	int ret;
317 
318 	ret = rdma_rw_map_sg(dev, sg, sg_cnt, dir);
319 	if (!ret)
320 		return -ENOMEM;
321 	sg_cnt = ret;
322 
323 	/*
324 	 * Skip to the S/G entry that sg_offset falls into:
325 	 */
326 	for (;;) {
327 		u32 len = sg_dma_len(sg);
328 
329 		if (sg_offset < len)
330 			break;
331 
332 		sg = sg_next(sg);
333 		sg_offset -= len;
334 		sg_cnt--;
335 	}
336 
337 	ret = -EIO;
338 	if (WARN_ON_ONCE(sg_cnt == 0))
339 		goto out_unmap_sg;
340 
341 	if (rdma_rw_io_needs_mr(qp->device, port_num, dir, sg_cnt)) {
342 		ret = rdma_rw_init_mr_wrs(ctx, qp, port_num, sg, sg_cnt,
343 				sg_offset, remote_addr, rkey, dir);
344 	} else if (sg_cnt > 1) {
345 		ret = rdma_rw_init_map_wrs(ctx, qp, sg, sg_cnt, sg_offset,
346 				remote_addr, rkey, dir);
347 	} else {
348 		ret = rdma_rw_init_single_wr(ctx, qp, sg, sg_offset,
349 				remote_addr, rkey, dir);
350 	}
351 
352 	if (ret < 0)
353 		goto out_unmap_sg;
354 	return ret;
355 
356 out_unmap_sg:
357 	rdma_rw_unmap_sg(dev, sg, sg_cnt, dir);
358 	return ret;
359 }
360 EXPORT_SYMBOL(rdma_rw_ctx_init);
361 
362 /**
363  * rdma_rw_ctx_signature_init - initialize a RW context with signature offload
364  * @ctx:	context to initialize
365  * @qp:		queue pair to operate on
366  * @port_num:	port num to which the connection is bound
367  * @sg:		scatterlist to READ/WRITE from/to
368  * @sg_cnt:	number of entries in @sg
369  * @prot_sg:	scatterlist to READ/WRITE protection information from/to
370  * @prot_sg_cnt: number of entries in @prot_sg
371  * @sig_attrs:	signature offloading algorithms
372  * @remote_addr:remote address to read/write (relative to @rkey)
373  * @rkey:	remote key to operate on
374  * @dir:	%DMA_TO_DEVICE for RDMA WRITE, %DMA_FROM_DEVICE for RDMA READ
375  *
376  * Returns the number of WQEs that will be needed on the workqueue if
377  * successful, or a negative error code.
378  */
379 int rdma_rw_ctx_signature_init(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
380 		u8 port_num, struct scatterlist *sg, u32 sg_cnt,
381 		struct scatterlist *prot_sg, u32 prot_sg_cnt,
382 		struct ib_sig_attrs *sig_attrs,
383 		u64 remote_addr, u32 rkey, enum dma_data_direction dir)
384 {
385 	struct ib_device *dev = qp->pd->device;
386 	u32 pages_per_mr = rdma_rw_fr_page_list_len(qp->pd->device,
387 						    qp->integrity_en);
388 	struct ib_rdma_wr *rdma_wr;
389 	int count = 0, ret;
390 
391 	if (sg_cnt > pages_per_mr || prot_sg_cnt > pages_per_mr) {
392 		pr_err("SG count too large: sg_cnt=%d, prot_sg_cnt=%d, pages_per_mr=%d\n",
393 		       sg_cnt, prot_sg_cnt, pages_per_mr);
394 		return -EINVAL;
395 	}
396 
397 	ret = rdma_rw_map_sg(dev, sg, sg_cnt, dir);
398 	if (!ret)
399 		return -ENOMEM;
400 	sg_cnt = ret;
401 
402 	if (prot_sg_cnt) {
403 		ret = rdma_rw_map_sg(dev, prot_sg, prot_sg_cnt, dir);
404 		if (!ret) {
405 			ret = -ENOMEM;
406 			goto out_unmap_sg;
407 		}
408 		prot_sg_cnt = ret;
409 	}
410 
411 	ctx->type = RDMA_RW_SIG_MR;
412 	ctx->nr_ops = 1;
413 	ctx->reg = kcalloc(1, sizeof(*ctx->reg), GFP_KERNEL);
414 	if (!ctx->reg) {
415 		ret = -ENOMEM;
416 		goto out_unmap_prot_sg;
417 	}
418 
419 	ctx->reg->mr = ib_mr_pool_get(qp, &qp->sig_mrs);
420 	if (!ctx->reg->mr) {
421 		ret = -EAGAIN;
422 		goto out_free_ctx;
423 	}
424 
425 	count += rdma_rw_inv_key(ctx->reg);
426 
427 	memcpy(ctx->reg->mr->sig_attrs, sig_attrs, sizeof(struct ib_sig_attrs));
428 
429 	ret = ib_map_mr_sg_pi(ctx->reg->mr, sg, sg_cnt, NULL, prot_sg,
430 			      prot_sg_cnt, NULL, SZ_4K);
431 	if (unlikely(ret)) {
432 		pr_err("failed to map PI sg (%d)\n", sg_cnt + prot_sg_cnt);
433 		goto out_destroy_sig_mr;
434 	}
435 
436 	ctx->reg->reg_wr.wr.opcode = IB_WR_REG_MR_INTEGRITY;
437 	ctx->reg->reg_wr.wr.wr_cqe = NULL;
438 	ctx->reg->reg_wr.wr.num_sge = 0;
439 	ctx->reg->reg_wr.wr.send_flags = 0;
440 	ctx->reg->reg_wr.access = IB_ACCESS_LOCAL_WRITE;
441 	if (rdma_protocol_iwarp(qp->device, port_num))
442 		ctx->reg->reg_wr.access |= IB_ACCESS_REMOTE_WRITE;
443 	ctx->reg->reg_wr.mr = ctx->reg->mr;
444 	ctx->reg->reg_wr.key = ctx->reg->mr->lkey;
445 	count++;
446 
447 	ctx->reg->sge.addr = ctx->reg->mr->iova;
448 	ctx->reg->sge.length = ctx->reg->mr->length;
449 	if (sig_attrs->wire.sig_type == IB_SIG_TYPE_NONE)
450 		ctx->reg->sge.length -= ctx->reg->mr->sig_attrs->meta_length;
451 
452 	rdma_wr = &ctx->reg->wr;
453 	rdma_wr->wr.sg_list = &ctx->reg->sge;
454 	rdma_wr->wr.num_sge = 1;
455 	rdma_wr->remote_addr = remote_addr;
456 	rdma_wr->rkey = rkey;
457 	if (dir == DMA_TO_DEVICE)
458 		rdma_wr->wr.opcode = IB_WR_RDMA_WRITE;
459 	else
460 		rdma_wr->wr.opcode = IB_WR_RDMA_READ;
461 	ctx->reg->reg_wr.wr.next = &rdma_wr->wr;
462 	count++;
463 
464 	return count;
465 
466 out_destroy_sig_mr:
467 	ib_mr_pool_put(qp, &qp->sig_mrs, ctx->reg->mr);
468 out_free_ctx:
469 	kfree(ctx->reg);
470 out_unmap_prot_sg:
471 	if (prot_sg_cnt)
472 		rdma_rw_unmap_sg(dev, prot_sg, prot_sg_cnt, dir);
473 out_unmap_sg:
474 	rdma_rw_unmap_sg(dev, sg, sg_cnt, dir);
475 	return ret;
476 }
477 EXPORT_SYMBOL(rdma_rw_ctx_signature_init);
478 
479 /*
480  * Now that we are going to post the WRs we can update the lkey and need_inval
481  * state on the MRs.  If we were doing this at init time, we would get double
482  * or missing invalidations if a context was initialized but not actually
483  * posted.
484  */
485 static void rdma_rw_update_lkey(struct rdma_rw_reg_ctx *reg, bool need_inval)
486 {
487 	reg->mr->need_inval = need_inval;
488 	ib_update_fast_reg_key(reg->mr, ib_inc_rkey(reg->mr->lkey));
489 	reg->reg_wr.key = reg->mr->lkey;
490 	reg->sge.lkey = reg->mr->lkey;
491 }
492 
493 /**
494  * rdma_rw_ctx_wrs - return chain of WRs for a RDMA READ or WRITE operation
495  * @ctx:	context to operate on
496  * @qp:		queue pair to operate on
497  * @port_num:	port num to which the connection is bound
498  * @cqe:	completion queue entry for the last WR
499  * @chain_wr:	WR to append to the posted chain
500  *
501  * Return the WR chain for the set of RDMA READ/WRITE operations described by
502  * @ctx, as well as any memory registration operations needed.  If @chain_wr
503  * is non-NULL the WR it points to will be appended to the chain of WRs posted.
504  * If @chain_wr is not set @cqe must be set so that the caller gets a
505  * completion notification.
506  */
507 struct ib_send_wr *rdma_rw_ctx_wrs(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
508 		u8 port_num, struct ib_cqe *cqe, struct ib_send_wr *chain_wr)
509 {
510 	struct ib_send_wr *first_wr, *last_wr;
511 	int i;
512 
513 	switch (ctx->type) {
514 	case RDMA_RW_SIG_MR:
515 	case RDMA_RW_MR:
516 		for (i = 0; i < ctx->nr_ops; i++) {
517 			rdma_rw_update_lkey(&ctx->reg[i],
518 				ctx->reg[i].wr.wr.opcode !=
519 					IB_WR_RDMA_READ_WITH_INV);
520 		}
521 
522 		if (ctx->reg[0].inv_wr.next)
523 			first_wr = &ctx->reg[0].inv_wr;
524 		else
525 			first_wr = &ctx->reg[0].reg_wr.wr;
526 		last_wr = &ctx->reg[ctx->nr_ops - 1].wr.wr;
527 		break;
528 	case RDMA_RW_MULTI_WR:
529 		first_wr = &ctx->map.wrs[0].wr;
530 		last_wr = &ctx->map.wrs[ctx->nr_ops - 1].wr;
531 		break;
532 	case RDMA_RW_SINGLE_WR:
533 		first_wr = &ctx->single.wr.wr;
534 		last_wr = &ctx->single.wr.wr;
535 		break;
536 	default:
537 		BUG();
538 	}
539 
540 	if (chain_wr) {
541 		last_wr->next = chain_wr;
542 	} else {
543 		last_wr->wr_cqe = cqe;
544 		last_wr->send_flags |= IB_SEND_SIGNALED;
545 	}
546 
547 	return first_wr;
548 }
549 EXPORT_SYMBOL(rdma_rw_ctx_wrs);
550 
551 /**
552  * rdma_rw_ctx_post - post a RDMA READ or RDMA WRITE operation
553  * @ctx:	context to operate on
554  * @qp:		queue pair to operate on
555  * @port_num:	port num to which the connection is bound
556  * @cqe:	completion queue entry for the last WR
557  * @chain_wr:	WR to append to the posted chain
558  *
559  * Post the set of RDMA READ/WRITE operations described by @ctx, as well as
560  * any memory registration operations needed.  If @chain_wr is non-NULL the
561  * WR it points to will be appended to the chain of WRs posted.  If @chain_wr
562  * is not set @cqe must be set so that the caller gets a completion
563  * notification.
564  */
565 int rdma_rw_ctx_post(struct rdma_rw_ctx *ctx, struct ib_qp *qp, u8 port_num,
566 		struct ib_cqe *cqe, struct ib_send_wr *chain_wr)
567 {
568 	struct ib_send_wr *first_wr;
569 
570 	first_wr = rdma_rw_ctx_wrs(ctx, qp, port_num, cqe, chain_wr);
571 	return ib_post_send(qp, first_wr, NULL);
572 }
573 EXPORT_SYMBOL(rdma_rw_ctx_post);
574 
575 /**
576  * rdma_rw_ctx_destroy - release all resources allocated by rdma_rw_ctx_init
577  * @ctx:	context to release
578  * @qp:		queue pair to operate on
579  * @port_num:	port num to which the connection is bound
580  * @sg:		scatterlist that was used for the READ/WRITE
581  * @sg_cnt:	number of entries in @sg
582  * @dir:	%DMA_TO_DEVICE for RDMA WRITE, %DMA_FROM_DEVICE for RDMA READ
583  */
584 void rdma_rw_ctx_destroy(struct rdma_rw_ctx *ctx, struct ib_qp *qp, u8 port_num,
585 		struct scatterlist *sg, u32 sg_cnt, enum dma_data_direction dir)
586 {
587 	int i;
588 
589 	switch (ctx->type) {
590 	case RDMA_RW_MR:
591 		for (i = 0; i < ctx->nr_ops; i++)
592 			ib_mr_pool_put(qp, &qp->rdma_mrs, ctx->reg[i].mr);
593 		kfree(ctx->reg);
594 		break;
595 	case RDMA_RW_MULTI_WR:
596 		kfree(ctx->map.wrs);
597 		kfree(ctx->map.sges);
598 		break;
599 	case RDMA_RW_SINGLE_WR:
600 		break;
601 	default:
602 		BUG();
603 		break;
604 	}
605 
606 	rdma_rw_unmap_sg(qp->pd->device, sg, sg_cnt, dir);
607 }
608 EXPORT_SYMBOL(rdma_rw_ctx_destroy);
609 
610 /**
611  * rdma_rw_ctx_destroy_signature - release all resources allocated by
612  *	rdma_rw_ctx_signature_init
613  * @ctx:	context to release
614  * @qp:		queue pair to operate on
615  * @port_num:	port num to which the connection is bound
616  * @sg:		scatterlist that was used for the READ/WRITE
617  * @sg_cnt:	number of entries in @sg
618  * @prot_sg:	scatterlist that was used for the READ/WRITE of the PI
619  * @prot_sg_cnt: number of entries in @prot_sg
620  * @dir:	%DMA_TO_DEVICE for RDMA WRITE, %DMA_FROM_DEVICE for RDMA READ
621  */
622 void rdma_rw_ctx_destroy_signature(struct rdma_rw_ctx *ctx, struct ib_qp *qp,
623 		u8 port_num, struct scatterlist *sg, u32 sg_cnt,
624 		struct scatterlist *prot_sg, u32 prot_sg_cnt,
625 		enum dma_data_direction dir)
626 {
627 	if (WARN_ON_ONCE(ctx->type != RDMA_RW_SIG_MR))
628 		return;
629 
630 	ib_mr_pool_put(qp, &qp->sig_mrs, ctx->reg->mr);
631 	kfree(ctx->reg);
632 
633 	if (prot_sg_cnt)
634 		rdma_rw_unmap_sg(qp->pd->device, prot_sg, prot_sg_cnt, dir);
635 	rdma_rw_unmap_sg(qp->pd->device, sg, sg_cnt, dir);
636 }
637 EXPORT_SYMBOL(rdma_rw_ctx_destroy_signature);
638 
639 /**
640  * rdma_rw_mr_factor - return number of MRs required for a payload
641  * @device:	device handling the connection
642  * @port_num:	port num to which the connection is bound
643  * @maxpages:	maximum payload pages per rdma_rw_ctx
644  *
645  * Returns the number of MRs the device requires to move @maxpayload
646  * bytes. The returned value is used during transport creation to
647  * compute max_rdma_ctxts and the size of the transport's Send and
648  * Send Completion Queues.
649  */
650 unsigned int rdma_rw_mr_factor(struct ib_device *device, u8 port_num,
651 			       unsigned int maxpages)
652 {
653 	unsigned int mr_pages;
654 
655 	if (rdma_rw_can_use_mr(device, port_num))
656 		mr_pages = rdma_rw_fr_page_list_len(device, false);
657 	else
658 		mr_pages = device->attrs.max_sge_rd;
659 	return DIV_ROUND_UP(maxpages, mr_pages);
660 }
661 EXPORT_SYMBOL(rdma_rw_mr_factor);
662 
663 void rdma_rw_init_qp(struct ib_device *dev, struct ib_qp_init_attr *attr)
664 {
665 	u32 factor;
666 
667 	WARN_ON_ONCE(attr->port_num == 0);
668 
669 	/*
670 	 * Each context needs at least one RDMA READ or WRITE WR.
671 	 *
672 	 * For some hardware we might need more, eventually we should ask the
673 	 * HCA driver for a multiplier here.
674 	 */
675 	factor = 1;
676 
677 	/*
678 	 * If the devices needs MRs to perform RDMA READ or WRITE operations,
679 	 * we'll need two additional MRs for the registrations and the
680 	 * invalidation.
681 	 */
682 	if (attr->create_flags & IB_QP_CREATE_INTEGRITY_EN ||
683 	    rdma_rw_can_use_mr(dev, attr->port_num))
684 		factor += 2;	/* inv + reg */
685 
686 	attr->cap.max_send_wr += factor * attr->cap.max_rdma_ctxs;
687 
688 	/*
689 	 * But maybe we were just too high in the sky and the device doesn't
690 	 * even support all we need, and we'll have to live with what we get..
691 	 */
692 	attr->cap.max_send_wr =
693 		min_t(u32, attr->cap.max_send_wr, dev->attrs.max_qp_wr);
694 }
695 
696 int rdma_rw_init_mrs(struct ib_qp *qp, struct ib_qp_init_attr *attr)
697 {
698 	struct ib_device *dev = qp->pd->device;
699 	u32 nr_mrs = 0, nr_sig_mrs = 0, max_num_sg = 0;
700 	int ret = 0;
701 
702 	if (attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) {
703 		nr_sig_mrs = attr->cap.max_rdma_ctxs;
704 		nr_mrs = attr->cap.max_rdma_ctxs;
705 		max_num_sg = rdma_rw_fr_page_list_len(dev, true);
706 	} else if (rdma_rw_can_use_mr(dev, attr->port_num)) {
707 		nr_mrs = attr->cap.max_rdma_ctxs;
708 		max_num_sg = rdma_rw_fr_page_list_len(dev, false);
709 	}
710 
711 	if (nr_mrs) {
712 		ret = ib_mr_pool_init(qp, &qp->rdma_mrs, nr_mrs,
713 				IB_MR_TYPE_MEM_REG,
714 				max_num_sg, 0);
715 		if (ret) {
716 			pr_err("%s: failed to allocated %d MRs\n",
717 				__func__, nr_mrs);
718 			return ret;
719 		}
720 	}
721 
722 	if (nr_sig_mrs) {
723 		ret = ib_mr_pool_init(qp, &qp->sig_mrs, nr_sig_mrs,
724 				IB_MR_TYPE_INTEGRITY, max_num_sg, max_num_sg);
725 		if (ret) {
726 			pr_err("%s: failed to allocated %d SIG MRs\n",
727 				__func__, nr_sig_mrs);
728 			goto out_free_rdma_mrs;
729 		}
730 	}
731 
732 	return 0;
733 
734 out_free_rdma_mrs:
735 	ib_mr_pool_destroy(qp, &qp->rdma_mrs);
736 	return ret;
737 }
738 
739 void rdma_rw_cleanup_mrs(struct ib_qp *qp)
740 {
741 	ib_mr_pool_destroy(qp, &qp->sig_mrs);
742 	ib_mr_pool_destroy(qp, &qp->rdma_mrs);
743 }
744