| 1 | /* |
| 2 | * Copyright 2010-2017 Intel Corporation. |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or modify |
| 5 | * it under the terms of the GNU General Public License, version 2, |
| 6 | * as published by the Free Software Foundation. |
| 7 | * |
| 8 | * This program is distributed in the hope that it will be useful, |
| 9 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 10 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 11 | * General Public License for more details. |
| 12 | * |
| 13 | * Disclaimer: The codes contained in these modules may be specific to |
| 14 | * the Intel Software Development Platform codenamed Knights Ferry, |
| 15 | * and the Intel product codenamed Knights Corner, and are not backward |
| 16 | * compatible with other Intel products. Additionally, Intel will NOT |
| 17 | * support the codes or instruction set in future products. |
| 18 | * |
| 19 | * Intel offers no warranty of any kind regarding the code. This code is |
| 20 | * licensed on an "AS IS" basis and Intel is not obligated to provide |
| 21 | * any support, assistance, installation, training, or other services |
| 22 | * of any kind. Intel is also not obligated to provide any updates, |
| 23 | * enhancements or extensions. Intel specifically disclaims any warranty |
| 24 | * of merchantability, non-infringement, fitness for any particular |
| 25 | * purpose, and any other warranty. |
| 26 | * |
| 27 | * Further, Intel disclaims all liability of any kind, including but |
| 28 | * not limited to liability for infringement of any proprietary rights, |
| 29 | * relating to the use of the code, even if Intel is notified of the |
| 30 | * possibility of such liability. Except as expressly stated in an Intel |
| 31 | * license agreement provided with this code and agreed upon with Intel, |
| 32 | * no license, express or implied, by estoppel or otherwise, to any |
| 33 | * intellectual property rights is granted herein. |
| 34 | */ |
| 35 | |
| 36 | #include "mic/micscif.h" |
| 37 | #include "mic/micscif_smpt.h" |
| 38 | #include "mic/mic_dma_api.h" |
| 39 | #include "mic/micscif_kmem_cache.h" |
| 40 | #include "mic/micscif_rma.h" |
| 41 | #include "mic/micscif_rma_list.h" |
| 42 | #if !defined(WINDOWS) && !defined(CONFIG_PREEMPT) |
| 43 | #include <linux/sched.h> |
| 44 | #endif |
| 45 | #include <linux/highmem.h> |
| 46 | #ifndef _MIC_SCIF_ |
| 47 | #include "mic_common.h" |
| 48 | #endif |
| 49 | #include <linux/nmi.h> |
| 50 | |
| 51 | static __always_inline |
| 52 | void *get_local_va(off_t off, struct reg_range_t *window, size_t len) |
| 53 | { |
| 54 | uint64_t page_nr = (off - window->offset) >> PAGE_SHIFT; |
| 55 | off_t page_off = off & ~PAGE_MASK; |
| 56 | void *va; |
| 57 | |
| 58 | if (RMA_WINDOW_SELF == window->type) { |
| 59 | struct page **pages = window->pinned_pages->pages; |
| 60 | va = (void *)((uint64_t) |
| 61 | (page_address(pages[page_nr])) | page_off); |
| 62 | } else { |
| 63 | dma_addr_t phys = micscif_get_dma_addr(window, off, NULL, NULL, NULL); |
| 64 | #ifdef CONFIG_ML1OM |
| 65 | if (RMA_ERROR_CODE == phys) |
| 66 | return NULL; |
| 67 | #endif |
| 68 | va = (void *)((uint64_t) (phys_to_virt(phys))); |
| 69 | } |
| 70 | return va; |
| 71 | } |
| 72 | |
| 73 | #ifdef _MIC_SCIF_ |
| 74 | static __always_inline |
| 75 | void *ioremap_remote(off_t off, struct reg_range_t *window, |
| 76 | size_t len, bool loopback, struct micscif_dev *dev, int *index, uint64_t *start_off) |
| 77 | { |
| 78 | void *ret; |
| 79 | dma_addr_t phys = micscif_get_dma_addr(window, off, NULL, index, start_off); |
| 80 | |
| 81 | #ifdef CONFIG_ML1OM |
| 82 | if (RMA_ERROR_CODE == phys) |
| 83 | return NULL; |
| 84 | #endif |
| 85 | if (!loopback) |
| 86 | ret = ioremap_nocache(phys, len); |
| 87 | else |
| 88 | ret = (void *)((uint64_t)phys_to_virt(phys)); |
| 89 | return ret; |
| 90 | } |
| 91 | |
| 92 | static __always_inline |
| 93 | void *ioremap_remote_gtt(off_t off, struct reg_range_t *window, |
| 94 | size_t len, bool loopback, struct micscif_dev *dev, int ch_num, struct mic_copy_work *work) |
| 95 | { |
| 96 | return ioremap_remote(off, window, len, loopback, dev, NULL, NULL); |
| 97 | } |
| 98 | #else |
| 99 | static __always_inline |
| 100 | void *ioremap_remote_gtt(off_t off, struct reg_range_t *window, |
| 101 | size_t len, bool loopback, struct micscif_dev *dev, int ch_num, struct mic_copy_work *work) |
| 102 | { |
| 103 | void *ret; |
| 104 | uint64_t page_nr = (off - window->offset) >> PAGE_SHIFT; |
| 105 | off_t page_off = off & ~PAGE_MASK; |
| 106 | if (!loopback) { |
| 107 | dma_addr_t phys = micscif_get_dma_addr(window, off, NULL, NULL, NULL); |
| 108 | /* Ideally there should be a helper to do the +/-1 */ |
| 109 | ret = get_per_dev_ctx(dev->sd_node - 1)->aper.va + phys; |
| 110 | } else { |
| 111 | struct page **pages = ((struct reg_range_t *) |
| 112 | (window->peer_window))->pinned_pages->pages; |
| 113 | ret = (void *)((uint64_t)phys_to_virt(page_to_phys(pages[page_nr])) |
| 114 | | page_off); |
| 115 | } |
| 116 | return ret; |
| 117 | } |
| 118 | |
| 119 | static __always_inline |
| 120 | void *ioremap_remote(off_t off, struct reg_range_t *window, |
| 121 | size_t len, bool loopback, struct micscif_dev *dev, int *index, uint64_t *start_off) |
| 122 | { |
| 123 | void *ret; |
| 124 | int page_nr = (int)((off - window->offset) >> PAGE_SHIFT); |
| 125 | off_t page_off = off & ~PAGE_MASK; |
| 126 | |
| 127 | if (!loopback) { |
| 128 | dma_addr_t phys; |
| 129 | mic_ctx_t *mic_ctx = get_per_dev_ctx(dev->sd_node - 1); |
| 130 | phys = micscif_get_dma_addr(window, off, NULL, index, start_off); |
| 131 | ret = mic_ctx->aper.va + phys; |
| 132 | } else { |
| 133 | struct page **pages = ((struct reg_range_t *) |
| 134 | (window->peer_window))->pinned_pages->pages; |
| 135 | ret = (void *)((uint64_t)phys_to_virt(page_to_phys(pages[page_nr])) |
| 136 | | page_off); |
| 137 | } |
| 138 | return ret; |
| 139 | } |
| 140 | #endif |
| 141 | |
| 142 | static __always_inline void |
| 143 | iounmap_remote(void *virt, size_t size, struct mic_copy_work *work) |
| 144 | { |
| 145 | #ifdef _MIC_SCIF_ |
| 146 | if (!work->loopback) |
| 147 | iounmap(virt); |
| 148 | #endif |
| 149 | } |
| 150 | |
| 151 | /* |
| 152 | * Takes care of ordering issue caused by |
| 153 | * 1. Hardware: Only in the case of cpu copy from host to card because of WC memory. |
| 154 | * 2. Software: If memcpy reorders copy instructions for optimization. This could happen |
| 155 | * at both host and card. |
| 156 | */ |
| 157 | static inline void ordered_memcpy(volatile char *dst, |
| 158 | const char *src, size_t count) |
| 159 | { |
| 160 | if (!count) |
| 161 | return; |
| 162 | |
| 163 | memcpy_toio(dst, src, --count); |
| 164 | wmb(); |
| 165 | *(dst + count) = *(src + count); |
| 166 | } |
| 167 | |
| 168 | static inline void micscif_unaligned_memcpy(volatile char *dst, |
| 169 | const char *src, size_t count, bool ordered) |
| 170 | { |
| 171 | if (unlikely(ordered)) |
| 172 | ordered_memcpy(dst, src, count); |
| 173 | else |
| 174 | memcpy_toio(dst, src, count); |
| 175 | } |
| 176 | |
| 177 | /* |
| 178 | * Copy between rma window and temporary buffer |
| 179 | */ |
| 180 | void micscif_rma_local_cpu_copy(uint64_t offset, struct reg_range_t *window, uint8_t *temp, size_t remaining_len, bool to_temp) |
| 181 | { |
| 182 | void *window_virt; |
| 183 | size_t loop_len; |
| 184 | int offset_in_page; |
| 185 | uint64_t end_offset; |
| 186 | struct list_head *item; |
| 187 | |
| 188 | BUG_ON(RMA_WINDOW_SELF != window->type); |
| 189 | |
| 190 | offset_in_page = offset & ~PAGE_MASK; |
| 191 | loop_len = PAGE_SIZE - offset_in_page; |
| 192 | |
| 193 | if (remaining_len < loop_len) |
| 194 | loop_len = remaining_len; |
| 195 | |
| 196 | if (!(window_virt = get_local_va(offset, window, loop_len))) |
| 197 | return; |
| 198 | if (to_temp) |
| 199 | memcpy(temp, window_virt, loop_len); |
| 200 | else |
| 201 | memcpy(window_virt, temp, loop_len); |
| 202 | |
| 203 | offset += loop_len; |
| 204 | temp += loop_len; |
| 205 | remaining_len -= loop_len; |
| 206 | |
| 207 | end_offset = window->offset + |
| 208 | (window->nr_pages << PAGE_SHIFT); |
| 209 | while (remaining_len) { |
| 210 | if (offset == end_offset) { |
| 211 | item = ( |
| 212 | &window->list_member)->next; |
| 213 | window = list_entry(item, |
| 214 | struct reg_range_t, |
| 215 | list_member); |
| 216 | end_offset = window->offset + |
| 217 | (window->nr_pages << PAGE_SHIFT); |
| 218 | } |
| 219 | |
| 220 | loop_len = min(PAGE_SIZE, remaining_len); |
| 221 | |
| 222 | if (!(window_virt = get_local_va(offset, window, loop_len))) |
| 223 | return; |
| 224 | |
| 225 | if (to_temp) |
| 226 | memcpy(temp, window_virt, loop_len); |
| 227 | else |
| 228 | memcpy(window_virt, temp, loop_len); |
| 229 | |
| 230 | offset += loop_len; |
| 231 | temp += loop_len; |
| 232 | remaining_len -= loop_len; |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | /* |
| 237 | * Comment this |
| 238 | * |
| 239 | */ |
| 240 | static int micscif_rma_list_dma_copy_unaligned(struct mic_copy_work *work, uint8_t *temp, struct dma_channel *chan, bool src_local) |
| 241 | { |
| 242 | struct dma_completion_cb *comp_cb = work->comp_cb; |
| 243 | dma_addr_t window_dma_addr, temp_dma_addr; |
| 244 | #ifndef _MIC_SCIF_ |
| 245 | dma_addr_t temp_phys = comp_cb->temp_phys; |
| 246 | #endif |
| 247 | size_t loop_len, nr_contig_bytes = 0, remaining_len = work->len; |
| 248 | int offset_in_page; |
| 249 | uint64_t end_offset = 0, offset = 0; |
| 250 | struct reg_range_t *window = NULL; |
| 251 | struct list_head *item = NULL; |
| 252 | int ret = 0; |
| 253 | void *window_virt_addr = NULL; |
| 254 | size_t tail_len = 0; |
| 255 | |
| 256 | if (src_local) { |
| 257 | offset = work->dst_offset; |
| 258 | window = work->dst_window; |
| 259 | } else { |
| 260 | offset = work->src_offset; |
| 261 | window = work->src_window; |
| 262 | } |
| 263 | |
| 264 | offset_in_page = offset & (L1_CACHE_BYTES - 1); |
| 265 | if (offset_in_page) { |
| 266 | loop_len = L1_CACHE_BYTES - offset_in_page; |
| 267 | loop_len = min(loop_len, remaining_len); |
| 268 | |
| 269 | if (!(window_virt_addr = ioremap_remote_gtt(offset, window, loop_len, |
| 270 | work->loopback, work->remote_dev, |
| 271 | get_chan_num(chan), work))) |
| 272 | return -ENOMEM; |
| 273 | |
| 274 | if (src_local) { |
| 275 | micscif_unaligned_memcpy(window_virt_addr, temp, loop_len, work->ordered && |
| 276 | !(remaining_len - loop_len)); |
| 277 | serializing_request(window_virt_addr); |
| 278 | } else { |
| 279 | memcpy_fromio(temp, window_virt_addr, loop_len); |
| 280 | serializing_request(temp); |
| 281 | } |
| 282 | #ifdef RMA_DEBUG |
| 283 | atomic_long_add_return(loop_len, &ms_info.rma_unaligned_cpu_cnt); |
| 284 | #endif |
| 285 | smp_mb(); |
| 286 | iounmap_remote(window_virt_addr, loop_len, work); |
| 287 | |
| 288 | offset += loop_len; |
| 289 | temp += loop_len; |
| 290 | #ifndef _MIC_SCIF_ |
| 291 | temp_phys += loop_len; |
| 292 | #endif |
| 293 | remaining_len -= loop_len; |
| 294 | } |
| 295 | |
| 296 | offset_in_page = offset & ~PAGE_MASK; |
| 297 | end_offset = window->offset + |
| 298 | (window->nr_pages << PAGE_SHIFT); |
| 299 | |
| 300 | tail_len = remaining_len & (L1_CACHE_BYTES - 1); |
| 301 | remaining_len -= tail_len; |
| 302 | while (remaining_len) { |
| 303 | if (offset == end_offset) { |
| 304 | item = (&window->list_member)->next; |
| 305 | window = list_entry(item, |
| 306 | struct reg_range_t, |
| 307 | list_member); |
| 308 | end_offset = window->offset + |
| 309 | (window->nr_pages << PAGE_SHIFT); |
| 310 | } |
| 311 | #ifndef _MIC_SCIF_ |
| 312 | temp_dma_addr = temp_phys; |
| 313 | #else |
| 314 | temp_dma_addr = (dma_addr_t)virt_to_phys(temp); |
| 315 | #endif |
| 316 | window_dma_addr = micscif_get_dma_addr(window, offset, &nr_contig_bytes, NULL, NULL); |
| 317 | |
| 318 | #ifdef CONFIG_ML1OM |
| 319 | if (RMA_ERROR_CODE == window_dma_addr) |
| 320 | return -ENXIO; |
| 321 | #endif |
| 322 | loop_len = min(nr_contig_bytes, remaining_len); |
| 323 | |
| 324 | if (src_local) { |
| 325 | if (unlikely(work->ordered && !tail_len && |
| 326 | !(remaining_len - loop_len) && |
| 327 | loop_len != L1_CACHE_BYTES)) { |
| 328 | /* |
| 329 | * Break up the last chunk of the transfer into two steps |
| 330 | * if there is no tail to gurantee DMA ordering. |
| 331 | * Passing DO_DMA_POLLING inserts a status update descriptor |
| 332 | * in step 1 which acts as a double sided synchronization |
| 333 | * fence for the DMA engine to ensure that the last cache line |
| 334 | * in step 2 is updated last. |
| 335 | */ |
| 336 | /* Step 1) DMA: Body Length - L1_CACHE_BYTES. */ |
| 337 | ret = do_dma(chan, DO_DMA_POLLING, temp_dma_addr, window_dma_addr, |
| 338 | loop_len - L1_CACHE_BYTES, NULL); |
| 339 | if (ret < 0) { |
| 340 | printk(KERN_ERR "%s %d Desc Prog Failed ret %d\n", |
| 341 | __func__, __LINE__, ret); |
| 342 | return ret; |
| 343 | } |
| 344 | offset += (loop_len - L1_CACHE_BYTES); |
| 345 | temp_dma_addr += (loop_len - L1_CACHE_BYTES); |
| 346 | window_dma_addr += (loop_len - L1_CACHE_BYTES); |
| 347 | remaining_len -= (loop_len - L1_CACHE_BYTES); |
| 348 | loop_len = remaining_len; |
| 349 | |
| 350 | /* Step 2) DMA: L1_CACHE_BYTES */ |
| 351 | ret = do_dma(chan, 0, temp_dma_addr, window_dma_addr, |
| 352 | loop_len, NULL); |
| 353 | if (ret < 0) { |
| 354 | printk(KERN_ERR "%s %d Desc Prog Failed ret %d\n", |
| 355 | __func__, __LINE__, ret); |
| 356 | return ret; |
| 357 | } |
| 358 | } else { |
| 359 | int flags = 0; |
| 360 | if (remaining_len == loop_len + L1_CACHE_BYTES) |
| 361 | flags = DO_DMA_POLLING; |
| 362 | ret = do_dma(chan, flags, temp_dma_addr, window_dma_addr, |
| 363 | loop_len, NULL); |
| 364 | } |
| 365 | } else { |
| 366 | ret = do_dma(chan, 0, window_dma_addr, temp_dma_addr, |
| 367 | loop_len, NULL); |
| 368 | } |
| 369 | if (ret < 0) { |
| 370 | printk(KERN_ERR "%s %d Desc Prog Failed ret %d\n", |
| 371 | __func__, __LINE__, ret); |
| 372 | return ret; |
| 373 | } |
| 374 | offset += loop_len; |
| 375 | temp += loop_len; |
| 376 | #ifndef _MIC_SCIF_ |
| 377 | temp_phys += loop_len; |
| 378 | #endif |
| 379 | remaining_len -= loop_len; |
| 380 | offset_in_page = 0; |
| 381 | } |
| 382 | if (tail_len) { |
| 383 | if (offset == end_offset) { |
| 384 | item = (&window->list_member)->next; |
| 385 | window = list_entry(item, |
| 386 | struct reg_range_t, |
| 387 | list_member); |
| 388 | end_offset = window->offset + |
| 389 | (window->nr_pages << PAGE_SHIFT); |
| 390 | } |
| 391 | if (!(window_virt_addr = ioremap_remote_gtt(offset, window, tail_len, |
| 392 | work->loopback, work->remote_dev, |
| 393 | get_chan_num(chan), work))) |
| 394 | return -ENOMEM; |
| 395 | |
| 396 | /* |
| 397 | * The CPU copy for the tail bytes must be initiated only once previous |
| 398 | * DMA transfers for this endpoint have completed to guarantee |
| 399 | * ordering. |
| 400 | */ |
| 401 | if (unlikely(work->ordered)) { |
| 402 | free_dma_channel(chan); |
| 403 | work->dma_chan_released = true; |
| 404 | if ((ret = drain_dma_intr(chan))) |
| 405 | return ret; |
| 406 | } |
| 407 | |
| 408 | if (src_local) { |
| 409 | micscif_unaligned_memcpy(window_virt_addr, temp, tail_len, work->ordered); |
| 410 | serializing_request(window_virt_addr); |
| 411 | } else { |
| 412 | memcpy_fromio(temp, window_virt_addr, tail_len); |
| 413 | serializing_request(temp); |
| 414 | } |
| 415 | #ifdef RMA_DEBUG |
| 416 | atomic_long_add_return(tail_len, &ms_info.rma_unaligned_cpu_cnt); |
| 417 | #endif |
| 418 | smp_mb(); |
| 419 | iounmap_remote(window_virt_addr, tail_len, work); |
| 420 | } |
| 421 | if (work->dma_chan_released) { |
| 422 | if ((ret = request_dma_channel(chan))) |
| 423 | return ret; |
| 424 | /* Callee frees the DMA channel lock, if it is held */ |
| 425 | work->dma_chan_released = false; |
| 426 | } |
| 427 | ret = do_dma(chan, DO_DMA_INTR, 0, 0, 0, comp_cb); |
| 428 | if (ret < 0) { |
| 429 | printk(KERN_ERR "%s %d Desc Prog Failed ret %d\n", |
| 430 | __func__, __LINE__, ret); |
| 431 | return ret; |
| 432 | } |
| 433 | return 0; |
| 434 | } |
| 435 | |
| 436 | static inline bool is_local_dma_addr(uint64_t addr) |
| 437 | { |
| 438 | #ifdef _MIC_SCIF_ |
| 439 | return (addr >> PAGE_SHIFT < num_physpages); |
| 440 | #else |
| 441 | return is_syspa(addr); |
| 442 | #endif |
| 443 | } |
| 444 | |
| 445 | /* |
| 446 | * micscif_rma_list_dma_copy_aligned: |
| 447 | * |
| 448 | * Traverse all the windows and perform DMA copy. |
| 449 | */ |
| 450 | static int micscif_rma_list_dma_copy_aligned(struct mic_copy_work *work, struct dma_channel *chan) |
| 451 | { |
| 452 | dma_addr_t src_dma_addr, dst_dma_addr; |
| 453 | size_t loop_len, remaining_len, tail_len, src_contig_bytes = 0, dst_contig_bytes = 0; |
| 454 | int src_cache_off, dst_cache_off, src_last_index = 0, dst_last_index = 0; |
| 455 | uint64_t end_src_offset, end_dst_offset; |
| 456 | void *src_virt, *dst_virt; |
| 457 | struct reg_range_t *src_window = work->src_window; |
| 458 | struct reg_range_t *dst_window = work->dst_window; |
| 459 | uint64_t src_offset = work->src_offset, dst_offset = work->dst_offset; |
| 460 | uint64_t src_start_offset = src_window->offset, dst_start_offset = dst_window->offset; |
| 461 | struct list_head *item; |
| 462 | int ret = 0; |
| 463 | |
| 464 | remaining_len = work->len; |
| 465 | |
| 466 | src_cache_off = src_offset & (L1_CACHE_BYTES - 1); |
| 467 | dst_cache_off = dst_offset & (L1_CACHE_BYTES - 1); |
| 468 | if (src_cache_off != dst_cache_off) { |
| 469 | BUG_ON(1); |
| 470 | } else if (src_cache_off != 0) { |
| 471 | /* Head */ |
| 472 | loop_len = L1_CACHE_BYTES - src_cache_off; |
| 473 | loop_len = min(loop_len, remaining_len); |
| 474 | src_dma_addr = micscif_get_dma_addr(src_window, src_offset, NULL, NULL, NULL); |
| 475 | dst_dma_addr = micscif_get_dma_addr(dst_window, dst_offset, NULL, NULL, NULL); |
| 476 | #ifdef CONFIG_ML1OM |
| 477 | if (RMA_ERROR_CODE == src_dma_addr) |
| 478 | return -ENXIO; |
| 479 | if (RMA_ERROR_CODE == dst_dma_addr) |
| 480 | return -ENXIO; |
| 481 | get_window_ref_count(src_window, 1); |
| 482 | get_window_ref_count(dst_window, 1); |
| 483 | #endif |
| 484 | if (is_local_dma_addr(src_dma_addr)) |
| 485 | src_virt = get_local_va(src_offset, src_window, loop_len); |
| 486 | else |
| 487 | src_virt = ioremap_remote_gtt(src_offset, src_window, |
| 488 | loop_len, work->loopback, |
| 489 | work->remote_dev, get_chan_num(chan), work); |
| 490 | if (!src_virt) { |
| 491 | #ifdef CONFIG_ML1OM |
| 492 | put_window_ref_count(src_window, 1); |
| 493 | put_window_ref_count(dst_window, 1); |
| 494 | #endif |
| 495 | return -ENOMEM; |
| 496 | } |
| 497 | if (is_local_dma_addr(dst_dma_addr)) |
| 498 | dst_virt = get_local_va(dst_offset, dst_window, loop_len); |
| 499 | else |
| 500 | dst_virt = ioremap_remote_gtt(dst_offset, dst_window, |
| 501 | loop_len, work->loopback, |
| 502 | work->remote_dev, get_chan_num(chan), work); |
| 503 | #ifdef CONFIG_ML1OM |
| 504 | put_window_ref_count(src_window, 1); |
| 505 | put_window_ref_count(dst_window, 1); |
| 506 | #endif |
| 507 | if (!dst_virt) { |
| 508 | if (!is_local_dma_addr(src_dma_addr)) |
| 509 | iounmap_remote(src_virt, loop_len, work); |
| 510 | return -ENOMEM; |
| 511 | } |
| 512 | if (is_local_dma_addr(src_dma_addr)){ |
| 513 | micscif_unaligned_memcpy(dst_virt, src_virt, loop_len, |
| 514 | remaining_len == loop_len ? work->ordered : false); |
| 515 | } |
| 516 | else{ |
| 517 | memcpy_fromio(dst_virt, src_virt, loop_len); |
| 518 | } |
| 519 | serializing_request(dst_virt); |
| 520 | smp_mb(); |
| 521 | if (!is_local_dma_addr(src_dma_addr)) |
| 522 | iounmap_remote(src_virt, loop_len, work); |
| 523 | if (!is_local_dma_addr(dst_dma_addr)) |
| 524 | iounmap_remote(dst_virt, loop_len, work); |
| 525 | src_offset += loop_len; |
| 526 | dst_offset += loop_len; |
| 527 | remaining_len -= loop_len; |
| 528 | } |
| 529 | |
| 530 | end_src_offset = src_window->offset + |
| 531 | (src_window->nr_pages << PAGE_SHIFT); |
| 532 | end_dst_offset = dst_window->offset + |
| 533 | (dst_window->nr_pages << PAGE_SHIFT); |
| 534 | tail_len = remaining_len & (L1_CACHE_BYTES - 1); |
| 535 | remaining_len -= tail_len; |
| 536 | while (remaining_len) { |
| 537 | if (src_offset == end_src_offset) { |
| 538 | item = (&src_window->list_member)->next; |
| 539 | src_window = list_entry(item, |
| 540 | struct reg_range_t, |
| 541 | list_member); |
| 542 | end_src_offset = src_window->offset + |
| 543 | (src_window->nr_pages << PAGE_SHIFT); |
| 544 | src_last_index = 0; |
| 545 | src_start_offset = src_window->offset; |
| 546 | } |
| 547 | if (dst_offset == end_dst_offset) { |
| 548 | item = (&dst_window->list_member)->next; |
| 549 | dst_window = list_entry(item, struct reg_range_t, list_member); |
| 550 | end_dst_offset = dst_window->offset + |
| 551 | (dst_window->nr_pages << PAGE_SHIFT); |
| 552 | dst_last_index = 0; |
| 553 | dst_start_offset = dst_window->offset; |
| 554 | } |
| 555 | |
| 556 | /* compute dma addresses for transfer */ |
| 557 | src_dma_addr = micscif_get_dma_addr(src_window, src_offset, &src_contig_bytes, &src_last_index, &src_start_offset); |
| 558 | dst_dma_addr = micscif_get_dma_addr(dst_window, dst_offset, &dst_contig_bytes, &dst_last_index, &dst_start_offset); |
| 559 | #ifdef CONFIG_ML1OM |
| 560 | if (RMA_ERROR_CODE == src_dma_addr) |
| 561 | return -ENXIO; |
| 562 | if (RMA_ERROR_CODE == dst_dma_addr) |
| 563 | return -ENXIO; |
| 564 | #endif |
| 565 | loop_len = min(src_contig_bytes, dst_contig_bytes); |
| 566 | loop_len = min(loop_len, remaining_len); |
| 567 | if (unlikely(work->ordered && !tail_len && |
| 568 | !(remaining_len - loop_len) && |
| 569 | loop_len != L1_CACHE_BYTES)) { |
| 570 | /* |
| 571 | * Break up the last chunk of the transfer into two steps |
| 572 | * if there is no tail to gurantee DMA ordering. |
| 573 | * Passing DO_DMA_POLLING inserts a status update descriptor |
| 574 | * in step 1 which acts as a double sided synchronization |
| 575 | * fence for the DMA engine to ensure that the last cache line |
| 576 | * in step 2 is updated last. |
| 577 | */ |
| 578 | /* Step 1) DMA: Body Length - L1_CACHE_BYTES. */ |
| 579 | ret = do_dma(chan, DO_DMA_POLLING, src_dma_addr, dst_dma_addr, |
| 580 | loop_len - L1_CACHE_BYTES, NULL); |
| 581 | if (ret < 0) { |
| 582 | printk(KERN_ERR "%s %d Desc Prog Failed ret %d\n", |
| 583 | __func__, __LINE__, ret); |
| 584 | return ret; |
| 585 | } |
| 586 | src_offset += (loop_len - L1_CACHE_BYTES); |
| 587 | dst_offset += (loop_len - L1_CACHE_BYTES); |
| 588 | src_dma_addr += (loop_len - L1_CACHE_BYTES); |
| 589 | dst_dma_addr += (loop_len - L1_CACHE_BYTES); |
| 590 | remaining_len -= (loop_len - L1_CACHE_BYTES); |
| 591 | loop_len = remaining_len; |
| 592 | |
| 593 | /* Step 2) DMA: L1_CACHE_BYTES */ |
| 594 | ret = do_dma(chan, 0, src_dma_addr, dst_dma_addr, |
| 595 | loop_len, NULL); |
| 596 | if (ret < 0) { |
| 597 | printk(KERN_ERR "%s %d Desc Prog Failed ret %d\n", |
| 598 | __func__, __LINE__, ret); |
| 599 | return ret; |
| 600 | } |
| 601 | } else { |
| 602 | int flags = 0; |
| 603 | if (remaining_len == loop_len + L1_CACHE_BYTES) |
| 604 | flags = DO_DMA_POLLING; |
| 605 | ret = do_dma(chan, flags, src_dma_addr, dst_dma_addr, |
| 606 | loop_len, NULL); |
| 607 | if (ret < 0) { |
| 608 | printk(KERN_ERR "%s %d Desc Prog Failed ret %d\n", |
| 609 | __func__, __LINE__, ret); |
| 610 | return ret; |
| 611 | } |
| 612 | } |
| 613 | src_offset += loop_len; |
| 614 | dst_offset += loop_len; |
| 615 | remaining_len -= loop_len; |
| 616 | } |
| 617 | #ifdef CONFIG_MK1OM |
| 618 | BUG_ON(remaining_len != 0); |
| 619 | #endif |
| 620 | #ifdef CONFIG_ML1OM |
| 621 | if (remaining_len) |
| 622 | return - ENXIO; |
| 623 | #endif |
| 624 | remaining_len = tail_len; |
| 625 | if (remaining_len) { |
| 626 | loop_len = remaining_len; |
| 627 | if (src_offset == end_src_offset) { |
| 628 | item = (&src_window->list_member)->next; |
| 629 | src_window = list_entry(item, |
| 630 | struct reg_range_t, |
| 631 | list_member); |
| 632 | } |
| 633 | if (dst_offset == end_dst_offset) { |
| 634 | item = (&dst_window->list_member)->next; |
| 635 | dst_window = list_entry(item, struct reg_range_t, list_member); |
| 636 | } |
| 637 | |
| 638 | src_dma_addr = micscif_get_dma_addr(src_window, src_offset, NULL, NULL, NULL); |
| 639 | dst_dma_addr = micscif_get_dma_addr(dst_window, dst_offset, NULL, NULL, NULL); |
| 640 | #ifdef CONFIG_ML1OM |
| 641 | if (RMA_ERROR_CODE == src_dma_addr) |
| 642 | return -ENXIO; |
| 643 | if (RMA_ERROR_CODE == dst_dma_addr) |
| 644 | return -ENXIO; |
| 645 | #endif |
| 646 | /* |
| 647 | * The CPU copy for the tail bytes must be initiated only once previous |
| 648 | * DMA transfers for this endpoint have completed to guarantee |
| 649 | * ordering. |
| 650 | */ |
| 651 | if (unlikely(work->ordered)) { |
| 652 | free_dma_channel(chan); |
| 653 | work->dma_chan_released = true; |
| 654 | if ((ret = drain_dma_poll(chan))) |
| 655 | return ret; |
| 656 | } |
| 657 | #ifdef CONFIG_ML1OM |
| 658 | get_window_ref_count(src_window, 1); |
| 659 | get_window_ref_count(dst_window, 1); |
| 660 | #endif |
| 661 | if (is_local_dma_addr(src_dma_addr)) |
| 662 | src_virt = get_local_va(src_offset, src_window, loop_len); |
| 663 | else |
| 664 | src_virt = ioremap_remote_gtt(src_offset, src_window, |
| 665 | loop_len, work->loopback, |
| 666 | work->remote_dev, get_chan_num(chan), work); |
| 667 | if (!src_virt) { |
| 668 | #ifdef CONFIG_ML1OM |
| 669 | put_window_ref_count(src_window, 1); |
| 670 | put_window_ref_count(dst_window, 1); |
| 671 | #endif |
| 672 | return -ENOMEM; |
| 673 | } |
| 674 | |
| 675 | if (is_local_dma_addr(dst_dma_addr)) |
| 676 | dst_virt = get_local_va(dst_offset, dst_window, loop_len); |
| 677 | else |
| 678 | dst_virt = ioremap_remote_gtt(dst_offset, dst_window, |
| 679 | loop_len, work->loopback, |
| 680 | work->remote_dev, get_chan_num(chan), work); |
| 681 | #ifdef CONFIG_ML1OM |
| 682 | put_window_ref_count(src_window, 1); |
| 683 | put_window_ref_count(dst_window, 1); |
| 684 | #endif |
| 685 | if (!dst_virt) { |
| 686 | if (!is_local_dma_addr(src_dma_addr)) |
| 687 | iounmap_remote(src_virt, loop_len, work); |
| 688 | return -ENOMEM; |
| 689 | } |
| 690 | |
| 691 | if (is_local_dma_addr(src_dma_addr)){ |
| 692 | micscif_unaligned_memcpy(dst_virt, src_virt, loop_len, work->ordered); |
| 693 | } |
| 694 | else{ |
| 695 | memcpy_fromio(dst_virt, src_virt, loop_len); |
| 696 | } |
| 697 | serializing_request(dst_virt); |
| 698 | smp_mb(); |
| 699 | if (!is_local_dma_addr(src_dma_addr)) |
| 700 | iounmap_remote(src_virt, loop_len, work); |
| 701 | |
| 702 | if (!is_local_dma_addr(dst_dma_addr)) |
| 703 | iounmap_remote(dst_virt, loop_len, work); |
| 704 | |
| 705 | remaining_len -= loop_len; |
| 706 | #ifdef CONFIG_MK1OM |
| 707 | BUG_ON(remaining_len != 0); |
| 708 | #endif |
| 709 | #ifdef CONFIG_ML1OM |
| 710 | if (remaining_len) |
| 711 | return - ENXIO; |
| 712 | #endif |
| 713 | } |
| 714 | |
| 715 | return ret; |
| 716 | } |
| 717 | |
| 718 | int micscif_rma_list_dma_copy_wrapper(struct endpt *epd, struct mic_copy_work *work, struct dma_channel *chan, off_t loffset) |
| 719 | { |
| 720 | int src_cache_off, dst_cache_off; |
| 721 | uint64_t src_offset = work->src_offset, dst_offset = work->dst_offset; |
| 722 | uint8_t *temp = NULL; |
| 723 | bool src_local = true, dst_local = false; |
| 724 | struct dma_completion_cb *comp_cb; |
| 725 | dma_addr_t src_dma_addr, dst_dma_addr; |
| 726 | #ifndef _MIC_SCIF_ |
| 727 | struct pci_dev *pdev; |
| 728 | #endif |
| 729 | |
| 730 | src_cache_off = src_offset & (L1_CACHE_BYTES - 1); |
| 731 | dst_cache_off = dst_offset & (L1_CACHE_BYTES - 1); |
| 732 | if (dst_cache_off == src_cache_off) |
| 733 | return micscif_rma_list_dma_copy_aligned(work, chan); |
| 734 | |
| 735 | if (work->loopback) { |
| 736 | #ifdef _MIC_SCIF_ |
| 737 | BUG_ON(micscif_rma_list_cpu_copy(work)); |
| 738 | return 0; |
| 739 | #else |
| 740 | BUG_ON(1); |
| 741 | #endif |
| 742 | } |
| 743 | |
| 744 | src_dma_addr = micscif_get_dma_addr(work->src_window, src_offset, NULL, NULL, NULL); |
| 745 | dst_dma_addr = micscif_get_dma_addr(work->dst_window, dst_offset, NULL, NULL, NULL); |
| 746 | |
| 747 | if (is_local_dma_addr(src_dma_addr)) |
| 748 | src_local = true; |
| 749 | else |
| 750 | src_local = false; |
| 751 | |
| 752 | if (is_local_dma_addr(dst_dma_addr)) |
| 753 | dst_local = true; |
| 754 | else |
| 755 | dst_local = false; |
| 756 | |
| 757 | dst_local = dst_local; |
| 758 | BUG_ON(work->len + (L1_CACHE_BYTES << 1) > KMEM_UNALIGNED_BUF_SIZE); |
| 759 | |
| 760 | /* Allocate dma_completion cb */ |
| 761 | if (!(comp_cb = kzalloc(sizeof(*comp_cb), GFP_KERNEL))) |
| 762 | goto error; |
| 763 | |
| 764 | work->comp_cb = comp_cb; |
| 765 | comp_cb->cb_cookie = (uint64_t)comp_cb; |
| 766 | comp_cb->dma_completion_func = &micscif_rma_completion_cb; |
| 767 | |
| 768 | if (work->len + (L1_CACHE_BYTES << 1) < KMEM_UNALIGNED_BUF_SIZE) { |
| 769 | comp_cb->is_cache = false; |
| 770 | if (!(temp = kmalloc(work->len + (L1_CACHE_BYTES << 1), GFP_KERNEL))) |
| 771 | goto free_comp_cb; |
| 772 | comp_cb->temp_buf_to_free = temp; |
| 773 | /* kmalloc(..) does not guarantee cache line alignment */ |
| 774 | if ((uint64_t)temp & (L1_CACHE_BYTES - 1)) |
| 775 | temp = (uint8_t*)ALIGN((uint64_t)temp, L1_CACHE_BYTES); |
| 776 | } else { |
| 777 | comp_cb->is_cache = true; |
| 778 | if (!(temp = micscif_kmem_cache_alloc())) |
| 779 | goto free_comp_cb; |
| 780 | comp_cb->temp_buf_to_free = temp; |
| 781 | } |
| 782 | |
| 783 | if (src_local) { |
| 784 | temp += dst_cache_off; |
| 785 | comp_cb->tmp_offset = dst_cache_off; |
| 786 | micscif_rma_local_cpu_copy(work->src_offset, work->src_window, temp, work->len, true); |
| 787 | } else { |
| 788 | comp_cb->dst_window = work->dst_window; |
| 789 | comp_cb->dst_offset = work->dst_offset; |
| 790 | work->src_offset = work->src_offset - src_cache_off; |
| 791 | comp_cb->len = work->len; |
| 792 | work->len = ALIGN(work->len + src_cache_off, L1_CACHE_BYTES); |
| 793 | comp_cb->header_padding = src_cache_off; |
| 794 | } |
| 795 | comp_cb->temp_buf = temp; |
| 796 | |
| 797 | #ifndef _MIC_SCIF_ |
| 798 | micscif_pci_dev(work->remote_dev->sd_node, &pdev); |
| 799 | comp_cb->temp_phys = mic_map_single(work->remote_dev->sd_node - 1, |
| 800 | pdev, temp, KMEM_UNALIGNED_BUF_SIZE); |
| 801 | |
| 802 | if (mic_map_error(comp_cb->temp_phys)) { |
| 803 | goto free_temp_buf; |
| 804 | } |
| 805 | |
| 806 | comp_cb->remote_node = work->remote_dev->sd_node; |
| 807 | #endif |
| 808 | if (0 > micscif_rma_list_dma_copy_unaligned(work, temp, chan, src_local)) |
| 809 | goto free_temp_buf; |
| 810 | if (!src_local) |
| 811 | work->fence_type = DO_DMA_INTR; |
| 812 | return 0; |
| 813 | free_temp_buf: |
| 814 | if (comp_cb->is_cache) |
| 815 | micscif_kmem_cache_free(comp_cb->temp_buf_to_free); |
| 816 | else |
| 817 | kfree(comp_cb->temp_buf_to_free); |
| 818 | free_comp_cb: |
| 819 | kfree(comp_cb); |
| 820 | error: |
| 821 | printk(KERN_ERR "Unable to malloc %s %d\n", __func__, __LINE__); |
| 822 | return -ENOMEM; |
| 823 | } |
| 824 | |
| 825 | #if !defined(WINDOWS) && !defined(CONFIG_PREEMPT) |
| 826 | static int softlockup_threshold = 60; |
| 827 | static void avert_softlockup(unsigned long data) |
| 828 | { |
| 829 | *(unsigned long*)data = 1; |
| 830 | } |
| 831 | |
| 832 | /* |
| 833 | * Add a timer to handle the case of hogging the cpu for |
| 834 | * time > softlockup_threshold. |
| 835 | * Add the timer every softlockup_threshold / 3 so that even if |
| 836 | * there is a huge delay in running our timer, we will still don't hit |
| 837 | * the softlockup case.(softlockup_tick() is run in hardirq() context while |
| 838 | * timers are run at softirq context) |
| 839 | * |
| 840 | */ |
| 841 | static inline void add_softlockup_timer(struct timer_list *timer, unsigned long *data) |
| 842 | { |
| 843 | setup_timer(timer, avert_softlockup, (unsigned long) data); |
| 844 | timer->expires = jiffies + usecs_to_jiffies(softlockup_threshold * 1000000 / 3); |
| 845 | add_timer(timer); |
| 846 | } |
| 847 | |
| 848 | static inline void del_softlockup_timer(struct timer_list *timer) |
| 849 | { |
| 850 | /* We need delete synchronously since the variable being touched by |
| 851 | * timer interrupt is on the stack |
| 852 | */ |
| 853 | del_timer_sync(timer); |
| 854 | } |
| 855 | #endif |
| 856 | |
| 857 | /* |
| 858 | * micscif_rma_list_cpu_copy: |
| 859 | * |
| 860 | * Traverse all the windows and perform CPU copy. |
| 861 | */ |
| 862 | int micscif_rma_list_cpu_copy(struct mic_copy_work *work) |
| 863 | { |
| 864 | void *src_virt, *dst_virt; |
| 865 | size_t loop_len, remaining_len; |
| 866 | int src_cache_off, dst_cache_off; |
| 867 | uint64_t src_offset = work->src_offset, dst_offset = work->dst_offset; |
| 868 | struct reg_range_t *src_window = work->src_window; |
| 869 | struct reg_range_t *dst_window = work->dst_window; |
| 870 | uint64_t end_src_offset, end_dst_offset; |
| 871 | struct list_head *item; |
| 872 | int srcchunk_ind = 0; |
| 873 | int dstchunk_ind = 0; |
| 874 | uint64_t src_start_offset, dst_start_offset; |
| 875 | int ret = 0; |
| 876 | #if !defined(WINDOWS) && !defined(CONFIG_PREEMPT) |
| 877 | unsigned long timer_fired = 0; |
| 878 | struct timer_list timer; |
| 879 | int cpu = smp_processor_id(); |
| 880 | add_softlockup_timer(&timer, &timer_fired); |
| 881 | #endif |
| 882 | |
| 883 | remaining_len = work->len; |
| 884 | src_start_offset = src_window->offset; |
| 885 | dst_start_offset = dst_window->offset; |
| 886 | |
| 887 | while (remaining_len) { |
| 888 | #if !defined(WINDOWS) && !defined(CONFIG_PREEMPT) |
| 889 | /* Ideally we should call schedule only if we didn't sleep |
| 890 | * in between. But there is no way to know that. |
| 891 | */ |
| 892 | if (timer_fired) { |
| 893 | timer_fired = 0; |
| 894 | if (smp_processor_id() == cpu) |
| 895 | touch_softlockup_watchdog(); |
| 896 | else |
| 897 | cpu = smp_processor_id(); |
| 898 | add_softlockup_timer(&timer, &timer_fired); |
| 899 | } |
| 900 | #endif |
| 901 | src_cache_off = src_offset & ~PAGE_MASK; |
| 902 | dst_cache_off = dst_offset & ~PAGE_MASK; |
| 903 | loop_len = PAGE_SIZE - |
| 904 | ((src_cache_off > dst_cache_off) ? |
| 905 | src_cache_off : dst_cache_off); |
| 906 | if (remaining_len < loop_len) |
| 907 | loop_len = remaining_len; |
| 908 | |
| 909 | if (RMA_WINDOW_SELF == src_window->type) |
| 910 | src_virt = get_local_va(src_offset, src_window, loop_len); |
| 911 | else |
| 912 | src_virt = ioremap_remote(src_offset, |
| 913 | src_window, loop_len, work->loopback, work->remote_dev, &srcchunk_ind, &src_start_offset); |
| 914 | if (!src_virt) { |
| 915 | ret = -ENOMEM; |
| 916 | goto error; |
| 917 | } |
| 918 | |
| 919 | if (RMA_WINDOW_SELF == dst_window->type) |
| 920 | dst_virt = get_local_va(dst_offset, dst_window, loop_len); |
| 921 | else |
| 922 | dst_virt = ioremap_remote(dst_offset, |
| 923 | dst_window, loop_len, work->loopback, work->remote_dev, &dstchunk_ind, &dst_start_offset); |
| 924 | if (!dst_virt) { |
| 925 | if (RMA_WINDOW_PEER == src_window->type) |
| 926 | iounmap_remote(src_virt, loop_len, work); |
| 927 | ret = -ENOMEM; |
| 928 | goto error; |
| 929 | } |
| 930 | |
| 931 | if (work->loopback) |
| 932 | memcpy(dst_virt, src_virt, loop_len); |
| 933 | else { |
| 934 | |
| 935 | if (RMA_WINDOW_SELF == src_window->type){ |
| 936 | memcpy_toio(dst_virt, src_virt, loop_len); |
| 937 | } |
| 938 | else{ |
| 939 | memcpy_fromio(dst_virt, src_virt, loop_len); |
| 940 | } |
| 941 | serializing_request(dst_virt); |
| 942 | smp_mb(); |
| 943 | } |
| 944 | if (RMA_WINDOW_PEER == src_window->type) |
| 945 | iounmap_remote(src_virt, loop_len, work); |
| 946 | |
| 947 | if (RMA_WINDOW_PEER == dst_window->type) |
| 948 | iounmap_remote(dst_virt, loop_len, work); |
| 949 | |
| 950 | src_offset += loop_len; |
| 951 | dst_offset += loop_len; |
| 952 | remaining_len -= loop_len; |
| 953 | if (remaining_len) { |
| 954 | end_src_offset = src_window->offset + |
| 955 | (src_window->nr_pages << PAGE_SHIFT); |
| 956 | end_dst_offset = dst_window->offset + |
| 957 | (dst_window->nr_pages << PAGE_SHIFT); |
| 958 | if (src_offset == end_src_offset) { |
| 959 | item = ( |
| 960 | &src_window->list_member)->next; |
| 961 | src_window = list_entry(item, |
| 962 | struct reg_range_t, |
| 963 | list_member); |
| 964 | srcchunk_ind = 0; |
| 965 | src_start_offset = src_window->offset; |
| 966 | } |
| 967 | if (dst_offset == end_dst_offset) { |
| 968 | item = ( |
| 969 | &dst_window->list_member)->next; |
| 970 | dst_window = list_entry(item, |
| 971 | struct reg_range_t, |
| 972 | list_member); |
| 973 | dstchunk_ind = 0; |
| 974 | dst_start_offset = dst_window->offset; |
| 975 | } |
| 976 | } |
| 977 | } |
| 978 | error: |
| 979 | #if !defined(WINDOWS) && !defined(CONFIG_PREEMPT) |
| 980 | del_softlockup_timer(&timer); |
| 981 | #endif |
| 982 | return ret; |
| 983 | } |