Merge branch 'for-2.6.25' of git://git.kernel.org/pub/scm/linux/kernel/git/olof/pasem...
[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 <linux/scatterlist.h>
40
41 #include "iscsi_iser.h"
42
43 #define ISER_KMALLOC_THRESHOLD 0x20000 /* 128K - kmalloc limit */
44
45 /**
46  * Decrements the reference count for the
47  * registered buffer & releases it
48  *
49  * returns 0 if released, 1 if deferred
50  */
51 int iser_regd_buff_release(struct iser_regd_buf *regd_buf)
52 {
53         struct ib_device *dev;
54
55         if ((atomic_read(&regd_buf->ref_count) == 0) ||
56             atomic_dec_and_test(&regd_buf->ref_count)) {
57                 /* if we used the dma mr, unreg is just NOP */
58                 if (regd_buf->reg.is_fmr)
59                         iser_unreg_mem(&regd_buf->reg);
60
61                 if (regd_buf->dma_addr) {
62                         dev = regd_buf->device->ib_device;
63                         ib_dma_unmap_single(dev,
64                                          regd_buf->dma_addr,
65                                          regd_buf->data_size,
66                                          regd_buf->direction);
67                 }
68                 /* else this regd buf is associated with task which we */
69                 /* dma_unmap_single/sg later */
70                 return 0;
71         } else {
72                 iser_dbg("Release deferred, regd.buff: 0x%p\n", regd_buf);
73                 return 1;
74         }
75 }
76
77 /**
78  * iser_reg_single - fills registered buffer descriptor with
79  *                   registration information
80  */
81 void iser_reg_single(struct iser_device *device,
82                      struct iser_regd_buf *regd_buf,
83                      enum dma_data_direction direction)
84 {
85         u64 dma_addr;
86
87         dma_addr = ib_dma_map_single(device->ib_device,
88                                      regd_buf->virt_addr,
89                                      regd_buf->data_size, direction);
90         BUG_ON(ib_dma_mapping_error(device->ib_device, dma_addr));
91
92         regd_buf->reg.lkey = device->mr->lkey;
93         regd_buf->reg.len  = regd_buf->data_size;
94         regd_buf->reg.va   = dma_addr;
95         regd_buf->reg.is_fmr = 0;
96
97         regd_buf->dma_addr  = dma_addr;
98         regd_buf->direction = direction;
99 }
100
101 /**
102  * iser_start_rdma_unaligned_sg
103  */
104 static int iser_start_rdma_unaligned_sg(struct iscsi_iser_cmd_task *iser_ctask,
105                                         enum iser_data_dir cmd_dir)
106 {
107         int dma_nents;
108         struct ib_device *dev;
109         char *mem = NULL;
110         struct iser_data_buf *data = &iser_ctask->data[cmd_dir];
111         unsigned long  cmd_data_len = data->data_len;
112
113         if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
114                 mem = (void *)__get_free_pages(GFP_NOIO,
115                       ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
116         else
117                 mem = kmalloc(cmd_data_len, GFP_NOIO);
118
119         if (mem == NULL) {
120                 iser_err("Failed to allocate mem size %d %d for copying sglist\n",
121                          data->size,(int)cmd_data_len);
122                 return -ENOMEM;
123         }
124
125         if (cmd_dir == ISER_DIR_OUT) {
126                 /* copy the unaligned sg the buffer which is used for RDMA */
127                 struct scatterlist *sgl = (struct scatterlist *)data->buf;
128                 struct scatterlist *sg;
129                 int i;
130                 char *p, *from;
131
132                 p = mem;
133                 for_each_sg(sgl, sg, data->size, i) {
134                         from = kmap_atomic(sg_page(sg), KM_USER0);
135                         memcpy(p,
136                                from + sg->offset,
137                                sg->length);
138                         kunmap_atomic(from, KM_USER0);
139                         p += sg->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 *sgl, *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                 sgl     = (struct scatterlist *)iser_ctask->data[ISER_DIR_IN].buf;
190                 sg_size = iser_ctask->data[ISER_DIR_IN].size;
191
192                 p = mem;
193                 for_each_sg(sgl, sg, sg_size, i) {
194                         to = kmap_atomic(sg_page(sg), KM_SOFTIRQ0);
195                         memcpy(to + sg->offset,
196                                p,
197                                sg->length);
198                         kunmap_atomic(to, KM_SOFTIRQ0);
199                         p += sg->length;
200                 }
201         }
202
203         cmd_data_len = iser_ctask->data[cmd_dir].data_len;
204
205         if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
206                 free_pages((unsigned long)mem_copy->copy_buf,
207                            ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
208         else
209                 kfree(mem_copy->copy_buf);
210
211         mem_copy->copy_buf = NULL;
212 }
213
214 /**
215  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
216  * and returns the length of resulting physical address array (may be less than
217  * the original due to possible compaction).
218  *
219  * we build a "page vec" under the assumption that the SG meets the RDMA
220  * alignment requirements. Other then the first and last SG elements, all
221  * the "internal" elements can be compacted into a list whose elements are
222  * dma addresses of physical pages. The code supports also the weird case
223  * where --few fragments of the same page-- are present in the SG as
224  * consecutive elements. Also, it handles one entry SG.
225  */
226 static int iser_sg_to_page_vec(struct iser_data_buf *data,
227                                struct iser_page_vec *page_vec,
228                                struct ib_device *ibdev)
229 {
230         struct scatterlist *sgl = (struct scatterlist *)data->buf;
231         struct scatterlist *sg;
232         u64 first_addr, last_addr, page;
233         int end_aligned;
234         unsigned int cur_page = 0;
235         unsigned long total_sz = 0;
236         int i;
237
238         /* compute the offset of first element */
239         page_vec->offset = (u64) sgl[0].offset & ~MASK_4K;
240
241         for_each_sg(sgl, sg, data->dma_nents, i) {
242                 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
243
244                 total_sz += dma_len;
245
246                 first_addr = ib_sg_dma_address(ibdev, sg);
247                 last_addr  = first_addr + dma_len;
248
249                 end_aligned   = !(last_addr  & ~MASK_4K);
250
251                 /* continue to collect page fragments till aligned or SG ends */
252                 while (!end_aligned && (i + 1 < data->dma_nents)) {
253                         sg = sg_next(sg);
254                         i++;
255                         dma_len = ib_sg_dma_len(ibdev, sg);
256                         total_sz += dma_len;
257                         last_addr = ib_sg_dma_address(ibdev, sg) + dma_len;
258                         end_aligned = !(last_addr  & ~MASK_4K);
259                 }
260
261                 /* handle the 1st page in the 1st DMA element */
262                 if (cur_page == 0) {
263                         page = first_addr & MASK_4K;
264                         page_vec->pages[cur_page] = page;
265                         cur_page++;
266                         page += SIZE_4K;
267                 } else
268                         page = first_addr;
269
270                 for (; page < last_addr; page += SIZE_4K) {
271                         page_vec->pages[cur_page] = page;
272                         cur_page++;
273                 }
274
275         }
276         page_vec->data_size = total_sz;
277         iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page);
278         return cur_page;
279 }
280
281 #define IS_4K_ALIGNED(addr)     ((((unsigned long)addr) & ~MASK_4K) == 0)
282
283 /**
284  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
285  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
286  * the number of entries which are aligned correctly. Supports the case where
287  * consecutive SG elements are actually fragments of the same physcial page.
288  */
289 static unsigned int iser_data_buf_aligned_len(struct iser_data_buf *data,
290                                               struct ib_device *ibdev)
291 {
292         struct scatterlist *sgl, *sg;
293         u64 end_addr, next_addr;
294         int i, cnt;
295         unsigned int ret_len = 0;
296
297         sgl = (struct scatterlist *)data->buf;
298
299         cnt = 0;
300         for_each_sg(sgl, sg, data->dma_nents, i) {
301                 /* iser_dbg("Checking sg iobuf [%d]: phys=0x%08lX "
302                    "offset: %ld sz: %ld\n", i,
303                    (unsigned long)sg_phys(sg),
304                    (unsigned long)sg->offset,
305                    (unsigned long)sg->length); */
306                 end_addr = ib_sg_dma_address(ibdev, sg) +
307                            ib_sg_dma_len(ibdev, sg);
308                 /* iser_dbg("Checking sg iobuf end address "
309                        "0x%08lX\n", end_addr); */
310                 if (i + 1 < data->dma_nents) {
311                         next_addr = ib_sg_dma_address(ibdev, sg_next(sg));
312                         /* are i, i+1 fragments of the same page? */
313                         if (end_addr == next_addr) {
314                                 cnt++;
315                                 continue;
316                         } else if (!IS_4K_ALIGNED(end_addr)) {
317                                 ret_len = cnt + 1;
318                                 break;
319                         }
320                 }
321                 cnt++;
322         }
323         if (i == data->dma_nents)
324                 ret_len = cnt;  /* loop ended */
325         iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
326                  ret_len, data->dma_nents, data);
327         return ret_len;
328 }
329
330 static void iser_data_buf_dump(struct iser_data_buf *data,
331                                struct ib_device *ibdev)
332 {
333         struct scatterlist *sgl = (struct scatterlist *)data->buf;
334         struct scatterlist *sg;
335         int i;
336
337         for_each_sg(sgl, sg, data->dma_nents, i)
338                 iser_err("sg[%d] dma_addr:0x%lX page:0x%p "
339                          "off:0x%x sz:0x%x dma_len:0x%x\n",
340                          i, (unsigned long)ib_sg_dma_address(ibdev, sg),
341                          sg_page(sg), sg->offset,
342                          sg->length, ib_sg_dma_len(ibdev, sg));
343 }
344
345 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
346 {
347         int i;
348
349         iser_err("page vec length %d data size %d\n",
350                  page_vec->length, page_vec->data_size);
351         for (i = 0; i < page_vec->length; i++)
352                 iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
353 }
354
355 static void iser_page_vec_build(struct iser_data_buf *data,
356                                 struct iser_page_vec *page_vec,
357                                 struct ib_device *ibdev)
358 {
359         int page_vec_len = 0;
360
361         page_vec->length = 0;
362         page_vec->offset = 0;
363
364         iser_dbg("Translating sg sz: %d\n", data->dma_nents);
365         page_vec_len = iser_sg_to_page_vec(data, page_vec, ibdev);
366         iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len);
367
368         page_vec->length = page_vec_len;
369
370         if (page_vec_len * SIZE_4K < page_vec->data_size) {
371                 iser_err("page_vec too short to hold this SG\n");
372                 iser_data_buf_dump(data, ibdev);
373                 iser_dump_page_vec(page_vec);
374                 BUG();
375         }
376 }
377
378 int iser_dma_map_task_data(struct iscsi_iser_cmd_task *iser_ctask,
379                             struct iser_data_buf       *data,
380                             enum   iser_data_dir       iser_dir,
381                             enum   dma_data_direction  dma_dir)
382 {
383         struct ib_device *dev;
384
385         iser_ctask->dir[iser_dir] = 1;
386         dev = iser_ctask->iser_conn->ib_conn->device->ib_device;
387
388         data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir);
389         if (data->dma_nents == 0) {
390                 iser_err("dma_map_sg failed!!!\n");
391                 return -EINVAL;
392         }
393         return 0;
394 }
395
396 void iser_dma_unmap_task_data(struct iscsi_iser_cmd_task *iser_ctask)
397 {
398         struct ib_device *dev;
399         struct iser_data_buf *data;
400
401         dev = iser_ctask->iser_conn->ib_conn->device->ib_device;
402
403         if (iser_ctask->dir[ISER_DIR_IN]) {
404                 data = &iser_ctask->data[ISER_DIR_IN];
405                 ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE);
406         }
407
408         if (iser_ctask->dir[ISER_DIR_OUT]) {
409                 data = &iser_ctask->data[ISER_DIR_OUT];
410                 ib_dma_unmap_sg(dev, data->buf, data->size, DMA_TO_DEVICE);
411         }
412 }
413
414 /**
415  * iser_reg_rdma_mem - Registers memory intended for RDMA,
416  * obtaining rkey and va
417  *
418  * returns 0 on success, errno code on failure
419  */
420 int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask,
421                       enum   iser_data_dir        cmd_dir)
422 {
423         struct iser_conn     *ib_conn = iser_ctask->iser_conn->ib_conn;
424         struct iser_device   *device = ib_conn->device;
425         struct ib_device     *ibdev = device->ib_device;
426         struct iser_data_buf *mem = &iser_ctask->data[cmd_dir];
427         struct iser_regd_buf *regd_buf;
428         int aligned_len;
429         int err;
430         int i;
431         struct scatterlist *sg;
432
433         regd_buf = &iser_ctask->rdma_regd[cmd_dir];
434
435         aligned_len = iser_data_buf_aligned_len(mem, ibdev);
436         if (aligned_len != mem->dma_nents) {
437                 iser_err("rdma alignment violation %d/%d aligned\n",
438                          aligned_len, mem->size);
439                 iser_data_buf_dump(mem, ibdev);
440
441                 /* unmap the command data before accessing it */
442                 iser_dma_unmap_task_data(iser_ctask);
443
444                 /* allocate copy buf, if we are writing, copy the */
445                 /* unaligned scatterlist, dma map the copy        */
446                 if (iser_start_rdma_unaligned_sg(iser_ctask, cmd_dir) != 0)
447                                 return -ENOMEM;
448                 mem = &iser_ctask->data_copy[cmd_dir];
449         }
450
451         /* if there a single dma entry, FMR is not needed */
452         if (mem->dma_nents == 1) {
453                 sg = (struct scatterlist *)mem->buf;
454
455                 regd_buf->reg.lkey = device->mr->lkey;
456                 regd_buf->reg.rkey = device->mr->rkey;
457                 regd_buf->reg.len  = ib_sg_dma_len(ibdev, &sg[0]);
458                 regd_buf->reg.va   = ib_sg_dma_address(ibdev, &sg[0]);
459                 regd_buf->reg.is_fmr = 0;
460
461                 iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X  "
462                          "va: 0x%08lX sz: %ld]\n",
463                          (unsigned int)regd_buf->reg.lkey,
464                          (unsigned int)regd_buf->reg.rkey,
465                          (unsigned long)regd_buf->reg.va,
466                          (unsigned long)regd_buf->reg.len);
467         } else { /* use FMR for multiple dma entries */
468                 iser_page_vec_build(mem, ib_conn->page_vec, ibdev);
469                 err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, &regd_buf->reg);
470                 if (err) {
471                         iser_data_buf_dump(mem, ibdev);
472                         iser_err("mem->dma_nents = %d (dlength = 0x%x)\n", mem->dma_nents,
473                                  ntoh24(iser_ctask->desc.iscsi_header.dlength));
474                         iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
475                                  ib_conn->page_vec->data_size, ib_conn->page_vec->length,
476                                  ib_conn->page_vec->offset);
477                         for (i=0 ; i<ib_conn->page_vec->length ; i++)
478                                 iser_err("page_vec[%d] = 0x%llx\n", i,
479                                          (unsigned long long) ib_conn->page_vec->pages[i]);
480                         return err;
481                 }
482         }
483
484         /* take a reference on this regd buf such that it will not be released *
485          * (eg in send dto completion) before we get the scsi response         */
486         atomic_inc(&regd_buf->ref_count);
487         return 0;
488 }