| 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 | #ifndef MICSCIF_RMA_H |
| 37 | #define MICSCIF_RMA_H |
| 38 | |
| 39 | #ifdef CONFIG_MMU_NOTIFIER |
| 40 | #include <linux/mmu_notifier.h> |
| 41 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 42 | #include <linux/huge_mm.h> |
| 43 | #endif |
| 44 | #ifdef CONFIG_HUGETLB_PAGE |
| 45 | #include <linux/hugetlb.h> |
| 46 | #endif |
| 47 | #endif |
| 48 | #include "scif.h" |
| 49 | #include <linux/errno.h> |
| 50 | #include <linux/hardirq.h> |
| 51 | #include <linux/types.h> |
| 52 | #include <linux/capability.h> |
| 53 | #include <linux/slab.h> |
| 54 | #include <linux/string.h> |
| 55 | #include <linux/gfp.h> |
| 56 | #include <linux/vmalloc.h> |
| 57 | #include <asm/io.h> |
| 58 | #include <linux/kernel.h> |
| 59 | #include <linux/mm_types.h> |
| 60 | #include <linux/jiffies.h> |
| 61 | #include <linux/timer.h> |
| 62 | #include <linux/irqflags.h> |
| 63 | #include <linux/time.h> |
| 64 | #include <linux/spinlock.h> |
| 65 | #include <linux/mutex.h> |
| 66 | #include <linux/semaphore.h> |
| 67 | #include <linux/kthread.h> |
| 68 | #include <linux/sched.h> |
| 69 | #include <linux/delay.h> |
| 70 | #include <linux/wait.h> |
| 71 | #include <asm/bug.h> |
| 72 | #include <linux/pci.h> |
| 73 | #include <linux/device.h> |
| 74 | #include <linux/fs.h> |
| 75 | #include <linux/list.h> |
| 76 | #include <linux/workqueue.h> |
| 77 | #include <linux/interrupt.h> |
| 78 | #include <asm/atomic.h> |
| 79 | #include <linux/netdevice.h> |
| 80 | #include <linux/debugfs.h> |
| 81 | #include "mic/micscif_kmem_cache.h" |
| 82 | |
| 83 | struct rma_mmu_notifier { |
| 84 | #ifdef CONFIG_MMU_NOTIFIER |
| 85 | struct mmu_notifier ep_mmu_notifier; |
| 86 | #endif |
| 87 | bool ep_mn_registered; |
| 88 | /* List of temp registration windows for self */ |
| 89 | struct list_head tc_reg_list; |
| 90 | struct mm_struct *mm; |
| 91 | struct endpt *ep; |
| 92 | struct list_head list_member; |
| 93 | }; |
| 94 | |
| 95 | /* Per Endpoint Remote Memory Access Information */ |
| 96 | struct endpt_rma_info { |
| 97 | /* List of registration windows for self */ |
| 98 | struct list_head reg_list; |
| 99 | /* List of registration windows for peer */ |
| 100 | struct list_head remote_reg_list; |
| 101 | /* Offset generator */ |
| 102 | struct va_gen_addr va_gen; |
| 103 | /* |
| 104 | * Synchronizes access to self/remote list and also |
| 105 | * protects the window from being destroyed while |
| 106 | * RMAs are in progress. |
| 107 | */ |
| 108 | struct mutex rma_lock; |
| 109 | /* |
| 110 | * Synchronizes access to temporary cached windows list |
| 111 | * for SCIF Registration Caching. |
| 112 | */ |
| 113 | spinlock_t tc_lock; |
| 114 | /* |
| 115 | * Synchronizes access to the list of MMU notifiers |
| 116 | * registered for this SCIF endpoint. |
| 117 | */ |
| 118 | struct mutex mmn_lock; |
| 119 | /* |
| 120 | * Synchronizes access to the SCIF registered address space |
| 121 | * offset generator. |
| 122 | */ |
| 123 | struct mutex va_lock; |
| 124 | /* |
| 125 | * Keeps track of number of outstanding temporary registered |
| 126 | * windows created by scif_vreadfrom/scif_vwriteto which have |
| 127 | * not been destroyed. tcw refers to the number of temporary |
| 128 | * cached windows and total number of pages pinned. |
| 129 | */ |
| 130 | atomic_t tw_refcount; |
| 131 | atomic_t tw_total_pages; |
| 132 | atomic_t tcw_refcount; |
| 133 | atomic_t tcw_total_pages; |
| 134 | /* |
| 135 | * MMU notifier so that we can destroy the windows when there is |
| 136 | * a change |
| 137 | */ |
| 138 | struct list_head mmn_list; |
| 139 | /* |
| 140 | * Keeps track of number of outstanding remote fence requests |
| 141 | * which have been received by the peer. |
| 142 | */ |
| 143 | int fence_refcount; |
| 144 | /* |
| 145 | * The close routine blocks on this wait queue to ensure that all |
| 146 | * remote fence requests have been serviced. |
| 147 | */ |
| 148 | wait_queue_head_t fence_wq; |
| 149 | /* |
| 150 | * DMA channel used for all DMA transfers for this endpoint. |
| 151 | */ |
| 152 | struct dma_channel *dma_chan; |
| 153 | /* Detect asynchronous list entry deletion */ |
| 154 | int async_list_del; |
| 155 | #ifdef _MIC_SCIF_ |
| 156 | /* Local P2P proxy DMA virtual address for SUD updates by peer */ |
| 157 | void *proxy_dma_va; |
| 158 | /* Local P2P proxy DMA physical address location for SUD updates */ |
| 159 | dma_addr_t proxy_dma_phys; |
| 160 | /* Remote P2P proxy DMA physical address location for SUD updates */ |
| 161 | dma_addr_t proxy_dma_peer_phys; |
| 162 | #endif |
| 163 | /* List of tasks which have remote memory mappings */ |
| 164 | struct list_head task_list; |
| 165 | }; |
| 166 | |
| 167 | /* Information used for tracking remote fence requests */ |
| 168 | struct fence_info { |
| 169 | /* State of this transfer */ |
| 170 | enum micscif_msg_state state; |
| 171 | |
| 172 | /* Fences wait on this queue */ |
| 173 | wait_queue_head_t wq; |
| 174 | |
| 175 | /* Used for storing the DMA mark */ |
| 176 | int dma_mark; |
| 177 | }; |
| 178 | |
| 179 | /* Per remote fence wait request */ |
| 180 | struct remote_fence_info { |
| 181 | /* The SCIF_WAIT message */ |
| 182 | struct nodemsg msg; |
| 183 | |
| 184 | struct list_head list_member; |
| 185 | }; |
| 186 | |
| 187 | /* Self or Peer window */ |
| 188 | enum rma_window_type { |
| 189 | RMA_WINDOW_SELF = 0x1, |
| 190 | RMA_WINDOW_PEER |
| 191 | }; |
| 192 | |
| 193 | /* The number of physical addresses that can be stored in a PAGE. */ |
| 194 | #define NR_PHYS_ADDR_IN_PAGE (PAGE_SIZE >> 3) |
| 195 | |
| 196 | /* |
| 197 | * Store an array of lookup offsets. Each offset in this array maps |
| 198 | * one 4K page containing 512 physical addresses i.e. 2MB. 512 such |
| 199 | * offsets in a 4K page will correspond to 1GB of registered address space. |
| 200 | */ |
| 201 | struct rma_lookup { |
| 202 | /* Array of offsets */ |
| 203 | dma_addr_t *lookup; |
| 204 | /* Offset used to map lookup array */ |
| 205 | dma_addr_t offset; |
| 206 | }; |
| 207 | |
| 208 | |
| 209 | /* |
| 210 | * A set of pinned pages obtained with scif_pin_pages() which could be part |
| 211 | * of multiple registered windows across different end points. |
| 212 | */ |
| 213 | struct scif_pinned_pages { |
| 214 | int64_t nr_pages; |
| 215 | int prot; |
| 216 | int map_flags; |
| 217 | atomic_t ref_count; |
| 218 | uint64_t magic; |
| 219 | /* |
| 220 | * Array of pointers to struct pages populated |
| 221 | * with get_user_pages(..) |
| 222 | */ |
| 223 | struct page **pages; |
| 224 | int *num_pages; |
| 225 | int64_t nr_contig_chunks; |
| 226 | /* Only for Hosts without THP but with Huge TLB FS Like SuSe11 SP1 */ |
| 227 | struct vm_area_struct **vma; |
| 228 | }; |
| 229 | |
| 230 | /* |
| 231 | * Information about a particular task which has remote memory mappings |
| 232 | * created via scif_mmap(..). |
| 233 | */ |
| 234 | struct rma_task_info { |
| 235 | /* |
| 236 | * Stores the pid struct of the grp_leader task structure which |
| 237 | * scif_mmap(..)'d the remote window. |
| 238 | */ |
| 239 | struct pid *pid; |
| 240 | int ref_count; |
| 241 | struct list_head list_member; |
| 242 | }; |
| 243 | |
| 244 | /* Registration Window for Self */ |
| 245 | struct reg_range_t { |
| 246 | int64_t nr_pages; |
| 247 | /* Number of contiguous physical chunks */ |
| 248 | int64_t nr_contig_chunks; |
| 249 | int prot; |
| 250 | int ref_count; |
| 251 | /* Cookie to detect corruption */ |
| 252 | uint64_t magic; |
| 253 | uint64_t offset; |
| 254 | /* va address that this window represents |
| 255 | * Useful for only for temp windows*/ |
| 256 | void *va_for_temp; |
| 257 | /* Used for temporary windows*/ |
| 258 | int dma_mark; |
| 259 | /* |
| 260 | * Pointer to EP. Useful for passing EP around |
| 261 | * with messages to avoid expensive list |
| 262 | * traversals. |
| 263 | */ |
| 264 | uint64_t ep; |
| 265 | |
| 266 | struct list_head list_member; |
| 267 | |
| 268 | enum rma_window_type type; |
| 269 | |
| 270 | /* |
| 271 | * Pointer to peer window. Useful for sending |
| 272 | * messages to peer without requiring an |
| 273 | * extra list traversal |
| 274 | */ |
| 275 | uint64_t peer_window; |
| 276 | |
| 277 | /* Unregistration state */ |
| 278 | enum micscif_msg_state unreg_state; |
| 279 | |
| 280 | /* |
| 281 | * True for temporary windows created via |
| 282 | * scif_vreadfrom/scif_vwriteto. |
| 283 | */ |
| 284 | bool temp; |
| 285 | |
| 286 | bool offset_freed; |
| 287 | |
| 288 | /* Local P2P proxy DMA physical address location for SUD updates */ |
| 289 | dma_addr_t proxy_dma_phys; |
| 290 | |
| 291 | union { |
| 292 | /* Self RAS */ |
| 293 | struct { |
| 294 | /* The set of pinned_pages backing this window */ |
| 295 | struct scif_pinned_pages *pinned_pages; |
| 296 | |
| 297 | /* Handle for sending ALLOC_REQ */ |
| 298 | struct allocmsg alloc_handle; |
| 299 | |
| 300 | /* Wait Queue for an registration (N)ACK */ |
| 301 | wait_queue_head_t regwq; |
| 302 | |
| 303 | /* Registration state */ |
| 304 | enum micscif_msg_state reg_state; |
| 305 | |
| 306 | /* Wait Queue for an unregistration (N)ACK */ |
| 307 | wait_queue_head_t unregwq; |
| 308 | }; |
| 309 | /* Peer RAS specific window elements */ |
| 310 | struct { |
| 311 | #ifdef CONFIG_ML1OM |
| 312 | /* Lookup for physical addresses used for mmap */ |
| 313 | struct rma_lookup phys_addr_lookup; |
| 314 | |
| 315 | /* Lookup for temp physical addresses used for mmap */ |
| 316 | struct rma_lookup temp_phys_addr_lookup; |
| 317 | |
| 318 | /* Mmap state */ |
| 319 | enum micscif_msg_state gttmap_state; |
| 320 | |
| 321 | /* Wait Queue for an unregistration (N)ACK */ |
| 322 | wait_queue_head_t gttmapwq; |
| 323 | |
| 324 | /* Ref count per page */ |
| 325 | int *page_ref_count; |
| 326 | #endif |
| 327 | /* Lookup for physical addresses used for DMA */ |
| 328 | struct rma_lookup dma_addr_lookup; |
| 329 | |
| 330 | /* Number of entries in lookup */ |
| 331 | int nr_lookup; |
| 332 | |
| 333 | /* Offset used to map the window by the peer */ |
| 334 | dma_addr_t mapped_offset; |
| 335 | |
| 336 | /* Ref count for tracking scif_get_pages */ |
| 337 | int get_put_ref_count; |
| 338 | }; |
| 339 | }; |
| 340 | #ifdef CONFIG_ML1OM |
| 341 | /* Array of physical addresses used for creating VtoP mappings */ |
| 342 | /* FIXME: these are phys_addr as seen by the peer node, node at the |
| 343 | * opposite end of the endpt |
| 344 | */ |
| 345 | dma_addr_t *phys_addr; |
| 346 | |
| 347 | /* Temporary array for storing physical addresses for performance */ |
| 348 | dma_addr_t *temp_phys_addr; |
| 349 | #endif |
| 350 | |
| 351 | /* Array of physical addresses used for Host & MIC initiated DMA */ |
| 352 | dma_addr_t *dma_addr; |
| 353 | |
| 354 | /* Array specifying number of pages for each physical address */ |
| 355 | int *num_pages; |
| 356 | struct mm_struct *mm; |
| 357 | } __attribute__ ((packed)); |
| 358 | |
| 359 | |
| 360 | #define RMA_MAGIC(x) BUG_ON(x->magic != SCIFEP_MAGIC) |
| 361 | |
| 362 | /* If this bit is set then the mark is a remote fence mark */ |
| 363 | #define SCIF_REMOTE_FENCE_BIT 30 |
| 364 | /* Magic value used to indicate a remote fence request */ |
| 365 | #define SCIF_REMOTE_FENCE (1ULL << SCIF_REMOTE_FENCE_BIT) |
| 366 | |
| 367 | enum rma_direction { |
| 368 | LOCAL_TO_REMOTE, |
| 369 | REMOTE_TO_LOCAL |
| 370 | }; |
| 371 | |
| 372 | /* Initialize RMA for this EP */ |
| 373 | int micscif_rma_ep_init(struct endpt *ep); |
| 374 | |
| 375 | /* Check if epd can be uninitialized */ |
| 376 | int micscif_rma_ep_can_uninit(struct endpt *ep); |
| 377 | |
| 378 | /* Obtain a new offset. Callee must grab RMA lock */ |
| 379 | int micscif_get_window_offset(struct endpt *ep, int flags, |
| 380 | uint64_t offset, size_t len, uint64_t *out_offset); |
| 381 | |
| 382 | /* Free offset. Callee must grab RMA lock */ |
| 383 | void micscif_free_window_offset(struct endpt *ep, |
| 384 | uint64_t offset, size_t len); |
| 385 | |
| 386 | /* Create self registration window */ |
| 387 | struct reg_range_t *micscif_create_window(struct endpt *ep, |
| 388 | int64_t nr_pages, uint64_t offset, bool temp); |
| 389 | |
| 390 | /* Create a set of pinned pages */ |
| 391 | struct scif_pinned_pages *micscif_create_pinned_pages(int nr_pages, int prot); |
| 392 | |
| 393 | /* Destroy a set of pinned pages */ |
| 394 | int micscif_destroy_pinned_pages(struct scif_pinned_pages *pages); |
| 395 | |
| 396 | /* Destroy self registration window.*/ |
| 397 | int micscif_destroy_window(struct endpt *ep, struct reg_range_t *window); |
| 398 | |
| 399 | int micscif_destroy_incomplete_window(struct endpt *ep, struct reg_range_t *window); |
| 400 | |
| 401 | /* Map pages of self window to Aperture/PCI */ |
| 402 | int micscif_map_window_pages(struct endpt *ep, struct reg_range_t *window, bool temp); |
| 403 | |
| 404 | /* Unregister a self window */ |
| 405 | int micscif_unregister_window(struct reg_range_t *window); |
| 406 | |
| 407 | /* Create remote registration window */ |
| 408 | struct reg_range_t *micscif_create_remote_window(struct endpt *ep, int nr_pages); |
| 409 | |
| 410 | /* Destroy remote registration window */ |
| 411 | void micscif_destroy_remote_window(struct endpt *ep, struct reg_range_t *window); |
| 412 | |
| 413 | int micscif_send_alloc_request(struct endpt *ep, struct reg_range_t *window); |
| 414 | |
| 415 | /* Prepare a remote registration window */ |
| 416 | int micscif_prep_remote_window(struct endpt *ep, struct reg_range_t *window); |
| 417 | |
| 418 | /* Create remote lookup entries for physical addresses */ |
| 419 | int micscif_create_remote_lookup(struct endpt *ep, struct reg_range_t *window); |
| 420 | |
| 421 | /* Destroy remote lookup entries for physical addresses */ |
| 422 | void micscif_destroy_remote_lookup(struct endpt *ep, struct reg_range_t *window); |
| 423 | |
| 424 | /* Send a SCIF_REGISTER message and wait for an ACK */ |
| 425 | int micscif_send_scif_register(struct endpt *ep, struct reg_range_t *window); |
| 426 | |
| 427 | /* Send a SCIF_UNREGISTER message */ |
| 428 | int micscif_send_scif_unregister(struct endpt *ep, struct reg_range_t *window); |
| 429 | |
| 430 | /* RMA copy API */ |
| 431 | int micscif_rma_copy(scif_epd_t epd, off_t loffset, void *addr, size_t len, |
| 432 | off_t roffset, int flags, enum rma_direction dir, bool last_chunk); |
| 433 | |
| 434 | /* Sends a remote fence mark request */ |
| 435 | int micscif_send_fence_mark(scif_epd_t epd, int *out_mark); |
| 436 | |
| 437 | /* Sends a remote fence wait request */ |
| 438 | int micscif_send_fence_wait(scif_epd_t epd, int mark); |
| 439 | |
| 440 | /* Sends a remote fence signal request */ |
| 441 | int micscif_send_fence_signal(scif_epd_t epd, off_t roff, uint64_t rval, |
| 442 | off_t loff, uint64_t lval, int flags); |
| 443 | |
| 444 | /* Setup a DMA mark for an endpoint */ |
| 445 | int micscif_fence_mark(scif_epd_t epd); |
| 446 | |
| 447 | void ep_unregister_mmu_notifier(struct endpt *ep); |
| 448 | #ifdef CONFIG_MMU_NOTIFIER |
| 449 | void micscif_mmu_notif_handler(struct work_struct *work); |
| 450 | #endif |
| 451 | |
| 452 | void micscif_rma_destroy_temp_windows(void); |
| 453 | void micscif_rma_destroy_tcw_ep(struct endpt *ep); |
| 454 | void micscif_rma_destroy_tcw_invalid(struct list_head *list); |
| 455 | |
| 456 | void micscif_rma_handle_remote_fences(void); |
| 457 | |
| 458 | /* Reserve a DMA channel for a particular endpoint */ |
| 459 | int micscif_reserve_dma_chan(struct endpt *ep); |
| 460 | |
| 461 | /* Program DMA SUD's after verifying the registered offset */ |
| 462 | int micscif_prog_signal(scif_epd_t epd, off_t offset, uint64_t val, |
| 463 | enum rma_window_type type); |
| 464 | |
| 465 | /* Kill any applications which have valid remote memory mappings */ |
| 466 | void micscif_kill_apps_with_mmaps(int node); |
| 467 | |
| 468 | /* Query if any applications have remote memory mappings */ |
| 469 | bool micscif_rma_do_apps_have_mmaps(int node); |
| 470 | |
| 471 | /* Get a reference to the current task which is creating a remote memory mapping */ |
| 472 | int micscif_rma_get_task(struct endpt *ep, int nr_pages); |
| 473 | |
| 474 | /* Release a reference to the current task which is destroying a remote memory mapping */ |
| 475 | void micscif_rma_put_task(struct endpt *ep, int nr_pages); |
| 476 | |
| 477 | /* Cleanup remote registration lists for zombie endpoints */ |
| 478 | void micscif_cleanup_rma_for_zombies(int node); |
| 479 | |
| 480 | #ifdef _MIC_SCIF_ |
| 481 | void micscif_teardown_proxy_dma(struct endpt *ep); |
| 482 | #endif |
| 483 | |
| 484 | static __always_inline |
| 485 | bool is_unaligned(off_t src_offset, off_t dst_offset) |
| 486 | { |
| 487 | src_offset = src_offset & (L1_CACHE_BYTES - 1); |
| 488 | dst_offset = dst_offset & (L1_CACHE_BYTES - 1); |
| 489 | if (src_offset == dst_offset) |
| 490 | return false; |
| 491 | else |
| 492 | return true; |
| 493 | } |
| 494 | |
| 495 | static __always_inline |
| 496 | int __scif_readfrom(scif_epd_t epd, off_t loffset, size_t len, |
| 497 | off_t roffset, int flags) |
| 498 | { |
| 499 | int err; |
| 500 | |
| 501 | pr_debug("SCIFAPI readfrom: ep %p loffset 0x%lx len 0x%lx" |
| 502 | " offset 0x%lx flags 0x%x\n", |
| 503 | epd, loffset, len, roffset, flags); |
| 504 | |
| 505 | if (is_unaligned(loffset, roffset)) { |
| 506 | while(len > MAX_UNALIGNED_BUF_SIZE) { |
| 507 | err = micscif_rma_copy(epd, loffset, NULL, |
| 508 | MAX_UNALIGNED_BUF_SIZE, |
| 509 | roffset, flags, REMOTE_TO_LOCAL, false); |
| 510 | if (err) |
| 511 | goto readfrom_err; |
| 512 | loffset += MAX_UNALIGNED_BUF_SIZE; |
| 513 | roffset += MAX_UNALIGNED_BUF_SIZE; |
| 514 | len -=MAX_UNALIGNED_BUF_SIZE; |
| 515 | } |
| 516 | } |
| 517 | err = micscif_rma_copy(epd, loffset, NULL, len, |
| 518 | roffset, flags, REMOTE_TO_LOCAL, true); |
| 519 | readfrom_err: |
| 520 | return err; |
| 521 | } |
| 522 | |
| 523 | static __always_inline |
| 524 | int __scif_writeto(scif_epd_t epd, off_t loffset, size_t len, |
| 525 | off_t roffset, int flags) |
| 526 | { |
| 527 | int err; |
| 528 | |
| 529 | pr_debug("SCIFAPI writeto: ep %p loffset 0x%lx len 0x%lx" |
| 530 | " roffset 0x%lx flags 0x%x\n", |
| 531 | epd, loffset, len, roffset, flags); |
| 532 | |
| 533 | if (is_unaligned(loffset, roffset)) { |
| 534 | while(len > MAX_UNALIGNED_BUF_SIZE) { |
| 535 | err = micscif_rma_copy(epd, loffset, NULL, |
| 536 | MAX_UNALIGNED_BUF_SIZE, |
| 537 | roffset, flags, LOCAL_TO_REMOTE, false); |
| 538 | if (err) |
| 539 | goto writeto_err; |
| 540 | loffset += MAX_UNALIGNED_BUF_SIZE; |
| 541 | roffset += MAX_UNALIGNED_BUF_SIZE; |
| 542 | len -= MAX_UNALIGNED_BUF_SIZE; |
| 543 | } |
| 544 | } |
| 545 | err = micscif_rma_copy(epd, loffset, NULL, len, |
| 546 | roffset, flags, LOCAL_TO_REMOTE, true); |
| 547 | writeto_err: |
| 548 | return err; |
| 549 | } |
| 550 | |
| 551 | static __always_inline |
| 552 | int __scif_vreadfrom(scif_epd_t epd, void *addr, size_t len, off_t roffset, int flags) |
| 553 | { |
| 554 | int err; |
| 555 | |
| 556 | pr_debug("SCIFAPI vreadfrom: ep %p addr %p len 0x%lx" |
| 557 | " roffset 0x%lx flags 0x%x\n", |
| 558 | epd, addr, len, roffset, flags); |
| 559 | |
| 560 | if (is_unaligned((off_t)addr, roffset)) { |
| 561 | if (len > MAX_UNALIGNED_BUF_SIZE) |
| 562 | flags &= ~SCIF_RMA_USECACHE; |
| 563 | |
| 564 | while(len > MAX_UNALIGNED_BUF_SIZE) { |
| 565 | err = micscif_rma_copy(epd, 0, addr, |
| 566 | MAX_UNALIGNED_BUF_SIZE, |
| 567 | roffset, flags, REMOTE_TO_LOCAL, false); |
| 568 | if (err) |
| 569 | goto vreadfrom_err; |
| 570 | addr = (void *)((uint64_t)addr + MAX_UNALIGNED_BUF_SIZE); |
| 571 | roffset += MAX_UNALIGNED_BUF_SIZE; |
| 572 | len -= MAX_UNALIGNED_BUF_SIZE; |
| 573 | } |
| 574 | } |
| 575 | err = micscif_rma_copy(epd, 0, addr, len, |
| 576 | roffset, flags, REMOTE_TO_LOCAL, true); |
| 577 | vreadfrom_err: |
| 578 | return err; |
| 579 | } |
| 580 | |
| 581 | static __always_inline |
| 582 | int __scif_vwriteto(scif_epd_t epd, void *addr, size_t len, off_t roffset, int flags) |
| 583 | { |
| 584 | int err; |
| 585 | |
| 586 | pr_debug("SCIFAPI vwriteto: ep %p addr %p len 0x%lx" |
| 587 | " roffset 0x%lx flags 0x%x\n", |
| 588 | epd, addr, len, roffset, flags); |
| 589 | |
| 590 | if (is_unaligned((off_t)addr, roffset)) { |
| 591 | if (len > MAX_UNALIGNED_BUF_SIZE) |
| 592 | flags &= ~SCIF_RMA_USECACHE; |
| 593 | |
| 594 | while(len > MAX_UNALIGNED_BUF_SIZE) { |
| 595 | err = micscif_rma_copy(epd, 0, addr, |
| 596 | MAX_UNALIGNED_BUF_SIZE, |
| 597 | roffset, flags, LOCAL_TO_REMOTE, false); |
| 598 | if (err) |
| 599 | goto vwriteto_err; |
| 600 | addr = (void *)((uint64_t)addr + MAX_UNALIGNED_BUF_SIZE); |
| 601 | roffset += MAX_UNALIGNED_BUF_SIZE; |
| 602 | len -= MAX_UNALIGNED_BUF_SIZE; |
| 603 | } |
| 604 | } |
| 605 | err = micscif_rma_copy(epd, 0, addr, len, |
| 606 | roffset, flags, LOCAL_TO_REMOTE, true); |
| 607 | vwriteto_err: |
| 608 | return err; |
| 609 | } |
| 610 | |
| 611 | void micscif_rma_completion_cb(uint64_t data); |
| 612 | |
| 613 | int micscif_pci_dev(uint16_t node, struct pci_dev **pdev); |
| 614 | #ifndef _MIC_SCIF_ |
| 615 | int micscif_pci_info(uint16_t node, struct scif_pci_info *dev); |
| 616 | #endif |
| 617 | |
| 618 | /* |
| 619 | * nr_pages in a 2MB page is specified via the top 12 bits in the |
| 620 | * physical address. |
| 621 | */ |
| 622 | |
| 623 | /* Check all parenthesis in these macros. See if putting in bottom makes sense? */ |
| 624 | #define RMA_HUGE_NR_PAGE_SHIFT ((52)) |
| 625 | #define RMA_HUGE_NR_PAGE_MASK (((0xFFFULL) << RMA_HUGE_NR_PAGE_SHIFT)) |
| 626 | #define RMA_GET_NR_PAGES(addr) ((addr) >> RMA_HUGE_NR_PAGE_SHIFT) |
| 627 | #define RMA_SET_NR_PAGES(addr, nr_pages) ((addr) = (((nr_pages) & 0xFFFULL) << RMA_HUGE_NR_PAGE_SHIFT) | ((uint64_t)(addr))) |
| 628 | #define RMA_GET_ADDR(addr) ((addr) & ~(RMA_HUGE_NR_PAGE_MASK)) |
| 629 | |
| 630 | extern bool mic_huge_page_enable; |
| 631 | |
| 632 | #define SCIF_HUGE_PAGE_SHIFT 21 |
| 633 | |
| 634 | /* |
| 635 | * micscif_is_huge_page: |
| 636 | * @page: A physical page. |
| 637 | */ |
| 638 | static __always_inline int |
| 639 | micscif_is_huge_page(struct scif_pinned_pages *pinned_pages, int index) |
| 640 | { |
| 641 | int huge = 0; |
| 642 | struct page *page = pinned_pages->pages[index]; |
| 643 | |
| 644 | if (compound_order(page) + PAGE_SHIFT == SCIF_HUGE_PAGE_SHIFT) |
| 645 | huge = 1; |
| 646 | if (huge) |
| 647 | ms_info.nr_2mb_pages++; |
| 648 | if (!mic_huge_page_enable) |
| 649 | huge = 0; |
| 650 | #ifdef RMA_DEBUG |
| 651 | WARN_ON(!page_count(page)); |
| 652 | WARN_ON(page_mapcount(page) < 0); |
| 653 | #endif |
| 654 | return huge; |
| 655 | } |
| 656 | |
| 657 | /* |
| 658 | * micscif_detect_large_page: |
| 659 | * @pinned_pages: A set of pinned pages. |
| 660 | */ |
| 661 | static __always_inline int |
| 662 | micscif_detect_large_page(struct scif_pinned_pages *pinned_pages, char *addr) |
| 663 | { |
| 664 | int i = 0, nr_pages, huge; |
| 665 | char *next_huge, *end; |
| 666 | char *end_addr = addr + (pinned_pages->nr_pages << PAGE_SHIFT); |
| 667 | |
| 668 | while (addr < end_addr) { |
| 669 | huge = micscif_is_huge_page(pinned_pages, i); |
| 670 | if (huge) { |
| 671 | next_huge = (char *)ALIGN( |
| 672 | (unsigned long)(addr + 1), |
| 673 | PMD_SIZE); |
| 674 | end = next_huge > end_addr ? end_addr : next_huge; |
| 675 | nr_pages = (int)((end - addr) >> PAGE_SHIFT); |
| 676 | pinned_pages->num_pages[i] = (int)nr_pages; |
| 677 | addr = end; |
| 678 | i += (int)nr_pages; |
| 679 | |
| 680 | } else { |
| 681 | pinned_pages->num_pages[i] = 1; |
| 682 | i++; |
| 683 | addr += PAGE_SIZE; |
| 684 | ms_info.nr_4k_pages++; |
| 685 | } |
| 686 | pinned_pages->nr_contig_chunks++; |
| 687 | } |
| 688 | return 0; |
| 689 | } |
| 690 | |
| 691 | /** |
| 692 | * micscif_set_nr_pages: |
| 693 | * @ep: end point |
| 694 | * @window: self registration window |
| 695 | * |
| 696 | * Set nr_pages in every entry of physical address/dma address array |
| 697 | * and also remove nr_pages information from physical addresses. |
| 698 | */ |
| 699 | static __always_inline void |
| 700 | micscif_set_nr_pages(struct micscif_dev *dev, struct reg_range_t *window) |
| 701 | { |
| 702 | int j; |
| 703 | #ifdef CONFIG_ML1OM |
| 704 | int l = 0, k; |
| 705 | #endif |
| 706 | |
| 707 | for (j = 0; j < window->nr_contig_chunks; j++) { |
| 708 | window->num_pages[j] = RMA_GET_NR_PAGES(window->dma_addr[j]); |
| 709 | if (window->num_pages[j]) |
| 710 | window->dma_addr[j] = RMA_GET_ADDR(window->dma_addr[j]); |
| 711 | else |
| 712 | break; |
| 713 | #ifdef CONFIG_ML1OM |
| 714 | for (k = 0; k < window->num_pages[j]; k++) |
| 715 | if (window->temp_phys_addr[j]) |
| 716 | window->phys_addr[l + k] = |
| 717 | RMA_GET_ADDR(window->temp_phys_addr[j]) + (k << PAGE_SHIFT); |
| 718 | l += window->num_pages[j]; |
| 719 | #endif |
| 720 | } |
| 721 | } |
| 722 | |
| 723 | #ifdef CONFIG_ML1OM |
| 724 | /* |
| 725 | * micscif_get_phys_addr: |
| 726 | * Obtain the phys_addr given the window and the offset. |
| 727 | * @window: Registered window. |
| 728 | * @off: Window offset. |
| 729 | */ |
| 730 | static __always_inline dma_addr_t |
| 731 | micscif_get_phys_addr(struct reg_range_t *window, uint64_t off) |
| 732 | { |
| 733 | int page_nr = (off - window->offset) >> PAGE_SHIFT; |
| 734 | off_t page_off = off & ~PAGE_MASK; |
| 735 | return window->phys_addr[page_nr] | page_off; |
| 736 | } |
| 737 | #endif |
| 738 | |
| 739 | #define RMA_ERROR_CODE (~(dma_addr_t)0x0) |
| 740 | |
| 741 | /* |
| 742 | * micscif_get_dma_addr: |
| 743 | * Obtain the dma_addr given the window and the offset. |
| 744 | * @window: Registered window. |
| 745 | * @off: Window offset. |
| 746 | * @nr_bytes: Return the number of contiguous bytes till next DMA addr index. |
| 747 | * @index: Return the index of the dma_addr array found. |
| 748 | * @start_off: start offset of index of the dma addr array found. |
| 749 | * The nr_bytes provides the callee an estimate of the maximum possible |
| 750 | * DMA xfer possible while the index/start_off provide faster lookups |
| 751 | * for the next iteration. |
| 752 | */ |
| 753 | static __always_inline dma_addr_t |
| 754 | micscif_get_dma_addr(struct reg_range_t *window, uint64_t off, size_t *nr_bytes, int *index, uint64_t *start_off) |
| 755 | { |
| 756 | if (window->nr_pages == window->nr_contig_chunks) { |
| 757 | int page_nr = (int)((off - window->offset) >> PAGE_SHIFT); |
| 758 | off_t page_off = off & ~PAGE_MASK; |
| 759 | if (nr_bytes) |
| 760 | *nr_bytes = PAGE_SIZE - page_off; |
| 761 | if (page_nr >= window->nr_pages) { |
| 762 | printk(KERN_ERR "%s dma_addr out of boundary\n", __FUNCTION__); |
| 763 | } |
| 764 | return window->dma_addr[page_nr] | page_off; |
| 765 | } else { |
| 766 | int i = index ? *index : 0; |
| 767 | uint64_t end; |
| 768 | uint64_t start = start_off ? *start_off : window->offset; |
| 769 | for (; i < window->nr_contig_chunks; i++) { |
| 770 | end = start + (window->num_pages[i] << PAGE_SHIFT); |
| 771 | if (off >= start && off < end) { |
| 772 | if (index) |
| 773 | *index = i; |
| 774 | if (start_off) |
| 775 | *start_off = start; |
| 776 | if (nr_bytes) |
| 777 | *nr_bytes = end - off; |
| 778 | return (window->dma_addr[i] + (off - start)); |
| 779 | } |
| 780 | start += (window->num_pages[i] << PAGE_SHIFT); |
| 781 | } |
| 782 | } |
| 783 | #ifdef CONFIG_MK1OM |
| 784 | printk(KERN_ERR "%s %d BUG. Addr not found? window %p off 0x%llx\n", __func__, __LINE__, window, off); |
| 785 | BUG_ON(1); |
| 786 | #endif |
| 787 | return RMA_ERROR_CODE; |
| 788 | } |
| 789 | |
| 790 | /* |
| 791 | * scif_memset: |
| 792 | * @va: kernel virtual address |
| 793 | * @c: The byte used to fill the memory |
| 794 | * @size: Buffer size |
| 795 | * |
| 796 | * Helper API which fills size bytes of memory pointed to by va with the |
| 797 | * constant byte c. This API fills the memory in chunks of 4GB - 1 bytes |
| 798 | * for a single invocation of memset(..) to work around a kernel bug in |
| 799 | * x86_64 @ https://bugzilla.kernel.org/show_bug.cgi?id=27732 |
| 800 | * where memset(..) does not do "ANY" work for size >= 4GB. |
| 801 | * This kernel bug has been fixed upstream in v3.2 via the commit |
| 802 | * titled "x86-64: Fix memset() to support sizes of 4Gb and above" |
| 803 | * but has not been backported to distributions like RHEL 6.3 yet. |
| 804 | */ |
| 805 | static __always_inline void scif_memset(char *va, int c, size_t size) |
| 806 | { |
| 807 | size_t loop_size; |
| 808 | const size_t four_gb = 4 * 1024 * 1024 * 1024ULL; |
| 809 | |
| 810 | while (size) { |
| 811 | loop_size = min(size, four_gb - 1); |
| 812 | memset(va, c, loop_size); |
| 813 | size -= loop_size; |
| 814 | va += loop_size; |
| 815 | } |
| 816 | } |
| 817 | |
| 818 | /* |
| 819 | * scif_zalloc: |
| 820 | * @size: Size of the allocation request. |
| 821 | * |
| 822 | * Helper API which attempts to allocate zeroed pages via |
| 823 | * __get_free_pages(..) first and then falls back on |
| 824 | * vmalloc(..) if that fails. This is required because |
| 825 | * vmalloc(..) is *slow*. |
| 826 | */ |
| 827 | static __always_inline void *scif_zalloc(size_t size) |
| 828 | { |
| 829 | void *ret; |
| 830 | size_t align = ALIGN(size, PAGE_SIZE); |
| 831 | |
| 832 | if (!align) |
| 833 | return NULL; |
| 834 | |
| 835 | if (align <= (1 << (MAX_ORDER + PAGE_SHIFT - 1))) |
| 836 | if ((ret = (void*)__get_free_pages(GFP_KERNEL | __GFP_ZERO, |
| 837 | get_order(align)))) |
| 838 | goto done; |
| 839 | if (!(ret = vmalloc(align))) |
| 840 | return NULL; |
| 841 | |
| 842 | /* TODO: Use vzalloc once kernel supports it */ |
| 843 | scif_memset(ret, 0, size); |
| 844 | done: |
| 845 | #ifdef RMA_DEBUG |
| 846 | atomic_long_add_return(align, &ms_info.rma_alloc_cnt); |
| 847 | #endif |
| 848 | return ret; |
| 849 | } |
| 850 | |
| 851 | /* |
| 852 | * scif_free: |
| 853 | * @addr: Address to be freed. |
| 854 | * @size: Size of the allocation. |
| 855 | * Helper API which frees memory allocated via scif_zalloc(). |
| 856 | */ |
| 857 | static __always_inline void scif_free(void *addr, size_t size) |
| 858 | { |
| 859 | size_t align = ALIGN(size, PAGE_SIZE); |
| 860 | |
| 861 | if (unlikely(is_vmalloc_addr(addr))) |
| 862 | vfree(addr); |
| 863 | else { |
| 864 | free_pages((unsigned long)addr, get_order(align)); |
| 865 | } |
| 866 | #ifdef RMA_DEBUG |
| 867 | WARN_ON(atomic_long_sub_return(align, &ms_info.rma_alloc_cnt) < 0); |
| 868 | #endif |
| 869 | } |
| 870 | |
| 871 | static __always_inline void |
| 872 | get_window_ref_count(struct reg_range_t *window, int64_t nr_pages) |
| 873 | { |
| 874 | window->ref_count += (int)nr_pages; |
| 875 | } |
| 876 | |
| 877 | static __always_inline void |
| 878 | put_window_ref_count(struct reg_range_t *window, int64_t nr_pages) |
| 879 | { |
| 880 | window->ref_count -= (int)nr_pages; |
| 881 | BUG_ON(window->nr_pages < 0); |
| 882 | } |
| 883 | |
| 884 | static __always_inline void |
| 885 | set_window_ref_count(struct reg_range_t *window, int64_t nr_pages) |
| 886 | { |
| 887 | window->ref_count = (int)nr_pages; |
| 888 | } |
| 889 | |
| 890 | /* Debug API's */ |
| 891 | void micscif_display_window(struct reg_range_t *window, const char *s, int line); |
| 892 | static inline struct mm_struct *__scif_acquire_mm(void) |
| 893 | { |
| 894 | if (mic_ulimit_check) { |
| 895 | #ifdef RMA_DEBUG |
| 896 | atomic_long_add_return(1, &ms_info.rma_mm_cnt); |
| 897 | #endif |
| 898 | return get_task_mm(current); |
| 899 | } |
| 900 | return NULL; |
| 901 | } |
| 902 | |
| 903 | static inline void __scif_release_mm(struct mm_struct *mm) |
| 904 | { |
| 905 | if (mic_ulimit_check && mm) { |
| 906 | #ifdef RMA_DEBUG |
| 907 | WARN_ON(atomic_long_sub_return(1, &ms_info.rma_mm_cnt) < 0); |
| 908 | #endif |
| 909 | mmput(mm); |
| 910 | } |
| 911 | } |
| 912 | |
| 913 | static inline int __scif_dec_pinned_vm_lock(struct mm_struct *mm, |
| 914 | int64_t nr_pages, bool try_lock) |
| 915 | { |
| 916 | if (mm && nr_pages && mic_ulimit_check) { |
| 917 | if (try_lock) { |
| 918 | if (!down_write_trylock(&mm->mmap_sem)) { |
| 919 | return -1; |
| 920 | } |
| 921 | } else { |
| 922 | down_write(&mm->mmap_sem); |
| 923 | } |
| 924 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3, 1, 0)) |
| 925 | mm->pinned_vm -= nr_pages; |
| 926 | #else |
| 927 | mm->locked_vm -= nr_pages; |
| 928 | #endif |
| 929 | up_write(&mm->mmap_sem); |
| 930 | } |
| 931 | return 0; |
| 932 | } |
| 933 | |
| 934 | static inline int __scif_check_inc_pinned_vm(struct mm_struct *mm, |
| 935 | int64_t nr_pages) |
| 936 | { |
| 937 | if (mm && mic_ulimit_check && nr_pages) { |
| 938 | unsigned long locked, lock_limit; |
| 939 | locked = nr_pages; |
| 940 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3, 1, 0)) |
| 941 | locked += mm->pinned_vm; |
| 942 | #else |
| 943 | locked += mm->locked_vm; |
| 944 | #endif |
| 945 | lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
| 946 | if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) { |
| 947 | pr_debug("locked(%lu) > lock_limit(%lu)\n", |
| 948 | locked, lock_limit); |
| 949 | return -ENOMEM; |
| 950 | } else { |
| 951 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3, 1, 0)) |
| 952 | mm->pinned_vm = locked; |
| 953 | #else |
| 954 | mm->locked_vm = locked; |
| 955 | #endif |
| 956 | } |
| 957 | } |
| 958 | return 0; |
| 959 | } |
| 960 | #endif |