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 */
iser_start_rdma_unaligned_sg(struct iscsi_iser_task * iser_task,struct iser_data_buf * data,struct iser_data_buf * data_copy,enum iser_data_dir cmd_dir)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 char *p, *from;
77
78 sgl = (struct scatterlist *)data->buf;
79 p = mem;
80 for_each_sg(sgl, sg, data->size, i) {
81 from = kmap_atomic(sg_page(sg));
82 memcpy(p,
83 from + sg->offset,
84 sg->length);
85 kunmap_atomic(from);
86 p += sg->length;
87 }
88 }
89
90 sg_init_one(&data_copy->sg_single, mem, cmd_data_len);
91 data_copy->buf = &data_copy->sg_single;
92 data_copy->size = 1;
93 data_copy->copy_buf = mem;
94
95 dma_nents = ib_dma_map_sg(dev, &data_copy->sg_single, 1,
96 (cmd_dir == ISER_DIR_OUT) ?
97 DMA_TO_DEVICE : DMA_FROM_DEVICE);
98 BUG_ON(dma_nents == 0);
99
100 data_copy->dma_nents = dma_nents;
101 data_copy->data_len = cmd_data_len;
102
103 return 0;
104 }
105
106 /**
107 * iser_finalize_rdma_unaligned_sg
108 */
109
iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task * iser_task,struct iser_data_buf * data,struct iser_data_buf * data_copy,enum iser_data_dir cmd_dir)110 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_task *iser_task,
111 struct iser_data_buf *data,
112 struct iser_data_buf *data_copy,
113 enum iser_data_dir cmd_dir)
114 {
115 struct ib_device *dev;
116 unsigned long cmd_data_len;
117
118 dev = iser_task->iser_conn->ib_conn.device->ib_device;
119
120 ib_dma_unmap_sg(dev, &data_copy->sg_single, 1,
121 (cmd_dir == ISER_DIR_OUT) ?
122 DMA_TO_DEVICE : DMA_FROM_DEVICE);
123
124 if (cmd_dir == ISER_DIR_IN) {
125 char *mem;
126 struct scatterlist *sgl, *sg;
127 unsigned char *p, *to;
128 unsigned int sg_size;
129 int i;
130
131 /* copy back read RDMA to unaligned sg */
132 mem = data_copy->copy_buf;
133
134 sgl = (struct scatterlist *)data->buf;
135 sg_size = data->size;
136
137 p = mem;
138 for_each_sg(sgl, sg, sg_size, i) {
139 to = kmap_atomic(sg_page(sg));
140 memcpy(to + sg->offset,
141 p,
142 sg->length);
143 kunmap_atomic(to);
144 p += sg->length;
145 }
146 }
147
148 cmd_data_len = data->data_len;
149
150 if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
151 free_pages((unsigned long)data_copy->copy_buf,
152 ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
153 else
154 kfree(data_copy->copy_buf);
155
156 data_copy->copy_buf = NULL;
157 }
158
159 #define IS_4K_ALIGNED(addr) ((((unsigned long)addr) & ~MASK_4K) == 0)
160
161 /**
162 * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
163 * and returns the length of resulting physical address array (may be less than
164 * the original due to possible compaction).
165 *
166 * we build a "page vec" under the assumption that the SG meets the RDMA
167 * alignment requirements. Other then the first and last SG elements, all
168 * the "internal" elements can be compacted into a list whose elements are
169 * dma addresses of physical pages. The code supports also the weird case
170 * where --few fragments of the same page-- are present in the SG as
171 * consecutive elements. Also, it handles one entry SG.
172 */
173
iser_sg_to_page_vec(struct iser_data_buf * data,struct ib_device * ibdev,u64 * pages,int * offset,int * data_size)174 static int iser_sg_to_page_vec(struct iser_data_buf *data,
175 struct ib_device *ibdev, u64 *pages,
176 int *offset, int *data_size)
177 {
178 struct scatterlist *sg, *sgl = (struct scatterlist *)data->buf;
179 u64 start_addr, end_addr, page, chunk_start = 0;
180 unsigned long total_sz = 0;
181 unsigned int dma_len;
182 int i, new_chunk, cur_page, last_ent = data->dma_nents - 1;
183
184 /* compute the offset of first element */
185 *offset = (u64) sgl[0].offset & ~MASK_4K;
186
187 new_chunk = 1;
188 cur_page = 0;
189 for_each_sg(sgl, sg, data->dma_nents, i) {
190 start_addr = ib_sg_dma_address(ibdev, sg);
191 if (new_chunk)
192 chunk_start = start_addr;
193 dma_len = ib_sg_dma_len(ibdev, sg);
194 end_addr = start_addr + dma_len;
195 total_sz += dma_len;
196
197 /* collect page fragments until aligned or end of SG list */
198 if (!IS_4K_ALIGNED(end_addr) && i < last_ent) {
199 new_chunk = 0;
200 continue;
201 }
202 new_chunk = 1;
203
204 /* address of the first page in the contiguous chunk;
205 masking relevant for the very first SG entry,
206 which might be unaligned */
207 page = chunk_start & MASK_4K;
208 do {
209 pages[cur_page++] = page;
210 page += SIZE_4K;
211 } while (page < end_addr);
212 }
213
214 *data_size = total_sz;
215 iser_dbg("page_vec->data_size:%d cur_page %d\n",
216 *data_size, cur_page);
217 return cur_page;
218 }
219
220
221 /**
222 * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
223 * for RDMA sub-list of a scatter-gather list of memory buffers, and returns
224 * the number of entries which are aligned correctly. Supports the case where
225 * consecutive SG elements are actually fragments of the same physcial page.
226 */
iser_data_buf_aligned_len(struct iser_data_buf * data,struct ib_device * ibdev)227 static int iser_data_buf_aligned_len(struct iser_data_buf *data,
228 struct ib_device *ibdev)
229 {
230 struct scatterlist *sgl, *sg, *next_sg = NULL;
231 u64 start_addr, end_addr;
232 int i, ret_len, start_check = 0;
233
234 if (data->dma_nents == 1)
235 return 1;
236
237 sgl = (struct scatterlist *)data->buf;
238 start_addr = ib_sg_dma_address(ibdev, sgl);
239
240 for_each_sg(sgl, sg, data->dma_nents, i) {
241 if (start_check && !IS_4K_ALIGNED(start_addr))
242 break;
243
244 next_sg = sg_next(sg);
245 if (!next_sg)
246 break;
247
248 end_addr = start_addr + ib_sg_dma_len(ibdev, sg);
249 start_addr = ib_sg_dma_address(ibdev, next_sg);
250
251 if (end_addr == start_addr) {
252 start_check = 0;
253 continue;
254 } else
255 start_check = 1;
256
257 if (!IS_4K_ALIGNED(end_addr))
258 break;
259 }
260 ret_len = (next_sg) ? i : i+1;
261 iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
262 ret_len, data->dma_nents, data);
263 return ret_len;
264 }
265
iser_data_buf_dump(struct iser_data_buf * data,struct ib_device * ibdev)266 static void iser_data_buf_dump(struct iser_data_buf *data,
267 struct ib_device *ibdev)
268 {
269 struct scatterlist *sgl = (struct scatterlist *)data->buf;
270 struct scatterlist *sg;
271 int i;
272
273 for_each_sg(sgl, sg, data->dma_nents, i)
274 iser_dbg("sg[%d] dma_addr:0x%lX page:0x%p "
275 "off:0x%x sz:0x%x dma_len:0x%x\n",
276 i, (unsigned long)ib_sg_dma_address(ibdev, sg),
277 sg_page(sg), sg->offset,
278 sg->length, ib_sg_dma_len(ibdev, sg));
279 }
280
iser_dump_page_vec(struct iser_page_vec * page_vec)281 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
282 {
283 int i;
284
285 iser_err("page vec length %d data size %d\n",
286 page_vec->length, page_vec->data_size);
287 for (i = 0; i < page_vec->length; i++)
288 iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
289 }
290
iser_page_vec_build(struct iser_data_buf * data,struct iser_page_vec * page_vec,struct ib_device * ibdev)291 static void iser_page_vec_build(struct iser_data_buf *data,
292 struct iser_page_vec *page_vec,
293 struct ib_device *ibdev)
294 {
295 int page_vec_len = 0;
296
297 page_vec->length = 0;
298 page_vec->offset = 0;
299
300 iser_dbg("Translating sg sz: %d\n", data->dma_nents);
301 page_vec_len = iser_sg_to_page_vec(data, ibdev, page_vec->pages,
302 &page_vec->offset,
303 &page_vec->data_size);
304 iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents, page_vec_len);
305
306 page_vec->length = page_vec_len;
307
308 if (page_vec_len * SIZE_4K < page_vec->data_size) {
309 iser_err("page_vec too short to hold this SG\n");
310 iser_data_buf_dump(data, ibdev);
311 iser_dump_page_vec(page_vec);
312 BUG();
313 }
314 }
315
iser_dma_map_task_data(struct iscsi_iser_task * iser_task,struct iser_data_buf * data,enum iser_data_dir iser_dir,enum dma_data_direction dma_dir)316 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
317 struct iser_data_buf *data,
318 enum iser_data_dir iser_dir,
319 enum dma_data_direction dma_dir)
320 {
321 struct ib_device *dev;
322
323 iser_task->dir[iser_dir] = 1;
324 dev = iser_task->iser_conn->ib_conn.device->ib_device;
325
326 data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir);
327 if (data->dma_nents == 0) {
328 iser_err("dma_map_sg failed!!!\n");
329 return -EINVAL;
330 }
331 return 0;
332 }
333
iser_dma_unmap_task_data(struct iscsi_iser_task * iser_task,struct iser_data_buf * data,enum dma_data_direction dir)334 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
335 struct iser_data_buf *data,
336 enum dma_data_direction dir)
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, dir);
342 }
343
fall_to_bounce_buf(struct iscsi_iser_task * iser_task,struct ib_device * ibdev,struct iser_data_buf * mem,struct iser_data_buf * mem_copy,enum iser_data_dir cmd_dir,int aligned_len)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 (cmd_dir == ISER_DIR_OUT) ?
363 DMA_TO_DEVICE : DMA_FROM_DEVICE);
364
365 /* allocate copy buf, if we are writing, copy the */
366 /* unaligned scatterlist, dma map the copy */
367 if (iser_start_rdma_unaligned_sg(iser_task, mem, mem_copy, cmd_dir) != 0)
368 return -ENOMEM;
369
370 return 0;
371 }
372
373 /**
374 * iser_reg_rdma_mem_fmr - Registers memory intended for RDMA,
375 * using FMR (if possible) obtaining rkey and va
376 *
377 * returns 0 on success, errno code on failure
378 */
iser_reg_rdma_mem_fmr(struct iscsi_iser_task * iser_task,enum iser_data_dir cmd_dir)379 int iser_reg_rdma_mem_fmr(struct iscsi_iser_task *iser_task,
380 enum iser_data_dir cmd_dir)
381 {
382 struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
383 struct iser_device *device = ib_conn->device;
384 struct ib_device *ibdev = device->ib_device;
385 struct iser_data_buf *mem = &iser_task->data[cmd_dir];
386 struct iser_regd_buf *regd_buf;
387 int aligned_len;
388 int err;
389 int i;
390 struct scatterlist *sg;
391
392 regd_buf = &iser_task->rdma_regd[cmd_dir];
393
394 aligned_len = iser_data_buf_aligned_len(mem, ibdev);
395 if (aligned_len != mem->dma_nents) {
396 err = fall_to_bounce_buf(iser_task, ibdev, mem,
397 &iser_task->data_copy[cmd_dir],
398 cmd_dir, aligned_len);
399 if (err) {
400 iser_err("failed to allocate bounce buffer\n");
401 return err;
402 }
403 mem = &iser_task->data_copy[cmd_dir];
404 }
405
406 /* if there a single dma entry, FMR is not needed */
407 if (mem->dma_nents == 1) {
408 sg = (struct scatterlist *)mem->buf;
409
410 regd_buf->reg.lkey = device->mr->lkey;
411 regd_buf->reg.rkey = device->mr->rkey;
412 regd_buf->reg.len = ib_sg_dma_len(ibdev, &sg[0]);
413 regd_buf->reg.va = ib_sg_dma_address(ibdev, &sg[0]);
414 regd_buf->reg.is_mr = 0;
415
416 iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X "
417 "va: 0x%08lX sz: %ld]\n",
418 (unsigned int)regd_buf->reg.lkey,
419 (unsigned int)regd_buf->reg.rkey,
420 (unsigned long)regd_buf->reg.va,
421 (unsigned long)regd_buf->reg.len);
422 } else { /* use FMR for multiple dma entries */
423 iser_page_vec_build(mem, ib_conn->fmr.page_vec, ibdev);
424 err = iser_reg_page_vec(ib_conn, ib_conn->fmr.page_vec,
425 ®d_buf->reg);
426 if (err && err != -EAGAIN) {
427 iser_data_buf_dump(mem, ibdev);
428 iser_err("mem->dma_nents = %d (dlength = 0x%x)\n",
429 mem->dma_nents,
430 ntoh24(iser_task->desc.iscsi_header.dlength));
431 iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
432 ib_conn->fmr.page_vec->data_size,
433 ib_conn->fmr.page_vec->length,
434 ib_conn->fmr.page_vec->offset);
435 for (i = 0; i < ib_conn->fmr.page_vec->length; i++)
436 iser_err("page_vec[%d] = 0x%llx\n", i,
437 (unsigned long long)ib_conn->fmr.page_vec->pages[i]);
438 }
439 if (err)
440 return err;
441 }
442 return 0;
443 }
444
445 static inline void
iser_set_dif_domain(struct scsi_cmnd * sc,struct ib_sig_attrs * sig_attrs,struct ib_sig_domain * domain)446 iser_set_dif_domain(struct scsi_cmnd *sc, struct ib_sig_attrs *sig_attrs,
447 struct ib_sig_domain *domain)
448 {
449 domain->sig_type = IB_SIG_TYPE_T10_DIF;
450 domain->sig.dif.pi_interval = sc->device->sector_size;
451 domain->sig.dif.ref_tag = scsi_get_lba(sc) & 0xffffffff;
452 /*
453 * At the moment we hard code those, but in the future
454 * we will take them from sc.
455 */
456 domain->sig.dif.apptag_check_mask = 0xffff;
457 domain->sig.dif.app_escape = true;
458 domain->sig.dif.ref_escape = true;
459 if (scsi_get_prot_type(sc) == SCSI_PROT_DIF_TYPE1 ||
460 scsi_get_prot_type(sc) == SCSI_PROT_DIF_TYPE2)
461 domain->sig.dif.ref_remap = true;
462 };
463
464 static int
iser_set_sig_attrs(struct scsi_cmnd * sc,struct ib_sig_attrs * sig_attrs)465 iser_set_sig_attrs(struct scsi_cmnd *sc, struct ib_sig_attrs *sig_attrs)
466 {
467 switch (scsi_get_prot_op(sc)) {
468 case SCSI_PROT_WRITE_INSERT:
469 case SCSI_PROT_READ_STRIP:
470 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
471 iser_set_dif_domain(sc, sig_attrs, &sig_attrs->wire);
472 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
473 break;
474 case SCSI_PROT_READ_INSERT:
475 case SCSI_PROT_WRITE_STRIP:
476 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
477 iser_set_dif_domain(sc, sig_attrs, &sig_attrs->mem);
478 /*
479 * At the moment we use this modparam to tell what is
480 * the memory bg_type, in the future we will take it
481 * from sc.
482 */
483 sig_attrs->mem.sig.dif.bg_type = iser_pi_guard ? IB_T10DIF_CSUM :
484 IB_T10DIF_CRC;
485 break;
486 case SCSI_PROT_READ_PASS:
487 case SCSI_PROT_WRITE_PASS:
488 iser_set_dif_domain(sc, sig_attrs, &sig_attrs->wire);
489 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
490 iser_set_dif_domain(sc, sig_attrs, &sig_attrs->mem);
491 /*
492 * At the moment we use this modparam to tell what is
493 * the memory bg_type, in the future we will take it
494 * from sc.
495 */
496 sig_attrs->mem.sig.dif.bg_type = iser_pi_guard ? IB_T10DIF_CSUM :
497 IB_T10DIF_CRC;
498 break;
499 default:
500 iser_err("Unsupported PI operation %d\n",
501 scsi_get_prot_op(sc));
502 return -EINVAL;
503 }
504
505 return 0;
506 }
507
508 static int
iser_set_prot_checks(struct scsi_cmnd * sc,u8 * mask)509 iser_set_prot_checks(struct scsi_cmnd *sc, u8 *mask)
510 {
511 switch (scsi_get_prot_type(sc)) {
512 case SCSI_PROT_DIF_TYPE0:
513 break;
514 case SCSI_PROT_DIF_TYPE1:
515 case SCSI_PROT_DIF_TYPE2:
516 *mask = ISER_CHECK_GUARD | ISER_CHECK_REFTAG;
517 break;
518 case SCSI_PROT_DIF_TYPE3:
519 *mask = ISER_CHECK_GUARD;
520 break;
521 default:
522 iser_err("Unsupported protection type %d\n",
523 scsi_get_prot_type(sc));
524 return -EINVAL;
525 }
526
527 return 0;
528 }
529
530 static int
iser_reg_sig_mr(struct iscsi_iser_task * iser_task,struct fast_reg_descriptor * desc,struct ib_sge * data_sge,struct ib_sge * prot_sge,struct ib_sge * sig_sge)531 iser_reg_sig_mr(struct iscsi_iser_task *iser_task,
532 struct fast_reg_descriptor *desc, struct ib_sge *data_sge,
533 struct ib_sge *prot_sge, struct ib_sge *sig_sge)
534 {
535 struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
536 struct iser_pi_context *pi_ctx = desc->pi_ctx;
537 struct ib_send_wr sig_wr, inv_wr;
538 struct ib_send_wr *bad_wr, *wr = NULL;
539 struct ib_sig_attrs sig_attrs;
540 int ret;
541 u32 key;
542
543 memset(&sig_attrs, 0, sizeof(sig_attrs));
544 ret = iser_set_sig_attrs(iser_task->sc, &sig_attrs);
545 if (ret)
546 goto err;
547
548 ret = iser_set_prot_checks(iser_task->sc, &sig_attrs.check_mask);
549 if (ret)
550 goto err;
551
552 if (!(desc->reg_indicators & ISER_SIG_KEY_VALID)) {
553 memset(&inv_wr, 0, sizeof(inv_wr));
554 inv_wr.opcode = IB_WR_LOCAL_INV;
555 inv_wr.wr_id = ISER_FASTREG_LI_WRID;
556 inv_wr.ex.invalidate_rkey = pi_ctx->sig_mr->rkey;
557 wr = &inv_wr;
558 /* Bump the key */
559 key = (u8)(pi_ctx->sig_mr->rkey & 0x000000FF);
560 ib_update_fast_reg_key(pi_ctx->sig_mr, ++key);
561 }
562
563 memset(&sig_wr, 0, sizeof(sig_wr));
564 sig_wr.opcode = IB_WR_REG_SIG_MR;
565 sig_wr.wr_id = ISER_FASTREG_LI_WRID;
566 sig_wr.sg_list = data_sge;
567 sig_wr.num_sge = 1;
568 sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
569 sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
570 if (scsi_prot_sg_count(iser_task->sc))
571 sig_wr.wr.sig_handover.prot = prot_sge;
572 sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE |
573 IB_ACCESS_REMOTE_READ |
574 IB_ACCESS_REMOTE_WRITE;
575
576 if (!wr)
577 wr = &sig_wr;
578 else
579 wr->next = &sig_wr;
580
581 ret = ib_post_send(ib_conn->qp, wr, &bad_wr);
582 if (ret) {
583 iser_err("reg_sig_mr failed, ret:%d\n", ret);
584 goto err;
585 }
586 desc->reg_indicators &= ~ISER_SIG_KEY_VALID;
587
588 sig_sge->lkey = pi_ctx->sig_mr->lkey;
589 sig_sge->addr = 0;
590 sig_sge->length = data_sge->length + prot_sge->length;
591 if (scsi_get_prot_op(iser_task->sc) == SCSI_PROT_WRITE_INSERT ||
592 scsi_get_prot_op(iser_task->sc) == SCSI_PROT_READ_STRIP) {
593 sig_sge->length += (data_sge->length /
594 iser_task->sc->device->sector_size) * 8;
595 }
596
597 iser_dbg("sig_sge: addr: 0x%llx length: %u lkey: 0x%x\n",
598 sig_sge->addr, sig_sge->length,
599 sig_sge->lkey);
600 err:
601 return ret;
602 }
603
iser_fast_reg_mr(struct iscsi_iser_task * iser_task,struct iser_regd_buf * regd_buf,struct iser_data_buf * mem,enum iser_reg_indicator ind,struct ib_sge * sge)604 static int iser_fast_reg_mr(struct iscsi_iser_task *iser_task,
605 struct iser_regd_buf *regd_buf,
606 struct iser_data_buf *mem,
607 enum iser_reg_indicator ind,
608 struct ib_sge *sge)
609 {
610 struct fast_reg_descriptor *desc = regd_buf->reg.mem_h;
611 struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
612 struct iser_device *device = ib_conn->device;
613 struct ib_device *ibdev = device->ib_device;
614 struct ib_mr *mr;
615 struct ib_fast_reg_page_list *frpl;
616 struct ib_send_wr fastreg_wr, inv_wr;
617 struct ib_send_wr *bad_wr, *wr = NULL;
618 u8 key;
619 int ret, offset, size, plen;
620
621 /* if there a single dma entry, dma mr suffices */
622 if (mem->dma_nents == 1) {
623 struct scatterlist *sg = (struct scatterlist *)mem->buf;
624
625 sge->lkey = device->mr->lkey;
626 sge->addr = ib_sg_dma_address(ibdev, &sg[0]);
627 sge->length = ib_sg_dma_len(ibdev, &sg[0]);
628
629 iser_dbg("Single DMA entry: lkey=0x%x, addr=0x%llx, length=0x%x\n",
630 sge->lkey, sge->addr, sge->length);
631 return 0;
632 }
633
634 if (ind == ISER_DATA_KEY_VALID) {
635 mr = desc->data_mr;
636 frpl = desc->data_frpl;
637 } else {
638 mr = desc->pi_ctx->prot_mr;
639 frpl = desc->pi_ctx->prot_frpl;
640 }
641
642 plen = iser_sg_to_page_vec(mem, device->ib_device, frpl->page_list,
643 &offset, &size);
644 if (plen * SIZE_4K < size) {
645 iser_err("fast reg page_list too short to hold this SG\n");
646 return -EINVAL;
647 }
648
649 if (!(desc->reg_indicators & ind)) {
650 memset(&inv_wr, 0, sizeof(inv_wr));
651 inv_wr.wr_id = ISER_FASTREG_LI_WRID;
652 inv_wr.opcode = IB_WR_LOCAL_INV;
653 inv_wr.ex.invalidate_rkey = mr->rkey;
654 wr = &inv_wr;
655 /* Bump the key */
656 key = (u8)(mr->rkey & 0x000000FF);
657 ib_update_fast_reg_key(mr, ++key);
658 }
659
660 /* Prepare FASTREG WR */
661 memset(&fastreg_wr, 0, sizeof(fastreg_wr));
662 fastreg_wr.wr_id = ISER_FASTREG_LI_WRID;
663 fastreg_wr.opcode = IB_WR_FAST_REG_MR;
664 fastreg_wr.wr.fast_reg.iova_start = frpl->page_list[0] + offset;
665 fastreg_wr.wr.fast_reg.page_list = frpl;
666 fastreg_wr.wr.fast_reg.page_list_len = plen;
667 fastreg_wr.wr.fast_reg.page_shift = SHIFT_4K;
668 fastreg_wr.wr.fast_reg.length = size;
669 fastreg_wr.wr.fast_reg.rkey = mr->rkey;
670 fastreg_wr.wr.fast_reg.access_flags = (IB_ACCESS_LOCAL_WRITE |
671 IB_ACCESS_REMOTE_WRITE |
672 IB_ACCESS_REMOTE_READ);
673
674 if (!wr)
675 wr = &fastreg_wr;
676 else
677 wr->next = &fastreg_wr;
678
679 ret = ib_post_send(ib_conn->qp, wr, &bad_wr);
680 if (ret) {
681 iser_err("fast registration failed, ret:%d\n", ret);
682 return ret;
683 }
684 desc->reg_indicators &= ~ind;
685
686 sge->lkey = mr->lkey;
687 sge->addr = frpl->page_list[0] + offset;
688 sge->length = size;
689
690 return ret;
691 }
692
693 /**
694 * iser_reg_rdma_mem_fastreg - Registers memory intended for RDMA,
695 * using Fast Registration WR (if possible) obtaining rkey and va
696 *
697 * returns 0 on success, errno code on failure
698 */
iser_reg_rdma_mem_fastreg(struct iscsi_iser_task * iser_task,enum iser_data_dir cmd_dir)699 int iser_reg_rdma_mem_fastreg(struct iscsi_iser_task *iser_task,
700 enum iser_data_dir cmd_dir)
701 {
702 struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn;
703 struct iser_device *device = ib_conn->device;
704 struct ib_device *ibdev = device->ib_device;
705 struct iser_data_buf *mem = &iser_task->data[cmd_dir];
706 struct iser_regd_buf *regd_buf = &iser_task->rdma_regd[cmd_dir];
707 struct fast_reg_descriptor *desc = NULL;
708 struct ib_sge data_sge;
709 int err, aligned_len;
710 unsigned long flags;
711
712 aligned_len = iser_data_buf_aligned_len(mem, ibdev);
713 if (aligned_len != mem->dma_nents) {
714 err = fall_to_bounce_buf(iser_task, ibdev, mem,
715 &iser_task->data_copy[cmd_dir],
716 cmd_dir, aligned_len);
717 if (err) {
718 iser_err("failed to allocate bounce buffer\n");
719 return err;
720 }
721 mem = &iser_task->data_copy[cmd_dir];
722 }
723
724 if (mem->dma_nents != 1 ||
725 scsi_get_prot_op(iser_task->sc) != SCSI_PROT_NORMAL) {
726 spin_lock_irqsave(&ib_conn->lock, flags);
727 desc = list_first_entry(&ib_conn->fastreg.pool,
728 struct fast_reg_descriptor, list);
729 list_del(&desc->list);
730 spin_unlock_irqrestore(&ib_conn->lock, flags);
731 regd_buf->reg.mem_h = desc;
732 }
733
734 err = iser_fast_reg_mr(iser_task, regd_buf, mem,
735 ISER_DATA_KEY_VALID, &data_sge);
736 if (err)
737 goto err_reg;
738
739 if (scsi_get_prot_op(iser_task->sc) != SCSI_PROT_NORMAL) {
740 struct ib_sge prot_sge, sig_sge;
741
742 memset(&prot_sge, 0, sizeof(prot_sge));
743 if (scsi_prot_sg_count(iser_task->sc)) {
744 mem = &iser_task->prot[cmd_dir];
745 aligned_len = iser_data_buf_aligned_len(mem, ibdev);
746 if (aligned_len != mem->dma_nents) {
747 err = fall_to_bounce_buf(iser_task, ibdev, mem,
748 &iser_task->prot_copy[cmd_dir],
749 cmd_dir, aligned_len);
750 if (err) {
751 iser_err("failed to allocate bounce buffer\n");
752 return err;
753 }
754 mem = &iser_task->prot_copy[cmd_dir];
755 }
756
757 err = iser_fast_reg_mr(iser_task, regd_buf, mem,
758 ISER_PROT_KEY_VALID, &prot_sge);
759 if (err)
760 goto err_reg;
761 }
762
763 err = iser_reg_sig_mr(iser_task, desc, &data_sge,
764 &prot_sge, &sig_sge);
765 if (err) {
766 iser_err("Failed to register signature mr\n");
767 return err;
768 }
769 desc->reg_indicators |= ISER_FASTREG_PROTECTED;
770
771 regd_buf->reg.lkey = sig_sge.lkey;
772 regd_buf->reg.rkey = desc->pi_ctx->sig_mr->rkey;
773 regd_buf->reg.va = sig_sge.addr;
774 regd_buf->reg.len = sig_sge.length;
775 regd_buf->reg.is_mr = 1;
776 } else {
777 if (desc) {
778 regd_buf->reg.rkey = desc->data_mr->rkey;
779 regd_buf->reg.is_mr = 1;
780 } else {
781 regd_buf->reg.rkey = device->mr->rkey;
782 regd_buf->reg.is_mr = 0;
783 }
784
785 regd_buf->reg.lkey = data_sge.lkey;
786 regd_buf->reg.va = data_sge.addr;
787 regd_buf->reg.len = data_sge.length;
788 }
789
790 return 0;
791 err_reg:
792 if (desc) {
793 spin_lock_irqsave(&ib_conn->lock, flags);
794 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
795 spin_unlock_irqrestore(&ib_conn->lock, flags);
796 }
797
798 return err;
799 }
800