1 /*
2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *	- Redistributions of source code must retain the above
16  *	  copyright notice, this list of conditions and the following
17  *	  disclaimer.
18  *
19  *	- Redistributions in binary form must reproduce the above
20  *	  copyright notice, this list of conditions and the following
21  *	  disclaimer in the documentation and/or other materials
22  *	  provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 #include <linux/module.h>
34 #include <linux/kernel.h>
35 #include <linux/slab.h>
36 #include <linux/mm.h>
37 #include <linux/highmem.h>
38 #include <linux/scatterlist.h>
39 
40 #include "iscsi_iser.h"
41 
42 #define ISER_KMALLOC_THRESHOLD 0x20000 /* 128K - kmalloc limit */
43 
44 /**
45  * iser_start_rdma_unaligned_sg
46  */
47 static int iser_start_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
48 					struct iser_data_buf *data,
49 					struct iser_data_buf *data_copy,
50 					enum iser_data_dir cmd_dir)
51 {
52 	struct ib_device *dev = iser_task->iser_conn->ib_conn.device->ib_device;
53 	struct scatterlist *sgl = (struct scatterlist *)data->buf;
54 	struct scatterlist *sg;
55 	char *mem = NULL;
56 	unsigned long  cmd_data_len = 0;
57 	int dma_nents, i;
58 
59 	for_each_sg(sgl, sg, data->size, i)
60 		cmd_data_len += ib_sg_dma_len(dev, sg);
61 
62 	if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
63 		mem = (void *)__get_free_pages(GFP_ATOMIC,
64 		      ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
65 	else
66 		mem = kmalloc(cmd_data_len, GFP_ATOMIC);
67 
68 	if (mem == NULL) {
69 		iser_err("Failed to allocate mem size %d %d for copying sglist\n",
70 			 data->size, (int)cmd_data_len);
71 		return -ENOMEM;
72 	}
73 
74 	if (cmd_dir == ISER_DIR_OUT) {
75 		/* copy the unaligned sg the buffer which is used for RDMA */
76 		int i;
77 		char *p, *from;
78 
79 		sgl = (struct scatterlist *)data->buf;
80 		p = mem;
81 		for_each_sg(sgl, sg, data->size, i) {
82 			from = kmap_atomic(sg_page(sg));
83 			memcpy(p,
84 			       from + sg->offset,
85 			       sg->length);
86 			kunmap_atomic(from);
87 			p += sg->length;
88 		}
89 	}
90 
91 	sg_init_one(&data_copy->sg_single, mem, cmd_data_len);
92 	data_copy->buf = &data_copy->sg_single;
93 	data_copy->size = 1;
94 	data_copy->copy_buf = mem;
95 
96 	dma_nents = ib_dma_map_sg(dev, &data_copy->sg_single, 1,
97 				  (cmd_dir == ISER_DIR_OUT) ?
98 				  DMA_TO_DEVICE : DMA_FROM_DEVICE);
99 	BUG_ON(dma_nents == 0);
100 
101 	data_copy->dma_nents = dma_nents;
102 	data_copy->data_len = cmd_data_len;
103 
104 	return 0;
105 }
106 
107 /**
108  * iser_finalize_rdma_unaligned_sg
109  */
110 
111 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
112 				     struct iser_data_buf *data,
113 				     struct iser_data_buf *data_copy,
114 				     enum iser_data_dir cmd_dir)
115 {
116 	struct ib_device *dev;
117 	unsigned long  cmd_data_len;
118 
119 	dev = iser_task->iser_conn->ib_conn.device->ib_device;
120 
121 	ib_dma_unmap_sg(dev, &data_copy->sg_single, 1,
122 			(cmd_dir == ISER_DIR_OUT) ?
123 			DMA_TO_DEVICE : DMA_FROM_DEVICE);
124 
125 	if (cmd_dir == ISER_DIR_IN) {
126 		char *mem;
127 		struct scatterlist *sgl, *sg;
128 		unsigned char *p, *to;
129 		unsigned int sg_size;
130 		int i;
131 
132 		/* copy back read RDMA to unaligned sg */
133 		mem = data_copy->copy_buf;
134 
135 		sgl = (struct scatterlist *)data->buf;
136 		sg_size = data->size;
137 
138 		p = mem;
139 		for_each_sg(sgl, sg, sg_size, i) {
140 			to = kmap_atomic(sg_page(sg));
141 			memcpy(to + sg->offset,
142 			       p,
143 			       sg->length);
144 			kunmap_atomic(to);
145 			p += sg->length;
146 		}
147 	}
148 
149 	cmd_data_len = data->data_len;
150 
151 	if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
152 		free_pages((unsigned long)data_copy->copy_buf,
153 			   ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
154 	else
155 		kfree(data_copy->copy_buf);
156 
157 	data_copy->copy_buf = NULL;
158 }
159 
160 #define IS_4K_ALIGNED(addr)	((((unsigned long)addr) & ~MASK_4K) == 0)
161 
162 /**
163  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
164  * and returns the length of resulting physical address array (may be less than
165  * the original due to possible compaction).
166  *
167  * we build a "page vec" under the assumption that the SG meets the RDMA
168  * alignment requirements. Other then the first and last SG elements, all
169  * the "internal" elements can be compacted into a list whose elements are
170  * dma addresses of physical pages. The code supports also the weird case
171  * where --few fragments of the same page-- are present in the SG as
172  * consecutive elements. Also, it handles one entry SG.
173  */
174 
175 static int iser_sg_to_page_vec(struct iser_data_buf *data,
176 			       struct ib_device *ibdev, u64 *pages,
177 			       int *offset, int *data_size)
178 {
179 	struct scatterlist *sg, *sgl = (struct scatterlist *)data->buf;
180 	u64 start_addr, end_addr, page, chunk_start = 0;
181 	unsigned long total_sz = 0;
182 	unsigned int dma_len;
183 	int i, new_chunk, cur_page, last_ent = data->dma_nents - 1;
184 
185 	/* compute the offset of first element */
186 	*offset = (u64) sgl[0].offset & ~MASK_4K;
187 
188 	new_chunk = 1;
189 	cur_page  = 0;
190 	for_each_sg(sgl, sg, data->dma_nents, i) {
191 		start_addr = ib_sg_dma_address(ibdev, sg);
192 		if (new_chunk)
193 			chunk_start = start_addr;
194 		dma_len = ib_sg_dma_len(ibdev, sg);
195 		end_addr = start_addr + dma_len;
196 		total_sz += dma_len;
197 
198 		/* collect page fragments until aligned or end of SG list */
199 		if (!IS_4K_ALIGNED(end_addr) && i < last_ent) {
200 			new_chunk = 0;
201 			continue;
202 		}
203 		new_chunk = 1;
204 
205 		/* address of the first page in the contiguous chunk;
206 		   masking relevant for the very first SG entry,
207 		   which might be unaligned */
208 		page = chunk_start & MASK_4K;
209 		do {
210 			pages[cur_page++] = page;
211 			page += SIZE_4K;
212 		} while (page < end_addr);
213 	}
214 
215 	*data_size = total_sz;
216 	iser_dbg("page_vec->data_size:%d cur_page %d\n",
217 		 *data_size, cur_page);
218 	return cur_page;
219 }
220 
221 
222 /**
223  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
224  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
225  * the number of entries which are aligned correctly. Supports the case where
226  * consecutive SG elements are actually fragments of the same physcial page.
227  */
228 static int iser_data_buf_aligned_len(struct iser_data_buf *data,
229 				      struct ib_device *ibdev)
230 {
231 	struct scatterlist *sgl, *sg, *next_sg = NULL;
232 	u64 start_addr, end_addr;
233 	int i, ret_len, start_check = 0;
234 
235 	if (data->dma_nents == 1)
236 		return 1;
237 
238 	sgl = (struct scatterlist *)data->buf;
239 	start_addr  = ib_sg_dma_address(ibdev, sgl);
240 
241 	for_each_sg(sgl, sg, data->dma_nents, i) {
242 		if (start_check && !IS_4K_ALIGNED(start_addr))
243 			break;
244 
245 		next_sg = sg_next(sg);
246 		if (!next_sg)
247 			break;
248 
249 		end_addr    = start_addr + ib_sg_dma_len(ibdev, sg);
250 		start_addr  = ib_sg_dma_address(ibdev, next_sg);
251 
252 		if (end_addr == start_addr) {
253 			start_check = 0;
254 			continue;
255 		} else
256 			start_check = 1;
257 
258 		if (!IS_4K_ALIGNED(end_addr))
259 			break;
260 	}
261 	ret_len = (next_sg) ? i : i+1;
262 	iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
263 		 ret_len, data->dma_nents, data);
264 	return ret_len;
265 }
266 
267 static void iser_data_buf_dump(struct iser_data_buf *data,
268 			       struct ib_device *ibdev)
269 {
270 	struct scatterlist *sgl = (struct scatterlist *)data->buf;
271 	struct scatterlist *sg;
272 	int i;
273 
274 	for_each_sg(sgl, sg, data->dma_nents, i)
275 		iser_dbg("sg[%d] dma_addr:0x%lX page:0x%p "
276 			 "off:0x%x sz:0x%x dma_len:0x%x\n",
277 			 i, (unsigned long)ib_sg_dma_address(ibdev, sg),
278 			 sg_page(sg), sg->offset,
279 			 sg->length, ib_sg_dma_len(ibdev, sg));
280 }
281 
282 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
283 {
284 	int i;
285 
286 	iser_err("page vec length %d data size %d\n",
287 		 page_vec->length, page_vec->data_size);
288 	for (i = 0; i < page_vec->length; i++)
289 		iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
290 }
291 
292 static void iser_page_vec_build(struct iser_data_buf *data,
293 				struct iser_page_vec *page_vec,
294 				struct ib_device *ibdev)
295 {
296 	int page_vec_len = 0;
297 
298 	page_vec->length = 0;
299 	page_vec->offset = 0;
300 
301 	iser_dbg("Translating sg sz: %d\n", data->dma_nents);
302 	page_vec_len = iser_sg_to_page_vec(data, ibdev, page_vec->pages,
303 					   &page_vec->offset,
304 					   &page_vec->data_size);
305 	iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents, page_vec_len);
306 
307 	page_vec->length = page_vec_len;
308 
309 	if (page_vec_len * SIZE_4K < page_vec->data_size) {
310 		iser_err("page_vec too short to hold this SG\n");
311 		iser_data_buf_dump(data, ibdev);
312 		iser_dump_page_vec(page_vec);
313 		BUG();
314 	}
315 }
316 
317 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
318 			    struct iser_data_buf *data,
319 			    enum iser_data_dir iser_dir,
320 			    enum dma_data_direction dma_dir)
321 {
322 	struct ib_device *dev;
323 
324 	iser_task->dir[iser_dir] = 1;
325 	dev = iser_task->iser_conn->ib_conn.device->ib_device;
326 
327 	data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir);
328 	if (data->dma_nents == 0) {
329 		iser_err("dma_map_sg failed!!!\n");
330 		return -EINVAL;
331 	}
332 	return 0;
333 }
334 
335 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
336 			      struct iser_data_buf *data)
337 {
338 	struct ib_device *dev;
339 
340 	dev = iser_task->iser_conn->ib_conn.device->ib_device;
341 	ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE);
342 }
343 
344 static int fall_to_bounce_buf(struct iscsi_iser_task *iser_task,
345 			      struct ib_device *ibdev,
346 			      struct iser_data_buf *mem,
347 			      struct iser_data_buf *mem_copy,
348 			      enum iser_data_dir cmd_dir,
349 			      int aligned_len)
350 {
351 	struct iscsi_conn    *iscsi_conn = iser_task->iser_conn->iscsi_conn;
352 
353 	iscsi_conn->fmr_unalign_cnt++;
354 	iser_warn("rdma alignment violation (%d/%d aligned) or FMR not supported\n",
355 		  aligned_len, mem->size);
356 
357 	if (iser_debug_level > 0)
358 		iser_data_buf_dump(mem, ibdev);
359 
360 	/* unmap the command data before accessing it */
361 	iser_dma_unmap_task_data(iser_task, mem);
362 
363 	/* allocate copy buf, if we are writing, copy the */
364 	/* unaligned scatterlist, dma map the copy        */
365 	if (iser_start_rdma_unaligned_sg(iser_task, mem, mem_copy, cmd_dir) != 0)
366 		return -ENOMEM;
367 
368 	return 0;
369 }
370 
371 /**
372  * iser_reg_rdma_mem_fmr - Registers memory intended for RDMA,
373  * using FMR (if possible) obtaining rkey and va
374  *
375  * returns 0 on success, errno code on failure
376  */
377 int iser_reg_rdma_mem_fmr(struct iscsi_iser_task *iser_task,
378 			  enum iser_data_dir cmd_dir)
379 {
380 	struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
381 	struct iser_device   *device = ib_conn->device;
382 	struct ib_device     *ibdev = device->ib_device;
383 	struct iser_data_buf *mem = &iser_task->data[cmd_dir];
384 	struct iser_regd_buf *regd_buf;
385 	int aligned_len;
386 	int err;
387 	int i;
388 	struct scatterlist *sg;
389 
390 	regd_buf = &iser_task->rdma_regd[cmd_dir];
391 
392 	aligned_len = iser_data_buf_aligned_len(mem, ibdev);
393 	if (aligned_len != mem->dma_nents) {
394 		err = fall_to_bounce_buf(iser_task, ibdev, mem,
395 					 &iser_task->data_copy[cmd_dir],
396 					 cmd_dir, aligned_len);
397 		if (err) {
398 			iser_err("failed to allocate bounce buffer\n");
399 			return err;
400 		}
401 		mem = &iser_task->data_copy[cmd_dir];
402 	}
403 
404 	/* if there a single dma entry, FMR is not needed */
405 	if (mem->dma_nents == 1) {
406 		sg = (struct scatterlist *)mem->buf;
407 
408 		regd_buf->reg.lkey = device->mr->lkey;
409 		regd_buf->reg.rkey = device->mr->rkey;
410 		regd_buf->reg.len  = ib_sg_dma_len(ibdev, &sg[0]);
411 		regd_buf->reg.va   = ib_sg_dma_address(ibdev, &sg[0]);
412 		regd_buf->reg.is_mr = 0;
413 
414 		iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X  "
415 			 "va: 0x%08lX sz: %ld]\n",
416 			 (unsigned int)regd_buf->reg.lkey,
417 			 (unsigned int)regd_buf->reg.rkey,
418 			 (unsigned long)regd_buf->reg.va,
419 			 (unsigned long)regd_buf->reg.len);
420 	} else { /* use FMR for multiple dma entries */
421 		iser_page_vec_build(mem, ib_conn->fmr.page_vec, ibdev);
422 		err = iser_reg_page_vec(ib_conn, ib_conn->fmr.page_vec,
423 					&regd_buf->reg);
424 		if (err && err != -EAGAIN) {
425 			iser_data_buf_dump(mem, ibdev);
426 			iser_err("mem->dma_nents = %d (dlength = 0x%x)\n",
427 				 mem->dma_nents,
428 				 ntoh24(iser_task->desc.iscsi_header.dlength));
429 			iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
430 				 ib_conn->fmr.page_vec->data_size,
431 				 ib_conn->fmr.page_vec->length,
432 				 ib_conn->fmr.page_vec->offset);
433 			for (i = 0; i < ib_conn->fmr.page_vec->length; i++)
434 				iser_err("page_vec[%d] = 0x%llx\n", i,
435 					 (unsigned long long)ib_conn->fmr.page_vec->pages[i]);
436 		}
437 		if (err)
438 			return err;
439 	}
440 	return 0;
441 }
442 
443 static inline void
444 iser_set_dif_domain(struct scsi_cmnd *sc, struct ib_sig_attrs *sig_attrs,
445 		    struct ib_sig_domain *domain)
446 {
447 	domain->sig_type = IB_SIG_TYPE_T10_DIF;
448 	domain->sig.dif.pi_interval = sc->device->sector_size;
449 	domain->sig.dif.ref_tag = scsi_get_lba(sc) & 0xffffffff;
450 	/*
451 	 * At the moment we hard code those, but in the future
452 	 * we will take them from sc.
453 	 */
454 	domain->sig.dif.apptag_check_mask = 0xffff;
455 	domain->sig.dif.app_escape = true;
456 	domain->sig.dif.ref_escape = true;
457 	if (scsi_get_prot_type(sc) == SCSI_PROT_DIF_TYPE1 ||
458 	    scsi_get_prot_type(sc) == SCSI_PROT_DIF_TYPE2)
459 		domain->sig.dif.ref_remap = true;
460 };
461 
462 static int
463 iser_set_sig_attrs(struct scsi_cmnd *sc, struct ib_sig_attrs *sig_attrs)
464 {
465 	switch (scsi_get_prot_op(sc)) {
466 	case SCSI_PROT_WRITE_INSERT:
467 	case SCSI_PROT_READ_STRIP:
468 		sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
469 		iser_set_dif_domain(sc, sig_attrs, &sig_attrs->wire);
470 		sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
471 		break;
472 	case SCSI_PROT_READ_INSERT:
473 	case SCSI_PROT_WRITE_STRIP:
474 		sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
475 		iser_set_dif_domain(sc, sig_attrs, &sig_attrs->mem);
476 		/*
477 		 * At the moment we use this modparam to tell what is
478 		 * the memory bg_type, in the future we will take it
479 		 * from sc.
480 		 */
481 		sig_attrs->mem.sig.dif.bg_type = iser_pi_guard ? IB_T10DIF_CSUM :
482 						 IB_T10DIF_CRC;
483 		break;
484 	case SCSI_PROT_READ_PASS:
485 	case SCSI_PROT_WRITE_PASS:
486 		iser_set_dif_domain(sc, sig_attrs, &sig_attrs->wire);
487 		sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
488 		iser_set_dif_domain(sc, sig_attrs, &sig_attrs->mem);
489 		/*
490 		 * At the moment we use this modparam to tell what is
491 		 * the memory bg_type, in the future we will take it
492 		 * from sc.
493 		 */
494 		sig_attrs->mem.sig.dif.bg_type = iser_pi_guard ? IB_T10DIF_CSUM :
495 						 IB_T10DIF_CRC;
496 		break;
497 	default:
498 		iser_err("Unsupported PI operation %d\n",
499 			 scsi_get_prot_op(sc));
500 		return -EINVAL;
501 	}
502 
503 	return 0;
504 }
505 
506 static int
507 iser_set_prot_checks(struct scsi_cmnd *sc, u8 *mask)
508 {
509 	switch (scsi_get_prot_type(sc)) {
510 	case SCSI_PROT_DIF_TYPE0:
511 		break;
512 	case SCSI_PROT_DIF_TYPE1:
513 	case SCSI_PROT_DIF_TYPE2:
514 		*mask = ISER_CHECK_GUARD | ISER_CHECK_REFTAG;
515 		break;
516 	case SCSI_PROT_DIF_TYPE3:
517 		*mask = ISER_CHECK_GUARD;
518 		break;
519 	default:
520 		iser_err("Unsupported protection type %d\n",
521 			 scsi_get_prot_type(sc));
522 		return -EINVAL;
523 	}
524 
525 	return 0;
526 }
527 
528 static int
529 iser_reg_sig_mr(struct iscsi_iser_task *iser_task,
530 		struct fast_reg_descriptor *desc, struct ib_sge *data_sge,
531 		struct ib_sge *prot_sge, struct ib_sge *sig_sge)
532 {
533 	struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
534 	struct iser_pi_context *pi_ctx = desc->pi_ctx;
535 	struct ib_send_wr sig_wr, inv_wr;
536 	struct ib_send_wr *bad_wr, *wr = NULL;
537 	struct ib_sig_attrs sig_attrs;
538 	int ret;
539 	u32 key;
540 
541 	memset(&sig_attrs, 0, sizeof(sig_attrs));
542 	ret = iser_set_sig_attrs(iser_task->sc, &sig_attrs);
543 	if (ret)
544 		goto err;
545 
546 	ret = iser_set_prot_checks(iser_task->sc, &sig_attrs.check_mask);
547 	if (ret)
548 		goto err;
549 
550 	if (!(desc->reg_indicators & ISER_SIG_KEY_VALID)) {
551 		memset(&inv_wr, 0, sizeof(inv_wr));
552 		inv_wr.opcode = IB_WR_LOCAL_INV;
553 		inv_wr.wr_id = ISER_FASTREG_LI_WRID;
554 		inv_wr.ex.invalidate_rkey = pi_ctx->sig_mr->rkey;
555 		wr = &inv_wr;
556 		/* Bump the key */
557 		key = (u8)(pi_ctx->sig_mr->rkey & 0x000000FF);
558 		ib_update_fast_reg_key(pi_ctx->sig_mr, ++key);
559 	}
560 
561 	memset(&sig_wr, 0, sizeof(sig_wr));
562 	sig_wr.opcode = IB_WR_REG_SIG_MR;
563 	sig_wr.wr_id = ISER_FASTREG_LI_WRID;
564 	sig_wr.sg_list = data_sge;
565 	sig_wr.num_sge = 1;
566 	sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
567 	sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
568 	if (scsi_prot_sg_count(iser_task->sc))
569 		sig_wr.wr.sig_handover.prot = prot_sge;
570 	sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE |
571 					      IB_ACCESS_REMOTE_READ |
572 					      IB_ACCESS_REMOTE_WRITE;
573 
574 	if (!wr)
575 		wr = &sig_wr;
576 	else
577 		wr->next = &sig_wr;
578 
579 	ret = ib_post_send(ib_conn->qp, wr, &bad_wr);
580 	if (ret) {
581 		iser_err("reg_sig_mr failed, ret:%d\n", ret);
582 		goto err;
583 	}
584 	desc->reg_indicators &= ~ISER_SIG_KEY_VALID;
585 
586 	sig_sge->lkey = pi_ctx->sig_mr->lkey;
587 	sig_sge->addr = 0;
588 	sig_sge->length = data_sge->length + prot_sge->length;
589 	if (scsi_get_prot_op(iser_task->sc) == SCSI_PROT_WRITE_INSERT ||
590 	    scsi_get_prot_op(iser_task->sc) == SCSI_PROT_READ_STRIP) {
591 		sig_sge->length += (data_sge->length /
592 				   iser_task->sc->device->sector_size) * 8;
593 	}
594 
595 	iser_dbg("sig_sge: addr: 0x%llx  length: %u lkey: 0x%x\n",
596 		 sig_sge->addr, sig_sge->length,
597 		 sig_sge->lkey);
598 err:
599 	return ret;
600 }
601 
602 static int iser_fast_reg_mr(struct iscsi_iser_task *iser_task,
603 			    struct iser_regd_buf *regd_buf,
604 			    struct iser_data_buf *mem,
605 			    enum iser_reg_indicator ind,
606 			    struct ib_sge *sge)
607 {
608 	struct fast_reg_descriptor *desc = regd_buf->reg.mem_h;
609 	struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
610 	struct iser_device *device = ib_conn->device;
611 	struct ib_device *ibdev = device->ib_device;
612 	struct ib_mr *mr;
613 	struct ib_fast_reg_page_list *frpl;
614 	struct ib_send_wr fastreg_wr, inv_wr;
615 	struct ib_send_wr *bad_wr, *wr = NULL;
616 	u8 key;
617 	int ret, offset, size, plen;
618 
619 	/* if there a single dma entry, dma mr suffices */
620 	if (mem->dma_nents == 1) {
621 		struct scatterlist *sg = (struct scatterlist *)mem->buf;
622 
623 		sge->lkey = device->mr->lkey;
624 		sge->addr   = ib_sg_dma_address(ibdev, &sg[0]);
625 		sge->length  = ib_sg_dma_len(ibdev, &sg[0]);
626 
627 		iser_dbg("Single DMA entry: lkey=0x%x, addr=0x%llx, length=0x%x\n",
628 			 sge->lkey, sge->addr, sge->length);
629 		return 0;
630 	}
631 
632 	if (ind == ISER_DATA_KEY_VALID) {
633 		mr = desc->data_mr;
634 		frpl = desc->data_frpl;
635 	} else {
636 		mr = desc->pi_ctx->prot_mr;
637 		frpl = desc->pi_ctx->prot_frpl;
638 	}
639 
640 	plen = iser_sg_to_page_vec(mem, device->ib_device, frpl->page_list,
641 				   &offset, &size);
642 	if (plen * SIZE_4K < size) {
643 		iser_err("fast reg page_list too short to hold this SG\n");
644 		return -EINVAL;
645 	}
646 
647 	if (!(desc->reg_indicators & ind)) {
648 		memset(&inv_wr, 0, sizeof(inv_wr));
649 		inv_wr.wr_id = ISER_FASTREG_LI_WRID;
650 		inv_wr.opcode = IB_WR_LOCAL_INV;
651 		inv_wr.ex.invalidate_rkey = mr->rkey;
652 		wr = &inv_wr;
653 		/* Bump the key */
654 		key = (u8)(mr->rkey & 0x000000FF);
655 		ib_update_fast_reg_key(mr, ++key);
656 	}
657 
658 	/* Prepare FASTREG WR */
659 	memset(&fastreg_wr, 0, sizeof(fastreg_wr));
660 	fastreg_wr.wr_id = ISER_FASTREG_LI_WRID;
661 	fastreg_wr.opcode = IB_WR_FAST_REG_MR;
662 	fastreg_wr.wr.fast_reg.iova_start = frpl->page_list[0] + offset;
663 	fastreg_wr.wr.fast_reg.page_list = frpl;
664 	fastreg_wr.wr.fast_reg.page_list_len = plen;
665 	fastreg_wr.wr.fast_reg.page_shift = SHIFT_4K;
666 	fastreg_wr.wr.fast_reg.length = size;
667 	fastreg_wr.wr.fast_reg.rkey = mr->rkey;
668 	fastreg_wr.wr.fast_reg.access_flags = (IB_ACCESS_LOCAL_WRITE  |
669 					       IB_ACCESS_REMOTE_WRITE |
670 					       IB_ACCESS_REMOTE_READ);
671 
672 	if (!wr)
673 		wr = &fastreg_wr;
674 	else
675 		wr->next = &fastreg_wr;
676 
677 	ret = ib_post_send(ib_conn->qp, wr, &bad_wr);
678 	if (ret) {
679 		iser_err("fast registration failed, ret:%d\n", ret);
680 		return ret;
681 	}
682 	desc->reg_indicators &= ~ind;
683 
684 	sge->lkey = mr->lkey;
685 	sge->addr = frpl->page_list[0] + offset;
686 	sge->length = size;
687 
688 	return ret;
689 }
690 
691 /**
692  * iser_reg_rdma_mem_fastreg - Registers memory intended for RDMA,
693  * using Fast Registration WR (if possible) obtaining rkey and va
694  *
695  * returns 0 on success, errno code on failure
696  */
697 int iser_reg_rdma_mem_fastreg(struct iscsi_iser_task *iser_task,
698 			      enum iser_data_dir cmd_dir)
699 {
700 	struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
701 	struct iser_device *device = ib_conn->device;
702 	struct ib_device *ibdev = device->ib_device;
703 	struct iser_data_buf *mem = &iser_task->data[cmd_dir];
704 	struct iser_regd_buf *regd_buf = &iser_task->rdma_regd[cmd_dir];
705 	struct fast_reg_descriptor *desc = NULL;
706 	struct ib_sge data_sge;
707 	int err, aligned_len;
708 	unsigned long flags;
709 
710 	aligned_len = iser_data_buf_aligned_len(mem, ibdev);
711 	if (aligned_len != mem->dma_nents) {
712 		err = fall_to_bounce_buf(iser_task, ibdev, mem,
713 					 &iser_task->data_copy[cmd_dir],
714 					 cmd_dir, aligned_len);
715 		if (err) {
716 			iser_err("failed to allocate bounce buffer\n");
717 			return err;
718 		}
719 		mem = &iser_task->data_copy[cmd_dir];
720 	}
721 
722 	if (mem->dma_nents != 1 ||
723 	    scsi_get_prot_op(iser_task->sc) != SCSI_PROT_NORMAL) {
724 		spin_lock_irqsave(&ib_conn->lock, flags);
725 		desc = list_first_entry(&ib_conn->fastreg.pool,
726 					struct fast_reg_descriptor, list);
727 		list_del(&desc->list);
728 		spin_unlock_irqrestore(&ib_conn->lock, flags);
729 		regd_buf->reg.mem_h = desc;
730 	}
731 
732 	err = iser_fast_reg_mr(iser_task, regd_buf, mem,
733 			       ISER_DATA_KEY_VALID, &data_sge);
734 	if (err)
735 		goto err_reg;
736 
737 	if (scsi_get_prot_op(iser_task->sc) != SCSI_PROT_NORMAL) {
738 		struct ib_sge prot_sge, sig_sge;
739 
740 		memset(&prot_sge, 0, sizeof(prot_sge));
741 		if (scsi_prot_sg_count(iser_task->sc)) {
742 			mem = &iser_task->prot[cmd_dir];
743 			aligned_len = iser_data_buf_aligned_len(mem, ibdev);
744 			if (aligned_len != mem->dma_nents) {
745 				err = fall_to_bounce_buf(iser_task, ibdev, mem,
746 							 &iser_task->prot_copy[cmd_dir],
747 							 cmd_dir, aligned_len);
748 				if (err) {
749 					iser_err("failed to allocate bounce buffer\n");
750 					return err;
751 				}
752 				mem = &iser_task->prot_copy[cmd_dir];
753 			}
754 
755 			err = iser_fast_reg_mr(iser_task, regd_buf, mem,
756 					       ISER_PROT_KEY_VALID, &prot_sge);
757 			if (err)
758 				goto err_reg;
759 		}
760 
761 		err = iser_reg_sig_mr(iser_task, desc, &data_sge,
762 				      &prot_sge, &sig_sge);
763 		if (err) {
764 			iser_err("Failed to register signature mr\n");
765 			return err;
766 		}
767 		desc->reg_indicators |= ISER_FASTREG_PROTECTED;
768 
769 		regd_buf->reg.lkey = sig_sge.lkey;
770 		regd_buf->reg.rkey = desc->pi_ctx->sig_mr->rkey;
771 		regd_buf->reg.va = sig_sge.addr;
772 		regd_buf->reg.len = sig_sge.length;
773 		regd_buf->reg.is_mr = 1;
774 	} else {
775 		if (desc) {
776 			regd_buf->reg.rkey = desc->data_mr->rkey;
777 			regd_buf->reg.is_mr = 1;
778 		} else {
779 			regd_buf->reg.rkey = device->mr->rkey;
780 			regd_buf->reg.is_mr = 0;
781 		}
782 
783 		regd_buf->reg.lkey = data_sge.lkey;
784 		regd_buf->reg.va = data_sge.addr;
785 		regd_buf->reg.len = data_sge.length;
786 	}
787 
788 	return 0;
789 err_reg:
790 	if (desc) {
791 		spin_lock_irqsave(&ib_conn->lock, flags);
792 		list_add_tail(&desc->list, &ib_conn->fastreg.pool);
793 		spin_unlock_irqrestore(&ib_conn->lock, flags);
794 	}
795 
796 	return err;
797 }
798