Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[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 device *dma_device;
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                         dma_device = regd_buf->device->ib_device->dma_device;
65                         dma_unmap_single(dma_device,
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         dma_addr_t dma_addr;
88
89         dma_addr  = dma_map_single(device->ib_device->dma_device,
90                                    regd_buf->virt_addr,
91                                    regd_buf->data_size, direction);
92         BUG_ON(dma_mapping_error(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 device *dma_device;
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                       long_log2(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         dma_device = iser_ctask->iser_conn->ib_conn->device->ib_device->dma_device;
151
152         if (cmd_dir == ISER_DIR_OUT)
153                 dma_nents = dma_map_sg(dma_device,
154                                        &iser_ctask->data_copy[cmd_dir].sg_single,
155                                        1, DMA_TO_DEVICE);
156         else
157                 dma_nents = dma_map_sg(dma_device,
158                                        &iser_ctask->data_copy[cmd_dir].sg_single,
159                                        1, DMA_FROM_DEVICE);
160
161         BUG_ON(dma_nents == 0);
162
163         iser_ctask->data_copy[cmd_dir].dma_nents = dma_nents;
164         return 0;
165 }
166
167 /**
168  * iser_finalize_rdma_unaligned_sg
169  */
170 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_cmd_task *iser_ctask,
171                                      enum iser_data_dir         cmd_dir)
172 {
173         struct device *dma_device;
174         struct iser_data_buf *mem_copy;
175         unsigned long  cmd_data_len;
176
177         dma_device = iser_ctask->iser_conn->ib_conn->device->ib_device->dma_device;
178         mem_copy   = &iser_ctask->data_copy[cmd_dir];
179
180         if (cmd_dir == ISER_DIR_OUT)
181                 dma_unmap_sg(dma_device, &mem_copy->sg_single, 1,
182                              DMA_TO_DEVICE);
183         else
184                 dma_unmap_sg(dma_device, &mem_copy->sg_single, 1,
185                              DMA_FROM_DEVICE);
186
187         if (cmd_dir == ISER_DIR_IN) {
188                 char *mem;
189                 struct scatterlist *sg;
190                 unsigned char *p, *to;
191                 unsigned int sg_size;
192                 int i;
193
194                 /* copy back read RDMA to unaligned sg */
195                 mem     = mem_copy->copy_buf;
196
197                 sg      = (struct scatterlist *)iser_ctask->data[ISER_DIR_IN].buf;
198                 sg_size = iser_ctask->data[ISER_DIR_IN].size;
199
200                 for (p = mem, i = 0; i < sg_size; i++){
201                         to = kmap_atomic(sg[i].page, KM_SOFTIRQ0);
202                         memcpy(to + sg[i].offset,
203                                p,
204                                sg[i].length);
205                         kunmap_atomic(to, KM_SOFTIRQ0);
206                         p += sg[i].length;
207                 }
208         }
209
210         cmd_data_len = iser_ctask->data[cmd_dir].data_len;
211
212         if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
213                 free_pages((unsigned long)mem_copy->copy_buf,
214                            long_log2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
215         else
216                 kfree(mem_copy->copy_buf);
217
218         mem_copy->copy_buf = NULL;
219 }
220
221 /**
222  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
223  * and returns the length of resulting physical address array (may be less than
224  * the original due to possible compaction).
225  *
226  * we build a "page vec" under the assumption that the SG meets the RDMA
227  * alignment requirements. Other then the first and last SG elements, all
228  * the "internal" elements can be compacted into a list whose elements are
229  * dma addresses of physical pages. The code supports also the weird case
230  * where --few fragments of the same page-- are present in the SG as
231  * consecutive elements. Also, it handles one entry SG.
232  */
233 static int iser_sg_to_page_vec(struct iser_data_buf *data,
234                                struct iser_page_vec *page_vec)
235 {
236         struct scatterlist *sg = (struct scatterlist *)data->buf;
237         dma_addr_t first_addr, last_addr, page;
238         int start_aligned, end_aligned;
239         unsigned int cur_page = 0;
240         unsigned long total_sz = 0;
241         int i;
242
243         /* compute the offset of first element */
244         page_vec->offset = (u64) sg[0].offset & ~MASK_4K;
245
246         for (i = 0; i < data->dma_nents; i++) {
247                 total_sz += sg_dma_len(&sg[i]);
248
249                 first_addr = sg_dma_address(&sg[i]);
250                 last_addr  = first_addr + sg_dma_len(&sg[i]);
251
252                 start_aligned = !(first_addr & ~MASK_4K);
253                 end_aligned   = !(last_addr  & ~MASK_4K);
254
255                 /* continue to collect page fragments till aligned or SG ends */
256                 while (!end_aligned && (i + 1 < data->dma_nents)) {
257                         i++;
258                         total_sz += sg_dma_len(&sg[i]);
259                         last_addr = sg_dma_address(&sg[i]) + sg_dma_len(&sg[i]);
260                         end_aligned = !(last_addr  & ~MASK_4K);
261                 }
262
263                 /* handle the 1st page in the 1st DMA element */
264                 if (cur_page == 0) {
265                         page = first_addr & MASK_4K;
266                         page_vec->pages[cur_page] = page;
267                         cur_page++;
268                         page += SIZE_4K;
269                 } else
270                         page = first_addr;
271
272                 for (; page < last_addr; page += SIZE_4K) {
273                         page_vec->pages[cur_page] = page;
274                         cur_page++;
275                 }
276
277         }
278         page_vec->data_size = total_sz;
279         iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page);
280         return cur_page;
281 }
282
283 #define IS_4K_ALIGNED(addr)     ((((unsigned long)addr) & ~MASK_4K) == 0)
284
285 /**
286  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
287  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
288  * the number of entries which are aligned correctly. Supports the case where
289  * consecutive SG elements are actually fragments of the same physcial page.
290  */
291 static unsigned int iser_data_buf_aligned_len(struct iser_data_buf *data)
292 {
293         struct scatterlist *sg;
294         dma_addr_t end_addr, next_addr;
295         int i, cnt;
296         unsigned int ret_len = 0;
297
298         sg = (struct scatterlist *)data->buf;
299
300         for (cnt = 0, i = 0; i < data->dma_nents; i++, cnt++) {
301                 /* iser_dbg("Checking sg iobuf [%d]: phys=0x%08lX "
302                    "offset: %ld sz: %ld\n", i,
303                    (unsigned long)page_to_phys(sg[i].page),
304                    (unsigned long)sg[i].offset,
305                    (unsigned long)sg[i].length); */
306                 end_addr = sg_dma_address(&sg[i]) +
307                            sg_dma_len(&sg[i]);
308                 /* iser_dbg("Checking sg iobuf end address "
309                        "0x%08lX\n", end_addr); */
310                 if (i + 1 < data->dma_nents) {
311                         next_addr = sg_dma_address(&sg[i+1]);
312                         /* are i, i+1 fragments of the same page? */
313                         if (end_addr == next_addr)
314                                 continue;
315                         else if (!IS_4K_ALIGNED(end_addr)) {
316                                 ret_len = cnt + 1;
317                                 break;
318                         }
319                 }
320         }
321         if (i == data->dma_nents)
322                 ret_len = cnt;  /* loop ended */
323         iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
324                  ret_len, data->dma_nents, data);
325         return ret_len;
326 }
327
328 static void iser_data_buf_dump(struct iser_data_buf *data)
329 {
330         struct scatterlist *sg = (struct scatterlist *)data->buf;
331         int i;
332
333         for (i = 0; i < data->dma_nents; i++)
334                 iser_err("sg[%d] dma_addr:0x%lX page:0x%p "
335                          "off:0x%x sz:0x%x dma_len:0x%x\n",
336                          i, (unsigned long)sg_dma_address(&sg[i]),
337                          sg[i].page, sg[i].offset,
338                          sg[i].length,sg_dma_len(&sg[i]));
339 }
340
341 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
342 {
343         int i;
344
345         iser_err("page vec length %d data size %d\n",
346                  page_vec->length, page_vec->data_size);
347         for (i = 0; i < page_vec->length; i++)
348                 iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
349 }
350
351 static void iser_page_vec_build(struct iser_data_buf *data,
352                                 struct iser_page_vec *page_vec)
353 {
354         int page_vec_len = 0;
355
356         page_vec->length = 0;
357         page_vec->offset = 0;
358
359         iser_dbg("Translating sg sz: %d\n", data->dma_nents);
360         page_vec_len = iser_sg_to_page_vec(data,page_vec);
361         iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len);
362
363         page_vec->length = page_vec_len;
364
365         if (page_vec_len * SIZE_4K < page_vec->data_size) {
366                 iser_err("page_vec too short to hold this SG\n");
367                 iser_data_buf_dump(data);
368                 iser_dump_page_vec(page_vec);
369                 BUG();
370         }
371 }
372
373 int iser_dma_map_task_data(struct iscsi_iser_cmd_task *iser_ctask,
374                             struct iser_data_buf       *data,
375                             enum   iser_data_dir       iser_dir,
376                             enum   dma_data_direction  dma_dir)
377 {
378         struct device *dma_device;
379
380         iser_ctask->dir[iser_dir] = 1;
381         dma_device =
382                 iser_ctask->iser_conn->ib_conn->device->ib_device->dma_device;
383
384         data->dma_nents = dma_map_sg(dma_device, data->buf, data->size, dma_dir);
385         if (data->dma_nents == 0) {
386                 iser_err("dma_map_sg failed!!!\n");
387                 return -EINVAL;
388         }
389         return 0;
390 }
391
392 void iser_dma_unmap_task_data(struct iscsi_iser_cmd_task *iser_ctask)
393 {
394         struct device  *dma_device;
395         struct iser_data_buf *data;
396
397         dma_device =
398                 iser_ctask->iser_conn->ib_conn->device->ib_device->dma_device;
399
400         if (iser_ctask->dir[ISER_DIR_IN]) {
401                 data = &iser_ctask->data[ISER_DIR_IN];
402                 dma_unmap_sg(dma_device, data->buf, data->size, DMA_FROM_DEVICE);
403         }
404
405         if (iser_ctask->dir[ISER_DIR_OUT]) {
406                 data = &iser_ctask->data[ISER_DIR_OUT];
407                 dma_unmap_sg(dma_device, data->buf, data->size, DMA_TO_DEVICE);
408         }
409 }
410
411 /**
412  * iser_reg_rdma_mem - Registers memory intended for RDMA,
413  * obtaining rkey and va
414  *
415  * returns 0 on success, errno code on failure
416  */
417 int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask,
418                       enum   iser_data_dir        cmd_dir)
419 {
420         struct iser_conn     *ib_conn = iser_ctask->iser_conn->ib_conn;
421         struct iser_device   *device = ib_conn->device;
422         struct iser_data_buf *mem = &iser_ctask->data[cmd_dir];
423         struct iser_regd_buf *regd_buf;
424         int aligned_len;
425         int err;
426         int i;
427         struct scatterlist *sg;
428
429         regd_buf = &iser_ctask->rdma_regd[cmd_dir];
430
431         aligned_len = iser_data_buf_aligned_len(mem);
432         if (aligned_len != mem->dma_nents) {
433                 iser_err("rdma alignment violation %d/%d aligned\n",
434                          aligned_len, mem->size);
435                 iser_data_buf_dump(mem);
436
437                 /* unmap the command data before accessing it */
438                 iser_dma_unmap_task_data(iser_ctask);
439
440                 /* allocate copy buf, if we are writing, copy the */
441                 /* unaligned scatterlist, dma map the copy        */
442                 if (iser_start_rdma_unaligned_sg(iser_ctask, cmd_dir) != 0)
443                                 return -ENOMEM;
444                 mem = &iser_ctask->data_copy[cmd_dir];
445         }
446
447         /* if there a single dma entry, FMR is not needed */
448         if (mem->dma_nents == 1) {
449                 sg = (struct scatterlist *)mem->buf;
450
451                 regd_buf->reg.lkey = device->mr->lkey;
452                 regd_buf->reg.rkey = device->mr->rkey;
453                 regd_buf->reg.len  = sg_dma_len(&sg[0]);
454                 regd_buf->reg.va   = sg_dma_address(&sg[0]);
455                 regd_buf->reg.is_fmr = 0;
456
457                 iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X  "
458                          "va: 0x%08lX sz: %ld]\n",
459                          (unsigned int)regd_buf->reg.lkey,
460                          (unsigned int)regd_buf->reg.rkey,
461                          (unsigned long)regd_buf->reg.va,
462                          (unsigned long)regd_buf->reg.len);
463         } else { /* use FMR for multiple dma entries */
464                 iser_page_vec_build(mem, ib_conn->page_vec);
465                 err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, &regd_buf->reg);
466                 if (err) {
467                         iser_data_buf_dump(mem);
468                         iser_err("mem->dma_nents = %d (dlength = 0x%x)\n", mem->dma_nents,
469                                  ntoh24(iser_ctask->desc.iscsi_header.dlength));
470                         iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n",
471                                  ib_conn->page_vec->data_size, ib_conn->page_vec->length,
472                                  ib_conn->page_vec->offset);
473                         for (i=0 ; i<ib_conn->page_vec->length ; i++)
474                                 iser_err("page_vec[%d] = 0x%llx\n", i,
475                                          (unsigned long long) ib_conn->page_vec->pages[i]);
476                         return err;
477                 }
478         }
479
480         /* take a reference on this regd buf such that it will not be released *
481          * (eg in send dto completion) before we get the scsi response         */
482         atomic_inc(&regd_buf->ref_count);
483         return 0;
484 }