| 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/micscif_kmem_cache.h" |
| 39 | #include "mic/micscif_rma_list.h" |
| 40 | #ifndef _MIC_SCIF_ |
| 41 | #include "mic_common.h" |
| 42 | #endif |
| 43 | #include "mic/mic_dma_api.h" |
| 44 | #include "mic/micscif_map.h" |
| 45 | |
| 46 | bool mic_reg_cache_enable = 0; |
| 47 | |
| 48 | bool mic_huge_page_enable = 1; |
| 49 | |
| 50 | #ifdef _MIC_SCIF_ |
| 51 | mic_dma_handle_t mic_dma_handle; |
| 52 | #endif |
| 53 | static inline |
| 54 | void micscif_rma_destroy_tcw(struct rma_mmu_notifier *mmn, |
| 55 | struct endpt *ep, bool inrange, |
| 56 | uint64_t start, uint64_t len); |
| 57 | #ifdef CONFIG_MMU_NOTIFIER |
| 58 | static void scif_mmu_notifier_release(struct mmu_notifier *mn, |
| 59 | struct mm_struct *mm); |
| 60 | static int scif_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, |
| 61 | struct mm_struct *mm, |
| 62 | unsigned long start, unsigned long end, |
| 63 | bool blockable); |
| 64 | static void scif_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, |
| 65 | struct mm_struct *mm, |
| 66 | unsigned long start, unsigned long end); |
| 67 | static const struct mmu_notifier_ops scif_mmu_notifier_ops = { |
| 68 | .release = scif_mmu_notifier_release, |
| 69 | .clear_flush_young = NULL, |
| 70 | .change_pte = NULL,/*TODO*/ |
| 71 | .invalidate_range_start = scif_mmu_notifier_invalidate_range_start, |
| 72 | .invalidate_range_end = scif_mmu_notifier_invalidate_range_end}; |
| 73 | |
| 74 | static void scif_mmu_notifier_release(struct mmu_notifier *mn, |
| 75 | struct mm_struct *mm) |
| 76 | { |
| 77 | struct endpt *ep; |
| 78 | struct rma_mmu_notifier *mmn; |
| 79 | mmn = container_of(mn, struct rma_mmu_notifier, ep_mmu_notifier); |
| 80 | ep = mmn->ep; |
| 81 | micscif_rma_destroy_tcw(mmn, ep, false, 0, 0); |
| 82 | pr_debug("%s\n", __func__); |
| 83 | return; |
| 84 | } |
| 85 | |
| 86 | static int scif_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, |
| 87 | struct mm_struct *mm, |
| 88 | unsigned long start, unsigned long end, |
| 89 | bool blockable) |
| 90 | { |
| 91 | struct endpt *ep; |
| 92 | struct rma_mmu_notifier *mmn; |
| 93 | mmn = container_of(mn, struct rma_mmu_notifier, ep_mmu_notifier); |
| 94 | ep = mmn->ep; |
| 95 | /* |
| 96 | * The kernel file `include/linux/mmu_notifier.h` states the following |
| 97 | * regarding the invalidate_range_start() callback: |
| 98 | * |
| 99 | * If blockable argument is set to false then the callback cannot |
| 100 | * sleep and has to return with -EAGAIN. 0 should be returned |
| 101 | * otherwise. |
| 102 | * |
| 103 | * The following function executes spin_lock_irqsave(), which feels like it |
| 104 | * qualifies as 'sleep'. However, returning -EAGAIN would require me to |
| 105 | * understand the location and function of all code that calls this |
| 106 | * callback. I do not yet have that understanding. |
| 107 | * |
| 108 | * For now, maintain the original behavior of calling |
| 109 | * micscif_rma_destroy_tcw() every time, accepting the spinlock. If this |
| 110 | * becomes problematic, either figure out all the code that can call this |
| 111 | * function and teach it to understand -EAGAIN, or investigate the `#ifdef |
| 112 | * CONFIG_MMU_NOTIFIER`. |
| 113 | * |
| 114 | * If you ended up here while tracking down a bug and pulling your hair |
| 115 | * out, sorry. :-( |
| 116 | */ |
| 117 | micscif_rma_destroy_tcw(mmn, ep, true, (uint64_t)start, (uint64_t)(end - start)); |
| 118 | pr_debug("%s start=%lx, end=%lx\n", __func__, start, end); |
| 119 | return 0; |
| 120 | } |
| 121 | |
| 122 | static void scif_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, |
| 123 | struct mm_struct *mm, |
| 124 | unsigned long start, unsigned long end) |
| 125 | { |
| 126 | /* Nothing to do here, everything needed was done in invalidate_range_start */ |
| 127 | pr_debug("%s\n", __func__); |
| 128 | return; |
| 129 | } |
| 130 | #endif |
| 131 | |
| 132 | #ifdef CONFIG_MMU_NOTIFIER |
| 133 | void ep_unregister_mmu_notifier(struct endpt *ep) |
| 134 | { |
| 135 | struct endpt_rma_info *rma = &ep->rma_info; |
| 136 | struct rma_mmu_notifier *mmn = NULL; |
| 137 | struct list_head *item, *tmp; |
| 138 | mutex_lock(&ep->rma_info.mmn_lock); |
| 139 | list_for_each_safe(item, tmp, &rma->mmn_list) { |
| 140 | mmn = list_entry(item, |
| 141 | struct rma_mmu_notifier, list_member); |
| 142 | mmu_notifier_unregister(&mmn->ep_mmu_notifier, mmn->mm); |
| 143 | #ifdef RMA_DEBUG |
| 144 | BUG_ON(atomic_long_sub_return(1, &ms_info.mmu_notif_cnt) < 0); |
| 145 | #endif |
| 146 | list_del(item); |
| 147 | kfree(mmn); |
| 148 | } |
| 149 | mutex_unlock(&ep->rma_info.mmn_lock); |
| 150 | } |
| 151 | |
| 152 | static void init_mmu_notifier(struct rma_mmu_notifier *mmn, struct mm_struct *mm, struct endpt *ep) |
| 153 | { |
| 154 | mmn->ep = ep; |
| 155 | mmn->mm = mm; |
| 156 | mmn->ep_mmu_notifier.ops = &scif_mmu_notifier_ops; |
| 157 | INIT_LIST_HEAD(&mmn->list_member); |
| 158 | INIT_LIST_HEAD(&mmn->tc_reg_list); |
| 159 | } |
| 160 | |
| 161 | static struct rma_mmu_notifier *find_mmu_notifier(struct mm_struct *mm, struct endpt_rma_info *rma) |
| 162 | { |
| 163 | struct rma_mmu_notifier *mmn; |
| 164 | struct list_head *item; |
| 165 | list_for_each(item, &rma->mmn_list) { |
| 166 | mmn = list_entry(item, |
| 167 | struct rma_mmu_notifier, list_member); |
| 168 | if (mmn->mm == mm) |
| 169 | return mmn; |
| 170 | } |
| 171 | return NULL; |
| 172 | } |
| 173 | #endif |
| 174 | |
| 175 | /** |
| 176 | * micscif_rma_ep_init: |
| 177 | * @ep: end point |
| 178 | * |
| 179 | * Initialize RMA per EP data structures. |
| 180 | */ |
| 181 | int micscif_rma_ep_init(struct endpt *ep) |
| 182 | { |
| 183 | int ret; |
| 184 | struct endpt_rma_info *rma = &ep->rma_info; |
| 185 | |
| 186 | mutex_init (&rma->rma_lock); |
| 187 | if ((ret = va_gen_init(&rma->va_gen, |
| 188 | VA_GEN_MIN, VA_GEN_RANGE)) < 0) |
| 189 | goto init_err; |
| 190 | spin_lock_init(&rma->tc_lock); |
| 191 | mutex_init (&rma->mmn_lock); |
| 192 | mutex_init (&rma->va_lock); |
| 193 | INIT_LIST_HEAD(&rma->reg_list); |
| 194 | INIT_LIST_HEAD(&rma->remote_reg_list); |
| 195 | atomic_set(&rma->tw_refcount, 0); |
| 196 | atomic_set(&rma->tw_total_pages, 0); |
| 197 | atomic_set(&rma->tcw_refcount, 0); |
| 198 | atomic_set(&rma->tcw_total_pages, 0); |
| 199 | init_waitqueue_head(&rma->fence_wq); |
| 200 | rma->fence_refcount = 0; |
| 201 | rma->async_list_del = 0; |
| 202 | rma->dma_chan = NULL; |
| 203 | INIT_LIST_HEAD(&rma->mmn_list); |
| 204 | INIT_LIST_HEAD(&rma->task_list); |
| 205 | init_err: |
| 206 | return ret; |
| 207 | } |
| 208 | |
| 209 | /** |
| 210 | * micscif_rma_ep_can_uninit: |
| 211 | * @ep: end point |
| 212 | * |
| 213 | * Returns 1 if an endpoint can be uninitialized and 0 otherwise. |
| 214 | */ |
| 215 | int micscif_rma_ep_can_uninit(struct endpt *ep) |
| 216 | { |
| 217 | int ret = 0; |
| 218 | |
| 219 | /* Destroy RMA Info only if both lists are empty */ |
| 220 | if (list_empty(&ep->rma_info.reg_list) && |
| 221 | list_empty(&ep->rma_info.remote_reg_list) && |
| 222 | #ifdef CONFIG_MMU_NOTIFIER |
| 223 | list_empty(&ep->rma_info.mmn_list) && |
| 224 | #endif |
| 225 | !atomic_read(&ep->rma_info.tw_refcount) && |
| 226 | !atomic_read(&ep->rma_info.tcw_refcount)) |
| 227 | ret = 1; |
| 228 | return ret; |
| 229 | } |
| 230 | |
| 231 | #ifdef _MIC_SCIF_ |
| 232 | /** |
| 233 | * __micscif_setup_proxy_dma: |
| 234 | * @ep: SCIF endpoint descriptor. |
| 235 | * |
| 236 | * Sets up data structures for P2P Proxy DMAs. |
| 237 | */ |
| 238 | static int __micscif_setup_proxy_dma(struct endpt *ep) |
| 239 | { |
| 240 | struct endpt_rma_info *rma = &ep->rma_info; |
| 241 | int err = 0; |
| 242 | uint64_t *tmp = NULL; |
| 243 | |
| 244 | mutex_lock(&rma->rma_lock); |
| 245 | if (is_p2p_scifdev(ep->remote_dev) && !rma->proxy_dma_va) { |
| 246 | if (!(tmp = scif_zalloc(PAGE_SIZE))) { |
| 247 | err = -ENOMEM; |
| 248 | goto error; |
| 249 | } |
| 250 | if ((err = map_virt_into_aperture(&rma->proxy_dma_phys, |
| 251 | tmp, |
| 252 | ep->remote_dev, PAGE_SIZE))) { |
| 253 | scif_free(tmp, PAGE_SIZE); |
| 254 | goto error; |
| 255 | } |
| 256 | *tmp = OP_IDLE; |
| 257 | rma->proxy_dma_va = tmp; |
| 258 | } |
| 259 | error: |
| 260 | mutex_unlock(&rma->rma_lock); |
| 261 | return err; |
| 262 | } |
| 263 | |
| 264 | static __always_inline int micscif_setup_proxy_dma(struct endpt *ep) |
| 265 | { |
| 266 | if (ep->rma_info.proxy_dma_va) |
| 267 | return 0; |
| 268 | |
| 269 | return __micscif_setup_proxy_dma(ep); |
| 270 | } |
| 271 | |
| 272 | /** |
| 273 | * micscif_teardown_proxy_dma: |
| 274 | * @ep: SCIF endpoint descriptor. |
| 275 | * |
| 276 | * Tears down data structures setup for P2P Proxy DMAs. |
| 277 | */ |
| 278 | void micscif_teardown_proxy_dma(struct endpt *ep) |
| 279 | { |
| 280 | struct endpt_rma_info *rma = &ep->rma_info; |
| 281 | mutex_lock(&rma->rma_lock); |
| 282 | if (rma->proxy_dma_va) { |
| 283 | unmap_from_aperture(rma->proxy_dma_phys, ep->remote_dev, PAGE_SIZE); |
| 284 | scif_free(rma->proxy_dma_va, PAGE_SIZE); |
| 285 | rma->proxy_dma_va = NULL; |
| 286 | } |
| 287 | mutex_unlock(&rma->rma_lock); |
| 288 | } |
| 289 | |
| 290 | /** |
| 291 | * micscif_proxy_dma: |
| 292 | * @ep: SCIF endpoint descriptor. |
| 293 | * @copy_work: DMA copy work information. |
| 294 | * |
| 295 | * This API does the following: |
| 296 | * 1) Sends the peer a SCIF Node QP message with the information |
| 297 | * required to program a proxy DMA to covert a P2P Read to a Write |
| 298 | * which will initiate a DMA transfer from the peer card to self. |
| 299 | * The reason for this special code path is KNF and KNC P2P read |
| 300 | * performance being much lower than P2P write performance on Crown |
| 301 | * Pass platforms. |
| 302 | * 2) Poll for an update of the known proxy dma VA to OP_COMPLETED |
| 303 | * via a SUD by the peer. |
| 304 | */ |
| 305 | static int micscif_proxy_dma(scif_epd_t epd, struct mic_copy_work *work) |
| 306 | { |
| 307 | struct endpt *ep = (struct endpt *)epd; |
| 308 | struct nodemsg msg; |
| 309 | unsigned long ts = jiffies; |
| 310 | struct endpt_rma_info *rma = &ep->rma_info; |
| 311 | int err; |
| 312 | volatile uint64_t *proxy_dma_va = rma->proxy_dma_va; |
| 313 | |
| 314 | mutex_lock(&ep->rma_info.rma_lock); |
| 315 | /* |
| 316 | * Bail out if there is a Proxy DMA already in progress |
| 317 | * for this endpoint. The callee will fallback on self |
| 318 | * DMAs upon an error. |
| 319 | */ |
| 320 | if (*proxy_dma_va != OP_IDLE) { |
| 321 | mutex_unlock(&ep->rma_info.rma_lock); |
| 322 | err = -EBUSY; |
| 323 | goto error; |
| 324 | } |
| 325 | *proxy_dma_va = OP_IN_PROGRESS; |
| 326 | mutex_unlock(&ep->rma_info.rma_lock); |
| 327 | |
| 328 | msg.src = ep->port; |
| 329 | msg.uop = work->ordered ? SCIF_PROXY_ORDERED_DMA : SCIF_PROXY_DMA; |
| 330 | msg.payload[0] = ep->remote_ep; |
| 331 | msg.payload[1] = work->src_offset; |
| 332 | msg.payload[2] = work->dst_offset; |
| 333 | msg.payload[3] = work->len; |
| 334 | |
| 335 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) |
| 336 | goto error_init_va; |
| 337 | |
| 338 | while (*proxy_dma_va != OP_COMPLETED) { |
| 339 | schedule(); |
| 340 | if (time_after(jiffies, |
| 341 | ts + NODE_ALIVE_TIMEOUT)) { |
| 342 | err = -EBUSY; |
| 343 | goto error_init_va; |
| 344 | } |
| 345 | } |
| 346 | err = 0; |
| 347 | error_init_va: |
| 348 | *proxy_dma_va = OP_IDLE; |
| 349 | error: |
| 350 | return err; |
| 351 | } |
| 352 | #endif |
| 353 | |
| 354 | /** |
| 355 | * micscif_create_pinned_pages: |
| 356 | * @nr_pages: number of pages in window |
| 357 | * @prot: read/write protection |
| 358 | * |
| 359 | * Allocate and prepare a set of pinned pages. |
| 360 | */ |
| 361 | struct scif_pinned_pages *micscif_create_pinned_pages(int nr_pages, int prot) |
| 362 | { |
| 363 | struct scif_pinned_pages *pinned_pages; |
| 364 | |
| 365 | might_sleep(); |
| 366 | if (!(pinned_pages = scif_zalloc(sizeof(*pinned_pages)))) |
| 367 | goto error; |
| 368 | |
| 369 | if (!(pinned_pages->pages = scif_zalloc(nr_pages * |
| 370 | sizeof(*(pinned_pages->pages))))) |
| 371 | goto error_free_pinned_pages; |
| 372 | |
| 373 | if (!(pinned_pages->num_pages = scif_zalloc(nr_pages * |
| 374 | sizeof(*(pinned_pages->num_pages))))) |
| 375 | goto error_free_pages; |
| 376 | |
| 377 | #if !defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HUGETLB_PAGE) && !defined(_MIC_SCIF_) |
| 378 | if (!(pinned_pages->vma = scif_zalloc(nr_pages * |
| 379 | sizeof(*(pinned_pages->vma))))) |
| 380 | goto error_free_num_pages; |
| 381 | #endif |
| 382 | |
| 383 | pinned_pages->prot = prot; |
| 384 | pinned_pages->magic = SCIFEP_MAGIC; |
| 385 | pinned_pages->nr_contig_chunks = 0; |
| 386 | return pinned_pages; |
| 387 | |
| 388 | #if !defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HUGETLB_PAGE) && !defined(_MIC_SCIF_) |
| 389 | error_free_num_pages: |
| 390 | scif_free(pinned_pages->num_pages, |
| 391 | pinned_pages->nr_pages * sizeof(*(pinned_pages->num_pages))); |
| 392 | #endif |
| 393 | error_free_pages: |
| 394 | scif_free(pinned_pages->pages, |
| 395 | pinned_pages->nr_pages * sizeof(*(pinned_pages->pages))); |
| 396 | error_free_pinned_pages: |
| 397 | scif_free(pinned_pages, sizeof(*pinned_pages)); |
| 398 | error: |
| 399 | return NULL; |
| 400 | } |
| 401 | |
| 402 | /** |
| 403 | * micscif_destroy_pinned_pages: |
| 404 | * @pinned_pages: A set of pinned pages. |
| 405 | * |
| 406 | * Deallocate resources for pinned pages. |
| 407 | */ |
| 408 | int micscif_destroy_pinned_pages(struct scif_pinned_pages *pinned_pages) |
| 409 | { |
| 410 | int j; |
| 411 | int writeable = pinned_pages->prot & SCIF_PROT_WRITE; |
| 412 | int kernel = SCIF_MAP_KERNEL & pinned_pages->map_flags; |
| 413 | |
| 414 | for (j = 0; j < pinned_pages->nr_pages; j++) { |
| 415 | if (pinned_pages->pages[j]) { |
| 416 | if (!kernel) { |
| 417 | if (writeable) |
| 418 | SetPageDirty(pinned_pages->pages[j]); |
| 419 | #ifdef RMA_DEBUG |
| 420 | BUG_ON(!page_count(pinned_pages->pages[j])); |
| 421 | BUG_ON(atomic_long_sub_return(1, &ms_info.rma_pin_cnt) < 0); |
| 422 | #endif |
| 423 | put_page(pinned_pages->pages[j]); |
| 424 | } |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | #if !defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_HUGETLB_PAGE) && !defined(_MIC_SCIF_) |
| 429 | scif_free(pinned_pages->vma, |
| 430 | pinned_pages->nr_pages * sizeof(*(pinned_pages->vma))); |
| 431 | #endif |
| 432 | scif_free(pinned_pages->pages, |
| 433 | pinned_pages->nr_pages * sizeof(*(pinned_pages->pages))); |
| 434 | scif_free(pinned_pages->num_pages, |
| 435 | pinned_pages->nr_pages * sizeof(*(pinned_pages->num_pages))); |
| 436 | scif_free(pinned_pages, sizeof(*pinned_pages)); |
| 437 | return 0; |
| 438 | } |
| 439 | |
| 440 | /* |
| 441 | * micscif_create_window: |
| 442 | * @ep: end point |
| 443 | * @pinned_pages: Set of pinned pages which wil back this window. |
| 444 | * @offset: offset hint |
| 445 | * |
| 446 | * Allocate and prepare a self registration window. |
| 447 | */ |
| 448 | struct reg_range_t *micscif_create_window(struct endpt *ep, |
| 449 | int64_t nr_pages, uint64_t offset, bool temp) |
| 450 | { |
| 451 | struct reg_range_t *window; |
| 452 | |
| 453 | might_sleep(); |
| 454 | if (!(window = scif_zalloc(sizeof(struct reg_range_t)))) |
| 455 | goto error; |
| 456 | |
| 457 | #ifdef CONFIG_ML1OM |
| 458 | if (!temp) { |
| 459 | if (!(window->phys_addr = scif_zalloc(nr_pages * |
| 460 | sizeof(*(window->phys_addr))))) |
| 461 | goto error_free_window; |
| 462 | |
| 463 | if (!(window->temp_phys_addr = scif_zalloc(nr_pages * |
| 464 | sizeof(*(window->temp_phys_addr))))) |
| 465 | goto error_free_window; |
| 466 | } |
| 467 | #endif |
| 468 | |
| 469 | if (!(window->dma_addr = scif_zalloc(nr_pages * |
| 470 | sizeof(*(window->dma_addr))))) |
| 471 | goto error_free_window; |
| 472 | |
| 473 | if (!(window->num_pages = scif_zalloc(nr_pages * |
| 474 | sizeof(*(window->num_pages))))) |
| 475 | goto error_free_window; |
| 476 | |
| 477 | window->offset = offset; |
| 478 | window->ep = (uint64_t)ep; |
| 479 | window->magic = SCIFEP_MAGIC; |
| 480 | window->reg_state = OP_IDLE; |
| 481 | init_waitqueue_head(&window->regwq); |
| 482 | window->unreg_state = OP_IDLE; |
| 483 | init_waitqueue_head(&window->unregwq); |
| 484 | INIT_LIST_HEAD(&window->list_member); |
| 485 | window->type = RMA_WINDOW_SELF; |
| 486 | window->temp = temp; |
| 487 | #ifdef _MIC_SCIF_ |
| 488 | micscif_setup_proxy_dma(ep); |
| 489 | #endif |
| 490 | return window; |
| 491 | |
| 492 | error_free_window: |
| 493 | if (window->dma_addr) |
| 494 | scif_free(window->dma_addr, nr_pages * sizeof(*(window->dma_addr))); |
| 495 | #ifdef CONFIG_ML1OM |
| 496 | if (window->temp_phys_addr) |
| 497 | scif_free(window->temp_phys_addr, nr_pages * sizeof(*(window->temp_phys_addr))); |
| 498 | if (window->phys_addr) |
| 499 | scif_free(window->phys_addr, nr_pages * sizeof(*(window->phys_addr))); |
| 500 | #endif |
| 501 | scif_free(window, sizeof(*window)); |
| 502 | error: |
| 503 | return NULL; |
| 504 | } |
| 505 | |
| 506 | /** |
| 507 | * micscif_destroy_incomplete_window: |
| 508 | * @ep: end point |
| 509 | * @window: registration window |
| 510 | * |
| 511 | * Deallocate resources for self window. |
| 512 | */ |
| 513 | int micscif_destroy_incomplete_window(struct endpt *ep, struct reg_range_t *window) |
| 514 | { |
| 515 | int err; |
| 516 | int64_t nr_pages = window->nr_pages; |
| 517 | struct allocmsg *alloc = &window->alloc_handle; |
| 518 | struct nodemsg msg; |
| 519 | |
| 520 | RMA_MAGIC(window); |
| 521 | retry: |
| 522 | err = wait_event_timeout(alloc->allocwq, alloc->state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); |
| 523 | if (!err && scifdev_alive(ep)) |
| 524 | goto retry; |
| 525 | |
| 526 | if (OP_COMPLETED == alloc->state) { |
| 527 | msg.uop = SCIF_FREE_VIRT; |
| 528 | msg.src = ep->port; |
| 529 | msg.payload[0] = ep->remote_ep; |
| 530 | msg.payload[1] = (uint64_t)window->alloc_handle.vaddr; |
| 531 | msg.payload[2] = (uint64_t)window; |
| 532 | msg.payload[3] = SCIF_REGISTER; |
| 533 | micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 534 | } |
| 535 | |
| 536 | micscif_free_window_offset(ep, window->offset, |
| 537 | window->nr_pages << PAGE_SHIFT); |
| 538 | if (window->dma_addr) |
| 539 | scif_free(window->dma_addr, nr_pages * |
| 540 | sizeof(*(window->dma_addr))); |
| 541 | if (window->num_pages) |
| 542 | scif_free(window->num_pages, nr_pages * |
| 543 | sizeof(*(window->num_pages))); |
| 544 | #ifdef CONFIG_ML1OM |
| 545 | if (window->phys_addr) |
| 546 | scif_free(window->phys_addr, window->nr_pages * |
| 547 | sizeof(*(window->phys_addr))); |
| 548 | if (window->temp_phys_addr) |
| 549 | scif_free(window->temp_phys_addr, nr_pages * |
| 550 | sizeof(*(window->temp_phys_addr))); |
| 551 | #endif |
| 552 | scif_free(window, sizeof(*window)); |
| 553 | return 0; |
| 554 | } |
| 555 | |
| 556 | /** |
| 557 | * micscif_destroy_window: |
| 558 | * @ep: end point |
| 559 | * @window: registration window |
| 560 | * |
| 561 | * Deallocate resources for self window. |
| 562 | */ |
| 563 | int micscif_destroy_window(struct endpt *ep, struct reg_range_t *window) |
| 564 | { |
| 565 | int j; |
| 566 | struct scif_pinned_pages *pinned_pages = window->pinned_pages; |
| 567 | int64_t nr_pages = window->nr_pages; |
| 568 | |
| 569 | might_sleep(); |
| 570 | RMA_MAGIC(window); |
| 571 | if (!window->temp && window->mm) { |
| 572 | __scif_dec_pinned_vm_lock(window->mm, window->nr_pages, 0); |
| 573 | __scif_release_mm(window->mm); |
| 574 | window->mm = NULL; |
| 575 | } |
| 576 | |
| 577 | if (!window->offset_freed) |
| 578 | micscif_free_window_offset(ep, window->offset, |
| 579 | window->nr_pages << PAGE_SHIFT); |
| 580 | for (j = 0; j < window->nr_contig_chunks; j++) { |
| 581 | if (window->dma_addr[j]) { |
| 582 | unmap_from_aperture( |
| 583 | window->dma_addr[j], |
| 584 | ep->remote_dev, |
| 585 | window->num_pages[j] << PAGE_SHIFT); |
| 586 | } |
| 587 | } |
| 588 | |
| 589 | /* |
| 590 | * Decrement references for this set of pinned pages from |
| 591 | * this window. |
| 592 | */ |
| 593 | j = atomic_sub_return((int32_t)pinned_pages->nr_pages, |
| 594 | &pinned_pages->ref_count); |
| 595 | BUG_ON(j < 0); |
| 596 | /* |
| 597 | * If the ref count for pinned_pages is zero then someone |
| 598 | * has already called scif_unpin_pages() for it and we should |
| 599 | * destroy the page cache. |
| 600 | */ |
| 601 | if (!j) |
| 602 | micscif_destroy_pinned_pages(window->pinned_pages); |
| 603 | if (window->dma_addr) |
| 604 | scif_free(window->dma_addr, nr_pages * |
| 605 | sizeof(*(window->dma_addr))); |
| 606 | if (window->num_pages) |
| 607 | scif_free(window->num_pages, nr_pages * |
| 608 | sizeof(*(window->num_pages))); |
| 609 | #ifdef CONFIG_ML1OM |
| 610 | if (window->phys_addr) |
| 611 | scif_free(window->phys_addr, window->nr_pages * |
| 612 | sizeof(*(window->phys_addr))); |
| 613 | if (window->temp_phys_addr) |
| 614 | scif_free(window->temp_phys_addr, nr_pages * |
| 615 | sizeof(*(window->temp_phys_addr))); |
| 616 | #endif |
| 617 | window->magic = 0; |
| 618 | scif_free(window, sizeof(*window)); |
| 619 | return 0; |
| 620 | } |
| 621 | |
| 622 | /** |
| 623 | * micscif_create_remote_lookup: |
| 624 | * @ep: end point |
| 625 | * @window: remote window |
| 626 | * |
| 627 | * Allocate and prepare lookup entries for the remote |
| 628 | * end to copy over the physical addresses. |
| 629 | * Returns 0 on success and appropriate errno on failure. |
| 630 | */ |
| 631 | int micscif_create_remote_lookup(struct endpt *ep, struct reg_range_t *window) |
| 632 | { |
| 633 | int i, j, err = 0; |
| 634 | int64_t nr_pages = window->nr_pages; |
| 635 | bool vmalloc_dma_phys; |
| 636 | #ifdef CONFIG_ML1OM |
| 637 | bool vmalloc_temp_phys = false; |
| 638 | bool vmalloc_phys = false; |
| 639 | #endif |
| 640 | might_sleep(); |
| 641 | |
| 642 | /* Map window */ |
| 643 | err = map_virt_into_aperture(&window->mapped_offset, |
| 644 | window, ep->remote_dev, sizeof(*window)); |
| 645 | if (err) |
| 646 | goto error_window; |
| 647 | |
| 648 | /* Compute the number of lookup entries. 21 == 2MB Shift */ |
| 649 | window->nr_lookup = ALIGN(nr_pages * PAGE_SIZE, |
| 650 | ((2) * 1024 * 1024)) >> 21; |
| 651 | |
| 652 | if (!(window->dma_addr_lookup.lookup = |
| 653 | scif_zalloc(window->nr_lookup * |
| 654 | sizeof(*(window->dma_addr_lookup.lookup))))) |
| 655 | goto error_window; |
| 656 | |
| 657 | /* Map DMA physical addess lookup array */ |
| 658 | err = map_virt_into_aperture(&window->dma_addr_lookup.offset, |
| 659 | window->dma_addr_lookup.lookup, ep->remote_dev, |
| 660 | window->nr_lookup * |
| 661 | sizeof(*window->dma_addr_lookup.lookup)); |
| 662 | if (err) |
| 663 | goto error_window; |
| 664 | |
| 665 | vmalloc_dma_phys = is_vmalloc_addr(&window->dma_addr[0]); |
| 666 | |
| 667 | #ifdef CONFIG_ML1OM |
| 668 | if (ep->remote_dev != &scif_dev[SCIF_HOST_NODE] && !is_self_scifdev(ep->remote_dev)) { |
| 669 | if (!(window->temp_phys_addr_lookup.lookup = |
| 670 | scif_zalloc(window->nr_lookup * |
| 671 | sizeof(*(window->temp_phys_addr_lookup.lookup))))) |
| 672 | goto error_window; |
| 673 | |
| 674 | /* Map physical addess lookup array */ |
| 675 | err = map_virt_into_aperture(&window->temp_phys_addr_lookup.offset, |
| 676 | window->temp_phys_addr_lookup.lookup, ep->remote_dev, |
| 677 | window->nr_lookup * |
| 678 | sizeof(*window->temp_phys_addr_lookup.lookup)); |
| 679 | if (err) |
| 680 | goto error_window; |
| 681 | |
| 682 | if (!(window->phys_addr_lookup.lookup = |
| 683 | scif_zalloc(window->nr_lookup * |
| 684 | sizeof(*(window->phys_addr_lookup.lookup))))) |
| 685 | goto error_window; |
| 686 | |
| 687 | /* Map physical addess lookup array */ |
| 688 | err = map_virt_into_aperture(&window->phys_addr_lookup.offset, |
| 689 | window->phys_addr_lookup.lookup, ep->remote_dev, |
| 690 | window->nr_lookup * |
| 691 | sizeof(*window->phys_addr_lookup.lookup)); |
| 692 | if (err) |
| 693 | goto error_window; |
| 694 | |
| 695 | vmalloc_phys = is_vmalloc_addr(&window->phys_addr[0]); |
| 696 | vmalloc_temp_phys = is_vmalloc_addr(&window->temp_phys_addr[0]); |
| 697 | } |
| 698 | #endif |
| 699 | |
| 700 | /* Now map each of the pages containing physical addresses */ |
| 701 | for (i = 0, j = 0; i < nr_pages; i += NR_PHYS_ADDR_IN_PAGE, j++) { |
| 702 | #ifdef CONFIG_ML1OM |
| 703 | if (ep->remote_dev != &scif_dev[SCIF_HOST_NODE] && !is_self_scifdev(ep->remote_dev)) { |
| 704 | err = map_page_into_aperture( |
| 705 | &window->temp_phys_addr_lookup.lookup[j], |
| 706 | vmalloc_temp_phys ? |
| 707 | vmalloc_to_page(&window->temp_phys_addr[i]) : |
| 708 | virt_to_page(&window->temp_phys_addr[i]), |
| 709 | ep->remote_dev); |
| 710 | if (err) |
| 711 | goto error_window; |
| 712 | |
| 713 | err = map_page_into_aperture( |
| 714 | &window->phys_addr_lookup.lookup[j], |
| 715 | vmalloc_phys ? |
| 716 | vmalloc_to_page(&window->phys_addr[i]) : |
| 717 | virt_to_page(&window->phys_addr[i]), |
| 718 | ep->remote_dev); |
| 719 | if (err) |
| 720 | goto error_window; |
| 721 | } |
| 722 | #endif |
| 723 | err = map_page_into_aperture( |
| 724 | &window->dma_addr_lookup.lookup[j], |
| 725 | vmalloc_dma_phys ? |
| 726 | vmalloc_to_page(&window->dma_addr[i]) : |
| 727 | virt_to_page(&window->dma_addr[i]), |
| 728 | ep->remote_dev); |
| 729 | if (err) |
| 730 | goto error_window; |
| 731 | } |
| 732 | return 0; |
| 733 | error_window: |
| 734 | return err; |
| 735 | } |
| 736 | |
| 737 | /** |
| 738 | * micscif_destroy_remote_lookup: |
| 739 | * @ep: end point |
| 740 | * @window: remote window |
| 741 | * |
| 742 | * Destroy lookup entries used for the remote |
| 743 | * end to copy over the physical addresses. |
| 744 | */ |
| 745 | void micscif_destroy_remote_lookup(struct endpt *ep, struct reg_range_t *window) |
| 746 | { |
| 747 | int i, j; |
| 748 | |
| 749 | RMA_MAGIC(window); |
| 750 | if (window->nr_lookup) { |
| 751 | for (i = 0, j = 0; i < window->nr_pages; |
| 752 | i += NR_PHYS_ADDR_IN_PAGE, j++) { |
| 753 | if (window->dma_addr_lookup.lookup && |
| 754 | window->dma_addr_lookup.lookup[j]) { |
| 755 | unmap_from_aperture( |
| 756 | window->dma_addr_lookup.lookup[j], |
| 757 | ep->remote_dev, PAGE_SIZE); |
| 758 | } |
| 759 | } |
| 760 | if (window->dma_addr_lookup.offset) { |
| 761 | unmap_from_aperture( |
| 762 | window->dma_addr_lookup.offset, |
| 763 | ep->remote_dev, window->nr_lookup * |
| 764 | sizeof(*window->dma_addr_lookup.lookup)); |
| 765 | } |
| 766 | if (window->dma_addr_lookup.lookup) |
| 767 | scif_free(window->dma_addr_lookup.lookup, window->nr_lookup * |
| 768 | sizeof(*(window->dma_addr_lookup.lookup))); |
| 769 | if (window->mapped_offset) { |
| 770 | unmap_from_aperture(window->mapped_offset, |
| 771 | ep->remote_dev, sizeof(*window)); |
| 772 | } |
| 773 | window->nr_lookup = 0; |
| 774 | } |
| 775 | } |
| 776 | |
| 777 | /** |
| 778 | * micscif_create_remote_window: |
| 779 | * @ep: end point |
| 780 | * @nr_pages: number of pages in window |
| 781 | * |
| 782 | * Allocate and prepare a remote registration window. |
| 783 | */ |
| 784 | struct reg_range_t *micscif_create_remote_window(struct endpt *ep, int nr_pages) |
| 785 | { |
| 786 | struct reg_range_t *window; |
| 787 | |
| 788 | might_sleep(); |
| 789 | if (!(window = scif_zalloc(sizeof(struct reg_range_t)))) |
| 790 | goto error_ret; |
| 791 | |
| 792 | window->magic = SCIFEP_MAGIC; |
| 793 | window->nr_pages = nr_pages; |
| 794 | |
| 795 | #if !defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 796 | if (!(window->page_ref_count = scif_zalloc(nr_pages * |
| 797 | sizeof(*(window->page_ref_count))))) |
| 798 | goto error_window; |
| 799 | #endif |
| 800 | |
| 801 | if (!(window->dma_addr = scif_zalloc(nr_pages * |
| 802 | sizeof(*(window->dma_addr))))) |
| 803 | goto error_window; |
| 804 | |
| 805 | if (!(window->num_pages = scif_zalloc(nr_pages * |
| 806 | sizeof(*(window->num_pages))))) |
| 807 | goto error_window; |
| 808 | |
| 809 | #ifdef CONFIG_ML1OM |
| 810 | if (!(window->phys_addr = scif_zalloc(nr_pages * |
| 811 | sizeof(*(window->phys_addr))))) |
| 812 | goto error_window; |
| 813 | |
| 814 | if (!(window->temp_phys_addr = scif_zalloc(nr_pages * |
| 815 | sizeof(*(window->temp_phys_addr))))) |
| 816 | goto error_window; |
| 817 | #endif |
| 818 | |
| 819 | if (micscif_create_remote_lookup(ep, window)) |
| 820 | goto error_window; |
| 821 | |
| 822 | window->ep = (uint64_t)ep; |
| 823 | window->type = RMA_WINDOW_PEER; |
| 824 | set_window_ref_count(window, nr_pages); |
| 825 | window->get_put_ref_count = 0; |
| 826 | window->unreg_state = OP_IDLE; |
| 827 | #if !defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 828 | window->gttmap_state = OP_IDLE; |
| 829 | init_waitqueue_head(&window->gttmapwq); |
| 830 | #endif |
| 831 | #ifdef _MIC_SCIF_ |
| 832 | micscif_setup_proxy_dma(ep); |
| 833 | window->proxy_dma_phys = ep->rma_info.proxy_dma_phys; |
| 834 | #endif |
| 835 | return window; |
| 836 | error_window: |
| 837 | micscif_destroy_remote_window(ep, window); |
| 838 | error_ret: |
| 839 | return NULL; |
| 840 | } |
| 841 | |
| 842 | /** |
| 843 | * micscif_destroy_remote_window: |
| 844 | * @ep: end point |
| 845 | * @window: remote registration window |
| 846 | * |
| 847 | * Deallocate resources for remote window. |
| 848 | */ |
| 849 | void micscif_destroy_remote_window(struct endpt *ep, struct reg_range_t *window) |
| 850 | { |
| 851 | RMA_MAGIC(window); |
| 852 | micscif_destroy_remote_lookup(ep, window); |
| 853 | if (window->dma_addr) |
| 854 | scif_free(window->dma_addr, window->nr_pages * |
| 855 | sizeof(*(window->dma_addr))); |
| 856 | if (window->num_pages) |
| 857 | scif_free(window->num_pages, window->nr_pages * |
| 858 | sizeof(*(window->num_pages))); |
| 859 | #ifdef CONFIG_ML1OM |
| 860 | if (window->phys_addr) |
| 861 | scif_free(window->phys_addr, window->nr_pages * |
| 862 | sizeof(*(window->phys_addr))); |
| 863 | if (window->temp_phys_addr) |
| 864 | scif_free(window->temp_phys_addr, window->nr_pages * |
| 865 | sizeof(*(window->temp_phys_addr))); |
| 866 | #endif |
| 867 | |
| 868 | #if !defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 869 | if (window->page_ref_count) |
| 870 | scif_free(window->page_ref_count, window->nr_pages * |
| 871 | sizeof(*(window->page_ref_count))); |
| 872 | #endif |
| 873 | window->magic = 0; |
| 874 | scif_free(window, sizeof(*window)); |
| 875 | } |
| 876 | |
| 877 | /** |
| 878 | * micscif_map_window_pages: |
| 879 | * @ep: end point |
| 880 | * @window: self registration window |
| 881 | * @tmp_wnd: is a temporary window? |
| 882 | * |
| 883 | * Map pages of a window into the aperture/PCI. |
| 884 | * Also compute physical addresses required for DMA. |
| 885 | */ |
| 886 | int micscif_map_window_pages(struct endpt *ep, struct reg_range_t *window, bool tmp_wnd) |
| 887 | { |
| 888 | int j, i, err = 0, nr_pages; |
| 889 | scif_pinned_pages_t pinned_pages; |
| 890 | |
| 891 | might_sleep(); |
| 892 | RMA_MAGIC(window); |
| 893 | |
| 894 | pinned_pages = window->pinned_pages; |
| 895 | for (j = 0, i = 0; j < window->nr_contig_chunks; j++, i += nr_pages) { |
| 896 | nr_pages = pinned_pages->num_pages[i]; |
| 897 | #ifdef _MIC_SCIF_ |
| 898 | #ifdef CONFIG_ML1OM |
| 899 | /* phys_addr[] holds addresses as seen from the remote node |
| 900 | * these addressed are then copied into the remote card's |
| 901 | * window structure |
| 902 | * when the remote node is the host and the card is knf |
| 903 | * these addresses are only created at the point of mapping |
| 904 | * the card physical address into gtt (for the KNC the |
| 905 | * the gtt code path returns the local address) |
| 906 | * when the remote node is loopback - the address remains |
| 907 | * the same |
| 908 | * when the remote node is a kn* - the base address of the local |
| 909 | * card as seen from the remote node is added in |
| 910 | */ |
| 911 | if (!tmp_wnd) { |
| 912 | if(ep->remote_dev != &scif_dev[SCIF_HOST_NODE]) { |
| 913 | if ((err = map_virt_into_aperture( |
| 914 | &window->temp_phys_addr[j], |
| 915 | phys_to_virt(page_to_phys(pinned_pages->pages[i])), |
| 916 | ep->remote_dev, |
| 917 | nr_pages << PAGE_SHIFT))) { |
| 918 | int k,l; |
| 919 | |
| 920 | for (l = k = 0; k < i; l++) { |
| 921 | nr_pages = pinned_pages->num_pages[k]; |
| 922 | window->temp_phys_addr[l] |
| 923 | &= ~RMA_HUGE_NR_PAGE_MASK; |
| 924 | unmap_from_aperture( |
| 925 | window->temp_phys_addr[l], |
| 926 | ep->remote_dev, |
| 927 | nr_pages << PAGE_SHIFT); |
| 928 | k += nr_pages; |
| 929 | window->temp_phys_addr[l] = 0; |
| 930 | } |
| 931 | return err; |
| 932 | } |
| 933 | if (!tmp_wnd) |
| 934 | RMA_SET_NR_PAGES(window->temp_phys_addr[j], nr_pages); |
| 935 | } |
| 936 | } |
| 937 | #endif |
| 938 | window->dma_addr[j] = |
| 939 | page_to_phys(pinned_pages->pages[i]); |
| 940 | if (!tmp_wnd) |
| 941 | RMA_SET_NR_PAGES(window->dma_addr[j], nr_pages); |
| 942 | #else |
| 943 | err = map_virt_into_aperture(&window->dma_addr[j], |
| 944 | phys_to_virt(page_to_phys(pinned_pages->pages[i])), |
| 945 | ep->remote_dev, nr_pages << PAGE_SHIFT); |
| 946 | if (err) |
| 947 | return err; |
| 948 | if (!tmp_wnd) |
| 949 | RMA_SET_NR_PAGES(window->dma_addr[j], nr_pages); |
| 950 | #endif |
| 951 | window->num_pages[j] = nr_pages; |
| 952 | } |
| 953 | return err; |
| 954 | } |
| 955 | |
| 956 | |
| 957 | /** |
| 958 | * micscif_unregister_window: |
| 959 | * @window: self registration window |
| 960 | * |
| 961 | * Send an unregistration request and wait for a response. |
| 962 | */ |
| 963 | int micscif_unregister_window(struct reg_range_t *window) |
| 964 | { |
| 965 | int err = 0; |
| 966 | struct endpt *ep = (struct endpt *)window->ep; |
| 967 | bool send_msg = false; |
| 968 | |
| 969 | might_sleep(); |
| 970 | BUG_ON(!mutex_is_locked(&ep->rma_info.rma_lock)); |
| 971 | |
| 972 | switch (window->unreg_state) { |
| 973 | case OP_IDLE: |
| 974 | { |
| 975 | window->unreg_state = OP_IN_PROGRESS; |
| 976 | send_msg = true; |
| 977 | /* fall through */ |
| 978 | } |
| 979 | case OP_IN_PROGRESS: |
| 980 | { |
| 981 | get_window_ref_count(window, 1); |
| 982 | mutex_unlock(&ep->rma_info.rma_lock); |
| 983 | if (send_msg && (err = micscif_send_scif_unregister(ep, window))) { |
| 984 | window->unreg_state = OP_COMPLETED; |
| 985 | goto done; |
| 986 | } |
| 987 | retry: |
| 988 | err = wait_event_timeout(window->unregwq, |
| 989 | window->unreg_state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); |
| 990 | if (!err && scifdev_alive(ep)) |
| 991 | goto retry; |
| 992 | if (!err) { |
| 993 | err = -ENODEV; |
| 994 | window->unreg_state = OP_COMPLETED; |
| 995 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 996 | } |
| 997 | if (err > 0) |
| 998 | err = 0; |
| 999 | done: |
| 1000 | mutex_lock(&ep->rma_info.rma_lock); |
| 1001 | put_window_ref_count(window, 1); |
| 1002 | break; |
| 1003 | } |
| 1004 | case OP_FAILED: |
| 1005 | { |
| 1006 | if (!scifdev_alive(ep)) { |
| 1007 | err = -ENODEV; |
| 1008 | window->unreg_state = OP_COMPLETED; |
| 1009 | } |
| 1010 | break; |
| 1011 | } |
| 1012 | case OP_COMPLETED: |
| 1013 | break; |
| 1014 | default: |
| 1015 | /* Invalid opcode? */ |
| 1016 | BUG_ON(1); |
| 1017 | } |
| 1018 | |
| 1019 | if (OP_COMPLETED == window->unreg_state && |
| 1020 | window->ref_count) |
| 1021 | put_window_ref_count(window, window->nr_pages); |
| 1022 | |
| 1023 | if (!window->ref_count) { |
| 1024 | atomic_inc(&ep->rma_info.tw_refcount); |
| 1025 | atomic_add_return((int32_t)window->nr_pages, &ep->rma_info.tw_total_pages); |
| 1026 | list_del(&window->list_member); |
| 1027 | micscif_free_window_offset(ep, window->offset, |
| 1028 | window->nr_pages << PAGE_SHIFT); |
| 1029 | window->offset_freed = true; |
| 1030 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1031 | if ((!!(window->pinned_pages->map_flags & SCIF_MAP_KERNEL)) |
| 1032 | && scifdev_alive(ep)) { |
| 1033 | drain_dma_intr(ep->rma_info.dma_chan); |
| 1034 | } else { |
| 1035 | if (!__scif_dec_pinned_vm_lock(window->mm, |
| 1036 | window->nr_pages, 1)) { |
| 1037 | __scif_release_mm(window->mm); |
| 1038 | window->mm = NULL; |
| 1039 | } |
| 1040 | } |
| 1041 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); |
| 1042 | mutex_lock(&ep->rma_info.rma_lock); |
| 1043 | } |
| 1044 | return err; |
| 1045 | } |
| 1046 | |
| 1047 | /** |
| 1048 | * micscif_send_alloc_request: |
| 1049 | * @ep: end point |
| 1050 | * @window: self registration window |
| 1051 | * |
| 1052 | * Send a remote window allocation request |
| 1053 | */ |
| 1054 | int micscif_send_alloc_request(struct endpt *ep, struct reg_range_t *window) |
| 1055 | { |
| 1056 | struct nodemsg msg; |
| 1057 | struct allocmsg *alloc = &window->alloc_handle; |
| 1058 | |
| 1059 | /* Set up the Alloc Handle */ |
| 1060 | alloc->uop = SCIF_REGISTER; |
| 1061 | alloc->state = OP_IN_PROGRESS; |
| 1062 | init_waitqueue_head(&alloc->allocwq); |
| 1063 | |
| 1064 | /* Send out an allocation request */ |
| 1065 | msg.uop = SCIF_ALLOC_REQ; |
| 1066 | msg.src = ep->port; |
| 1067 | msg.payload[0] = ep->remote_ep; |
| 1068 | msg.payload[1] = window->nr_pages; |
| 1069 | msg.payload[2] = (uint64_t)&window->alloc_handle; |
| 1070 | msg.payload[3] = SCIF_REGISTER; |
| 1071 | return micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 1072 | } |
| 1073 | |
| 1074 | /** |
| 1075 | * micscif_prep_remote_window: |
| 1076 | * @ep: end point |
| 1077 | * @window: self registration window |
| 1078 | * |
| 1079 | * Send a remote window allocation request, wait for an allocation response, |
| 1080 | * prepare the remote window and notify the peer to unmap it once done. |
| 1081 | */ |
| 1082 | int micscif_prep_remote_window(struct endpt *ep, struct reg_range_t *window) |
| 1083 | { |
| 1084 | struct nodemsg msg; |
| 1085 | struct reg_range_t *remote_window; |
| 1086 | struct allocmsg *alloc = &window->alloc_handle; |
| 1087 | dma_addr_t *dma_phys_lookup, *tmp; |
| 1088 | int i = 0, j = 0; |
| 1089 | int nr_contig_chunks, loop_nr_contig_chunks, remaining_nr_contig_chunks, nr_lookup; |
| 1090 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 1091 | dma_addr_t *phys_lookup = 0; |
| 1092 | #endif |
| 1093 | int err, map_err; |
| 1094 | |
| 1095 | nr_contig_chunks = remaining_nr_contig_chunks = (int)window->nr_contig_chunks; |
| 1096 | |
| 1097 | if ((map_err = micscif_map_window_pages(ep, window, false))) { |
| 1098 | printk(KERN_ERR "%s %d map_err %d\n", __func__, __LINE__, map_err); |
| 1099 | } |
| 1100 | retry: |
| 1101 | /* Now wait for the response */ |
| 1102 | err = wait_event_timeout(alloc->allocwq, alloc->state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); |
| 1103 | if (!err && scifdev_alive(ep)) |
| 1104 | goto retry; |
| 1105 | |
| 1106 | if (!err) |
| 1107 | err = -ENODEV; |
| 1108 | |
| 1109 | if (err > 0) |
| 1110 | err = 0; |
| 1111 | else |
| 1112 | return err; |
| 1113 | |
| 1114 | /* Bail out. The remote end rejected this request */ |
| 1115 | if (OP_FAILED == alloc->state) |
| 1116 | return -ENOMEM; |
| 1117 | |
| 1118 | if (map_err) { |
| 1119 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, map_err); |
| 1120 | msg.uop = SCIF_FREE_VIRT; |
| 1121 | msg.src = ep->port; |
| 1122 | msg.payload[0] = ep->remote_ep; |
| 1123 | msg.payload[1] = (uint64_t)window->alloc_handle.vaddr; |
| 1124 | msg.payload[2] = (uint64_t)window; |
| 1125 | msg.payload[3] = SCIF_REGISTER; |
| 1126 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) |
| 1127 | err = -ENOTCONN; |
| 1128 | else |
| 1129 | err = map_err; |
| 1130 | return err; |
| 1131 | } |
| 1132 | |
| 1133 | |
| 1134 | remote_window = scif_ioremap(alloc->phys_addr, |
| 1135 | sizeof(*window), ep->remote_dev); |
| 1136 | |
| 1137 | RMA_MAGIC(remote_window); |
| 1138 | |
| 1139 | /* Compute the number of lookup entries. 21 == 2MB Shift */ |
| 1140 | nr_lookup = ALIGN(nr_contig_chunks * PAGE_SIZE, ((2) * 1024 * 1024)) >> 21; |
| 1141 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 1142 | if (is_p2p_scifdev(ep->remote_dev)) |
| 1143 | phys_lookup = scif_ioremap(remote_window->temp_phys_addr_lookup.offset, |
| 1144 | nr_lookup * |
| 1145 | sizeof(*remote_window->temp_phys_addr_lookup.lookup), |
| 1146 | ep->remote_dev); |
| 1147 | #endif |
| 1148 | |
| 1149 | dma_phys_lookup = scif_ioremap(remote_window->dma_addr_lookup.offset, |
| 1150 | nr_lookup * |
| 1151 | sizeof(*remote_window->dma_addr_lookup.lookup), |
| 1152 | ep->remote_dev); |
| 1153 | |
| 1154 | while (remaining_nr_contig_chunks) { |
| 1155 | loop_nr_contig_chunks = min(remaining_nr_contig_chunks, (int)NR_PHYS_ADDR_IN_PAGE); |
| 1156 | /* #1/2 - Copy physical addresses over to the remote side */ |
| 1157 | |
| 1158 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 1159 | /* If the remote dev is self or is any node except the host |
| 1160 | * its OK to copy the bus address to the remote window |
| 1161 | * in the case of the host (for KNF only) the bus address |
| 1162 | * is generated at the time of mmap(..) into card memory |
| 1163 | * and does not exist at this time |
| 1164 | */ |
| 1165 | /* Note: |
| 1166 | * the phys_addr[] holds MIC address for remote cards |
| 1167 | * -> GTT offset for the host (KNF) |
| 1168 | * -> local address for the host (KNC) |
| 1169 | * -> local address for loopback |
| 1170 | * this is done in map_window_pages(..) except for GTT |
| 1171 | * offset for KNF |
| 1172 | */ |
| 1173 | if (is_p2p_scifdev(ep->remote_dev)) { |
| 1174 | tmp = scif_ioremap(phys_lookup[j], |
| 1175 | loop_nr_contig_chunks * sizeof(*window->temp_phys_addr), |
| 1176 | ep->remote_dev); |
| 1177 | memcpy_toio(tmp, &window->temp_phys_addr[i], |
| 1178 | loop_nr_contig_chunks * sizeof(*window->temp_phys_addr)); |
| 1179 | serializing_request(tmp); |
| 1180 | smp_mb(); |
| 1181 | scif_iounmap(tmp, PAGE_SIZE, ep->remote_dev); |
| 1182 | } |
| 1183 | #endif |
| 1184 | /* #2/2 - Copy DMA addresses (addresses that are fed into the DMA engine) |
| 1185 | * We transfer bus addresses which are then converted into a MIC physical |
| 1186 | * address on the remote side if it is a MIC, if the remote node is a host |
| 1187 | * we transfer the MIC physical address |
| 1188 | */ |
| 1189 | tmp = scif_ioremap( |
| 1190 | dma_phys_lookup[j], |
| 1191 | loop_nr_contig_chunks * sizeof(*window->dma_addr), |
| 1192 | ep->remote_dev); |
| 1193 | #ifdef _MIC_SCIF_ |
| 1194 | if (is_p2p_scifdev(ep->remote_dev)) { |
| 1195 | /* knf: |
| 1196 | * send the address as mapped through the GTT (the remote node's |
| 1197 | * base address for this node is already added in) |
| 1198 | * knc: |
| 1199 | * add remote node's base address for this node to convert it |
| 1200 | * into a MIC address |
| 1201 | */ |
| 1202 | int m; |
| 1203 | dma_addr_t dma_addr; |
| 1204 | for (m = 0; m < loop_nr_contig_chunks; m++) { |
| 1205 | #ifdef CONFIG_ML1OM |
| 1206 | dma_addr = window->temp_phys_addr[i + m]; |
| 1207 | #else |
| 1208 | dma_addr = window->dma_addr[i + m] + |
| 1209 | ep->remote_dev->sd_base_addr; |
| 1210 | #endif |
| 1211 | writeq(dma_addr, &tmp[m]); |
| 1212 | } |
| 1213 | } else |
| 1214 | /* Host node or loopback - transfer DMA addresses as is, this is |
| 1215 | * the same as a MIC physical address (we use the dma_addr |
| 1216 | * and not the phys_addr array since the phys_addr is only setup |
| 1217 | * if there is a mmap() request from the host) |
| 1218 | */ |
| 1219 | memcpy_toio(tmp, &window->dma_addr[i], |
| 1220 | loop_nr_contig_chunks * sizeof(*window->dma_addr)); |
| 1221 | #else |
| 1222 | /* Transfer the physical address array - this is the MIC address |
| 1223 | * as seen by the card |
| 1224 | */ |
| 1225 | memcpy_toio(tmp, &window->dma_addr[i], |
| 1226 | loop_nr_contig_chunks * sizeof(*window->dma_addr)); |
| 1227 | #endif |
| 1228 | remaining_nr_contig_chunks -= loop_nr_contig_chunks; |
| 1229 | i += loop_nr_contig_chunks; |
| 1230 | j++; |
| 1231 | serializing_request(tmp); |
| 1232 | smp_mb(); |
| 1233 | scif_iounmap(tmp, PAGE_SIZE, ep->remote_dev); |
| 1234 | } |
| 1235 | |
| 1236 | /* Prepare the remote window for the peer */ |
| 1237 | remote_window->peer_window = (uint64_t)window; |
| 1238 | remote_window->offset = window->offset; |
| 1239 | remote_window->prot = window->prot; |
| 1240 | remote_window->nr_contig_chunks = nr_contig_chunks; |
| 1241 | #ifdef _MIC_SCIF_ |
| 1242 | if (!ep->rma_info.proxy_dma_peer_phys) |
| 1243 | ep->rma_info.proxy_dma_peer_phys = remote_window->proxy_dma_phys; |
| 1244 | #endif |
| 1245 | #if defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 1246 | if (is_p2p_scifdev(ep->remote_dev)) |
| 1247 | scif_iounmap(phys_lookup, |
| 1248 | nr_lookup * |
| 1249 | sizeof(*remote_window->temp_phys_addr_lookup.lookup), |
| 1250 | ep->remote_dev); |
| 1251 | #endif |
| 1252 | scif_iounmap(dma_phys_lookup, |
| 1253 | nr_lookup * |
| 1254 | sizeof(*remote_window->dma_addr_lookup.lookup), |
| 1255 | ep->remote_dev); |
| 1256 | scif_iounmap(remote_window, sizeof(*remote_window), ep->remote_dev); |
| 1257 | window->peer_window = (uint64_t)alloc->vaddr; |
| 1258 | return err; |
| 1259 | } |
| 1260 | |
| 1261 | /** |
| 1262 | * micscif_send_scif_register: |
| 1263 | * @ep: end point |
| 1264 | * @window: self registration window |
| 1265 | * |
| 1266 | * Send a SCIF_REGISTER message if EP is connected and wait for a |
| 1267 | * SCIF_REGISTER_(N)ACK message else send a SCIF_FREE_VIRT |
| 1268 | * message so that the peer can free its remote window allocated earlier. |
| 1269 | */ |
| 1270 | int micscif_send_scif_register(struct endpt *ep, struct reg_range_t *window) |
| 1271 | { |
| 1272 | int err = 0; |
| 1273 | struct nodemsg msg; |
| 1274 | |
| 1275 | msg.src = ep->port; |
| 1276 | msg.payload[0] = ep->remote_ep; |
| 1277 | msg.payload[1] = (uint64_t)window->alloc_handle.vaddr; |
| 1278 | msg.payload[2] = (uint64_t)window; |
| 1279 | if (SCIFEP_CONNECTED == ep->state) { |
| 1280 | msg.uop = SCIF_REGISTER; |
| 1281 | window->reg_state = OP_IN_PROGRESS; |
| 1282 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) { |
| 1283 | micscif_set_nr_pages(ep->remote_dev, window); |
| 1284 | retry: |
| 1285 | err = wait_event_timeout(window->regwq, |
| 1286 | window->reg_state != OP_IN_PROGRESS, NODE_ALIVE_TIMEOUT); |
| 1287 | if (!err && scifdev_alive(ep)) |
| 1288 | goto retry; |
| 1289 | if (!err) |
| 1290 | err = -ENODEV; |
| 1291 | if (err > 0) |
| 1292 | err = 0; |
| 1293 | if (OP_FAILED == window->reg_state) |
| 1294 | err = -ENOTCONN; |
| 1295 | } else { |
| 1296 | micscif_set_nr_pages(ep->remote_dev, window); |
| 1297 | } |
| 1298 | } else { |
| 1299 | msg.uop = SCIF_FREE_VIRT; |
| 1300 | msg.payload[3] = SCIF_REGISTER; |
| 1301 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) |
| 1302 | err = -ENOTCONN; |
| 1303 | micscif_set_nr_pages(ep->remote_dev, window); |
| 1304 | } |
| 1305 | return err; |
| 1306 | } |
| 1307 | |
| 1308 | /** |
| 1309 | * micscif_send_scif_unregister: |
| 1310 | * @ep: end point |
| 1311 | * @window: self registration window |
| 1312 | * |
| 1313 | * Send a SCIF_UNREGISTER message. |
| 1314 | */ |
| 1315 | int micscif_send_scif_unregister(struct endpt *ep, struct reg_range_t *window) |
| 1316 | { |
| 1317 | struct nodemsg msg; |
| 1318 | |
| 1319 | RMA_MAGIC(window); |
| 1320 | msg.uop = SCIF_UNREGISTER; |
| 1321 | msg.src = ep->port; |
| 1322 | msg.payload[0] = (uint64_t)window->alloc_handle.vaddr; |
| 1323 | msg.payload[1] = (uint64_t)window; |
| 1324 | return micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 1325 | } |
| 1326 | |
| 1327 | /** |
| 1328 | * micscif_get_window_offset: |
| 1329 | * @epd: end point descriptor |
| 1330 | * @flags: flags |
| 1331 | * @offset: offset hint |
| 1332 | * @len: length of range |
| 1333 | * @out_offset: computed offset returned by reference. |
| 1334 | * |
| 1335 | * Compute/Claim a new offset for this EP. The callee is supposed to grab |
| 1336 | * the RMA mutex before calling this API. |
| 1337 | */ |
| 1338 | int micscif_get_window_offset(struct endpt *ep, int flags, |
| 1339 | uint64_t offset, size_t len, uint64_t *out_offset) |
| 1340 | { |
| 1341 | uint64_t computed_offset; |
| 1342 | int err = 0; |
| 1343 | |
| 1344 | might_sleep(); |
| 1345 | mutex_lock(&ep->rma_info.va_lock); |
| 1346 | if (flags & SCIF_MAP_FIXED) { |
| 1347 | computed_offset = va_gen_claim(&ep->rma_info.va_gen, |
| 1348 | (uint64_t)offset, len); |
| 1349 | if (INVALID_VA_GEN_ADDRESS == computed_offset) |
| 1350 | err = -EADDRINUSE; |
| 1351 | } else { |
| 1352 | computed_offset = va_gen_alloc(&ep->rma_info.va_gen, |
| 1353 | len, PAGE_SIZE); |
| 1354 | if (INVALID_VA_GEN_ADDRESS == computed_offset) |
| 1355 | err = -ENOMEM; |
| 1356 | } |
| 1357 | *out_offset = computed_offset; |
| 1358 | mutex_unlock(&ep->rma_info.va_lock); |
| 1359 | return err; |
| 1360 | } |
| 1361 | |
| 1362 | /** |
| 1363 | * micscif_free_window_offset: |
| 1364 | * @offset: offset hint |
| 1365 | * @len: length of range |
| 1366 | * |
| 1367 | * Free offset for this EP. The callee is supposed to grab |
| 1368 | * the RMA mutex before calling this API. |
| 1369 | */ |
| 1370 | void micscif_free_window_offset(struct endpt *ep, |
| 1371 | uint64_t offset, size_t len) |
| 1372 | { |
| 1373 | mutex_lock(&ep->rma_info.va_lock); |
| 1374 | va_gen_free(&ep->rma_info.va_gen, offset, len); |
| 1375 | mutex_unlock(&ep->rma_info.va_lock); |
| 1376 | } |
| 1377 | |
| 1378 | /** |
| 1379 | * scif_register_temp: |
| 1380 | * @epd: End Point Descriptor. |
| 1381 | * @addr: virtual address to/from which to copy |
| 1382 | * @len: length of range to copy |
| 1383 | * @out_offset: computed offset returned by reference. |
| 1384 | * @out_window: allocated registered window returned by reference. |
| 1385 | * |
| 1386 | * Create a temporary registered window. The peer will not know about this |
| 1387 | * window. This API is used for scif_vreadfrom()/scif_vwriteto() API's. |
| 1388 | */ |
| 1389 | static int |
| 1390 | micscif_register_temp(scif_epd_t epd, void *addr, size_t len, int prot, |
| 1391 | off_t *out_offset, struct reg_range_t **out_window) |
| 1392 | { |
| 1393 | struct endpt *ep = (struct endpt *)epd; |
| 1394 | int err; |
| 1395 | scif_pinned_pages_t pinned_pages; |
| 1396 | size_t aligned_len; |
| 1397 | |
| 1398 | aligned_len = ALIGN(len, PAGE_SIZE); |
| 1399 | |
| 1400 | if ((err = __scif_pin_pages((void *)((uint64_t)addr & |
| 1401 | PAGE_MASK), |
| 1402 | aligned_len, &prot, 0, &pinned_pages))) |
| 1403 | return err; |
| 1404 | |
| 1405 | pinned_pages->prot = prot; |
| 1406 | |
| 1407 | /* Compute the offset for this registration */ |
| 1408 | if ((err = micscif_get_window_offset(ep, 0, 0, |
| 1409 | aligned_len, (uint64_t *)out_offset))) |
| 1410 | goto error_unpin; |
| 1411 | |
| 1412 | /* Allocate and prepare self registration window */ |
| 1413 | if (!(*out_window = micscif_create_window(ep, aligned_len >> PAGE_SHIFT, |
| 1414 | *out_offset, true))) { |
| 1415 | micscif_free_window_offset(ep, *out_offset, aligned_len); |
| 1416 | err = -ENOMEM; |
| 1417 | goto error_unpin; |
| 1418 | } |
| 1419 | |
| 1420 | (*out_window)->pinned_pages = pinned_pages; |
| 1421 | (*out_window)->nr_pages = pinned_pages->nr_pages; |
| 1422 | (*out_window)->nr_contig_chunks = pinned_pages->nr_contig_chunks; |
| 1423 | (*out_window)->prot = pinned_pages->prot; |
| 1424 | |
| 1425 | (*out_window)->va_for_temp = (void*)((uint64_t)addr & PAGE_MASK); |
| 1426 | if ((err = micscif_map_window_pages(ep, *out_window, true))) { |
| 1427 | /* Something went wrong! Rollback */ |
| 1428 | micscif_destroy_window(ep, *out_window); |
| 1429 | *out_window = NULL; |
| 1430 | } else |
| 1431 | *out_offset |= ((uint64_t)addr & ~PAGE_MASK); |
| 1432 | |
| 1433 | return err; |
| 1434 | error_unpin: |
| 1435 | if (err) |
| 1436 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 1437 | scif_unpin_pages(pinned_pages); |
| 1438 | return err; |
| 1439 | } |
| 1440 | |
| 1441 | /** |
| 1442 | * micscif_rma_completion_cb: |
| 1443 | * @data: RMA cookie |
| 1444 | * |
| 1445 | * RMA interrupt completion callback. |
| 1446 | */ |
| 1447 | void micscif_rma_completion_cb(uint64_t data) |
| 1448 | { |
| 1449 | struct dma_completion_cb *comp_cb = (struct dma_completion_cb *)data; |
| 1450 | #ifndef _MIC_SCIF_ |
| 1451 | struct pci_dev *pdev; |
| 1452 | #endif |
| 1453 | |
| 1454 | /* Free DMA Completion CB. */ |
| 1455 | if (comp_cb && comp_cb->temp_buf) { |
| 1456 | if (comp_cb->dst_window) { |
| 1457 | micscif_rma_local_cpu_copy(comp_cb->dst_offset, |
| 1458 | comp_cb->dst_window, comp_cb->temp_buf + comp_cb->header_padding, |
| 1459 | comp_cb->len, false); |
| 1460 | } |
| 1461 | #ifndef _MIC_SCIF_ |
| 1462 | micscif_pci_dev(comp_cb->remote_node, &pdev); |
| 1463 | mic_ctx_unmap_single(get_per_dev_ctx(comp_cb->remote_node - 1), |
| 1464 | comp_cb->temp_phys, KMEM_UNALIGNED_BUF_SIZE); |
| 1465 | #endif |
| 1466 | if (comp_cb->is_cache) |
| 1467 | micscif_kmem_cache_free(comp_cb->temp_buf_to_free); |
| 1468 | else |
| 1469 | kfree(comp_cb->temp_buf_to_free); |
| 1470 | } |
| 1471 | kfree(comp_cb); |
| 1472 | } |
| 1473 | |
| 1474 | static void __micscif_rma_destroy_tcw_ep(struct endpt *ep); |
| 1475 | static |
| 1476 | bool micscif_rma_tc_can_cache(struct endpt *ep, size_t cur_bytes) |
| 1477 | { |
| 1478 | if ((cur_bytes >> PAGE_SHIFT) > ms_info.mi_rma_tc_limit) |
| 1479 | return false; |
| 1480 | if ((atomic_read(&ep->rma_info.tcw_total_pages) |
| 1481 | + (cur_bytes >> PAGE_SHIFT)) > |
| 1482 | ms_info.mi_rma_tc_limit) { |
| 1483 | printk(KERN_ALERT "%s %d total=%d, current=%zu reached max\n", |
| 1484 | __func__, __LINE__, |
| 1485 | atomic_read(&ep->rma_info.tcw_total_pages), |
| 1486 | (1 + (cur_bytes >> PAGE_SHIFT))); |
| 1487 | micscif_rma_destroy_tcw_invalid(&ms_info.mi_rma_tc); |
| 1488 | __micscif_rma_destroy_tcw_ep(ep); |
| 1489 | } |
| 1490 | return true; |
| 1491 | } |
| 1492 | |
| 1493 | /** |
| 1494 | * micscif_rma_copy: |
| 1495 | * @epd: end point descriptor. |
| 1496 | * @loffset: offset in local registered address space to/from which to copy |
| 1497 | * @addr: user virtual address to/from which to copy |
| 1498 | * @len: length of range to copy |
| 1499 | * @roffset: offset in remote registered address space to/from which to copy |
| 1500 | * @flags: flags |
| 1501 | * @dir: LOCAL->REMOTE or vice versa. |
| 1502 | * |
| 1503 | * Validate parameters, check if src/dst registered ranges requested for copy |
| 1504 | * are valid and initiate either CPU or DMA copy. |
| 1505 | */ |
| 1506 | int micscif_rma_copy(scif_epd_t epd, off_t loffset, void *addr, size_t len, |
| 1507 | off_t roffset, int flags, enum rma_direction dir, bool last_chunk) |
| 1508 | { |
| 1509 | struct endpt *ep = (struct endpt *)epd; |
| 1510 | struct micscif_rma_req remote_req; |
| 1511 | struct micscif_rma_req req; |
| 1512 | struct reg_range_t *window = NULL; |
| 1513 | struct reg_range_t *remote_window = NULL; |
| 1514 | struct mic_copy_work copy_work; |
| 1515 | bool loopback; |
| 1516 | int err = 0; |
| 1517 | struct dma_channel *chan; |
| 1518 | struct rma_mmu_notifier *mmn = NULL; |
| 1519 | bool insert_window = false; |
| 1520 | bool cache = false; |
| 1521 | |
| 1522 | if ((err = verify_epd(ep))) |
| 1523 | return err; |
| 1524 | |
| 1525 | if (flags && !(flags & (SCIF_RMA_USECPU | SCIF_RMA_USECACHE | SCIF_RMA_SYNC | SCIF_RMA_ORDERED))) |
| 1526 | return -EINVAL; |
| 1527 | |
| 1528 | if (!len) |
| 1529 | return -EINVAL; |
| 1530 | loopback = is_self_scifdev(ep->remote_dev) ? true : false; |
| 1531 | copy_work.fence_type = ((flags & SCIF_RMA_SYNC) && last_chunk) ? DO_DMA_POLLING : 0; |
| 1532 | copy_work.ordered = !!((flags & SCIF_RMA_ORDERED) && last_chunk); |
| 1533 | |
| 1534 | #ifdef CONFIG_MMU_NOTIFIER |
| 1535 | if (!mic_reg_cache_enable) |
| 1536 | flags &= ~SCIF_RMA_USECACHE; |
| 1537 | #else |
| 1538 | flags &= ~SCIF_RMA_USECACHE; |
| 1539 | #endif |
| 1540 | #ifndef _MIC_SCIF_ |
| 1541 | #ifdef CONFIG_ML1OM |
| 1542 | /* Use DMA Copies even if CPU copy is requested on KNF MIC from Host */ |
| 1543 | if (flags & SCIF_RMA_USECPU) { |
| 1544 | flags &= ~SCIF_RMA_USECPU; |
| 1545 | if (last_chunk) |
| 1546 | copy_work.fence_type = DO_DMA_POLLING; |
| 1547 | } |
| 1548 | #endif |
| 1549 | /* Use CPU for Host<->Host Copies */ |
| 1550 | if (loopback) { |
| 1551 | flags |= SCIF_RMA_USECPU; |
| 1552 | copy_work.fence_type = 0x0; |
| 1553 | } |
| 1554 | #endif |
| 1555 | |
| 1556 | cache = flags & SCIF_RMA_USECACHE; |
| 1557 | |
| 1558 | /* Trying to wrap around */ |
| 1559 | if ((loffset && (loffset + (off_t)len < loffset)) || |
| 1560 | (roffset + (off_t)len < roffset)) |
| 1561 | return -EINVAL; |
| 1562 | |
| 1563 | remote_req.out_window = &remote_window; |
| 1564 | remote_req.offset = roffset; |
| 1565 | remote_req.nr_bytes = len; |
| 1566 | /* |
| 1567 | * If transfer is from local to remote then the remote window |
| 1568 | * must be writeable and vice versa. |
| 1569 | */ |
| 1570 | remote_req.prot = LOCAL_TO_REMOTE == dir ? VM_WRITE : VM_READ; |
| 1571 | remote_req.type = WINDOW_PARTIAL; |
| 1572 | remote_req.head = &ep->rma_info.remote_reg_list; |
| 1573 | |
| 1574 | #ifdef CONFIG_MMU_NOTIFIER |
| 1575 | if (addr && cache) { |
| 1576 | mutex_lock(&ep->rma_info.mmn_lock); |
| 1577 | mmn = find_mmu_notifier(current->mm, &ep->rma_info); |
| 1578 | if (!mmn) { |
| 1579 | mmn = kzalloc(sizeof(*mmn), GFP_KERNEL); |
| 1580 | if (!mmn) { |
| 1581 | mutex_unlock(&ep->rma_info.mmn_lock); |
| 1582 | return -ENOMEM; |
| 1583 | } |
| 1584 | init_mmu_notifier(mmn, current->mm, ep); |
| 1585 | if (mmu_notifier_register(&mmn->ep_mmu_notifier, current->mm)) { |
| 1586 | mutex_unlock(&ep->rma_info.mmn_lock); |
| 1587 | kfree(mmn); |
| 1588 | return -EBUSY; |
| 1589 | } |
| 1590 | #ifdef RMA_DEBUG |
| 1591 | atomic_long_add_return(1, &ms_info.mmu_notif_cnt); |
| 1592 | #endif |
| 1593 | list_add(&mmn->list_member, &ep->rma_info.mmn_list); |
| 1594 | } |
| 1595 | mutex_unlock(&ep->rma_info.mmn_lock); |
| 1596 | } |
| 1597 | #endif |
| 1598 | |
| 1599 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 1600 | #ifdef _MIC_SCIF_ |
| 1601 | if (!(flags & SCIF_RMA_USECPU)) { |
| 1602 | /* |
| 1603 | * Proxy the DMA only for P2P reads with transfer size |
| 1604 | * greater than proxy DMA threshold. scif_vreadfrom(..) |
| 1605 | * and scif_vwriteto(..) is not supported since the peer |
| 1606 | * does not have the page lists required to perform the |
| 1607 | * proxy DMA. |
| 1608 | */ |
| 1609 | if (ep->remote_dev->sd_proxy_dma_reads && |
| 1610 | !addr && dir == REMOTE_TO_LOCAL && |
| 1611 | ep->rma_info.proxy_dma_va && |
| 1612 | len >= ms_info.mi_proxy_dma_threshold) { |
| 1613 | copy_work.len = len; |
| 1614 | copy_work.src_offset = roffset; |
| 1615 | copy_work.dst_offset = loffset; |
| 1616 | /* Fall through if there were errors */ |
| 1617 | if (!(err = micscif_proxy_dma(epd, ©_work))) |
| 1618 | goto error; |
| 1619 | } |
| 1620 | } |
| 1621 | #endif |
| 1622 | mutex_lock(&ep->rma_info.rma_lock); |
| 1623 | if (addr) { |
| 1624 | req.out_window = &window; |
| 1625 | req.nr_bytes = ALIGN(len + ((uint64_t)addr & ~PAGE_MASK), PAGE_SIZE); |
| 1626 | if (mmn) |
| 1627 | req.head = &mmn->tc_reg_list; |
| 1628 | req.va_for_temp = (void*)((uint64_t)addr & PAGE_MASK); |
| 1629 | req.prot = (LOCAL_TO_REMOTE == dir ? VM_READ : VM_WRITE | VM_READ); |
| 1630 | /* Does a valid local window exist? */ |
| 1631 | |
| 1632 | pr_debug("%s %d req.va_for_temp %p addr %p req.nr_bytes 0x%lx len 0x%lx\n", |
| 1633 | __func__, __LINE__, req.va_for_temp, addr, req.nr_bytes, len); |
| 1634 | spin_lock(&ep->rma_info.tc_lock); |
| 1635 | if (!mmn || (err = micscif_query_tcw(ep, &req))) { |
| 1636 | pr_debug("%s %d err %d req.va_for_temp %p addr %p req.nr_bytes 0x%lx len 0x%lx\n", |
| 1637 | __func__, __LINE__, err, req.va_for_temp, addr, req.nr_bytes, len); |
| 1638 | spin_unlock(&ep->rma_info.tc_lock); |
| 1639 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1640 | if (cache) |
| 1641 | if (!micscif_rma_tc_can_cache(ep, req.nr_bytes)) |
| 1642 | cache = false; |
| 1643 | if ((err = micscif_register_temp(epd, req.va_for_temp, req.nr_bytes, |
| 1644 | req.prot, |
| 1645 | &loffset, &window))) { |
| 1646 | goto error; |
| 1647 | } |
| 1648 | mutex_lock(&ep->rma_info.rma_lock); |
| 1649 | pr_debug("New temp window created addr %p\n", addr); |
| 1650 | if (cache) { |
| 1651 | atomic_inc(&ep->rma_info.tcw_refcount); |
| 1652 | atomic_add_return((int32_t)window->nr_pages, &ep->rma_info.tcw_total_pages); |
| 1653 | if (mmn) { |
| 1654 | spin_lock(&ep->rma_info.tc_lock); |
| 1655 | micscif_insert_tcw(window, &mmn->tc_reg_list); |
| 1656 | spin_unlock(&ep->rma_info.tc_lock); |
| 1657 | } |
| 1658 | } |
| 1659 | insert_window = true; |
| 1660 | } else { |
| 1661 | spin_unlock(&ep->rma_info.tc_lock); |
| 1662 | pr_debug("window found for addr %p\n", addr); |
| 1663 | BUG_ON(window->va_for_temp > addr); |
| 1664 | } |
| 1665 | loffset = window->offset + ((uint64_t)addr - (uint64_t)window->va_for_temp); |
| 1666 | pr_debug("%s %d addr %p loffset 0x%lx window->nr_pages 0x%llx" |
| 1667 | " window->va_for_temp %p\n", __func__, __LINE__, |
| 1668 | addr, loffset, window->nr_pages, window->va_for_temp); |
| 1669 | RMA_MAGIC(window); |
| 1670 | } |
| 1671 | |
| 1672 | /* Does a valid remote window exist? */ |
| 1673 | if ((err = micscif_query_window(&remote_req))) { |
| 1674 | pr_debug("%s %d err %d roffset 0x%lx len 0x%lx\n", |
| 1675 | __func__, __LINE__, err, roffset, len); |
| 1676 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1677 | goto error; |
| 1678 | } |
| 1679 | RMA_MAGIC(remote_window); |
| 1680 | if (!addr) { |
| 1681 | req.out_window = &window; |
| 1682 | req.offset = loffset; |
| 1683 | /* |
| 1684 | * If transfer is from local to remote then the self window |
| 1685 | * must be readable and vice versa. |
| 1686 | */ |
| 1687 | req.prot = LOCAL_TO_REMOTE == dir ? VM_READ : VM_WRITE; |
| 1688 | req.nr_bytes = len; |
| 1689 | req.type = WINDOW_PARTIAL; |
| 1690 | req.head = &ep->rma_info.reg_list; |
| 1691 | /* Does a valid local window exist? */ |
| 1692 | if ((err = micscif_query_window(&req))) { |
| 1693 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 1694 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1695 | goto error; |
| 1696 | } |
| 1697 | RMA_MAGIC(window); |
| 1698 | } |
| 1699 | |
| 1700 | /* |
| 1701 | * Preprare copy_work for submitting work to the DMA kernel thread |
| 1702 | * or CPU copy routine. |
| 1703 | */ |
| 1704 | copy_work.len = len; |
| 1705 | copy_work.loopback = loopback; |
| 1706 | copy_work.remote_dev = ep->remote_dev; |
| 1707 | copy_work.dma_chan_released = false; |
| 1708 | if (LOCAL_TO_REMOTE == dir) { |
| 1709 | copy_work.src_offset = loffset; |
| 1710 | copy_work.src_window = window; |
| 1711 | copy_work.dst_offset = roffset; |
| 1712 | copy_work.dst_window = remote_window; |
| 1713 | } else { |
| 1714 | copy_work.src_offset = roffset; |
| 1715 | copy_work.src_window = remote_window; |
| 1716 | copy_work.dst_offset = loffset; |
| 1717 | copy_work.dst_window = window; |
| 1718 | } |
| 1719 | |
| 1720 | if (!(flags & SCIF_RMA_USECPU)) { |
| 1721 | chan = ep->rma_info.dma_chan; |
| 1722 | if ((err = request_dma_channel(chan))) { |
| 1723 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1724 | goto error; |
| 1725 | } |
| 1726 | err = micscif_rma_list_dma_copy_wrapper(epd, ©_work, |
| 1727 | chan, loffset); |
| 1728 | if (!copy_work.dma_chan_released) |
| 1729 | free_dma_channel(chan); |
| 1730 | } |
| 1731 | if (flags & SCIF_RMA_USECPU) { |
| 1732 | /* Initiate synchronous CPU copy */ |
| 1733 | micscif_rma_list_cpu_copy(©_work); |
| 1734 | } |
| 1735 | if (insert_window && !cache) { |
| 1736 | atomic_inc(&ep->rma_info.tw_refcount); |
| 1737 | atomic_add_return((int32_t)window->nr_pages, &ep->rma_info.tw_total_pages); |
| 1738 | } |
| 1739 | |
| 1740 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1741 | |
| 1742 | if (last_chunk) { |
| 1743 | if (DO_DMA_POLLING == copy_work.fence_type) |
| 1744 | err = drain_dma_poll(ep->rma_info.dma_chan); |
| 1745 | else if (DO_DMA_INTR == copy_work.fence_type) |
| 1746 | err = drain_dma_intr(ep->rma_info.dma_chan); |
| 1747 | } |
| 1748 | |
| 1749 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 1750 | if (insert_window && !cache) |
| 1751 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); |
| 1752 | return err; |
| 1753 | error: |
| 1754 | if (err) { |
| 1755 | if (addr && window && !cache) |
| 1756 | micscif_destroy_window(ep, window); |
| 1757 | printk(KERN_ERR "%s %d err %d len 0x%lx\n", __func__, __LINE__, err, len); |
| 1758 | } |
| 1759 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 1760 | return err; |
| 1761 | } |
| 1762 | |
| 1763 | /** |
| 1764 | * micscif_send_fence_mark: |
| 1765 | * @epd: end point descriptor. |
| 1766 | * @out_mark: Output DMA mark reported by peer. |
| 1767 | * |
| 1768 | * Send a remote fence mark request. |
| 1769 | */ |
| 1770 | int micscif_send_fence_mark(scif_epd_t epd, int *out_mark) |
| 1771 | { |
| 1772 | int err; |
| 1773 | struct nodemsg msg; |
| 1774 | struct fence_info *fence_req; |
| 1775 | struct endpt *ep = (struct endpt *)epd; |
| 1776 | |
| 1777 | if (!(fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL))) { |
| 1778 | err = -ENOMEM; |
| 1779 | goto error; |
| 1780 | } |
| 1781 | |
| 1782 | fence_req->state = OP_IN_PROGRESS; |
| 1783 | init_waitqueue_head(&fence_req->wq); |
| 1784 | |
| 1785 | msg.src = ep->port; |
| 1786 | msg.uop = SCIF_MARK; |
| 1787 | msg.payload[0] = ep->remote_ep; |
| 1788 | msg.payload[1] = (uint64_t)fence_req; |
| 1789 | |
| 1790 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) |
| 1791 | goto error; |
| 1792 | |
| 1793 | retry: |
| 1794 | err = wait_event_timeout(fence_req->wq, |
| 1795 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); |
| 1796 | if (!err && scifdev_alive(ep)) |
| 1797 | goto retry; |
| 1798 | if (!err) |
| 1799 | err = -ENODEV; |
| 1800 | if (err > 0) |
| 1801 | err = 0; |
| 1802 | if (err < 0) { |
| 1803 | mutex_lock(&ep->rma_info.rma_lock); |
| 1804 | if (OP_IN_PROGRESS == fence_req->state) |
| 1805 | fence_req->state = OP_FAILED; |
| 1806 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1807 | } |
| 1808 | if (OP_COMPLETED == fence_req->state) |
| 1809 | *out_mark = SCIF_REMOTE_FENCE | fence_req->dma_mark; |
| 1810 | |
| 1811 | if (OP_FAILED == fence_req->state && !err) |
| 1812 | err = -ENOMEM; |
| 1813 | mutex_lock(&ep->rma_info.rma_lock); |
| 1814 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1815 | kfree(fence_req); |
| 1816 | error: |
| 1817 | return err; |
| 1818 | } |
| 1819 | |
| 1820 | /** |
| 1821 | * micscif_send_fence_wait: |
| 1822 | * @epd: end point descriptor. |
| 1823 | * @mark: DMA mark to wait for. |
| 1824 | * |
| 1825 | * Send a remote fence wait request. |
| 1826 | */ |
| 1827 | int micscif_send_fence_wait(scif_epd_t epd, int mark) |
| 1828 | { |
| 1829 | int err; |
| 1830 | struct nodemsg msg; |
| 1831 | struct fence_info *fence_req; |
| 1832 | struct endpt *ep = (struct endpt *)epd; |
| 1833 | |
| 1834 | if (!(fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL))) { |
| 1835 | err = -ENOMEM; |
| 1836 | goto error; |
| 1837 | } |
| 1838 | |
| 1839 | fence_req->state = OP_IN_PROGRESS; |
| 1840 | init_waitqueue_head(&fence_req->wq); |
| 1841 | |
| 1842 | msg.src = ep->port; |
| 1843 | msg.uop = SCIF_WAIT; |
| 1844 | msg.payload[0] = ep->remote_ep; |
| 1845 | msg.payload[1] = (uint64_t)fence_req; |
| 1846 | msg.payload[2] = mark; |
| 1847 | |
| 1848 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) |
| 1849 | goto error; |
| 1850 | retry: |
| 1851 | err = wait_event_timeout(fence_req->wq, |
| 1852 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); |
| 1853 | if (!err && scifdev_alive(ep)) |
| 1854 | goto retry; |
| 1855 | if (!err) |
| 1856 | err = -ENODEV; |
| 1857 | if (err > 0) |
| 1858 | err = 0; |
| 1859 | if (err < 0) { |
| 1860 | mutex_lock(&ep->rma_info.rma_lock); |
| 1861 | if (OP_IN_PROGRESS == fence_req->state) |
| 1862 | fence_req->state = OP_FAILED; |
| 1863 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1864 | } |
| 1865 | if (OP_FAILED == fence_req->state && !err) |
| 1866 | err = -ENOMEM; |
| 1867 | mutex_lock(&ep->rma_info.rma_lock); |
| 1868 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1869 | kfree(fence_req); |
| 1870 | error: |
| 1871 | return err; |
| 1872 | } |
| 1873 | |
| 1874 | /** |
| 1875 | * micscif_send_fence_signal: |
| 1876 | * @epd - endpoint descriptor |
| 1877 | * @loff - local offset |
| 1878 | * @lval - local value to write to loffset |
| 1879 | * @roff - remote offset |
| 1880 | * @rval - remote value to write to roffset |
| 1881 | * @flags - flags |
| 1882 | * |
| 1883 | * Sends a remote fence signal request |
| 1884 | */ |
| 1885 | int micscif_send_fence_signal(scif_epd_t epd, off_t roff, uint64_t rval, |
| 1886 | off_t loff, uint64_t lval, int flags) |
| 1887 | { |
| 1888 | int err = 0; |
| 1889 | struct nodemsg msg; |
| 1890 | struct fence_info *fence_req; |
| 1891 | struct endpt *ep = (struct endpt *)epd; |
| 1892 | |
| 1893 | if (!(fence_req = kmalloc(sizeof(*fence_req), GFP_KERNEL))) { |
| 1894 | err = -ENOMEM; |
| 1895 | goto error; |
| 1896 | } |
| 1897 | |
| 1898 | fence_req->state = OP_IN_PROGRESS; |
| 1899 | init_waitqueue_head(&fence_req->wq); |
| 1900 | |
| 1901 | msg.src = ep->port; |
| 1902 | if (flags & SCIF_SIGNAL_LOCAL) { |
| 1903 | msg.uop = SCIF_SIG_LOCAL; |
| 1904 | msg.payload[0] = ep->remote_ep; |
| 1905 | msg.payload[1] = roff; |
| 1906 | msg.payload[2] = rval; |
| 1907 | msg.payload[3] = (uint64_t)fence_req; |
| 1908 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) |
| 1909 | goto error_free; |
| 1910 | retry1: |
| 1911 | err = wait_event_timeout(fence_req->wq, |
| 1912 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); |
| 1913 | if (!err && scifdev_alive(ep)) |
| 1914 | goto retry1; |
| 1915 | if (!err) |
| 1916 | err = -ENODEV; |
| 1917 | if (err > 0) |
| 1918 | err = 0; |
| 1919 | if (err < 0) { |
| 1920 | mutex_lock(&ep->rma_info.rma_lock); |
| 1921 | if (OP_IN_PROGRESS == fence_req->state) |
| 1922 | fence_req->state = OP_FAILED; |
| 1923 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1924 | } |
| 1925 | if (OP_FAILED == fence_req->state && !err) { |
| 1926 | err = -ENXIO; |
| 1927 | goto error_free; |
| 1928 | } |
| 1929 | } |
| 1930 | fence_req->state = OP_IN_PROGRESS; |
| 1931 | |
| 1932 | if (flags & SCIF_SIGNAL_REMOTE) { |
| 1933 | msg.uop = SCIF_SIG_REMOTE; |
| 1934 | msg.payload[0] = ep->remote_ep; |
| 1935 | msg.payload[1] = loff; |
| 1936 | msg.payload[2] = lval; |
| 1937 | msg.payload[3] = (uint64_t)fence_req; |
| 1938 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) |
| 1939 | goto error_free; |
| 1940 | retry2: |
| 1941 | err = wait_event_timeout(fence_req->wq, |
| 1942 | (OP_IN_PROGRESS != fence_req->state), NODE_ALIVE_TIMEOUT); |
| 1943 | if (!err && scifdev_alive(ep)) |
| 1944 | goto retry2; |
| 1945 | if (!err) |
| 1946 | err = -ENODEV; |
| 1947 | if (err > 0) |
| 1948 | err = 0; |
| 1949 | if (err < 0) { |
| 1950 | mutex_lock(&ep->rma_info.rma_lock); |
| 1951 | if (OP_IN_PROGRESS == fence_req->state) |
| 1952 | fence_req->state = OP_FAILED; |
| 1953 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1954 | } |
| 1955 | if (OP_FAILED == fence_req->state && !err) { |
| 1956 | err = -ENXIO; |
| 1957 | goto error_free; |
| 1958 | } |
| 1959 | } |
| 1960 | error_free: |
| 1961 | mutex_lock(&ep->rma_info.rma_lock); |
| 1962 | mutex_unlock(&ep->rma_info.rma_lock); |
| 1963 | kfree(fence_req); |
| 1964 | error: |
| 1965 | return err; |
| 1966 | } |
| 1967 | |
| 1968 | /* |
| 1969 | * micscif_fence_mark: |
| 1970 | * |
| 1971 | * @epd - endpoint descriptor |
| 1972 | * Set up a mark for this endpoint and return the value of the mark. |
| 1973 | */ |
| 1974 | int micscif_fence_mark(scif_epd_t epd) |
| 1975 | { |
| 1976 | int mark = 0; |
| 1977 | struct endpt *ep = (struct endpt *)epd; |
| 1978 | struct dma_channel *chan = ep->rma_info.dma_chan; |
| 1979 | |
| 1980 | if ((mark = request_dma_channel(chan))) |
| 1981 | goto error; |
| 1982 | |
| 1983 | mark = program_dma_mark(chan); |
| 1984 | |
| 1985 | free_dma_channel(chan); |
| 1986 | error: |
| 1987 | return mark; |
| 1988 | } |
| 1989 | |
| 1990 | /** |
| 1991 | * micscif_rma_destroy_temp_windows: |
| 1992 | * |
| 1993 | * This routine destroys temporary registered windows created |
| 1994 | * by scif_vreadfrom() and scif_vwriteto(). |
| 1995 | */ |
| 1996 | void micscif_rma_destroy_temp_windows(void) |
| 1997 | { |
| 1998 | struct list_head *item, *tmp; |
| 1999 | struct reg_range_t *window; |
| 2000 | struct endpt *ep; |
| 2001 | struct dma_channel *chan; |
| 2002 | might_sleep(); |
| 2003 | restart: |
| 2004 | spin_lock(&ms_info.mi_rmalock); |
| 2005 | list_for_each_safe(item, tmp, &ms_info.mi_rma) { |
| 2006 | window = list_entry(item, |
| 2007 | struct reg_range_t, list_member); |
| 2008 | ep = (struct endpt *)window->ep; |
| 2009 | chan = ep->rma_info.dma_chan; |
| 2010 | |
| 2011 | list_del(&window->list_member); |
| 2012 | spin_unlock(&ms_info.mi_rmalock); |
| 2013 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2014 | if (!chan || |
| 2015 | !scifdev_alive(ep) || |
| 2016 | (!is_current_dma_mark(chan, window->dma_mark) && |
| 2017 | is_dma_mark_processed(chan, window->dma_mark)) || |
| 2018 | !drain_dma_intr(chan)) { |
| 2019 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2020 | /* Remove window from global list */ |
| 2021 | window->unreg_state = OP_COMPLETED; |
| 2022 | } else { |
| 2023 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2024 | /* DMA engine hung ?? */ |
| 2025 | printk(KERN_ERR "%s %d DMA channel %d hung ep->state %d " |
| 2026 | "window->dma_mark 0x%x channel_mark 0x%x\n", |
| 2027 | __func__, __LINE__, get_chan_num(chan), |
| 2028 | ep->sd_state, window->dma_mark, get_dma_mark(chan)); |
| 2029 | WARN_ON(1); |
| 2030 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); |
| 2031 | goto restart; |
| 2032 | } |
| 2033 | |
| 2034 | if (OP_COMPLETED == window->unreg_state) { |
| 2035 | BUG_ON(atomic_sub_return((int32_t)window->nr_pages, |
| 2036 | &ep->rma_info.tw_total_pages) < 0); |
| 2037 | if (RMA_WINDOW_SELF == window->type) |
| 2038 | micscif_destroy_window(ep, window); |
| 2039 | else |
| 2040 | micscif_destroy_remote_window(ep, window); |
| 2041 | BUG_ON(atomic_dec_return( |
| 2042 | &ep->rma_info.tw_refcount) < 0); |
| 2043 | } |
| 2044 | goto restart; |
| 2045 | } |
| 2046 | spin_unlock(&ms_info.mi_rmalock); |
| 2047 | } |
| 2048 | |
| 2049 | /** |
| 2050 | * micscif_rma_destroy_tcw: |
| 2051 | * |
| 2052 | * This routine destroys temporary registered windows created |
| 2053 | * by scif_vreadfrom() and scif_vwriteto(). |
| 2054 | */ |
| 2055 | static |
| 2056 | void __micscif_rma_destroy_tcw(struct rma_mmu_notifier *mmn, |
| 2057 | struct endpt *ep, bool inrange, |
| 2058 | uint64_t start, uint64_t len) |
| 2059 | { |
| 2060 | struct list_head *item, *tmp; |
| 2061 | struct reg_range_t *window; |
| 2062 | uint64_t start_va, end_va; |
| 2063 | uint64_t end = start + len; |
| 2064 | list_for_each_safe(item, tmp, &mmn->tc_reg_list) { |
| 2065 | window = list_entry(item, |
| 2066 | struct reg_range_t, list_member); |
| 2067 | ep = (struct endpt *)window->ep; |
| 2068 | if (inrange) { |
| 2069 | if (0 == len) |
| 2070 | break; |
| 2071 | start_va = (uint64_t)window->va_for_temp; |
| 2072 | end_va = start_va+ (window->nr_pages << PAGE_SHIFT); |
| 2073 | if (start < start_va) { |
| 2074 | if (end <= start_va) { |
| 2075 | break; |
| 2076 | } else { |
| 2077 | } |
| 2078 | |
| 2079 | } else { |
| 2080 | if (start >= end_va) { |
| 2081 | continue; |
| 2082 | } else { |
| 2083 | } |
| 2084 | } |
| 2085 | } |
| 2086 | __micscif_rma_destroy_tcw_helper(window); |
| 2087 | } |
| 2088 | } |
| 2089 | |
| 2090 | static inline |
| 2091 | void micscif_rma_destroy_tcw(struct rma_mmu_notifier *mmn, |
| 2092 | struct endpt *ep, bool inrange, |
| 2093 | uint64_t start, uint64_t len) |
| 2094 | { |
| 2095 | unsigned long sflags; |
| 2096 | |
| 2097 | spin_lock_irqsave(&ep->rma_info.tc_lock, sflags); |
| 2098 | __micscif_rma_destroy_tcw(mmn, ep, inrange, start, len); |
| 2099 | spin_unlock_irqrestore(&ep->rma_info.tc_lock, sflags); |
| 2100 | } |
| 2101 | |
| 2102 | static void __micscif_rma_destroy_tcw_ep(struct endpt *ep) |
| 2103 | { |
| 2104 | struct list_head *item, *tmp; |
| 2105 | struct rma_mmu_notifier *mmn; |
| 2106 | spin_lock(&ep->rma_info.tc_lock); |
| 2107 | list_for_each_safe(item, tmp, &ep->rma_info.mmn_list) { |
| 2108 | mmn = list_entry(item, |
| 2109 | struct rma_mmu_notifier, list_member); |
| 2110 | __micscif_rma_destroy_tcw(mmn, ep, false, 0, 0); |
| 2111 | } |
| 2112 | spin_unlock(&ep->rma_info.tc_lock); |
| 2113 | } |
| 2114 | |
| 2115 | void micscif_rma_destroy_tcw_ep(struct endpt *ep) |
| 2116 | { |
| 2117 | struct list_head *item, *tmp; |
| 2118 | struct rma_mmu_notifier *mmn; |
| 2119 | list_for_each_safe(item, tmp, &ep->rma_info.mmn_list) { |
| 2120 | mmn = list_entry(item, |
| 2121 | struct rma_mmu_notifier, list_member); |
| 2122 | micscif_rma_destroy_tcw(mmn, ep, false, 0, 0); |
| 2123 | } |
| 2124 | } |
| 2125 | |
| 2126 | /** |
| 2127 | * micscif_rma_destroy_tcw: |
| 2128 | * |
| 2129 | * This routine destroys temporary registered windows created |
| 2130 | * by scif_vreadfrom() and scif_vwriteto(). |
| 2131 | */ |
| 2132 | void micscif_rma_destroy_tcw_invalid(struct list_head *list) |
| 2133 | { |
| 2134 | struct list_head *item, *tmp; |
| 2135 | struct reg_range_t *window; |
| 2136 | struct endpt *ep; |
| 2137 | struct dma_channel *chan; |
| 2138 | might_sleep(); |
| 2139 | restart: |
| 2140 | spin_lock(&ms_info.mi_rmalock); |
| 2141 | list_for_each_safe(item, tmp, list) { |
| 2142 | window = list_entry(item, |
| 2143 | struct reg_range_t, list_member); |
| 2144 | ep = (struct endpt *)window->ep; |
| 2145 | chan = ep->rma_info.dma_chan; |
| 2146 | list_del(&window->list_member); |
| 2147 | spin_unlock(&ms_info.mi_rmalock); |
| 2148 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2149 | mutex_lock(&ep->rma_info.rma_lock); |
| 2150 | if (!chan || |
| 2151 | !scifdev_alive(ep) || |
| 2152 | (!is_current_dma_mark(chan, window->dma_mark) && |
| 2153 | is_dma_mark_processed(chan, window->dma_mark)) || |
| 2154 | !drain_dma_intr(chan)) { |
| 2155 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2156 | BUG_ON(atomic_sub_return((int32_t)window->nr_pages, |
| 2157 | &ep->rma_info.tcw_total_pages) < 0); |
| 2158 | micscif_destroy_window(ep, window); |
| 2159 | BUG_ON(atomic_dec_return( |
| 2160 | &ep->rma_info.tcw_refcount) < 0); |
| 2161 | } else { |
| 2162 | /* DMA engine hung ?? */ |
| 2163 | printk(KERN_ERR "%s %d DMA channel %d hung ep->state %d " |
| 2164 | "window->dma_mark 0x%x channel_mark 0x%x\n", |
| 2165 | __func__, __LINE__, get_chan_num(chan), |
| 2166 | ep->sd_state, window->dma_mark, get_dma_mark(chan)); |
| 2167 | WARN_ON(1); |
| 2168 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2169 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2170 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); |
| 2171 | goto restart; |
| 2172 | } |
| 2173 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2174 | goto restart; |
| 2175 | } |
| 2176 | spin_unlock(&ms_info.mi_rmalock); |
| 2177 | } |
| 2178 | |
| 2179 | /** |
| 2180 | * micscif_rma_handle_remote_fences: |
| 2181 | * |
| 2182 | * This routine services remote fence requests. |
| 2183 | */ |
| 2184 | void micscif_rma_handle_remote_fences(void) |
| 2185 | { |
| 2186 | struct list_head *item, *tmp; |
| 2187 | struct remote_fence_info *fence; |
| 2188 | struct endpt *ep; |
| 2189 | int mark; |
| 2190 | |
| 2191 | might_sleep(); |
| 2192 | mutex_lock(&ms_info.mi_fencelock); |
| 2193 | list_for_each_safe(item, tmp, &ms_info.mi_fence) { |
| 2194 | fence = list_entry(item, |
| 2195 | struct remote_fence_info, list_member); |
| 2196 | /* Remove fence from global list */ |
| 2197 | list_del(&fence->list_member); |
| 2198 | |
| 2199 | /* Initiate the fence operation */ |
| 2200 | ep = (struct endpt *)fence->msg.payload[0]; |
| 2201 | mark = (int)fence->msg.payload[2]; |
| 2202 | BUG_ON(!(mark & SCIF_REMOTE_FENCE)); |
| 2203 | if (dma_mark_wait(ep->rma_info.dma_chan, |
| 2204 | mark & ~SCIF_REMOTE_FENCE, false)) { |
| 2205 | printk(KERN_ERR "%s %d err\n", __func__, __LINE__); |
| 2206 | fence->msg.uop = SCIF_WAIT_NACK; |
| 2207 | } else { |
| 2208 | fence->msg.uop = SCIF_WAIT_ACK; |
| 2209 | } |
| 2210 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2211 | fence->msg.payload[0] = ep->remote_ep; |
| 2212 | /* No error handling for Notification messages. */ |
| 2213 | micscif_nodeqp_send(ep->remote_dev, &fence->msg, ep); |
| 2214 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2215 | kfree(fence); |
| 2216 | /* |
| 2217 | * Decrement ref count and wake up |
| 2218 | * any thread blocked in the EP close routine waiting |
| 2219 | * for all such remote fence requests to complete. |
| 2220 | */ |
| 2221 | ep->rma_info.fence_refcount--; |
| 2222 | wake_up(&ep->rma_info.fence_wq); |
| 2223 | } |
| 2224 | mutex_unlock(&ms_info.mi_fencelock); |
| 2225 | } |
| 2226 | |
| 2227 | #ifdef CONFIG_MMU_NOTIFIER |
| 2228 | void micscif_mmu_notif_handler(struct work_struct *work) |
| 2229 | { |
| 2230 | struct list_head *pos, *tmpq; |
| 2231 | struct endpt *ep; |
| 2232 | restart: |
| 2233 | micscif_rma_destroy_tcw_invalid(&ms_info.mi_rma_tc); |
| 2234 | spin_lock(&ms_info.mi_rmalock); |
| 2235 | list_for_each_safe(pos, tmpq, &ms_info.mi_mmu_notif_cleanup) { |
| 2236 | ep = list_entry(pos, struct endpt, mmu_list); |
| 2237 | list_del(&ep->mmu_list); |
| 2238 | spin_unlock(&ms_info.mi_rmalock); |
| 2239 | BUG_ON(list_empty(&ep->rma_info.mmn_list)); |
| 2240 | |
| 2241 | micscif_rma_destroy_tcw_ep(ep); |
| 2242 | ep_unregister_mmu_notifier(ep); |
| 2243 | queue_work(ms_info.mi_misc_wq, &ms_info.mi_misc_work); |
| 2244 | goto restart; |
| 2245 | } |
| 2246 | spin_unlock(&ms_info.mi_rmalock); |
| 2247 | } |
| 2248 | #endif |
| 2249 | |
| 2250 | /** |
| 2251 | * micscif_reserve_dma_chan: |
| 2252 | * @ep: Endpoint Descriptor. |
| 2253 | * |
| 2254 | * This routine reserves a DMA channel for a particular |
| 2255 | * endpoint. All DMA transfers for an endpoint are always |
| 2256 | * programmed on the same DMA channel. |
| 2257 | */ |
| 2258 | int micscif_reserve_dma_chan(struct endpt *ep) |
| 2259 | { |
| 2260 | int err = 0; |
| 2261 | #ifndef _MIC_SCIF_ |
| 2262 | /* |
| 2263 | * Host Loopback cannot use DMA by design and hence |
| 2264 | * reserving DMA channels is a nop. |
| 2265 | */ |
| 2266 | if (is_self_scifdev(ep->remote_dev)) |
| 2267 | return 0; |
| 2268 | #endif |
| 2269 | mutex_lock(&ep->rma_info.rma_lock); |
| 2270 | if (!ep->rma_info.dma_chan) { |
| 2271 | struct dma_channel **chan = &ep->rma_info.dma_chan; |
| 2272 | unsigned long ts = jiffies; |
| 2273 | #ifndef _MIC_SCIF_ |
| 2274 | mic_ctx_t *mic_ctx = |
| 2275 | get_per_dev_ctx(ep->remote_dev->sd_node - 1); |
| 2276 | BUG_ON(!ep->remote_dev->sd_node); |
| 2277 | #endif |
| 2278 | while (true) { |
| 2279 | if (!(err = allocate_dma_channel((struct mic_dma_ctx_t *) |
| 2280 | #ifdef _MIC_SCIF_ |
| 2281 | mic_dma_handle, |
| 2282 | #else |
| 2283 | mic_ctx->dma_handle, |
| 2284 | #endif |
| 2285 | chan))) |
| 2286 | break; |
| 2287 | schedule(); |
| 2288 | if (time_after(jiffies, |
| 2289 | ts + NODE_ALIVE_TIMEOUT)) { |
| 2290 | err = -EBUSY; |
| 2291 | goto error; |
| 2292 | } |
| 2293 | } |
| 2294 | mic_dma_thread_free_chan(*chan); |
| 2295 | } |
| 2296 | error: |
| 2297 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2298 | return err; |
| 2299 | } |
| 2300 | |
| 2301 | /* |
| 2302 | * micscif_prog_signal: |
| 2303 | * @epd - Endpoint Descriptor |
| 2304 | * @offset - registered address |
| 2305 | * @val - Value to be programmed in SUD. |
| 2306 | * @type - Type of the window. |
| 2307 | * |
| 2308 | * Program a status update descriptor adter ensuring that the offset |
| 2309 | * provided is indeed valid. |
| 2310 | */ |
| 2311 | int micscif_prog_signal(scif_epd_t epd, off_t offset, uint64_t val, |
| 2312 | enum rma_window_type type) |
| 2313 | { |
| 2314 | struct endpt *ep = (struct endpt *)epd; |
| 2315 | struct dma_channel *chan = ep->rma_info.dma_chan; |
| 2316 | struct reg_range_t *window = NULL; |
| 2317 | struct micscif_rma_req req; |
| 2318 | int err; |
| 2319 | dma_addr_t phys; |
| 2320 | |
| 2321 | mutex_lock(&ep->rma_info.rma_lock); |
| 2322 | req.out_window = &window; |
| 2323 | req.offset = offset; |
| 2324 | req.nr_bytes = sizeof(uint64_t); |
| 2325 | req.prot = SCIF_PROT_WRITE; |
| 2326 | req.type = WINDOW_SINGLE; |
| 2327 | if (RMA_WINDOW_SELF == type) |
| 2328 | req.head = &ep->rma_info.reg_list; |
| 2329 | else |
| 2330 | req.head = &ep->rma_info.remote_reg_list; |
| 2331 | /* Does a valid window exist? */ |
| 2332 | if ((err = micscif_query_window(&req))) { |
| 2333 | printk(KERN_ERR "%s %d err %d\n", |
| 2334 | __func__, __LINE__, err); |
| 2335 | goto unlock_ret; |
| 2336 | } |
| 2337 | RMA_MAGIC(window); |
| 2338 | |
| 2339 | #ifndef _MIC_SCIF_ |
| 2340 | if (unlikely(is_self_scifdev(ep->remote_dev))) { |
| 2341 | void *dst_virt; |
| 2342 | if (RMA_WINDOW_SELF == type) |
| 2343 | dst_virt = get_local_va(offset, window, |
| 2344 | sizeof(uint32_t)); |
| 2345 | else { |
| 2346 | struct page **pages = ((struct reg_range_t *) |
| 2347 | (window->peer_window))->pinned_pages->pages; |
| 2348 | int page_nr = (int) ( (offset - window->offset) >> PAGE_SHIFT ); |
| 2349 | off_t page_off = offset & ~PAGE_MASK; |
| 2350 | dst_virt = (void *)((uint64_t)phys_to_virt(page_to_phys( |
| 2351 | pages[page_nr])) | page_off); |
| 2352 | } |
| 2353 | *(uint64_t*)dst_virt = val; |
| 2354 | goto unlock_ret; |
| 2355 | } |
| 2356 | #endif |
| 2357 | phys = micscif_get_dma_addr(window, offset, NULL, NULL, NULL); |
| 2358 | if ((err = request_dma_channel(chan))) |
| 2359 | goto unlock_ret; |
| 2360 | err = do_status_update(chan, phys, val); |
| 2361 | free_dma_channel(chan); |
| 2362 | unlock_ret: |
| 2363 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2364 | return err; |
| 2365 | } |
| 2366 | |
| 2367 | /* |
| 2368 | * __micscif_kill_apps_with_mmaps: |
| 2369 | * @ep - The SCIF endpoint |
| 2370 | * |
| 2371 | * Kill the applications which have valid remote memory mappings |
| 2372 | * created via scif_mmap(..). |
| 2373 | */ |
| 2374 | static void __micscif_kill_apps_with_mmaps(struct endpt *ep) |
| 2375 | { |
| 2376 | struct list_head *item; |
| 2377 | struct rma_task_info *info; |
| 2378 | |
| 2379 | spin_lock(&ep->lock); |
| 2380 | list_for_each(item, &ep->rma_info.task_list) { |
| 2381 | info = list_entry(item, struct rma_task_info, list_member); |
| 2382 | kill_pid(info->pid, SIGKILL, 1); |
| 2383 | pr_debug("%s ep %p pid %p ref %d\n", |
| 2384 | __func__, ep, info->pid, info->ref_count); |
| 2385 | } |
| 2386 | spin_unlock(&ep->lock); |
| 2387 | } |
| 2388 | |
| 2389 | /* |
| 2390 | * _micscif_kill_apps_with_mmaps: |
| 2391 | * @node - remote node id. |
| 2392 | * @head - head of the list of endpoints to kill. |
| 2393 | * |
| 2394 | * Traverse the list of endpoints for a particular remote node and |
| 2395 | * kill applications with valid remote memory mappings. |
| 2396 | */ |
| 2397 | static void _micscif_kill_apps_with_mmaps(int node, struct list_head *head) |
| 2398 | { |
| 2399 | struct endpt *ep; |
| 2400 | unsigned long sflags; |
| 2401 | struct list_head *item; |
| 2402 | |
| 2403 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 2404 | list_for_each(item, head) { |
| 2405 | ep = list_entry(item, struct endpt, list); |
| 2406 | if (ep->remote_dev->sd_node == node) |
| 2407 | __micscif_kill_apps_with_mmaps(ep); |
| 2408 | } |
| 2409 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 2410 | } |
| 2411 | |
| 2412 | /* |
| 2413 | * micscif_kill_apps_with_mmaps: |
| 2414 | * @node - remote node id. |
| 2415 | * |
| 2416 | * Wrapper for killing applications with valid remote memory mappings |
| 2417 | * for a particular node. This API is called by peer nodes as part of |
| 2418 | * handling a lost node. |
| 2419 | */ |
| 2420 | void micscif_kill_apps_with_mmaps(int node) |
| 2421 | { |
| 2422 | _micscif_kill_apps_with_mmaps(node, &ms_info.mi_connected); |
| 2423 | _micscif_kill_apps_with_mmaps(node, &ms_info.mi_disconnected); |
| 2424 | } |
| 2425 | |
| 2426 | /* |
| 2427 | * micscif_query_apps_with_mmaps: |
| 2428 | * @node - remote node id. |
| 2429 | * @head - head of the list of endpoints to query. |
| 2430 | * |
| 2431 | * Query if any applications for a remote node have valid remote memory |
| 2432 | * mappings. |
| 2433 | */ |
| 2434 | static bool micscif_query_apps_with_mmaps(int node, struct list_head *head) |
| 2435 | { |
| 2436 | struct endpt *ep; |
| 2437 | unsigned long sflags; |
| 2438 | struct list_head *item; |
| 2439 | bool ret = false; |
| 2440 | |
| 2441 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 2442 | list_for_each(item, head) { |
| 2443 | ep = list_entry(item, struct endpt, list); |
| 2444 | if (ep->remote_dev->sd_node == node && |
| 2445 | !list_empty(&ep->rma_info.task_list)) { |
| 2446 | ret = true; |
| 2447 | break; |
| 2448 | } |
| 2449 | } |
| 2450 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 2451 | return ret; |
| 2452 | } |
| 2453 | |
| 2454 | /* |
| 2455 | * micscif_rma_do_apps_have_mmaps: |
| 2456 | * @node - remote node id. |
| 2457 | * |
| 2458 | * Wrapper for querying if any applications have remote memory mappings |
| 2459 | * for a particular node. |
| 2460 | */ |
| 2461 | bool micscif_rma_do_apps_have_mmaps(int node) |
| 2462 | { |
| 2463 | return (micscif_query_apps_with_mmaps(node, &ms_info.mi_connected) || |
| 2464 | micscif_query_apps_with_mmaps(node, &ms_info.mi_disconnected)); |
| 2465 | } |
| 2466 | |
| 2467 | /* |
| 2468 | * __micscif_cleanup_rma_for_zombies: |
| 2469 | * @ep - The SCIF endpoint |
| 2470 | * |
| 2471 | * This API is only called while handling a lost node: |
| 2472 | * a) Remote node is dead. |
| 2473 | * b) All endpoints with remote memory mappings have been killed. |
| 2474 | * So we can traverse the remote_reg_list without any locks. Since |
| 2475 | * the window has not yet been unregistered we can drop the ref count |
| 2476 | * and queue it to the cleanup thread. |
| 2477 | */ |
| 2478 | static void __micscif_cleanup_rma_for_zombies(struct endpt *ep) |
| 2479 | { |
| 2480 | struct list_head *pos, *tmp; |
| 2481 | struct reg_range_t *window; |
| 2482 | |
| 2483 | list_for_each_safe(pos, tmp, &ep->rma_info.remote_reg_list) { |
| 2484 | window = list_entry(pos, struct reg_range_t, list_member); |
| 2485 | /* If unregistration is complete then why is it on the list? */ |
| 2486 | WARN_ON(window->unreg_state == OP_COMPLETED); |
| 2487 | if (window->ref_count) |
| 2488 | put_window_ref_count(window, window->nr_pages); |
| 2489 | if (!window->ref_count) { |
| 2490 | atomic_inc(&ep->rma_info.tw_refcount); |
| 2491 | atomic_add_return((int32_t)window->nr_pages, |
| 2492 | &ep->rma_info.tw_total_pages); |
| 2493 | list_del(&window->list_member); |
| 2494 | micscif_queue_for_cleanup(window, &ms_info.mi_rma); |
| 2495 | } |
| 2496 | } |
| 2497 | } |
| 2498 | |
| 2499 | /* |
| 2500 | * micscif_cleanup_rma_for_zombies: |
| 2501 | * @node - remote node id. |
| 2502 | * |
| 2503 | * Cleanup remote registration lists for zombie endpoints. |
| 2504 | */ |
| 2505 | void micscif_cleanup_rma_for_zombies(int node) |
| 2506 | { |
| 2507 | struct endpt *ep; |
| 2508 | unsigned long sflags; |
| 2509 | struct list_head *item; |
| 2510 | |
| 2511 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); |
| 2512 | list_for_each(item, &ms_info.mi_zombie) { |
| 2513 | ep = list_entry(item, struct endpt, list); |
| 2514 | if (ep->remote_dev && ep->remote_dev->sd_node == node) { |
| 2515 | /* |
| 2516 | * If the zombie endpoint remote node matches the lost |
| 2517 | * node then the scifdev should not be alive. |
| 2518 | */ |
| 2519 | WARN_ON(scifdev_alive(ep)); |
| 2520 | __micscif_cleanup_rma_for_zombies(ep); |
| 2521 | } |
| 2522 | } |
| 2523 | spin_unlock_irqrestore(&ms_info.mi_eplock, sflags); |
| 2524 | } |
| 2525 | |
| 2526 | /* |
| 2527 | * micscif_rma_get_task: |
| 2528 | * |
| 2529 | * Store the parent task struct and bump up the number of remote mappings. |
| 2530 | * If this is the first remote memory mapping for this endpoint then |
| 2531 | * create a new rma_task_info entry in the epd task list. |
| 2532 | */ |
| 2533 | int micscif_rma_get_task(struct endpt *ep, int nr_pages) |
| 2534 | { |
| 2535 | struct list_head *item; |
| 2536 | struct rma_task_info *info; |
| 2537 | int err = 0; |
| 2538 | |
| 2539 | spin_lock(&ep->lock); |
| 2540 | list_for_each(item, &ep->rma_info.task_list) { |
| 2541 | info = list_entry(item, struct rma_task_info, list_member); |
| 2542 | if (info->pid == task_tgid(current)) { |
| 2543 | info->ref_count += nr_pages; |
| 2544 | pr_debug("%s ep %p existing pid %p ref %d\n", |
| 2545 | __func__, ep, info->pid, info->ref_count); |
| 2546 | goto unlock; |
| 2547 | } |
| 2548 | } |
| 2549 | spin_unlock(&ep->lock); |
| 2550 | |
| 2551 | /* A new task is mapping this window. Create a new entry */ |
| 2552 | if (!(info = kzalloc(sizeof(*info), GFP_KERNEL))) { |
| 2553 | err = -ENOMEM; |
| 2554 | goto done; |
| 2555 | } |
| 2556 | |
| 2557 | info->pid = get_pid(task_tgid(current)); |
| 2558 | info->ref_count = nr_pages; |
| 2559 | pr_debug("%s ep %p new pid %p ref %d\n", |
| 2560 | __func__, ep, info->pid, info->ref_count); |
| 2561 | spin_lock(&ep->lock); |
| 2562 | list_add_tail(&info->list_member, &ep->rma_info.task_list); |
| 2563 | unlock: |
| 2564 | spin_unlock(&ep->lock); |
| 2565 | done: |
| 2566 | return err; |
| 2567 | } |
| 2568 | |
| 2569 | /* |
| 2570 | * micscif_rma_put_task: |
| 2571 | * |
| 2572 | * Bump down the number of remote mappings. if the ref count for this |
| 2573 | * particular task drops to zero then remove the rma_task_info from |
| 2574 | * the epd task list. |
| 2575 | */ |
| 2576 | void micscif_rma_put_task(struct endpt *ep, int nr_pages) |
| 2577 | { |
| 2578 | struct list_head *item; |
| 2579 | struct rma_task_info *info; |
| 2580 | |
| 2581 | spin_lock(&ep->lock); |
| 2582 | list_for_each(item, &ep->rma_info.task_list) { |
| 2583 | info = list_entry(item, struct rma_task_info, list_member); |
| 2584 | if (info->pid == task_tgid(current)) { |
| 2585 | info->ref_count -= nr_pages; |
| 2586 | pr_debug("%s ep %p pid %p ref %d\n", |
| 2587 | __func__, ep, info->pid, info->ref_count); |
| 2588 | if (!info->ref_count) { |
| 2589 | list_del(&info->list_member); |
| 2590 | put_pid(info->pid); |
| 2591 | kfree(info); |
| 2592 | } |
| 2593 | goto done; |
| 2594 | } |
| 2595 | } |
| 2596 | /* Why was the task not found? This is a bug. */ |
| 2597 | WARN_ON(1); |
| 2598 | done: |
| 2599 | spin_unlock(&ep->lock); |
| 2600 | return; |
| 2601 | } |
| 2602 | |
| 2603 | /* Only debug API's below */ |
| 2604 | void micscif_display_window(struct reg_range_t *window, const char *s, int line) |
| 2605 | { |
| 2606 | int j; |
| 2607 | |
| 2608 | printk("%s %d window %p type %d temp %d offset 0x%llx" |
| 2609 | " nr_pages 0x%llx nr_contig_chunks 0x%llx" |
| 2610 | " prot %d ref_count %d magic 0x%llx peer_window 0x%llx" |
| 2611 | " unreg_state 0x%x va_for_temp %p\n", |
| 2612 | s, line, window, window->type, window->temp, |
| 2613 | window->offset, window->nr_pages, window->nr_contig_chunks, |
| 2614 | window->prot, window->ref_count, window->magic, |
| 2615 | window->peer_window, window->unreg_state, window->va_for_temp); |
| 2616 | |
| 2617 | for (j = 0; j < window->nr_contig_chunks; j++) |
| 2618 | pr_debug("page[%d] = dma_addr 0x%llx num_pages 0x%x\n", |
| 2619 | j, |
| 2620 | window->dma_addr[j], |
| 2621 | window->num_pages[j]); |
| 2622 | |
| 2623 | if (RMA_WINDOW_SELF == window->type && window->pinned_pages) |
| 2624 | for (j = 0; j < window->nr_pages; j++) |
| 2625 | pr_debug("page[%d] = pinned_pages %p address %p\n", |
| 2626 | j, window->pinned_pages->pages[j], |
| 2627 | page_address(window->pinned_pages->pages[j])); |
| 2628 | |
| 2629 | #ifdef CONFIG_ML1OM |
| 2630 | if (window->temp_phys_addr) |
| 2631 | for (j = 0; j < window->nr_contig_chunks; j++) |
| 2632 | pr_debug("page[%d] = temp_phys_addr 0x%llx\n", |
| 2633 | j, window->temp_phys_addr[j]); |
| 2634 | if (window->phys_addr) |
| 2635 | for (j = 0; j < window->nr_pages; j++) |
| 2636 | pr_debug("page[%d] = phys_addr 0x%llx\n", |
| 2637 | j, window->phys_addr[j]); |
| 2638 | #endif |
| 2639 | RMA_MAGIC(window); |
| 2640 | } |