IB/iser: Use the new verbs DMA mapping functions
[linux-2.6] / drivers / infiniband / ulp / iser / iser_memory.c
1 /*
2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  * $Id: iser_memory.c 6964 2006-05-07 11:11:43Z ogerlitz $
33  */
34 #include <linux/module.h>
35 #include <linux/kernel.h>
36 #include <linux/slab.h>
37 #include <linux/mm.h>
38 #include <linux/highmem.h>
39 #include <asm/io.h>
40 #include <asm/scatterlist.h>
41 #include <linux/scatterlist.h>
42
43 #include "iscsi_iser.h"
44
45 #define ISER_KMALLOC_THRESHOLD 0x20000 /* 128K - kmalloc limit */
46
47 /**
48  * Decrements the reference count for the
49  * registered buffer & releases it
50  *
51  * returns 0 if released, 1 if deferred
52  */
53 int iser_regd_buff_release(struct iser_regd_buf *regd_buf)
54 {
55         struct ib_device *dev;
56
57         if ((atomic_read(&regd_buf->ref_count) == 0) ||
58             atomic_dec_and_test(&regd_buf->ref_count)) {
59                 /* if we used the dma mr, unreg is just NOP */
60                 if (regd_buf->reg.is_fmr)
61                         iser_unreg_mem(&regd_buf->reg);
62
63                 if (regd_buf->dma_addr) {
64                         dev = regd_buf->device->ib_device;
65                         ib_dma_unmap_single(dev,
66                                          regd_buf->dma_addr,
67                                          regd_buf->data_size,
68                                          regd_buf->direction);
69                 }
70                 /* else this regd buf is associated with task which we */
71                 /* dma_unmap_single/sg later */
72                 return 0;
73         } else {
74                 iser_dbg("Release deferred, regd.buff: 0x%p\n", regd_buf);
75                 return 1;
76         }
77 }
78
79 /**
80  * iser_reg_single - fills registered buffer descriptor with
81  *                   registration information
82  */
83 void iser_reg_single(struct iser_device *device,
84                      struct iser_regd_buf *regd_buf,
85                      enum dma_data_direction direction)
86 {
87         u64 dma_addr;
88
89         dma_addr = ib_dma_map_single(device->ib_device,
90                                      regd_buf->virt_addr,
91                                      regd_buf->data_size, direction);
92         BUG_ON(ib_dma_mapping_error(device->ib_device, dma_addr));
93
94         regd_buf->reg.lkey = device->mr->lkey;
95         regd_buf->reg.len  = regd_buf->data_size;
96         regd_buf->reg.va   = dma_addr;
97         regd_buf->reg.is_fmr = 0;
98
99         regd_buf->dma_addr  = dma_addr;
100         regd_buf->direction = direction;
101 }
102
103 /**
104  * iser_start_rdma_unaligned_sg
105  */
106 int iser_start_rdma_unaligned_sg(struct iscsi_iser_cmd_task  *iser_ctask,
107                                  enum iser_data_dir cmd_dir)
108 {
109         int dma_nents;
110         struct ib_device *dev;
111         char *mem = NULL;
112         struct iser_data_buf *data = &iser_ctask->data[cmd_dir];
113         unsigned long  cmd_data_len = data->data_len;
114
115         if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
116                 mem = (void *)__get_free_pages(GFP_NOIO,
117                       ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
118         else
119                 mem = kmalloc(cmd_data_len, GFP_NOIO);
120
121         if (mem == NULL) {
122                 iser_err("Failed to allocate mem size %d %d for copying sglist\n",
123                          data->size,(int)cmd_data_len);
124                 return -ENOMEM;
125         }
126
127         if (cmd_dir == ISER_DIR_OUT) {
128                 /* copy the unaligned sg the buffer which is used for RDMA */
129                 struct scatterlist *sg = (struct scatterlist *)data->buf;
130                 int i;
131                 char *p, *from;
132
133                 for (p = mem, i = 0; i < data->size; i++) {
134                         from = kmap_atomic(sg[i].page, KM_USER0);
135                         memcpy(p,
136                                from + sg[i].offset,
137                                sg[i].length);
138                         kunmap_atomic(from, KM_USER0);
139                         p += sg[i].length;
140                 }
141         }
142
143         sg_init_one(&iser_ctask->data_copy[cmd_dir].sg_single, mem, cmd_data_len);
144         iser_ctask->data_copy[cmd_dir].buf  =
145                 &iser_ctask->data_copy[cmd_dir].sg_single;
146         iser_ctask->data_copy[cmd_dir].size = 1;
147
148         iser_ctask->data_copy[cmd_dir].copy_buf  = mem;
149
150         dev = iser_ctask->iser_conn->ib_conn->device->ib_device;
151         dma_nents = ib_dma_map_sg(dev,
152                                   &iser_ctask->data_copy[cmd_dir].sg_single,
153                                   1,
154                                   (cmd_dir == ISER_DIR_OUT) ?
155                                   DMA_TO_DEVICE : DMA_FROM_DEVICE);
156         BUG_ON(dma_nents == 0);
157
158         iser_ctask->data_copy[cmd_dir].dma_nents = dma_nents;
159         return 0;
160 }
161
162 /**
163  * iser_finalize_rdma_unaligned_sg
164  */
165 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_cmd_task *iser_ctask,
166                                      enum iser_data_dir         cmd_dir)
167 {
168         struct ib_device *dev;
169         struct iser_data_buf *mem_copy;
170         unsigned long  cmd_data_len;
171
172         dev = iser_ctask->iser_conn->ib_conn->device->ib_device;
173         mem_copy = &iser_ctask->data_copy[cmd_dir];
174
175         ib_dma_unmap_sg(dev, &mem_copy->sg_single, 1,
176                         (cmd_dir == ISER_DIR_OUT) ?
177                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
178
179         if (cmd_dir == ISER_DIR_IN) {
180                 char *mem;
181                 struct scatterlist *sg;
182                 unsigned char *p, *to;
183                 unsigned int sg_size;
184                 int i;
185
186                 /* copy back read RDMA to unaligned sg */
187                 mem     = mem_copy->copy_buf;
188
189                 sg      = (struct scatterlist *)iser_ctask->data[ISER_DIR_IN].buf;
190                 sg_size = iser_ctask->data[ISER_DIR_IN].size;
191
192                 for (p = mem, i = 0; i < sg_size; i++){
193                         to = kmap_atomic(sg[i].page, KM_SOFTIRQ0);
194                         memcpy(to + sg[i].offset,
195                                p,
196                                sg[i].length);
197                         kunmap_atomic(to, KM_SOFTIRQ0);
198                         p += sg[i].length;
199                 }
200         }
201
202         cmd_data_len = iser_ctask->data[cmd_dir].data_len;
203
204         if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
205                 free_pages((unsigned long)mem_copy->copy_buf,
206                            ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
207         else
208                 kfree(mem_copy->copy_buf);
209
210         mem_copy->copy_buf = NULL;
211 }
212
213 /**
214  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
215  * and returns the length of resulting physical address array (may be less than
216  * the original due to possible compaction).
217  *
218  * we build a "page vec" under the assumption that the SG meets the RDMA
219  * alignment requirements. Other then the first and last SG elements, all
220  * the "internal" elements can be compacted into a list whose elements are
221  * dma addresses of physical pages. The code supports also the weird case
222  * where --few fragments of the same page-- are present in the SG as
223  * consecutive elements. Also, it handles one entry SG.
224  */
225 static int iser_sg_to_page_vec(struct iser_data_buf *data,
226                                struct iser_page_vec *page_vec,
227                                struct ib_device *ibdev)
228 {
229         struct scatterlist *sg = (struct scatterlist *)data->buf;
230         u64 first_addr, last_addr, page;
231         int end_aligned;
232         unsigned int cur_page = 0;
233         unsigned long total_sz = 0;
234         int i;
235
236         /* compute the offset of first element */
237         page_vec->offset = (u64) sg[0].offset & ~MASK_4K;
238
239         for (i = 0; i < data->dma_nents; i++) {
240                 unsigned int dma_len = ib_sg_dma_len(ibdev, &sg[i]);
241
242                 total_sz += dma_len;
243
244                 first_addr = ib_sg_dma_address(ibdev, &sg[i]);
245                 last_addr  = first_addr + dma_len;
246
247                 end_aligned   = !(last_addr  & ~MASK_4K);
248
249                 /* continue to collect page fragments till aligned or SG ends */
250                 while (!end_aligned && (i + 1 < data->dma_nents)) {
251                         i++;
252                         dma_len = ib_sg_dma_len(ibdev, &sg[i]);
253                         total_sz += dma_len;
254                         last_addr = ib_sg_dma_address(ibdev, &sg[i]) + dma_len;
255                         end_aligned = !(last_addr  & ~MASK_4K);
256                 }
257
258                 /* handle the 1st page in the 1st DMA element */
259                 if (cur_page == 0) {
260                         page = first_addr & MASK_4K;
261                         page_vec->pages[cur_page] = page;
262                         cur_page++;
263                         page += SIZE_4K;
264                 } else
265                         page = first_addr;
266
267                 for (; page < last_addr; page += SIZE_4K) {
268                         page_vec->pages[cur_page] = page;
269                         cur_page++;
270                 }
271
272         }
273         page_vec->data_size = total_sz;
274         iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page);
275         return cur_page;
276 }
277
278 #define IS_4K_ALIGNED(addr)     ((((unsigned long)addr) & ~MASK_4K) == 0)
279
280 /**
281  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
282  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
283  * the number of entries which are aligned correctly. Supports the case where
284  * consecutive SG elements are actually fragments of the same physcial page.
285  */
286 static unsigned int iser_data_buf_aligned_len(struct iser_data_buf *data,
287                                               struct ib_device *ibdev)
288 {
289         struct scatterlist *sg;
290         u64 end_addr, next_addr;
291         int i, cnt;
292         unsigned int ret_len = 0;
293
294         sg = (struct scatterlist *)data->buf;
295
296         for (cnt = 0, i = 0; i < data->dma_nents; i++, cnt++) {
297                 /* iser_dbg("Checking sg iobuf [%d]: phys=0x%08lX "
298                    "offset: %ld sz: %ld\n", i,
299                    (unsigned long)page_to_phys(sg[i].page),
300                    (unsigned long)sg[i].offset,
301                    (unsigned long)sg[i].length); */
302                 end_addr = ib_sg_dma_address(ibdev, &sg[i]) +
303                            ib_sg_dma_len(ibdev, &sg[i]);
304                 /* iser_dbg("Checking sg iobuf end address "
305                        "0x%08lX\n", end_addr); */
306                 if (i + 1 < data->dma_nents) {
307                         next_addr = ib_sg_dma_address(ibdev, &sg[i+1]);
308                         /* are i, i+1 fragments of the same page? */
309                         if (end_addr == next_addr)
310                                 continue;
311                         else if (!IS_4K_ALIGNED(end_addr)) {
312                                 ret_len = cnt + 1;
313                                 break;
314                         }
315                 }
316         }
317         if (i == data->dma_nents)
318                 ret_len = cnt;  /* loop ended */
319         iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
320                  ret_len, data->dma_nents, data);
321         return ret_len;
322 }
323
324 static void iser_data_buf_dump(struct iser_data_buf *data,
325                                struct ib_device *ibdev)
326 {
327         struct scatterlist *sg = (struct scatterlist *)data->buf;
328         int i;
329
330         for (i = 0; i < data->dma_nents; i++)
331                 iser_err("sg[%d] dma_addr:0x%lX page:0x%p "
332                          "off:0x%x sz:0x%x dma_len:0x%x\n",
333                          i, (unsigned long)ib_sg_dma_address(ibdev, &sg[i]),
334                          sg[i].page, sg[i].offset,
335                          sg[i].length, ib_sg_dma_len(ibdev, &sg[i]));
336 }
337
338 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
339 {
340         int i;
341
342         iser_err("page vec length %d data size %d\n",
343                  page_vec->length, page_vec->data_size);
344         for (i = 0; i < page_vec->length; i++)
345                 iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
346 }
347
348 static void iser_page_vec_build(struct iser_data_buf *data,
349                                 struct iser_page_vec *page_vec,
350                                 struct ib_device *ibdev)
351 {
352         int page_vec_len = 0;
353
354         page_vec->length = 0;
355         page_vec->offset = 0;
356
357         iser_dbg("Translating sg sz: %d\n", data->dma_nents);
358         page_vec_len = iser_sg_to_page_vec(data, page_vec, ibdev);
359         iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len);
360
361         page_vec->length = page_vec_len;
362
363         if (page_vec_len * SIZE_4K < page_vec->data_size) {
364                 iser_err("page_vec too short to hold this SG\n");
365                 iser_data_buf_dump(data, ibdev);
366                 iser_dump_page_vec(page_vec);
367                 BUG();
368         }
369 }
370
371 int iser_dma_map_task_data(struct iscsi_iser_cmd_task *iser_ctask,
372                             struct iser_data_buf       *data,
373                             enum   iser_data_dir       iser_dir,
374                             enum   dma_data_direction  dma_dir)
375 {
376         struct ib_device *dev;
377
378         iser_ctask->dir[iser_dir] = 1;
379         dev = iser_ctask->iser_conn->ib_conn->device->ib_device;
380
381         data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir);
382         if (data->dma_nents == 0) {
383                 iser_err("dma_map_sg failed!!!\n");
384                 return -EINVAL;
385         }
386         return 0;
387 }
388
389 void iser_dma_unmap_task_data(struct iscsi_iser_cmd_task *iser_ctask)
390 {
391         struct ib_device *dev;
392         struct iser_data_buf *data;
393
394         dev = iser_ctask->iser_conn->ib_conn->device->ib_device;
395
396         if (iser_ctask->dir[ISER_DIR_IN]) {
397                 data = &iser_ctask->data[ISER_DIR_IN];
398                 ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE);
399         }
400
401         if (iser_ctask->dir[ISER_DIR_OUT]) {
402                 data = &iser_ctask->data[ISER_DIR_OUT];
403                 ib_dma_unmap_sg(dev, data->buf, data->size, DMA_TO_DEVICE);
404         }
405 }
406
407 /**
408  * iser_reg_rdma_mem - Registers memory intended for RDMA,
409  * obtaining rkey and va
410  *
411  * returns 0 on success, errno code on failure
412  */
413 int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask,
414                       enum   iser_data_dir        cmd_dir)
415 {
416         struct iser_conn     *ib_conn = iser_ctask->iser_conn->ib_conn;
417         struct iser_device   *device = ib_conn->device;
418         struct ib_device     *ibdev = device->ib_device;
419         struct iser_data_buf *mem = &iser_ctask->data[cmd_dir];
420         struct iser_regd_buf *regd_buf;
421         int aligned_len;
422         int err;
423         int i;
424         struct scatterlist *sg;
425
426         regd_buf = &iser_ctask->rdma_regd[cmd_dir];
427
428         aligned_len = iser_data_buf_aligned_len(mem, ibdev);
429         if (aligned_len != mem->dma_nents) {
430                 iser_err("rdma alignment violation %d/%d aligned\n",
431                          aligned_len, mem->size);
432                 iser_data_buf_dump(mem, ibdev);
433
434                 /* unmap the command data before accessing it */
435                 iser_dma_unmap_task_data(iser_ctask);
436
437                 /* allocate copy buf, if we are writing, copy the */
438                 /* unaligned scatterlist, dma map the copy        */
439                 if (iser_start_rdma_unaligned_sg(iser_ctask, cmd_dir) != 0)
440                                 return -ENOMEM;
441                 mem = &iser_ctask->data_copy[cmd_dir];
442         }
443
444         /* if there a single dma entry, FMR is not needed */
445         if (mem->dma_nents == 1) {
446                 sg = (struct scatterlist *)mem->buf;
447
448                 regd_buf->reg.lkey = device->mr->lkey;
449                 regd_buf->reg.rkey = device->mr->rkey;
450                 regd_buf->reg.len  = ib_sg_dma_len(ibdev, &sg[0]);
451                 regd_buf->reg.va   = ib_sg_dma_address(ibdev, &sg[0]);
452                 regd_buf->reg.is_fmr = 0;
453
454                 iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X  "
455                          "va: 0x%08lX sz: %ld]\n",
456                          (unsigned int)regd_buf->reg.lkey,
457                          (unsigned int)regd_buf->reg.rkey,
458                          (unsigned long)regd_buf->reg.va,
459                          (unsigned long)regd_buf->reg.len);
460         } else { /* use FMR for multiple dma entries */
461                 iser_page_vec_build(mem, ib_conn->page_vec, ibdev);
462                 err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, &regd_buf->reg);
463                 if (err) {
464                         iser_data_buf_dump(mem, ibdev);
465                         iser_err("mem->dma_nents = %d (dlength = 0x%x)\n", mem->dma_nents,
466                                  ntoh24(iser_ctask->desc.iscsi_header.dlength));
467                         iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
468                                  ib_conn->page_vec->data_size, ib_conn->page_vec->length,
469                                  ib_conn->page_vec->offset);
470                         for (i=0 ; i<ib_conn->page_vec->length ; i++)
471                                 iser_err("page_vec[%d] = 0x%llx\n", i,
472                                          (unsigned long long) ib_conn->page_vec->pages[i]);
473                         return err;
474                 }
475         }
476
477         /* take a reference on this regd buf such that it will not be released *
478          * (eg in send dto completion) before we get the scsi response         */
479         atomic_inc(&regd_buf->ref_count);
480         return 0;
481 }