| 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_H |
| 37 | #define MICSCIF_H |
| 38 | |
| 39 | #include <linux/errno.h> |
| 40 | #include <linux/hardirq.h> |
| 41 | #include <linux/types.h> |
| 42 | #include <linux/capability.h> |
| 43 | #include <linux/slab.h> |
| 44 | #include <linux/string.h> |
| 45 | #include <linux/gfp.h> |
| 46 | #include <linux/vmalloc.h> |
| 47 | #include <asm/io.h> |
| 48 | #include <linux/kernel.h> |
| 49 | #include <linux/mm_types.h> |
| 50 | #include <linux/jiffies.h> |
| 51 | #include <linux/timer.h> |
| 52 | #include <linux/irqflags.h> |
| 53 | #include <linux/time.h> |
| 54 | #include <linux/spinlock.h> |
| 55 | #include <linux/mutex.h> |
| 56 | #include <linux/semaphore.h> |
| 57 | #include <linux/kthread.h> |
| 58 | #include <linux/sched.h> |
| 59 | #include <linux/delay.h> |
| 60 | #include <linux/wait.h> |
| 61 | #include <asm/bug.h> |
| 62 | #include <linux/pci.h> |
| 63 | #include <linux/device.h> |
| 64 | #include <linux/fs.h> |
| 65 | #include <linux/list.h> |
| 66 | #include <linux/workqueue.h> |
| 67 | #include <linux/interrupt.h> |
| 68 | #include <asm/atomic.h> |
| 69 | #include <linux/netdevice.h> |
| 70 | #include <linux/debugfs.h> |
| 71 | |
| 72 | #ifdef _MODULE_SCIF_ |
| 73 | #include <linux/mman.h> |
| 74 | #include <linux/pagemap.h> |
| 75 | #include <asm/uaccess.h> |
| 76 | #include <linux/poll.h> |
| 77 | #include <linux/mmzone.h> |
| 78 | #include <linux/version.h> |
| 79 | #endif /* MODULE_SCIF */ |
| 80 | |
| 81 | #include <linux/notifier.h> |
| 82 | #include "scif.h" |
| 83 | #include "mic/micbaseaddressdefine.h" |
| 84 | #include "mic/micsboxdefine.h" |
| 85 | |
| 86 | /* The test runs in a separate thread context from the bottom |
| 87 | * half that processes messages from the card and setup p2p |
| 88 | * when these run concurrently, p2p messages get lost since they |
| 89 | * may be consumed by the test thread |
| 90 | */ |
| 91 | //#define ENABLE_TEST // Used to enable testing at board connect |
| 92 | #ifdef MIC_IS_EMULATION |
| 93 | #define TEST_LOOP 2 |
| 94 | #else |
| 95 | #define TEST_LOOP 2000 |
| 96 | #endif |
| 97 | |
| 98 | //#define P2P_HACK 0 |
| 99 | #include "scif.h" |
| 100 | #include "scif_ioctl.h" |
| 101 | |
| 102 | #define SCIF_READY_MAGIC_NUM 0x1eedfee0 |
| 103 | |
| 104 | #ifndef SCIF_MAJOR |
| 105 | #define SCIF_MAJOR 0 /* Use dynamic major number by default */ |
| 106 | #endif |
| 107 | |
| 108 | #define SCIF_HOST_NODE 0 // By default the host is always node zero |
| 109 | |
| 110 | #define SCIF_RMA_TEMP_CACHE_LIMIT 0x20000 |
| 111 | /* |
| 112 | * The overhead for proxying a P2P DMA read to convert it to |
| 113 | * a DMA write by sending a SCIF Node QP message has been |
| 114 | * seen to be higher than programming a P2P DMA Read on self |
| 115 | * for transfer sizes less than the PROXY_DMA_THRESHOLD. |
| 116 | * The minimum threshold is different for Jaketown versus |
| 117 | * Ivytown and tuned for best DMA performance. |
| 118 | */ |
| 119 | #define SCIF_PROXY_DMA_THRESHOLD_JKT (32 * 1024ULL) |
| 120 | #define SCIF_PROXY_DMA_THRESHOLD_IVT (1024 * 1024ULL) |
| 121 | |
| 122 | //#define RMA_DEBUG 0 |
| 123 | |
| 124 | /* Pre-defined L1_CACHE_SHIFT is 6 on RH and 7 on Suse */ |
| 125 | #undef L1_CACHE_SHIFT |
| 126 | #define L1_CACHE_SHIFT 6 |
| 127 | #undef L1_CACHE_BYTES |
| 128 | #define L1_CACHE_BYTES (1 << L1_CACHE_SHIFT) |
| 129 | |
| 130 | #define MI_EPLOCK_HELD (true) |
| 131 | #define MAX_RDMASR 8 |
| 132 | |
| 133 | // Device wide SCIF information |
| 134 | struct micscif_info { |
| 135 | uint32_t mi_nodeid; // Node ID this node is to others. |
| 136 | |
| 137 | struct mutex mi_conflock; // Configuration lock (used in p2p setup) |
| 138 | uint32_t mi_maxid; // Max known board ID |
| 139 | uint32_t mi_total; // Total number of running interfaces |
| 140 | uint32_t mi_nr_zombies; // Keep track of the number of zombie EP. |
| 141 | unsigned long mi_mask; // bit mask of online scif interfaces |
| 142 | uint64_t mi_nr_ioremap; // Keep track of number of ioremap() calls on the host |
| 143 | // to decide when to purge aliases for performance. |
| 144 | spinlock_t mi_eplock; |
| 145 | spinlock_t mi_connlock; |
| 146 | spinlock_t mi_rmalock; // Synchronize access to list of temporary registered |
| 147 | // windows to be destroyed. |
| 148 | struct mutex mi_fencelock; // Synchronize access to list of remote fences requested. |
| 149 | struct mutex mi_event_cblock; |
| 150 | spinlock_t mi_nb_connect_lock; |
| 151 | |
| 152 | struct list_head mi_uaccept; // List of user acceptreq waiting for acceptreg |
| 153 | struct list_head mi_listen; // List of listening end points |
| 154 | struct list_head mi_zombie; // List of zombie end points with pending RMA's. |
| 155 | struct list_head mi_connected; // List of end points in connected state |
| 156 | struct list_head mi_disconnected; // List of end points in disconnected state |
| 157 | struct list_head mi_rma; // List of temporary registered windows to be destroyed. |
| 158 | struct list_head mi_rma_tc; // List of temporary |
| 159 | // registered & cached windows |
| 160 | // to be destroyed. |
| 161 | struct list_head mi_fence; // List of remote fence requests. |
| 162 | struct list_head mi_event_cb; /* List of event handlers registered */ |
| 163 | struct list_head mi_nb_connect_list; |
| 164 | #ifdef CONFIG_MMU_NOTIFIER |
| 165 | struct list_head mi_mmu_notif_cleanup; |
| 166 | #endif |
| 167 | struct notifier_block mi_panic_nb; |
| 168 | #ifndef _MIC_SCIF_ |
| 169 | /* The host needs to keep track of node dependencies in form of graph. |
| 170 | * This will need to be dynamically grown to support hotplug. |
| 171 | */ |
| 172 | uint32_t **mi_depmtrx; |
| 173 | /* |
| 174 | * Wait queue used for blocking while waiting for nodes |
| 175 | * to respond for disconnect message sent from host. |
| 176 | */ |
| 177 | wait_queue_head_t mi_disconn_wq; |
| 178 | /* stus of node remove operation*/ |
| 179 | uint64_t mi_disconnect_status; |
| 180 | atomic_long_t mi_unique_msgid; |
| 181 | #endif |
| 182 | /* |
| 183 | * Watchdog timeout on the host. Timer expiry will result in the host |
| 184 | * treating the remote node as a lost node. Default value is |
| 185 | * DEFAULT_WATCHDOG_TO and can be modified to a value greater than 1 |
| 186 | * second via SCIF sysfs watchdog_to entry. |
| 187 | */ |
| 188 | int mi_watchdog_to; // Watchdog timeout |
| 189 | int mi_watchdog_enabled; // Watchdog timeout enabled |
| 190 | int mi_watchdog_auto_reboot; // Watchdog auto reboot enabled |
| 191 | struct workqueue_struct *mi_misc_wq; // Workqueue for miscellaneous SCIF tasks. |
| 192 | struct work_struct mi_misc_work; |
| 193 | #ifdef CONFIG_MMU_NOTIFIER |
| 194 | struct workqueue_struct *mi_mmu_notif_wq; // Workqueue for MMU notifier cleanup tasks. |
| 195 | struct work_struct mi_mmu_notif_work; |
| 196 | #endif |
| 197 | int nr_gtt_entries; // GTT Debug Counter to detect leaks |
| 198 | uint64_t nr_2mb_pages; // Debug Counter for number of 2mb pages. |
| 199 | uint64_t nr_4k_pages; // Debug Counter for number of 4K pages |
| 200 | uint8_t en_msg_log; |
| 201 | wait_queue_head_t mi_exitwq; |
| 202 | unsigned long mi_rma_tc_limit; |
| 203 | uint64_t mi_proxy_dma_threshold; |
| 204 | #ifdef RMA_DEBUG |
| 205 | atomic_long_t rma_mm_cnt; |
| 206 | atomic_long_t rma_unaligned_cpu_cnt; |
| 207 | atomic_long_t rma_alloc_cnt; |
| 208 | atomic_long_t rma_pin_cnt; |
| 209 | #ifdef CONFIG_MMU_NOTIFIER |
| 210 | atomic_long_t mmu_notif_cnt; |
| 211 | #endif |
| 212 | #endif |
| 213 | #ifdef _MIC_SCIF_ |
| 214 | int mi_intr_rcnt[MAX_RDMASR]; // Ref count to track SCIF Interrupt Handlers |
| 215 | #endif |
| 216 | struct workqueue_struct *mi_conn_wq; |
| 217 | struct work_struct mi_conn_work; |
| 218 | }; |
| 219 | |
| 220 | extern struct micscif_info ms_info; |
| 221 | |
| 222 | #define SCIF_NODE_MAGIC_BIT 63 |
| 223 | /* Magic value used to indicate a remote idle node without grabbing any locks */ |
| 224 | #define SCIF_NODE_IDLE (1ULL << SCIF_NODE_MAGIC_BIT) |
| 225 | |
| 226 | enum scif_state { |
| 227 | SCIFDEV_NOTPRESENT, |
| 228 | SCIFDEV_INIT, |
| 229 | SCIFDEV_RUNNING, |
| 230 | SCIFDEV_SLEEPING, |
| 231 | SCIFDEV_STOPPING, |
| 232 | SCIFDEV_STOPPED |
| 233 | }; |
| 234 | |
| 235 | extern bool mic_p2p_enable; |
| 236 | extern bool mic_p2p_proxy_enable; |
| 237 | extern bool mic_reg_cache_enable; |
| 238 | extern bool mic_ulimit_check; |
| 239 | /* p2p mapping from node id to peer id */ |
| 240 | struct scif_p2p_info { |
| 241 | int ppi_peer_id; |
| 242 | struct scatterlist *ppi_sg[2]; |
| 243 | uint64_t sg_nentries[2]; // no of entries in scatterlists |
| 244 | dma_addr_t ppi_pa[2]; // one for mmio; one for aper |
| 245 | dma_addr_t ppi_mic_addr[2]; // one for mmio; one for aper |
| 246 | uint64_t ppi_len[2]; |
| 247 | #define PPI_MMIO 0 |
| 248 | #define PPI_APER 1 |
| 249 | enum scif_state ppi_disc_state; //Disconnection state of this peer node. |
| 250 | struct list_head ppi_list; |
| 251 | }; |
| 252 | |
| 253 | /* one per remote node */ |
| 254 | struct micscif_dev { |
| 255 | uint16_t sd_node; |
| 256 | enum scif_state sd_state; |
| 257 | volatile void *mm_sbox; |
| 258 | uint64_t sd_base_addr; /* Remote node's base bus addr |
| 259 | * for the local node's aperture |
| 260 | */ |
| 261 | #ifndef _MIC_SCIF_ |
| 262 | struct list_head sd_p2p; /* List of bus addresses for |
| 263 | * other nodes, these are allocated |
| 264 | * by the host driver and are |
| 265 | * valid only on the host node |
| 266 | */ |
| 267 | struct delayed_work sd_watchdog_work; |
| 268 | wait_queue_head_t sd_watchdog_wq; |
| 269 | struct workqueue_struct *sd_ln_wq; |
| 270 | char sd_ln_wqname[16]; |
| 271 | #endif |
| 272 | |
| 273 | int n_qpairs; /* FIXME: |
| 274 | * This is always set to 1, |
| 275 | */ |
| 276 | |
| 277 | struct micscif_qp *qpairs; /* Same FIXME as above |
| 278 | * There is single qp established |
| 279 | * with this remote node |
| 280 | */ |
| 281 | |
| 282 | struct workqueue_struct *sd_intr_wq; /* sd_intr_wq & sd_intr_bh |
| 283 | * together constitute the workqueue |
| 284 | * infrastructure needed to |
| 285 | * run the bottom half handler |
| 286 | * for messages received from |
| 287 | * this node |
| 288 | */ |
| 289 | char sd_intr_wqname[16]; |
| 290 | struct work_struct sd_intr_bh; |
| 291 | unsigned int sd_intr_handle; |
| 292 | uint32_t sd_rdmasr; |
| 293 | struct workqueue_struct *sd_loopb_wq; |
| 294 | char sd_loopb_wqname[16]; |
| 295 | struct work_struct sd_loopb_work; |
| 296 | struct list_head sd_loopb_recv_q; |
| 297 | /* Lock to synchronize remote node state transitions */ |
| 298 | struct mutex sd_lock; |
| 299 | /* |
| 300 | * Global Ref count per SCIF device tracking all SCIF API's which |
| 301 | * might communicate across PCIe. |
| 302 | */ |
| 303 | atomic_long_t scif_ref_cnt; |
| 304 | /* |
| 305 | * Global Ref count per SCIF device tracking scif_mmap()/ |
| 306 | * scif_get_pages(). sd_lock protects scif_map_ref_cnt |
| 307 | * hence it does not need to be an atomic operation. Note that |
| 308 | * scif_mmap()/scif_get_pages() is not in the critical |
| 309 | * perf path. |
| 310 | */ |
| 311 | int scif_map_ref_cnt; |
| 312 | /* |
| 313 | * Wait queue used for blocking while waiting for nodes |
| 314 | * to wake up or to be removed. |
| 315 | */ |
| 316 | wait_queue_head_t sd_wq; |
| 317 | uint64_t sd_wait_status; |
| 318 | #ifdef _MIC_SCIF_ |
| 319 | wait_queue_head_t sd_p2p_wq; |
| 320 | bool sd_proxy_dma_reads; |
| 321 | struct delayed_work sd_p2p_dwork; |
| 322 | int sd_p2p_retry; |
| 323 | #endif |
| 324 | /* |
| 325 | * The NUMA node the peer is attached to on the host. |
| 326 | */ |
| 327 | int sd_numa_node; |
| 328 | /* |
| 329 | * Waitqueue for blocking while waiting for remote memory |
| 330 | * mappings to drop to zero. |
| 331 | */ |
| 332 | wait_queue_head_t sd_mmap_wq; |
| 333 | |
| 334 | /* When a nodeqp message is received, this is set. |
| 335 | * And it is reset by the watchdog time */ |
| 336 | atomic_t sd_node_alive; |
| 337 | int num_active_conn; |
| 338 | #ifdef ENABLE_TEST |
| 339 | struct workqueue_struct *producer; |
| 340 | struct workqueue_struct *consumer; |
| 341 | char producer_name[16]; |
| 342 | char consumer_name[16]; |
| 343 | struct work_struct producer_work; |
| 344 | struct work_struct consumer_work; |
| 345 | int count; |
| 346 | int test_done; |
| 347 | #endif // ENABLE_TEST |
| 348 | }; |
| 349 | |
| 350 | extern struct micscif_dev scif_dev[]; |
| 351 | |
| 352 | #include "mic/micscif_nodeqp.h" |
| 353 | #include "mic/micscif_nm.h" |
| 354 | #include "mic/micscif_smpt.h" |
| 355 | #include "mic/micscif_va_gen.h" |
| 356 | #include "mic/mic_dma_api.h" |
| 357 | #include "mic/mic_dma_lib.h" |
| 358 | #include "mic/micscif_rma.h" |
| 359 | #include "mic/micscif_rma_list.h" |
| 360 | |
| 361 | /* |
| 362 | * data structure used to sync SCIF_GET_NODE_INFO messaging |
| 363 | */ |
| 364 | struct get_node_info { |
| 365 | enum micscif_msg_state state; |
| 366 | wait_queue_head_t wq; |
| 367 | }; |
| 368 | |
| 369 | static inline uint64_t align_low(uint64_t data, uint32_t granularity) |
| 370 | { |
| 371 | return ALIGN(data - (granularity - 1), granularity); |
| 372 | } |
| 373 | |
| 374 | #define SCIF_MIN(a, b) (((a) < (b)) ? (a) : (b)) |
| 375 | #define SCIF_MAX(a, b) (((a) > (b)) ? (a) : (b)) |
| 376 | |
| 377 | enum endptstate { |
| 378 | SCIFEP_CLOSED, // Internal state |
| 379 | SCIFEP_UNBOUND, // External state |
| 380 | SCIFEP_BOUND, // External state |
| 381 | SCIFEP_LISTENING, // External state |
| 382 | SCIFEP_CONNECTED, // External state |
| 383 | SCIFEP_CONNECTING, // Internal state |
| 384 | SCIFEP_MAPPING, // Internal state |
| 385 | SCIFEP_CLOSING, // Internal state |
| 386 | SCIFEP_CLLISTEN, // Internal state |
| 387 | SCIFEP_DISCONNECTED, // Internal state |
| 388 | SCIFEP_ZOMBIE // Internal state |
| 389 | }; |
| 390 | |
| 391 | extern char *scif_ep_states[]; |
| 392 | |
| 393 | // Used for coordinating connection accept sequence. This is the data structure |
| 394 | // for the conlist in the endpoint. |
| 395 | struct conreq { |
| 396 | struct nodemsg msg; |
| 397 | struct list_head list; |
| 398 | }; |
| 399 | |
| 400 | /* Size of the RB for the Node QP */ |
| 401 | #define NODE_QP_SIZE 0x10000 |
| 402 | /* Size of the RB for the Endpoint QP */ |
| 403 | #define ENDPT_QP_SIZE 0x1000 |
| 404 | |
| 405 | struct endpt_qp_info { |
| 406 | /* Qpair for this endpoint */ |
| 407 | struct micscif_qp *qp; |
| 408 | /* |
| 409 | * Physical addr of the QP for Host or |
| 410 | * GTT offset of the QP for MIC. |
| 411 | * Required for unmapping the QP during close. |
| 412 | */ |
| 413 | dma_addr_t qp_offset; |
| 414 | /* |
| 415 | * Payload in a SCIF_CNCT_GNT message containing the |
| 416 | * physical address of the remote_qp. |
| 417 | */ |
| 418 | dma_addr_t cnct_gnt_payload; |
| 419 | }; |
| 420 | |
| 421 | #define SCIFEP_MAGIC 0x5c1f000000005c1f |
| 422 | |
| 423 | struct endpt { |
| 424 | volatile enum endptstate state; |
| 425 | spinlock_t lock; |
| 426 | |
| 427 | struct scif_portID port; |
| 428 | struct scif_portID peer; |
| 429 | |
| 430 | int backlog; |
| 431 | |
| 432 | struct endpt_qp_info qp_info; |
| 433 | struct endpt_rma_info rma_info; |
| 434 | /* |
| 435 | * scifdev used by this endpt to communicate with remote node. |
| 436 | */ |
| 437 | struct micscif_dev *remote_dev; |
| 438 | uint64_t remote_ep; |
| 439 | /* |
| 440 | * Keep track of number of connection requests. |
| 441 | */ |
| 442 | int conreqcnt; |
| 443 | /* |
| 444 | * Cache remote SCIF device state. |
| 445 | */ |
| 446 | enum scif_state sd_state; |
| 447 | /* |
| 448 | * True if the endpoint was created |
| 449 | * via scif_accept(..). |
| 450 | */ |
| 451 | bool accepted_ep; |
| 452 | /* |
| 453 | * Open file information used to match the id passed |
| 454 | * in with the flush routine. |
| 455 | */ |
| 456 | struct files_struct *files; |
| 457 | /* |
| 458 | * Reference count for functions using this endpoint. |
| 459 | */ |
| 460 | struct kref ref_count; |
| 461 | struct list_head conlist; |
| 462 | wait_queue_head_t conwq; |
| 463 | wait_queue_head_t disconwq; |
| 464 | wait_queue_head_t diswq; |
| 465 | wait_queue_head_t sendwq; |
| 466 | wait_queue_head_t recvwq; |
| 467 | struct mutex sendlock; |
| 468 | struct mutex recvlock; |
| 469 | struct list_head list; |
| 470 | |
| 471 | #ifdef CONFIG_MMU_NOTIFIER |
| 472 | struct list_head mmu_list; |
| 473 | #endif |
| 474 | |
| 475 | struct list_head li_accept; /* pending ACCEPTREG */ |
| 476 | int acceptcnt; /* pending ACCEPTREG cnt */ |
| 477 | struct list_head liacceptlist; /* link to listen accept */ |
| 478 | struct list_head miacceptlist; /* link to mi_uaccept */ |
| 479 | struct endpt *listenep; /* associated listen ep */ |
| 480 | |
| 481 | /* Non-blocking connect */ |
| 482 | struct work_struct conn_work; |
| 483 | struct scif_portID conn_port; |
| 484 | int conn_err; |
| 485 | int conn_async_state; |
| 486 | wait_queue_head_t conn_pend_wq; |
| 487 | struct list_head conn_list; |
| 488 | }; |
| 489 | |
| 490 | static __always_inline void |
| 491 | micscif_queue_for_cleanup(struct reg_range_t *window, struct list_head *list) |
| 492 | { |
| 493 | struct endpt *ep = (struct endpt *)window->ep; |
| 494 | INIT_LIST_HEAD(&window->list_member); |
| 495 | window->dma_mark = get_dma_mark(ep->rma_info.dma_chan); |
| 496 | spin_lock(&ms_info.mi_rmalock); |
| 497 | list_add_tail(&window->list_member, list); |
| 498 | spin_unlock(&ms_info.mi_rmalock); |
| 499 | queue_work(ms_info.mi_misc_wq, &ms_info.mi_misc_work); |
| 500 | } |
| 501 | |
| 502 | static __always_inline void |
| 503 | __micscif_rma_destroy_tcw_helper(struct reg_range_t *window) |
| 504 | { |
| 505 | list_del(&window->list_member); |
| 506 | micscif_queue_for_cleanup(window, &ms_info.mi_rma_tc); |
| 507 | } |
| 508 | |
| 509 | void print_ep_state(struct endpt *ep, char *label); |
| 510 | |
| 511 | // Function prototypes needed by Unix/Linux drivers linking to scif |
| 512 | int scif_fdopen(struct file *f); |
| 513 | int scif_fdclose(struct file *f); |
| 514 | int scif_process_ioctl(struct file *f, unsigned int cmd, uint64_t arg); |
| 515 | int micscif_mmap(struct file *file, struct vm_area_struct *vma); |
| 516 | int scif_mmap(struct vm_area_struct *vma, scif_epd_t epd); |
| 517 | void scif_munmap(struct vm_area_struct *vma); |
| 518 | void scif_proc_init(void); |
| 519 | void scif_proc_cleanup(void); |
| 520 | int scif_user_send(scif_epd_t epd, void *msg, int len, int flags); |
| 521 | int scif_user_recv(scif_epd_t epd, void *msg, int len, int flags); |
| 522 | int __scif_pin_pages(void *addr, size_t len, int *out_prot, |
| 523 | int map_flags, scif_pinned_pages_t *pages); |
| 524 | scif_epd_t __scif_open(void); |
| 525 | int __scif_bind(scif_epd_t epd, uint16_t pn); |
| 526 | int __scif_listen(scif_epd_t epd, int backlog); |
| 527 | int __scif_connect(scif_epd_t epd, struct scif_portID *dst, bool non_block); |
| 528 | int __scif_accept(scif_epd_t epd, struct scif_portID *peer, scif_epd_t |
| 529 | *newepd, int flags); |
| 530 | int __scif_close(scif_epd_t epd); |
| 531 | int __scif_send(scif_epd_t epd, void *msg, int len, int flags); |
| 532 | int __scif_recv(scif_epd_t epd, void *msg, int len, int flags); |
| 533 | off_t __scif_register(scif_epd_t epd, void *addr, size_t len, off_t offset, |
| 534 | int prot_flags, int map_flags); |
| 535 | int __scif_unregister(scif_epd_t epd, off_t offset, size_t len); |
| 536 | int __scif_readfrom(scif_epd_t epd, off_t loffset, size_t len, off_t |
| 537 | roffset, int rma_flags); |
| 538 | int __scif_writeto(scif_epd_t epd, off_t loffset, size_t len, off_t |
| 539 | roffset, int rma_flags); |
| 540 | int __scif_fence_mark(scif_epd_t epd, int flags, int *mark); |
| 541 | int __scif_fence_wait(scif_epd_t epd, int mark); |
| 542 | int __scif_fence_signal(scif_epd_t epd, off_t loff, uint64_t lval, off_t roff, |
| 543 | uint64_t rval, int flags); |
| 544 | off_t __scif_register_pinned_pages(scif_epd_t epd, |
| 545 | scif_pinned_pages_t pinned_pages, off_t offset, int map_flags); |
| 546 | int __scif_get_pages(scif_epd_t epd, off_t offset, size_t len, |
| 547 | struct scif_range **pages); |
| 548 | int __scif_put_pages(struct scif_range *pages); |
| 549 | int __scif_flush(scif_epd_t epd); |
| 550 | |
| 551 | void micscif_misc_handler(struct work_struct *work); |
| 552 | void micscif_conn_handler(struct work_struct *work); |
| 553 | |
| 554 | uint16_t rsrv_scif_port(uint16_t port); |
| 555 | uint16_t get_scif_port(void); |
| 556 | void put_scif_port(uint16_t port); |
| 557 | |
| 558 | void micscif_send_exit(void); |
| 559 | |
| 560 | void scif_ref_rel(struct kref *kref_count); |
| 561 | |
| 562 | #ifdef _MODULE_SCIF_ |
| 563 | unsigned int micscif_poll(struct file *f, poll_table *wait); |
| 564 | unsigned int scif_pollfd(struct file *f, poll_table *wait, scif_epd_t epd); |
| 565 | unsigned int __scif_pollfd(struct file *f, poll_table *wait, struct endpt *ep); |
| 566 | int micscif_flush(struct file *f, fl_owner_t id); |
| 567 | #endif |
| 568 | |
| 569 | #ifdef _MIC_SCIF_ |
| 570 | void mic_debug_init(void); |
| 571 | void micscif_get_node_info(void); |
| 572 | void scif_poll_qp_state(struct work_struct *work); |
| 573 | #endif |
| 574 | void mic_debug_uninit(void); |
| 575 | |
| 576 | #define serializing_request(x) ((void)*(volatile uint8_t*)(x)) |
| 577 | |
| 578 | // State list helper functions. |
| 579 | // Each of these functions must be called with the end point lock unlocked. If |
| 580 | // the end point is found on the list the end point returned will have its lock |
| 581 | // set and sflags will return the value to be used to do an unlock_irqrestore |
| 582 | // at the end of the calling function. |
| 583 | static inline struct endpt * |
| 584 | micscif_find_listen_ep(uint16_t port, unsigned long *sflags) |
| 585 | { |
| 586 | struct endpt *ep = NULL; |
| 587 | struct list_head *pos, *tmpq; |
| 588 | unsigned long flags; |
| 589 | |
| 590 | spin_lock_irqsave(&ms_info.mi_eplock, flags); |
| 591 | list_for_each_safe(pos, tmpq, &ms_info.mi_listen) { |
| 592 | ep = list_entry(pos, struct endpt, list); |
| 593 | if (ep->port.port == port) { |
| 594 | *sflags = flags; |
| 595 | spin_lock(&ep->lock); |
| 596 | spin_unlock(&ms_info.mi_eplock); |
| 597 | return ep; |
| 598 | } |
| 599 | } |
| 600 | spin_unlock_irqrestore(&ms_info.mi_eplock, flags); |
| 601 | return (struct endpt *)NULL; |
| 602 | } |
| 603 | |
| 604 | // Must be called with end point locked |
| 605 | static inline struct conreq * |
| 606 | miscscif_get_connection_request(struct endpt *ep, uint64_t payload) |
| 607 | { |
| 608 | struct conreq *conreq; |
| 609 | struct list_head *pos, *tmpq; |
| 610 | |
| 611 | list_for_each_safe(pos, tmpq, &ep->conlist) { |
| 612 | conreq = list_entry(pos, struct conreq, list); |
| 613 | if (conreq->msg.payload[0] == payload) { |
| 614 | list_del(pos); |
| 615 | ep->conreqcnt--; |
| 616 | return conreq; |
| 617 | } |
| 618 | } |
| 619 | return (struct conreq *)NULL; |
| 620 | } |
| 621 | |
| 622 | // There is no requirement for the callee to have the end point |
| 623 | // locked like other API's above. |
| 624 | static inline void |
| 625 | micscif_remove_zombie_ep(struct endpt *ep) |
| 626 | { |
| 627 | struct list_head *pos, *tmpq; |
| 628 | unsigned long sflags; |
| 629 | struct endpt *tmpep; |
| 630 | |
| 631 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); |
| 632 | list_for_each_safe(pos, tmpq, &ms_info.mi_zombie) { |
| 633 | tmpep = list_entry(pos, struct endpt, list); |
| 634 | if (tmpep == ep) { |
| 635 | list_del(pos); |
| 636 | ms_info.mi_nr_zombies--; |
| 637 | } |
| 638 | } |
| 639 | spin_unlock_irqrestore(&ms_info.mi_eplock, sflags); |
| 640 | } |
| 641 | |
| 642 | static inline void |
| 643 | micscif_cleanup_zombie_epd(void) |
| 644 | { |
| 645 | struct list_head *pos, *tmpq; |
| 646 | unsigned long sflags; |
| 647 | struct endpt *ep; |
| 648 | |
| 649 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); |
| 650 | list_for_each_safe(pos, tmpq, &ms_info.mi_zombie) { |
| 651 | ep = list_entry(pos, struct endpt, list); |
| 652 | if (micscif_rma_ep_can_uninit(ep)) { |
| 653 | list_del(pos); |
| 654 | ms_info.mi_nr_zombies--; |
| 655 | va_gen_destroy(&ep->rma_info.va_gen); |
| 656 | kfree(ep); |
| 657 | } |
| 658 | } |
| 659 | spin_unlock_irqrestore(&ms_info.mi_eplock, sflags); |
| 660 | } |
| 661 | |
| 662 | #define SCIF_WAKE_UP_SEND (1 << 1) |
| 663 | #define SCIF_WAKE_UP_RECV (1 << 2) |
| 664 | |
| 665 | /** |
| 666 | * scif_wakeup_ep() - Wake up all clients based on the type |
| 667 | * requested i.e. threads blocked in scif_send(..) and/or scif_recv(..). |
| 668 | */ |
| 669 | static inline void |
| 670 | scif_wakeup_ep(int type) |
| 671 | { |
| 672 | struct endpt *ep; |
| 673 | unsigned long sflags; |
| 674 | struct list_head *pos, *tmpq; |
| 675 | |
| 676 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 677 | list_for_each_safe(pos, tmpq, &ms_info.mi_connected) { |
| 678 | ep = list_entry(pos, struct endpt, list); |
| 679 | if (type & SCIF_WAKE_UP_SEND) |
| 680 | wake_up_interruptible(&ep->sendwq); |
| 681 | if (type & SCIF_WAKE_UP_RECV) |
| 682 | wake_up_interruptible(&ep->recvwq); |
| 683 | } |
| 684 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 685 | } |
| 686 | |
| 687 | /* |
| 688 | * is_self_scifdev: |
| 689 | * @dev: The remote SCIF Device |
| 690 | * |
| 691 | * Returns true if the SCIF Device passed is the self aka Loopback SCIF device. |
| 692 | */ |
| 693 | static inline int is_self_scifdev(struct micscif_dev *dev) |
| 694 | { |
| 695 | return dev->sd_node == ms_info.mi_nodeid; |
| 696 | } |
| 697 | |
| 698 | /* |
| 699 | * is_p2p_scifdev: |
| 700 | * @dev: The remote SCIF Device |
| 701 | * |
| 702 | * Returns true if the SCIF Device is a MIC Peer to Peer SCIF device. |
| 703 | */ |
| 704 | static inline bool is_p2p_scifdev(struct micscif_dev *dev) |
| 705 | { |
| 706 | #ifdef _MIC_SCIF_ |
| 707 | return dev != &scif_dev[SCIF_HOST_NODE] && !is_self_scifdev(dev); |
| 708 | #else |
| 709 | return false; |
| 710 | #endif |
| 711 | } |
| 712 | |
| 713 | /* |
| 714 | * get_conn_count: |
| 715 | * @dev: The remote SCIF Device |
| 716 | * |
| 717 | * Increments the number of active SCIF connections. Callee is expected |
| 718 | * to synchronize calling this API with put_conn_count. |
| 719 | */ |
| 720 | static __always_inline void |
| 721 | get_conn_count(struct micscif_dev *dev) |
| 722 | { |
| 723 | dev->num_active_conn++; |
| 724 | } |
| 725 | |
| 726 | /* |
| 727 | * put_conn_count: |
| 728 | * @dev: The remote SCIF Device |
| 729 | * |
| 730 | * Decrements the number of active connections. Callee is expected |
| 731 | * to synchronize calling this API with get_conn_count. |
| 732 | */ |
| 733 | static __always_inline void |
| 734 | put_conn_count(struct micscif_dev *dev) |
| 735 | { |
| 736 | dev->num_active_conn--; |
| 737 | BUG_ON(dev->num_active_conn < 0); |
| 738 | } |
| 739 | |
| 740 | /* |
| 741 | * get_kref_count: |
| 742 | * epd: SCIF endpoint |
| 743 | * |
| 744 | * Increments kmod endpoint reference count. Callee is expected |
| 745 | * to synchronize calling this API with put_kref_count. |
| 746 | */ |
| 747 | static __always_inline void |
| 748 | get_kref_count(scif_epd_t epd) |
| 749 | { |
| 750 | kref_get(&(epd->ref_count)); |
| 751 | } |
| 752 | |
| 753 | /* |
| 754 | * put_kref_count: |
| 755 | * epd: SCIF endpoint |
| 756 | * |
| 757 | * Decrements kmod endpoint reference count. Callee is expected |
| 758 | * to synchronize calling this API with get_kref_count. |
| 759 | */ |
| 760 | static __always_inline void |
| 761 | put_kref_count(scif_epd_t epd) |
| 762 | { |
| 763 | kref_put(&(epd->ref_count), scif_ref_rel); |
| 764 | } |
| 765 | |
| 766 | /* |
| 767 | * is_scifdev_alive: |
| 768 | * @dev: The remote SCIF Device |
| 769 | * |
| 770 | * Returns true if the remote SCIF Device is running or sleeping for |
| 771 | * this endpoint. |
| 772 | */ |
| 773 | static inline int scifdev_alive(struct endpt *ep) |
| 774 | { |
| 775 | return (((SCIFDEV_RUNNING == ep->remote_dev->sd_state) || |
| 776 | (SCIFDEV_SLEEPING == ep->remote_dev->sd_state)) && |
| 777 | SCIFDEV_RUNNING == ep->sd_state); |
| 778 | } |
| 779 | |
| 780 | /* |
| 781 | * verify_epd: |
| 782 | * ep: SCIF endpoint |
| 783 | * |
| 784 | * Checks several generic error conditions and returns the |
| 785 | * appropiate error. |
| 786 | */ |
| 787 | static inline int verify_epd(struct endpt *ep) |
| 788 | { |
| 789 | if (ep->state == SCIFEP_DISCONNECTED) |
| 790 | return -ECONNRESET; |
| 791 | |
| 792 | if (ep->state != SCIFEP_CONNECTED) |
| 793 | return -ENOTCONN; |
| 794 | |
| 795 | if (!scifdev_alive(ep)) |
| 796 | return -ENODEV; |
| 797 | |
| 798 | return 0; |
| 799 | } |
| 800 | |
| 801 | /** |
| 802 | * scif_invalidate_ep() - Set remote SCIF device state for all connected |
| 803 | * and disconnected endpoints for a particular node to SCIFDEV_STOPPED, |
| 804 | * change endpoint state to disconnected and wake up all send/recv/con |
| 805 | * waitqueues. |
| 806 | */ |
| 807 | static inline void |
| 808 | scif_invalidate_ep(int node) |
| 809 | { |
| 810 | struct endpt *ep; |
| 811 | unsigned long sflags; |
| 812 | struct list_head *pos, *tmpq; |
| 813 | |
| 814 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 815 | list_for_each_safe(pos, tmpq, &ms_info.mi_disconnected) { |
| 816 | ep = list_entry(pos, struct endpt, list); |
| 817 | if (ep->remote_dev->sd_node == node) { |
| 818 | spin_lock(&ep->lock); |
| 819 | ep->sd_state = SCIFDEV_STOPPED; |
| 820 | spin_unlock(&ep->lock); |
| 821 | } |
| 822 | } |
| 823 | list_for_each_safe(pos, tmpq, &ms_info.mi_connected) { |
| 824 | ep = list_entry(pos, struct endpt, list); |
| 825 | if (ep->remote_dev->sd_node == node) { |
| 826 | list_del(pos); |
| 827 | put_conn_count(ep->remote_dev); |
| 828 | spin_lock(&ep->lock); |
| 829 | ep->state = SCIFEP_DISCONNECTED; |
| 830 | list_add_tail(&ep->list, &ms_info.mi_disconnected); |
| 831 | ep->sd_state = SCIFDEV_STOPPED; |
| 832 | wake_up_interruptible(&ep->sendwq); |
| 833 | wake_up_interruptible(&ep->recvwq); |
| 834 | wake_up_interruptible(&ep->conwq); |
| 835 | spin_unlock(&ep->lock); |
| 836 | } |
| 837 | } |
| 838 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 839 | flush_workqueue(ms_info.mi_conn_wq); |
| 840 | } |
| 841 | |
| 842 | /* |
| 843 | * Only Debug Functions Below |
| 844 | */ |
| 845 | #define SCIF_CRUMB pr_debug("%s %d\n", __func__, __LINE__) |
| 846 | |
| 847 | static inline void |
| 848 | micscif_display_all_zombie_ep(void) |
| 849 | { |
| 850 | struct list_head *pos, *tmpq; |
| 851 | unsigned long sflags; |
| 852 | struct endpt *ep; |
| 853 | |
| 854 | pr_debug("Zombie Info Start\n"); |
| 855 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); |
| 856 | list_for_each_safe(pos, tmpq, &ms_info.mi_zombie) { |
| 857 | ep = list_entry(pos, struct endpt, list); |
| 858 | if (!list_empty(&ep->rma_info.reg_list)) |
| 859 | micscif_display_all_windows(&ep->rma_info.reg_list); |
| 860 | if (!list_empty(&ep->rma_info.remote_reg_list)) |
| 861 | micscif_display_all_windows( |
| 862 | &ep->rma_info.remote_reg_list); |
| 863 | } |
| 864 | spin_unlock_irqrestore(&ms_info.mi_eplock, sflags); |
| 865 | pr_debug("Zombie Info End\n"); |
| 866 | } |
| 867 | |
| 868 | static inline void dump_ep(scif_epd_t epd, const char *func, int line) |
| 869 | { |
| 870 | struct endpt *ep = (struct endpt *)epd; |
| 871 | pr_debug("%s %d state %d lock %p port.node 0x%x" |
| 872 | "port.port 0x%x peer.node 0x%x peer.port 0x%x backlog %d qp %p" |
| 873 | "qp_offset 0x%llx cnct_gnt_payload 0x%llx remote_dev %p\n", |
| 874 | func, line, ep->state, &ep->lock, ep->port.node, |
| 875 | ep->port.port, ep->peer.node, ep->peer.port, ep->backlog, |
| 876 | ep->qp_info.qp, ep->qp_info.qp_offset, |
| 877 | ep->qp_info.cnct_gnt_payload, ep->remote_dev); |
| 878 | } |
| 879 | |
| 880 | static inline void dump_qp(volatile struct micscif_qp *qp, const char *func, int line) |
| 881 | { |
| 882 | pr_debug("%s %d qp %p local_buf 0x%llx" |
| 883 | " local_qp 0x%llx remote_buf 0x%llx remote_qp %p ep 0x%llx\n", |
| 884 | func, line, qp, qp->local_buf, |
| 885 | qp->local_qp, qp->remote_buf, qp->remote_qp, qp->ep); |
| 886 | } |
| 887 | |
| 888 | static inline void dump_rb(struct micscif_rb *rb, const char *func, int line) |
| 889 | { |
| 890 | pr_debug("%s %d rb %p rb_base %p *read_ptr 0x%x" |
| 891 | " *write_ptr 0x%x size 0x%x" |
| 892 | " cro 0x%x cwo 0x%x ocro 0x%x ocwo 0x%x\n", |
| 893 | func, line, rb, rb->rb_base, *rb->read_ptr, |
| 894 | *rb->write_ptr, rb->size, rb->current_read_offset, |
| 895 | rb->current_write_offset, |
| 896 | rb->old_current_read_offset, |
| 897 | rb->old_current_write_offset); |
| 898 | } |
| 899 | |
| 900 | #endif /* MICSCIF_H */ |