Loading...
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/kernel.h>
34#include <linux/slab.h>
35#include <linux/mm.h>
36#include <linux/highmem.h>
37#include <linux/scatterlist.h>
38
39#include "iscsi_iser.h"
40
41void iser_reg_comp(struct ib_cq *cq, struct ib_wc *wc)
42{
43 iser_err_comp(wc, "memreg");
44}
45
46static struct iser_fr_desc *iser_reg_desc_get_fr(struct ib_conn *ib_conn)
47{
48 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
49 struct iser_fr_desc *desc;
50 unsigned long flags;
51
52 spin_lock_irqsave(&fr_pool->lock, flags);
53 desc = list_first_entry(&fr_pool->list,
54 struct iser_fr_desc, list);
55 list_del(&desc->list);
56 spin_unlock_irqrestore(&fr_pool->lock, flags);
57
58 return desc;
59}
60
61static void iser_reg_desc_put_fr(struct ib_conn *ib_conn,
62 struct iser_fr_desc *desc)
63{
64 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
65 unsigned long flags;
66
67 spin_lock_irqsave(&fr_pool->lock, flags);
68 list_add(&desc->list, &fr_pool->list);
69 spin_unlock_irqrestore(&fr_pool->lock, flags);
70}
71
72int iser_dma_map_task_data(struct iscsi_iser_task *iser_task,
73 enum iser_data_dir iser_dir,
74 enum dma_data_direction dma_dir)
75{
76 struct iser_data_buf *data = &iser_task->data[iser_dir];
77 struct ib_device *dev;
78
79 iser_task->dir[iser_dir] = 1;
80 dev = iser_task->iser_conn->ib_conn.device->ib_device;
81
82 data->dma_nents = ib_dma_map_sg(dev, data->sg, data->size, dma_dir);
83 if (unlikely(data->dma_nents == 0)) {
84 iser_err("dma_map_sg failed!!!\n");
85 return -EINVAL;
86 }
87
88 if (scsi_prot_sg_count(iser_task->sc)) {
89 struct iser_data_buf *pdata = &iser_task->prot[iser_dir];
90
91 pdata->dma_nents = ib_dma_map_sg(dev, pdata->sg, pdata->size, dma_dir);
92 if (unlikely(pdata->dma_nents == 0)) {
93 iser_err("protection dma_map_sg failed!!!\n");
94 goto out_unmap;
95 }
96 }
97
98 return 0;
99
100out_unmap:
101 ib_dma_unmap_sg(dev, data->sg, data->size, dma_dir);
102 return -EINVAL;
103}
104
105
106void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task,
107 enum iser_data_dir iser_dir,
108 enum dma_data_direction dma_dir)
109{
110 struct iser_data_buf *data = &iser_task->data[iser_dir];
111 struct ib_device *dev;
112
113 dev = iser_task->iser_conn->ib_conn.device->ib_device;
114 ib_dma_unmap_sg(dev, data->sg, data->size, dma_dir);
115
116 if (scsi_prot_sg_count(iser_task->sc)) {
117 struct iser_data_buf *pdata = &iser_task->prot[iser_dir];
118
119 ib_dma_unmap_sg(dev, pdata->sg, pdata->size, dma_dir);
120 }
121}
122
123static int iser_reg_dma(struct iser_device *device, struct iser_data_buf *mem,
124 struct iser_mem_reg *reg)
125{
126 struct scatterlist *sg = mem->sg;
127
128 reg->sge.lkey = device->pd->local_dma_lkey;
129 /*
130 * FIXME: rework the registration code path to differentiate
131 * rkey/lkey use cases
132 */
133
134 if (device->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
135 reg->rkey = device->pd->unsafe_global_rkey;
136 else
137 reg->rkey = 0;
138 reg->sge.addr = sg_dma_address(&sg[0]);
139 reg->sge.length = sg_dma_len(&sg[0]);
140
141 iser_dbg("Single DMA entry: lkey=0x%x, rkey=0x%x, addr=0x%llx,"
142 " length=0x%x\n", reg->sge.lkey, reg->rkey,
143 reg->sge.addr, reg->sge.length);
144
145 return 0;
146}
147
148void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
149 enum iser_data_dir cmd_dir)
150{
151 struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
152 struct iser_fr_desc *desc;
153 struct ib_mr_status mr_status;
154
155 desc = reg->desc;
156 if (!desc)
157 return;
158
159 /*
160 * The signature MR cannot be invalidated and reused without checking.
161 * libiscsi calls the check_protection transport handler only if
162 * SCSI-Response is received. And the signature MR is not checked if
163 * the task is completed for some other reason like a timeout or error
164 * handling. That's why we must check the signature MR here before
165 * putting it to the free pool.
166 */
167 if (unlikely(desc->sig_protected)) {
168 desc->sig_protected = false;
169 ib_check_mr_status(desc->rsc.sig_mr, IB_MR_CHECK_SIG_STATUS,
170 &mr_status);
171 }
172 iser_reg_desc_put_fr(&iser_task->iser_conn->ib_conn, reg->desc);
173 reg->desc = NULL;
174}
175
176static void iser_set_dif_domain(struct scsi_cmnd *sc,
177 struct ib_sig_domain *domain)
178{
179 domain->sig_type = IB_SIG_TYPE_T10_DIF;
180 domain->sig.dif.pi_interval = scsi_prot_interval(sc);
181 domain->sig.dif.ref_tag = t10_pi_ref_tag(scsi_cmd_to_rq(sc));
182 /*
183 * At the moment we hard code those, but in the future
184 * we will take them from sc.
185 */
186 domain->sig.dif.apptag_check_mask = 0xffff;
187 domain->sig.dif.app_escape = true;
188 domain->sig.dif.ref_escape = true;
189 if (sc->prot_flags & SCSI_PROT_REF_INCREMENT)
190 domain->sig.dif.ref_remap = true;
191}
192
193static int iser_set_sig_attrs(struct scsi_cmnd *sc,
194 struct ib_sig_attrs *sig_attrs)
195{
196 switch (scsi_get_prot_op(sc)) {
197 case SCSI_PROT_WRITE_INSERT:
198 case SCSI_PROT_READ_STRIP:
199 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
200 iser_set_dif_domain(sc, &sig_attrs->wire);
201 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
202 break;
203 case SCSI_PROT_READ_INSERT:
204 case SCSI_PROT_WRITE_STRIP:
205 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
206 iser_set_dif_domain(sc, &sig_attrs->mem);
207 sig_attrs->mem.sig.dif.bg_type = sc->prot_flags & SCSI_PROT_IP_CHECKSUM ?
208 IB_T10DIF_CSUM : IB_T10DIF_CRC;
209 break;
210 case SCSI_PROT_READ_PASS:
211 case SCSI_PROT_WRITE_PASS:
212 iser_set_dif_domain(sc, &sig_attrs->wire);
213 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
214 iser_set_dif_domain(sc, &sig_attrs->mem);
215 sig_attrs->mem.sig.dif.bg_type = sc->prot_flags & SCSI_PROT_IP_CHECKSUM ?
216 IB_T10DIF_CSUM : IB_T10DIF_CRC;
217 break;
218 default:
219 iser_err("Unsupported PI operation %d\n",
220 scsi_get_prot_op(sc));
221 return -EINVAL;
222 }
223
224 return 0;
225}
226
227static inline void iser_set_prot_checks(struct scsi_cmnd *sc, u8 *mask)
228{
229 *mask = 0;
230 if (sc->prot_flags & SCSI_PROT_REF_CHECK)
231 *mask |= IB_SIG_CHECK_REFTAG;
232 if (sc->prot_flags & SCSI_PROT_GUARD_CHECK)
233 *mask |= IB_SIG_CHECK_GUARD;
234}
235
236static inline void iser_inv_rkey(struct ib_send_wr *inv_wr, struct ib_mr *mr,
237 struct ib_cqe *cqe, struct ib_send_wr *next_wr)
238{
239 inv_wr->opcode = IB_WR_LOCAL_INV;
240 inv_wr->wr_cqe = cqe;
241 inv_wr->ex.invalidate_rkey = mr->rkey;
242 inv_wr->send_flags = 0;
243 inv_wr->num_sge = 0;
244 inv_wr->next = next_wr;
245}
246
247static int iser_reg_sig_mr(struct iscsi_iser_task *iser_task,
248 struct iser_data_buf *mem,
249 struct iser_data_buf *sig_mem,
250 struct iser_reg_resources *rsc,
251 struct iser_mem_reg *sig_reg)
252{
253 struct iser_tx_desc *tx_desc = &iser_task->desc;
254 struct ib_cqe *cqe = &iser_task->iser_conn->ib_conn.reg_cqe;
255 struct ib_mr *mr = rsc->sig_mr;
256 struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
257 struct ib_reg_wr *wr = &tx_desc->reg_wr;
258 int ret;
259
260 memset(sig_attrs, 0, sizeof(*sig_attrs));
261 ret = iser_set_sig_attrs(iser_task->sc, sig_attrs);
262 if (ret)
263 goto err;
264
265 iser_set_prot_checks(iser_task->sc, &sig_attrs->check_mask);
266
267 if (rsc->sig_mr->need_inval)
268 iser_inv_rkey(&tx_desc->inv_wr, mr, cqe, &wr->wr);
269
270 ib_update_fast_reg_key(mr, ib_inc_rkey(mr->rkey));
271
272 ret = ib_map_mr_sg_pi(mr, mem->sg, mem->dma_nents, NULL,
273 sig_mem->sg, sig_mem->dma_nents, NULL, SZ_4K);
274 if (unlikely(ret)) {
275 iser_err("failed to map PI sg (%d)\n",
276 mem->dma_nents + sig_mem->dma_nents);
277 goto err;
278 }
279
280 memset(wr, 0, sizeof(*wr));
281 wr->wr.next = &tx_desc->send_wr;
282 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
283 wr->wr.wr_cqe = cqe;
284 wr->wr.num_sge = 0;
285 wr->wr.send_flags = 0;
286 wr->mr = mr;
287 wr->key = mr->rkey;
288 wr->access = IB_ACCESS_LOCAL_WRITE |
289 IB_ACCESS_REMOTE_READ |
290 IB_ACCESS_REMOTE_WRITE;
291 rsc->sig_mr->need_inval = true;
292
293 sig_reg->sge.lkey = mr->lkey;
294 sig_reg->rkey = mr->rkey;
295 sig_reg->sge.addr = mr->iova;
296 sig_reg->sge.length = mr->length;
297
298 iser_dbg("lkey=0x%x rkey=0x%x addr=0x%llx length=%u\n",
299 sig_reg->sge.lkey, sig_reg->rkey, sig_reg->sge.addr,
300 sig_reg->sge.length);
301err:
302 return ret;
303}
304
305static int iser_fast_reg_mr(struct iscsi_iser_task *iser_task,
306 struct iser_data_buf *mem,
307 struct iser_reg_resources *rsc,
308 struct iser_mem_reg *reg)
309{
310 struct iser_tx_desc *tx_desc = &iser_task->desc;
311 struct ib_cqe *cqe = &iser_task->iser_conn->ib_conn.reg_cqe;
312 struct ib_mr *mr = rsc->mr;
313 struct ib_reg_wr *wr = &tx_desc->reg_wr;
314 int n;
315
316 if (rsc->mr->need_inval)
317 iser_inv_rkey(&tx_desc->inv_wr, mr, cqe, &wr->wr);
318
319 ib_update_fast_reg_key(mr, ib_inc_rkey(mr->rkey));
320
321 n = ib_map_mr_sg(mr, mem->sg, mem->dma_nents, NULL, SZ_4K);
322 if (unlikely(n != mem->dma_nents)) {
323 iser_err("failed to map sg (%d/%d)\n",
324 n, mem->dma_nents);
325 return n < 0 ? n : -EINVAL;
326 }
327
328 wr->wr.next = &tx_desc->send_wr;
329 wr->wr.opcode = IB_WR_REG_MR;
330 wr->wr.wr_cqe = cqe;
331 wr->wr.send_flags = 0;
332 wr->wr.num_sge = 0;
333 wr->mr = mr;
334 wr->key = mr->rkey;
335 wr->access = IB_ACCESS_LOCAL_WRITE |
336 IB_ACCESS_REMOTE_WRITE |
337 IB_ACCESS_REMOTE_READ;
338
339 rsc->mr->need_inval = true;
340
341 reg->sge.lkey = mr->lkey;
342 reg->rkey = mr->rkey;
343 reg->sge.addr = mr->iova;
344 reg->sge.length = mr->length;
345
346 iser_dbg("lkey=0x%x rkey=0x%x addr=0x%llx length=0x%x\n",
347 reg->sge.lkey, reg->rkey, reg->sge.addr, reg->sge.length);
348
349 return 0;
350}
351
352int iser_reg_mem_fastreg(struct iscsi_iser_task *task,
353 enum iser_data_dir dir,
354 bool all_imm)
355{
356 struct ib_conn *ib_conn = &task->iser_conn->ib_conn;
357 struct iser_device *device = ib_conn->device;
358 struct iser_data_buf *mem = &task->data[dir];
359 struct iser_mem_reg *reg = &task->rdma_reg[dir];
360 struct iser_fr_desc *desc;
361 bool use_dma_key;
362 int err;
363
364 use_dma_key = mem->dma_nents == 1 && (all_imm || !iser_always_reg) &&
365 scsi_get_prot_op(task->sc) == SCSI_PROT_NORMAL;
366 if (use_dma_key)
367 return iser_reg_dma(device, mem, reg);
368
369 desc = iser_reg_desc_get_fr(ib_conn);
370 if (scsi_get_prot_op(task->sc) == SCSI_PROT_NORMAL) {
371 err = iser_fast_reg_mr(task, mem, &desc->rsc, reg);
372 if (unlikely(err))
373 goto err_reg;
374 } else {
375 err = iser_reg_sig_mr(task, mem, &task->prot[dir],
376 &desc->rsc, reg);
377 if (unlikely(err))
378 goto err_reg;
379
380 desc->sig_protected = true;
381 }
382
383 reg->desc = desc;
384
385 return 0;
386
387err_reg:
388 iser_reg_desc_put_fr(ib_conn, desc);
389
390 return err;
391}
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 */
47static 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->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
111void 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->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
175static 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 */
228static 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
267static 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
282static 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
292static 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
317int 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->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
335void 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->ib_conn->device->ib_device;
341 ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE);
342}
343
344static 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->ib_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 */
377int iser_reg_rdma_mem_fmr(struct iscsi_iser_task *iser_task,
378 enum iser_data_dir cmd_dir)
379{
380 struct iser_conn *ib_conn = iser_task->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 ®d_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
443static inline enum ib_t10_dif_type
444scsi2ib_prot_type(unsigned char prot_type)
445{
446 switch (prot_type) {
447 case SCSI_PROT_DIF_TYPE0:
448 return IB_T10DIF_NONE;
449 case SCSI_PROT_DIF_TYPE1:
450 return IB_T10DIF_TYPE1;
451 case SCSI_PROT_DIF_TYPE2:
452 return IB_T10DIF_TYPE2;
453 case SCSI_PROT_DIF_TYPE3:
454 return IB_T10DIF_TYPE3;
455 default:
456 return IB_T10DIF_NONE;
457 }
458}
459
460
461static int
462iser_set_sig_attrs(struct scsi_cmnd *sc, struct ib_sig_attrs *sig_attrs)
463{
464 unsigned char scsi_ptype = scsi_get_prot_type(sc);
465
466 sig_attrs->mem.sig_type = IB_SIG_TYPE_T10_DIF;
467 sig_attrs->wire.sig_type = IB_SIG_TYPE_T10_DIF;
468 sig_attrs->mem.sig.dif.pi_interval = sc->device->sector_size;
469 sig_attrs->wire.sig.dif.pi_interval = sc->device->sector_size;
470
471 switch (scsi_get_prot_op(sc)) {
472 case SCSI_PROT_WRITE_INSERT:
473 case SCSI_PROT_READ_STRIP:
474 sig_attrs->mem.sig.dif.type = IB_T10DIF_NONE;
475 sig_attrs->wire.sig.dif.type = scsi2ib_prot_type(scsi_ptype);
476 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
477 sig_attrs->wire.sig.dif.ref_tag = scsi_get_lba(sc) &
478 0xffffffff;
479 break;
480 case SCSI_PROT_READ_INSERT:
481 case SCSI_PROT_WRITE_STRIP:
482 sig_attrs->mem.sig.dif.type = scsi2ib_prot_type(scsi_ptype);
483 sig_attrs->mem.sig.dif.bg_type = IB_T10DIF_CRC;
484 sig_attrs->mem.sig.dif.ref_tag = scsi_get_lba(sc) &
485 0xffffffff;
486 sig_attrs->wire.sig.dif.type = IB_T10DIF_NONE;
487 break;
488 case SCSI_PROT_READ_PASS:
489 case SCSI_PROT_WRITE_PASS:
490 sig_attrs->mem.sig.dif.type = scsi2ib_prot_type(scsi_ptype);
491 sig_attrs->mem.sig.dif.bg_type = IB_T10DIF_CRC;
492 sig_attrs->mem.sig.dif.ref_tag = scsi_get_lba(sc) &
493 0xffffffff;
494 sig_attrs->wire.sig.dif.type = scsi2ib_prot_type(scsi_ptype);
495 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC;
496 sig_attrs->wire.sig.dif.ref_tag = scsi_get_lba(sc) &
497 0xffffffff;
498 break;
499 default:
500 iser_err("Unsupported PI operation %d\n",
501 scsi_get_prot_op(sc));
502 return -EINVAL;
503 }
504 return 0;
505}
506
507
508static int
509iser_set_prot_checks(struct scsi_cmnd *sc, u8 *mask)
510{
511 switch (scsi_get_prot_type(sc)) {
512 case SCSI_PROT_DIF_TYPE0:
513 *mask = 0x0;
514 break;
515 case SCSI_PROT_DIF_TYPE1:
516 case SCSI_PROT_DIF_TYPE2:
517 *mask = ISER_CHECK_GUARD | ISER_CHECK_REFTAG;
518 break;
519 case SCSI_PROT_DIF_TYPE3:
520 *mask = ISER_CHECK_GUARD;
521 break;
522 default:
523 iser_err("Unsupported protection type %d\n",
524 scsi_get_prot_type(sc));
525 return -EINVAL;
526 }
527
528 return 0;
529}
530
531static int
532iser_reg_sig_mr(struct iscsi_iser_task *iser_task,
533 struct fast_reg_descriptor *desc, struct ib_sge *data_sge,
534 struct ib_sge *prot_sge, struct ib_sge *sig_sge)
535{
536 struct iser_conn *ib_conn = iser_task->ib_conn;
537 struct iser_pi_context *pi_ctx = desc->pi_ctx;
538 struct ib_send_wr sig_wr, inv_wr;
539 struct ib_send_wr *bad_wr, *wr = NULL;
540 struct ib_sig_attrs sig_attrs;
541 int ret;
542 u32 key;
543
544 memset(&sig_attrs, 0, sizeof(sig_attrs));
545 ret = iser_set_sig_attrs(iser_task->sc, &sig_attrs);
546 if (ret)
547 goto err;
548
549 ret = iser_set_prot_checks(iser_task->sc, &sig_attrs.check_mask);
550 if (ret)
551 goto err;
552
553 if (!(desc->reg_indicators & ISER_SIG_KEY_VALID)) {
554 memset(&inv_wr, 0, sizeof(inv_wr));
555 inv_wr.opcode = IB_WR_LOCAL_INV;
556 inv_wr.wr_id = ISER_FASTREG_LI_WRID;
557 inv_wr.ex.invalidate_rkey = pi_ctx->sig_mr->rkey;
558 wr = &inv_wr;
559 /* Bump the key */
560 key = (u8)(pi_ctx->sig_mr->rkey & 0x000000FF);
561 ib_update_fast_reg_key(pi_ctx->sig_mr, ++key);
562 }
563
564 memset(&sig_wr, 0, sizeof(sig_wr));
565 sig_wr.opcode = IB_WR_REG_SIG_MR;
566 sig_wr.wr_id = ISER_FASTREG_LI_WRID;
567 sig_wr.sg_list = data_sge;
568 sig_wr.num_sge = 1;
569 sig_wr.wr.sig_handover.sig_attrs = &sig_attrs;
570 sig_wr.wr.sig_handover.sig_mr = pi_ctx->sig_mr;
571 if (scsi_prot_sg_count(iser_task->sc))
572 sig_wr.wr.sig_handover.prot = prot_sge;
573 sig_wr.wr.sig_handover.access_flags = IB_ACCESS_LOCAL_WRITE |
574 IB_ACCESS_REMOTE_READ |
575 IB_ACCESS_REMOTE_WRITE;
576
577 if (!wr)
578 wr = &sig_wr;
579 else
580 wr->next = &sig_wr;
581
582 ret = ib_post_send(ib_conn->qp, wr, &bad_wr);
583 if (ret) {
584 iser_err("reg_sig_mr failed, ret:%d\n", ret);
585 goto err;
586 }
587 desc->reg_indicators &= ~ISER_SIG_KEY_VALID;
588
589 sig_sge->lkey = pi_ctx->sig_mr->lkey;
590 sig_sge->addr = 0;
591 sig_sge->length = data_sge->length + prot_sge->length;
592 if (scsi_get_prot_op(iser_task->sc) == SCSI_PROT_WRITE_INSERT ||
593 scsi_get_prot_op(iser_task->sc) == SCSI_PROT_READ_STRIP) {
594 sig_sge->length += (data_sge->length /
595 iser_task->sc->device->sector_size) * 8;
596 }
597
598 iser_dbg("sig_sge: addr: 0x%llx length: %u lkey: 0x%x\n",
599 sig_sge->addr, sig_sge->length,
600 sig_sge->lkey);
601err:
602 return ret;
603}
604
605static int iser_fast_reg_mr(struct iscsi_iser_task *iser_task,
606 struct iser_regd_buf *regd_buf,
607 struct iser_data_buf *mem,
608 enum iser_reg_indicator ind,
609 struct ib_sge *sge)
610{
611 struct fast_reg_descriptor *desc = regd_buf->reg.mem_h;
612 struct iser_conn *ib_conn = iser_task->ib_conn;
613 struct iser_device *device = ib_conn->device;
614 struct ib_device *ibdev = device->ib_device;
615 struct ib_mr *mr;
616 struct ib_fast_reg_page_list *frpl;
617 struct ib_send_wr fastreg_wr, inv_wr;
618 struct ib_send_wr *bad_wr, *wr = NULL;
619 u8 key;
620 int ret, offset, size, plen;
621
622 /* if there a single dma entry, dma mr suffices */
623 if (mem->dma_nents == 1) {
624 struct scatterlist *sg = (struct scatterlist *)mem->buf;
625
626 sge->lkey = device->mr->lkey;
627 sge->addr = ib_sg_dma_address(ibdev, &sg[0]);
628 sge->length = ib_sg_dma_len(ibdev, &sg[0]);
629
630 iser_dbg("Single DMA entry: lkey=0x%x, addr=0x%llx, length=0x%x\n",
631 sge->lkey, sge->addr, sge->length);
632 return 0;
633 }
634
635 if (ind == ISER_DATA_KEY_VALID) {
636 mr = desc->data_mr;
637 frpl = desc->data_frpl;
638 } else {
639 mr = desc->pi_ctx->prot_mr;
640 frpl = desc->pi_ctx->prot_frpl;
641 }
642
643 plen = iser_sg_to_page_vec(mem, device->ib_device, frpl->page_list,
644 &offset, &size);
645 if (plen * SIZE_4K < size) {
646 iser_err("fast reg page_list too short to hold this SG\n");
647 return -EINVAL;
648 }
649
650 if (!(desc->reg_indicators & ind)) {
651 memset(&inv_wr, 0, sizeof(inv_wr));
652 inv_wr.wr_id = ISER_FASTREG_LI_WRID;
653 inv_wr.opcode = IB_WR_LOCAL_INV;
654 inv_wr.ex.invalidate_rkey = mr->rkey;
655 wr = &inv_wr;
656 /* Bump the key */
657 key = (u8)(mr->rkey & 0x000000FF);
658 ib_update_fast_reg_key(mr, ++key);
659 }
660
661 /* Prepare FASTREG WR */
662 memset(&fastreg_wr, 0, sizeof(fastreg_wr));
663 fastreg_wr.wr_id = ISER_FASTREG_LI_WRID;
664 fastreg_wr.opcode = IB_WR_FAST_REG_MR;
665 fastreg_wr.wr.fast_reg.iova_start = frpl->page_list[0] + offset;
666 fastreg_wr.wr.fast_reg.page_list = frpl;
667 fastreg_wr.wr.fast_reg.page_list_len = plen;
668 fastreg_wr.wr.fast_reg.page_shift = SHIFT_4K;
669 fastreg_wr.wr.fast_reg.length = size;
670 fastreg_wr.wr.fast_reg.rkey = mr->rkey;
671 fastreg_wr.wr.fast_reg.access_flags = (IB_ACCESS_LOCAL_WRITE |
672 IB_ACCESS_REMOTE_WRITE |
673 IB_ACCESS_REMOTE_READ);
674
675 if (!wr)
676 wr = &fastreg_wr;
677 else
678 wr->next = &fastreg_wr;
679
680 ret = ib_post_send(ib_conn->qp, wr, &bad_wr);
681 if (ret) {
682 iser_err("fast registration failed, ret:%d\n", ret);
683 return ret;
684 }
685 desc->reg_indicators &= ~ind;
686
687 sge->lkey = mr->lkey;
688 sge->addr = frpl->page_list[0] + offset;
689 sge->length = size;
690
691 return ret;
692}
693
694/**
695 * iser_reg_rdma_mem_fastreg - Registers memory intended for RDMA,
696 * using Fast Registration WR (if possible) obtaining rkey and va
697 *
698 * returns 0 on success, errno code on failure
699 */
700int iser_reg_rdma_mem_fastreg(struct iscsi_iser_task *iser_task,
701 enum iser_data_dir cmd_dir)
702{
703 struct iser_conn *ib_conn = iser_task->ib_conn;
704 struct iser_device *device = ib_conn->device;
705 struct ib_device *ibdev = device->ib_device;
706 struct iser_data_buf *mem = &iser_task->data[cmd_dir];
707 struct iser_regd_buf *regd_buf = &iser_task->rdma_regd[cmd_dir];
708 struct fast_reg_descriptor *desc = NULL;
709 struct ib_sge data_sge;
710 int err, aligned_len;
711 unsigned long flags;
712
713 aligned_len = iser_data_buf_aligned_len(mem, ibdev);
714 if (aligned_len != mem->dma_nents) {
715 err = fall_to_bounce_buf(iser_task, ibdev, mem,
716 &iser_task->data_copy[cmd_dir],
717 cmd_dir, aligned_len);
718 if (err) {
719 iser_err("failed to allocate bounce buffer\n");
720 return err;
721 }
722 mem = &iser_task->data_copy[cmd_dir];
723 }
724
725 if (mem->dma_nents != 1 ||
726 scsi_get_prot_op(iser_task->sc) != SCSI_PROT_NORMAL) {
727 spin_lock_irqsave(&ib_conn->lock, flags);
728 desc = list_first_entry(&ib_conn->fastreg.pool,
729 struct fast_reg_descriptor, list);
730 list_del(&desc->list);
731 spin_unlock_irqrestore(&ib_conn->lock, flags);
732 regd_buf->reg.mem_h = desc;
733 }
734
735 err = iser_fast_reg_mr(iser_task, regd_buf, mem,
736 ISER_DATA_KEY_VALID, &data_sge);
737 if (err)
738 goto err_reg;
739
740 if (scsi_get_prot_op(iser_task->sc) != SCSI_PROT_NORMAL) {
741 struct ib_sge prot_sge, sig_sge;
742
743 memset(&prot_sge, 0, sizeof(prot_sge));
744 if (scsi_prot_sg_count(iser_task->sc)) {
745 mem = &iser_task->prot[cmd_dir];
746 aligned_len = iser_data_buf_aligned_len(mem, ibdev);
747 if (aligned_len != mem->dma_nents) {
748 err = fall_to_bounce_buf(iser_task, ibdev, mem,
749 &iser_task->prot_copy[cmd_dir],
750 cmd_dir, aligned_len);
751 if (err) {
752 iser_err("failed to allocate bounce buffer\n");
753 return err;
754 }
755 mem = &iser_task->prot_copy[cmd_dir];
756 }
757
758 err = iser_fast_reg_mr(iser_task, regd_buf, mem,
759 ISER_PROT_KEY_VALID, &prot_sge);
760 if (err)
761 goto err_reg;
762 }
763
764 err = iser_reg_sig_mr(iser_task, desc, &data_sge,
765 &prot_sge, &sig_sge);
766 if (err) {
767 iser_err("Failed to register signature mr\n");
768 return err;
769 }
770 desc->reg_indicators |= ISER_FASTREG_PROTECTED;
771
772 regd_buf->reg.lkey = sig_sge.lkey;
773 regd_buf->reg.rkey = desc->pi_ctx->sig_mr->rkey;
774 regd_buf->reg.va = sig_sge.addr;
775 regd_buf->reg.len = sig_sge.length;
776 regd_buf->reg.is_mr = 1;
777 } else {
778 if (desc) {
779 regd_buf->reg.rkey = desc->data_mr->rkey;
780 regd_buf->reg.is_mr = 1;
781 } else {
782 regd_buf->reg.rkey = device->mr->rkey;
783 regd_buf->reg.is_mr = 0;
784 }
785
786 regd_buf->reg.lkey = data_sge.lkey;
787 regd_buf->reg.va = data_sge.addr;
788 regd_buf->reg.len = data_sge.length;
789 }
790
791 return 0;
792err_reg:
793 if (desc) {
794 spin_lock_irqsave(&ib_conn->lock, flags);
795 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
796 spin_unlock_irqrestore(&ib_conn->lock, flags);
797 }
798
799 return err;
800}