Merge branch 'for-linus' of master.kernel.org:/pub/scm/linux/kernel/git/roland/infiniband
[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 <asm/io.h>
39 #include <asm/scatterlist.h>
40 #include <linux/scatterlist.h>
41
42 #include "iscsi_iser.h"
43
44 #define ISER_KMALLOC_THRESHOLD 0x20000 /* 128K - kmalloc limit */
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 device *dma_device;
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.rkey != 0)
59                         iser_unreg_mem(&regd_buf->reg);
60
61                 if (regd_buf->dma_addr) {
62                         dma_device = regd_buf->device->ib_device->dma_device;
63                         dma_unmap_single(dma_device,
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         dma_addr_t dma_addr;
86
87         dma_addr  = dma_map_single(device->ib_device->dma_device,
88                                    regd_buf->virt_addr,
89                                    regd_buf->data_size, direction);
90         BUG_ON(dma_mapping_error(dma_addr));
91
92         regd_buf->reg.lkey = device->mr->lkey;
93         regd_buf->reg.rkey = 0; /* indicate there's no need to unreg */
94         regd_buf->reg.len  = regd_buf->data_size;
95         regd_buf->reg.va   = dma_addr;
96
97         regd_buf->dma_addr  = dma_addr;
98         regd_buf->direction = direction;
99 }
100
101 /**
102  * iser_start_rdma_unaligned_sg
103  */
104 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 device *dma_device;
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                       long_log2(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 *sg = (struct scatterlist *)data->buf;
128                 int i;
129                 char *p, *from;
130
131                 for (p = mem, i = 0; i < data->size; i++) {
132                         from = kmap_atomic(sg[i].page, KM_USER0);
133                         memcpy(p,
134                                from + sg[i].offset,
135                                sg[i].length);
136                         kunmap_atomic(from, KM_USER0);
137                         p += sg[i].length;
138                 }
139         }
140
141         sg_init_one(&iser_ctask->data_copy[cmd_dir].sg_single, mem, cmd_data_len);
142         iser_ctask->data_copy[cmd_dir].buf  =
143                 &iser_ctask->data_copy[cmd_dir].sg_single;
144         iser_ctask->data_copy[cmd_dir].size = 1;
145
146         iser_ctask->data_copy[cmd_dir].copy_buf  = mem;
147
148         dma_device = iser_ctask->iser_conn->ib_conn->device->ib_device->dma_device;
149
150         if (cmd_dir == ISER_DIR_OUT)
151                 dma_nents = dma_map_sg(dma_device,
152                                        &iser_ctask->data_copy[cmd_dir].sg_single,
153                                        1, DMA_TO_DEVICE);
154         else
155                 dma_nents = dma_map_sg(dma_device,
156                                        &iser_ctask->data_copy[cmd_dir].sg_single,
157                                        1, DMA_FROM_DEVICE);
158
159         BUG_ON(dma_nents == 0);
160
161         iser_ctask->data_copy[cmd_dir].dma_nents = dma_nents;
162         return 0;
163 }
164
165 /**
166  * iser_finalize_rdma_unaligned_sg
167  */
168 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_cmd_task *iser_ctask,
169                                      enum iser_data_dir         cmd_dir)
170 {
171         struct device *dma_device;
172         struct iser_data_buf *mem_copy;
173         unsigned long  cmd_data_len;
174
175         dma_device = iser_ctask->iser_conn->ib_conn->device->ib_device->dma_device;
176         mem_copy   = &iser_ctask->data_copy[cmd_dir];
177
178         if (cmd_dir == ISER_DIR_OUT)
179                 dma_unmap_sg(dma_device, &mem_copy->sg_single, 1,
180                              DMA_TO_DEVICE);
181         else
182                 dma_unmap_sg(dma_device, &mem_copy->sg_single, 1,
183                              DMA_FROM_DEVICE);
184
185         if (cmd_dir == ISER_DIR_IN) {
186                 char *mem;
187                 struct scatterlist *sg;
188                 unsigned char *p, *to;
189                 unsigned int sg_size;
190                 int i;
191
192                 /* copy back read RDMA to unaligned sg */
193                 mem     = mem_copy->copy_buf;
194
195                 sg      = (struct scatterlist *)iser_ctask->data[ISER_DIR_IN].buf;
196                 sg_size = iser_ctask->data[ISER_DIR_IN].size;
197
198                 for (p = mem, i = 0; i < sg_size; i++){
199                         to = kmap_atomic(sg[i].page, KM_SOFTIRQ0);
200                         memcpy(to + sg[i].offset,
201                                p,
202                                sg[i].length);
203                         kunmap_atomic(to, KM_SOFTIRQ0);
204                         p += sg[i].length;
205                 }
206         }
207
208         cmd_data_len = iser_ctask->data[cmd_dir].data_len;
209
210         if (cmd_data_len > ISER_KMALLOC_THRESHOLD)
211                 free_pages((unsigned long)mem_copy->copy_buf,
212                            long_log2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT);
213         else
214                 kfree(mem_copy->copy_buf);
215
216         mem_copy->copy_buf = NULL;
217 }
218
219 /**
220  * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses
221  * and returns the length of resulting physical address array (may be less than
222  * the original due to possible compaction).
223  *
224  * we build a "page vec" under the assumption that the SG meets the RDMA
225  * alignment requirements. Other then the first and last SG elements, all
226  * the "internal" elements can be compacted into a list whose elements are
227  * dma addresses of physical pages. The code supports also the weird case
228  * where --few fragments of the same page-- are present in the SG as
229  * consecutive elements. Also, it handles one entry SG.
230  */
231 static int iser_sg_to_page_vec(struct iser_data_buf *data,
232                                struct iser_page_vec *page_vec)
233 {
234         struct scatterlist *sg = (struct scatterlist *)data->buf;
235         dma_addr_t first_addr, last_addr, page;
236         int start_aligned, end_aligned;
237         unsigned int cur_page = 0;
238         unsigned long total_sz = 0;
239         int i;
240
241         /* compute the offset of first element */
242         page_vec->offset = (u64) sg[0].offset;
243
244         for (i = 0; i < data->dma_nents; i++) {
245                 total_sz += sg_dma_len(&sg[i]);
246
247                 first_addr = sg_dma_address(&sg[i]);
248                 last_addr  = first_addr + sg_dma_len(&sg[i]);
249
250                 start_aligned = !(first_addr & ~PAGE_MASK);
251                 end_aligned   = !(last_addr  & ~PAGE_MASK);
252
253                 /* continue to collect page fragments till aligned or SG ends */
254                 while (!end_aligned && (i + 1 < data->dma_nents)) {
255                         i++;
256                         total_sz += sg_dma_len(&sg[i]);
257                         last_addr = sg_dma_address(&sg[i]) + sg_dma_len(&sg[i]);
258                         end_aligned = !(last_addr  & ~PAGE_MASK);
259                 }
260
261                 first_addr = first_addr & PAGE_MASK;
262
263                 for (page = first_addr; page < last_addr; page += PAGE_SIZE)
264                         page_vec->pages[cur_page++] = page;
265
266         }
267         page_vec->data_size = total_sz;
268         iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page);
269         return cur_page;
270 }
271
272 #define MASK_4K                 ((1UL << 12) - 1) /* 0xFFF */
273 #define IS_4K_ALIGNED(addr)     ((((unsigned long)addr) & MASK_4K) == 0)
274
275 /**
276  * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned
277  * for RDMA sub-list of a scatter-gather list of memory buffers, and  returns
278  * the number of entries which are aligned correctly. Supports the case where
279  * consecutive SG elements are actually fragments of the same physcial page.
280  */
281 static unsigned int iser_data_buf_aligned_len(struct iser_data_buf *data)
282 {
283         struct scatterlist *sg;
284         dma_addr_t end_addr, next_addr;
285         int i, cnt;
286         unsigned int ret_len = 0;
287
288         sg = (struct scatterlist *)data->buf;
289
290         for (cnt = 0, i = 0; i < data->dma_nents; i++, cnt++) {
291                 /* iser_dbg("Checking sg iobuf [%d]: phys=0x%08lX "
292                    "offset: %ld sz: %ld\n", i,
293                    (unsigned long)page_to_phys(sg[i].page),
294                    (unsigned long)sg[i].offset,
295                    (unsigned long)sg[i].length); */
296                 end_addr = sg_dma_address(&sg[i]) +
297                            sg_dma_len(&sg[i]);
298                 /* iser_dbg("Checking sg iobuf end address "
299                        "0x%08lX\n", end_addr); */
300                 if (i + 1 < data->dma_nents) {
301                         next_addr = sg_dma_address(&sg[i+1]);
302                         /* are i, i+1 fragments of the same page? */
303                         if (end_addr == next_addr)
304                                 continue;
305                         else if (!IS_4K_ALIGNED(end_addr)) {
306                                 ret_len = cnt + 1;
307                                 break;
308                         }
309                 }
310         }
311         if (i == data->dma_nents)
312                 ret_len = cnt;  /* loop ended */
313         iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n",
314                  ret_len, data->dma_nents, data);
315         return ret_len;
316 }
317
318 static void iser_data_buf_dump(struct iser_data_buf *data)
319 {
320         struct scatterlist *sg = (struct scatterlist *)data->buf;
321         int i;
322
323         for (i = 0; i < data->size; i++)
324                 iser_err("sg[%d] dma_addr:0x%lX page:0x%p "
325                          "off:%d sz:%d dma_len:%d\n",
326                          i, (unsigned long)sg_dma_address(&sg[i]),
327                          sg[i].page, sg[i].offset,
328                          sg[i].length,sg_dma_len(&sg[i]));
329 }
330
331 static void iser_dump_page_vec(struct iser_page_vec *page_vec)
332 {
333         int i;
334
335         iser_err("page vec length %d data size %d\n",
336                  page_vec->length, page_vec->data_size);
337         for (i = 0; i < page_vec->length; i++)
338                 iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]);
339 }
340
341 static void iser_page_vec_build(struct iser_data_buf *data,
342                                 struct iser_page_vec *page_vec)
343 {
344         int page_vec_len = 0;
345
346         page_vec->length = 0;
347         page_vec->offset = 0;
348
349         iser_dbg("Translating sg sz: %d\n", data->dma_nents);
350         page_vec_len = iser_sg_to_page_vec(data,page_vec);
351         iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len);
352
353         page_vec->length = page_vec_len;
354
355         if (page_vec_len * PAGE_SIZE < page_vec->data_size) {
356                 iser_err("page_vec too short to hold this SG\n");
357                 iser_data_buf_dump(data);
358                 iser_dump_page_vec(page_vec);
359                 BUG();
360         }
361 }
362
363 /**
364  * iser_reg_rdma_mem - Registers memory intended for RDMA,
365  * obtaining rkey and va
366  *
367  * returns 0 on success, errno code on failure
368  */
369 int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask,
370                       enum   iser_data_dir        cmd_dir)
371 {
372         struct iser_conn     *ib_conn = iser_ctask->iser_conn->ib_conn;
373         struct iser_data_buf *mem = &iser_ctask->data[cmd_dir];
374         struct iser_regd_buf *regd_buf;
375         int aligned_len;
376         int err;
377
378         regd_buf = &iser_ctask->rdma_regd[cmd_dir];
379
380         aligned_len = iser_data_buf_aligned_len(mem);
381         if (aligned_len != mem->size) {
382                 iser_err("rdma alignment violation %d/%d aligned\n",
383                          aligned_len, mem->size);
384                 iser_data_buf_dump(mem);
385                 /* allocate copy buf, if we are writing, copy the */
386                 /* unaligned scatterlist, dma map the copy        */
387                 if (iser_start_rdma_unaligned_sg(iser_ctask, cmd_dir) != 0)
388                                 return -ENOMEM;
389                 mem = &iser_ctask->data_copy[cmd_dir];
390         }
391
392         iser_page_vec_build(mem, ib_conn->page_vec);
393         err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, &regd_buf->reg);
394         if (err)
395                 return err;
396
397         /* take a reference on this regd buf such that it will not be released *
398          * (eg in send dto completion) before we get the scsi response         */
399         atomic_inc(&regd_buf->ref_count);
400         return 0;
401 }