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(®d_buf->ref_count) == 0) || 56 atomic_dec_and_test(®d_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(®d_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 *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 dev = iser_ctask->iser_conn->ib_conn->device->ib_device; 149 dma_nents = ib_dma_map_sg(dev, 150 &iser_ctask->data_copy[cmd_dir].sg_single, 151 1, 152 (cmd_dir == ISER_DIR_OUT) ? 153 DMA_TO_DEVICE : DMA_FROM_DEVICE); 154 BUG_ON(dma_nents == 0); 155 156 iser_ctask->data_copy[cmd_dir].dma_nents = dma_nents; 157 return 0; 158 } 159 160 /** 161 * iser_finalize_rdma_unaligned_sg 162 */ 163 void iser_finalize_rdma_unaligned_sg(struct iscsi_iser_cmd_task *iser_ctask, 164 enum iser_data_dir cmd_dir) 165 { 166 struct ib_device *dev; 167 struct iser_data_buf *mem_copy; 168 unsigned long cmd_data_len; 169 170 dev = iser_ctask->iser_conn->ib_conn->device->ib_device; 171 mem_copy = &iser_ctask->data_copy[cmd_dir]; 172 173 ib_dma_unmap_sg(dev, &mem_copy->sg_single, 1, 174 (cmd_dir == ISER_DIR_OUT) ? 175 DMA_TO_DEVICE : DMA_FROM_DEVICE); 176 177 if (cmd_dir == ISER_DIR_IN) { 178 char *mem; 179 struct scatterlist *sg; 180 unsigned char *p, *to; 181 unsigned int sg_size; 182 int i; 183 184 /* copy back read RDMA to unaligned sg */ 185 mem = mem_copy->copy_buf; 186 187 sg = (struct scatterlist *)iser_ctask->data[ISER_DIR_IN].buf; 188 sg_size = iser_ctask->data[ISER_DIR_IN].size; 189 190 for (p = mem, i = 0; i < sg_size; i++){ 191 to = kmap_atomic(sg[i].page, KM_SOFTIRQ0); 192 memcpy(to + sg[i].offset, 193 p, 194 sg[i].length); 195 kunmap_atomic(to, KM_SOFTIRQ0); 196 p += sg[i].length; 197 } 198 } 199 200 cmd_data_len = iser_ctask->data[cmd_dir].data_len; 201 202 if (cmd_data_len > ISER_KMALLOC_THRESHOLD) 203 free_pages((unsigned long)mem_copy->copy_buf, 204 ilog2(roundup_pow_of_two(cmd_data_len)) - PAGE_SHIFT); 205 else 206 kfree(mem_copy->copy_buf); 207 208 mem_copy->copy_buf = NULL; 209 } 210 211 /** 212 * iser_sg_to_page_vec - Translates scatterlist entries to physical addresses 213 * and returns the length of resulting physical address array (may be less than 214 * the original due to possible compaction). 215 * 216 * we build a "page vec" under the assumption that the SG meets the RDMA 217 * alignment requirements. Other then the first and last SG elements, all 218 * the "internal" elements can be compacted into a list whose elements are 219 * dma addresses of physical pages. The code supports also the weird case 220 * where --few fragments of the same page-- are present in the SG as 221 * consecutive elements. Also, it handles one entry SG. 222 */ 223 static int iser_sg_to_page_vec(struct iser_data_buf *data, 224 struct iser_page_vec *page_vec, 225 struct ib_device *ibdev) 226 { 227 struct scatterlist *sg = (struct scatterlist *)data->buf; 228 u64 first_addr, last_addr, page; 229 int end_aligned; 230 unsigned int cur_page = 0; 231 unsigned long total_sz = 0; 232 int i; 233 234 /* compute the offset of first element */ 235 page_vec->offset = (u64) sg[0].offset & ~MASK_4K; 236 237 for (i = 0; i < data->dma_nents; i++) { 238 unsigned int dma_len = ib_sg_dma_len(ibdev, &sg[i]); 239 240 total_sz += dma_len; 241 242 first_addr = ib_sg_dma_address(ibdev, &sg[i]); 243 last_addr = first_addr + dma_len; 244 245 end_aligned = !(last_addr & ~MASK_4K); 246 247 /* continue to collect page fragments till aligned or SG ends */ 248 while (!end_aligned && (i + 1 < data->dma_nents)) { 249 i++; 250 dma_len = ib_sg_dma_len(ibdev, &sg[i]); 251 total_sz += dma_len; 252 last_addr = ib_sg_dma_address(ibdev, &sg[i]) + dma_len; 253 end_aligned = !(last_addr & ~MASK_4K); 254 } 255 256 /* handle the 1st page in the 1st DMA element */ 257 if (cur_page == 0) { 258 page = first_addr & MASK_4K; 259 page_vec->pages[cur_page] = page; 260 cur_page++; 261 page += SIZE_4K; 262 } else 263 page = first_addr; 264 265 for (; page < last_addr; page += SIZE_4K) { 266 page_vec->pages[cur_page] = page; 267 cur_page++; 268 } 269 270 } 271 page_vec->data_size = total_sz; 272 iser_dbg("page_vec->data_size:%d cur_page %d\n", page_vec->data_size,cur_page); 273 return cur_page; 274 } 275 276 #define IS_4K_ALIGNED(addr) ((((unsigned long)addr) & ~MASK_4K) == 0) 277 278 /** 279 * iser_data_buf_aligned_len - Tries to determine the maximal correctly aligned 280 * for RDMA sub-list of a scatter-gather list of memory buffers, and returns 281 * the number of entries which are aligned correctly. Supports the case where 282 * consecutive SG elements are actually fragments of the same physcial page. 283 */ 284 static unsigned int iser_data_buf_aligned_len(struct iser_data_buf *data, 285 struct ib_device *ibdev) 286 { 287 struct scatterlist *sg; 288 u64 end_addr, next_addr; 289 int i, cnt; 290 unsigned int ret_len = 0; 291 292 sg = (struct scatterlist *)data->buf; 293 294 for (cnt = 0, i = 0; i < data->dma_nents; i++, cnt++) { 295 /* iser_dbg("Checking sg iobuf [%d]: phys=0x%08lX " 296 "offset: %ld sz: %ld\n", i, 297 (unsigned long)page_to_phys(sg[i].page), 298 (unsigned long)sg[i].offset, 299 (unsigned long)sg[i].length); */ 300 end_addr = ib_sg_dma_address(ibdev, &sg[i]) + 301 ib_sg_dma_len(ibdev, &sg[i]); 302 /* iser_dbg("Checking sg iobuf end address " 303 "0x%08lX\n", end_addr); */ 304 if (i + 1 < data->dma_nents) { 305 next_addr = ib_sg_dma_address(ibdev, &sg[i+1]); 306 /* are i, i+1 fragments of the same page? */ 307 if (end_addr == next_addr) 308 continue; 309 else if (!IS_4K_ALIGNED(end_addr)) { 310 ret_len = cnt + 1; 311 break; 312 } 313 } 314 } 315 if (i == data->dma_nents) 316 ret_len = cnt; /* loop ended */ 317 iser_dbg("Found %d aligned entries out of %d in sg:0x%p\n", 318 ret_len, data->dma_nents, data); 319 return ret_len; 320 } 321 322 static void iser_data_buf_dump(struct iser_data_buf *data, 323 struct ib_device *ibdev) 324 { 325 struct scatterlist *sg = (struct scatterlist *)data->buf; 326 int i; 327 328 for (i = 0; i < data->dma_nents; i++) 329 iser_err("sg[%d] dma_addr:0x%lX page:0x%p " 330 "off:0x%x sz:0x%x dma_len:0x%x\n", 331 i, (unsigned long)ib_sg_dma_address(ibdev, &sg[i]), 332 sg[i].page, sg[i].offset, 333 sg[i].length, ib_sg_dma_len(ibdev, &sg[i])); 334 } 335 336 static void iser_dump_page_vec(struct iser_page_vec *page_vec) 337 { 338 int i; 339 340 iser_err("page vec length %d data size %d\n", 341 page_vec->length, page_vec->data_size); 342 for (i = 0; i < page_vec->length; i++) 343 iser_err("%d %lx\n",i,(unsigned long)page_vec->pages[i]); 344 } 345 346 static void iser_page_vec_build(struct iser_data_buf *data, 347 struct iser_page_vec *page_vec, 348 struct ib_device *ibdev) 349 { 350 int page_vec_len = 0; 351 352 page_vec->length = 0; 353 page_vec->offset = 0; 354 355 iser_dbg("Translating sg sz: %d\n", data->dma_nents); 356 page_vec_len = iser_sg_to_page_vec(data, page_vec, ibdev); 357 iser_dbg("sg len %d page_vec_len %d\n", data->dma_nents,page_vec_len); 358 359 page_vec->length = page_vec_len; 360 361 if (page_vec_len * SIZE_4K < page_vec->data_size) { 362 iser_err("page_vec too short to hold this SG\n"); 363 iser_data_buf_dump(data, ibdev); 364 iser_dump_page_vec(page_vec); 365 BUG(); 366 } 367 } 368 369 int iser_dma_map_task_data(struct iscsi_iser_cmd_task *iser_ctask, 370 struct iser_data_buf *data, 371 enum iser_data_dir iser_dir, 372 enum dma_data_direction dma_dir) 373 { 374 struct ib_device *dev; 375 376 iser_ctask->dir[iser_dir] = 1; 377 dev = iser_ctask->iser_conn->ib_conn->device->ib_device; 378 379 data->dma_nents = ib_dma_map_sg(dev, data->buf, data->size, dma_dir); 380 if (data->dma_nents == 0) { 381 iser_err("dma_map_sg failed!!!\n"); 382 return -EINVAL; 383 } 384 return 0; 385 } 386 387 void iser_dma_unmap_task_data(struct iscsi_iser_cmd_task *iser_ctask) 388 { 389 struct ib_device *dev; 390 struct iser_data_buf *data; 391 392 dev = iser_ctask->iser_conn->ib_conn->device->ib_device; 393 394 if (iser_ctask->dir[ISER_DIR_IN]) { 395 data = &iser_ctask->data[ISER_DIR_IN]; 396 ib_dma_unmap_sg(dev, data->buf, data->size, DMA_FROM_DEVICE); 397 } 398 399 if (iser_ctask->dir[ISER_DIR_OUT]) { 400 data = &iser_ctask->data[ISER_DIR_OUT]; 401 ib_dma_unmap_sg(dev, data->buf, data->size, DMA_TO_DEVICE); 402 } 403 } 404 405 /** 406 * iser_reg_rdma_mem - Registers memory intended for RDMA, 407 * obtaining rkey and va 408 * 409 * returns 0 on success, errno code on failure 410 */ 411 int iser_reg_rdma_mem(struct iscsi_iser_cmd_task *iser_ctask, 412 enum iser_data_dir cmd_dir) 413 { 414 struct iser_conn *ib_conn = iser_ctask->iser_conn->ib_conn; 415 struct iser_device *device = ib_conn->device; 416 struct ib_device *ibdev = device->ib_device; 417 struct iser_data_buf *mem = &iser_ctask->data[cmd_dir]; 418 struct iser_regd_buf *regd_buf; 419 int aligned_len; 420 int err; 421 int i; 422 struct scatterlist *sg; 423 424 regd_buf = &iser_ctask->rdma_regd[cmd_dir]; 425 426 aligned_len = iser_data_buf_aligned_len(mem, ibdev); 427 if (aligned_len != mem->dma_nents) { 428 iser_err("rdma alignment violation %d/%d aligned\n", 429 aligned_len, mem->size); 430 iser_data_buf_dump(mem, ibdev); 431 432 /* unmap the command data before accessing it */ 433 iser_dma_unmap_task_data(iser_ctask); 434 435 /* allocate copy buf, if we are writing, copy the */ 436 /* unaligned scatterlist, dma map the copy */ 437 if (iser_start_rdma_unaligned_sg(iser_ctask, cmd_dir) != 0) 438 return -ENOMEM; 439 mem = &iser_ctask->data_copy[cmd_dir]; 440 } 441 442 /* if there a single dma entry, FMR is not needed */ 443 if (mem->dma_nents == 1) { 444 sg = (struct scatterlist *)mem->buf; 445 446 regd_buf->reg.lkey = device->mr->lkey; 447 regd_buf->reg.rkey = device->mr->rkey; 448 regd_buf->reg.len = ib_sg_dma_len(ibdev, &sg[0]); 449 regd_buf->reg.va = ib_sg_dma_address(ibdev, &sg[0]); 450 regd_buf->reg.is_fmr = 0; 451 452 iser_dbg("PHYSICAL Mem.register: lkey: 0x%08X rkey: 0x%08X " 453 "va: 0x%08lX sz: %ld]\n", 454 (unsigned int)regd_buf->reg.lkey, 455 (unsigned int)regd_buf->reg.rkey, 456 (unsigned long)regd_buf->reg.va, 457 (unsigned long)regd_buf->reg.len); 458 } else { /* use FMR for multiple dma entries */ 459 iser_page_vec_build(mem, ib_conn->page_vec, ibdev); 460 err = iser_reg_page_vec(ib_conn, ib_conn->page_vec, ®d_buf->reg); 461 if (err) { 462 iser_data_buf_dump(mem, ibdev); 463 iser_err("mem->dma_nents = %d (dlength = 0x%x)\n", mem->dma_nents, 464 ntoh24(iser_ctask->desc.iscsi_header.dlength)); 465 iser_err("page_vec: data_size = 0x%x, length = %d, offset = 0x%x\n", 466 ib_conn->page_vec->data_size, ib_conn->page_vec->length, 467 ib_conn->page_vec->offset); 468 for (i=0 ; i<ib_conn->page_vec->length ; i++) 469 iser_err("page_vec[%d] = 0x%llx\n", i, 470 (unsigned long long) ib_conn->page_vec->pages[i]); 471 return err; 472 } 473 } 474 475 /* take a reference on this regd buf such that it will not be released * 476 * (eg in send dto completion) before we get the scsi response */ 477 atomic_inc(®d_buf->ref_count); 478 return 0; 479 } 480