| 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 <linux/poll.h> |
| 37 | #include <linux/time.h> |
| 38 | #include <linux/ktime.h> |
| 39 | #include <linux/sched.h> |
| 40 | #include <linux/kref.h> |
| 41 | #include <linux/module.h> |
| 42 | #include "scif.h" |
| 43 | #include "mic/micscif.h" |
| 44 | #ifndef _MIC_SCIF_ |
| 45 | #include "mic_common.h" |
| 46 | #endif |
| 47 | #include "mic/micscif_map.h" |
| 48 | |
| 49 | #define SCIF_MAP_ULIMIT 0x40 |
| 50 | |
| 51 | bool mic_ulimit_check = 0; |
| 52 | |
| 53 | char *scif_ep_states[] = { |
| 54 | "Closed", |
| 55 | "Unbound", |
| 56 | "Bound", |
| 57 | "Listening", |
| 58 | "Connected", |
| 59 | "Connecting", |
| 60 | "Mapping", |
| 61 | "Closing", |
| 62 | "Close Listening", |
| 63 | "Disconnected", |
| 64 | "Zombie"}; |
| 65 | |
| 66 | enum conn_async_state { |
| 67 | ASYNC_CONN_IDLE = 1, /* ep setup for async connect */ |
| 68 | ASYNC_CONN_INPROGRESS, /* async connect in progress */ |
| 69 | ASYNC_CONN_FLUSH_WORK /* async work flush in progress */ |
| 70 | }; |
| 71 | |
| 72 | /** |
| 73 | * scif_open() - Create a SCIF end point |
| 74 | * |
| 75 | * Create a SCIF end point and set the state to UNBOUND. This function |
| 76 | * returns the address of the end point data structure. |
| 77 | */ |
| 78 | scif_epd_t |
| 79 | __scif_open(void) |
| 80 | { |
| 81 | struct endpt *ep; |
| 82 | |
| 83 | might_sleep(); |
| 84 | if ((ep = (struct endpt *)kzalloc(sizeof(struct endpt), GFP_KERNEL)) == NULL) { |
| 85 | printk(KERN_ERR "SCIFAPI open: kzalloc fail on scif end point descriptor\n"); |
| 86 | goto err_ep_alloc; |
| 87 | } |
| 88 | |
| 89 | if ((ep->qp_info.qp = (struct micscif_qp *) |
| 90 | kzalloc(sizeof(struct micscif_qp), GFP_KERNEL)) == NULL) { |
| 91 | printk(KERN_ERR "SCIFAPI open: kzalloc fail on scif end point queue pointer\n"); |
| 92 | goto err_qp_alloc; |
| 93 | } |
| 94 | |
| 95 | spin_lock_init(&ep->lock); |
| 96 | mutex_init (&ep->sendlock); |
| 97 | mutex_init (&ep->recvlock); |
| 98 | |
| 99 | if (micscif_rma_ep_init(ep) < 0) { |
| 100 | printk(KERN_ERR "SCIFAPI _open: RMA EP Init failed\n"); |
| 101 | goto err_rma_init; |
| 102 | } |
| 103 | |
| 104 | ep->state = SCIFEP_UNBOUND; |
| 105 | pr_debug("SCIFAPI open: ep %p success\n", ep); |
| 106 | return (scif_epd_t)ep; |
| 107 | |
| 108 | err_rma_init: |
| 109 | kfree(ep->qp_info.qp); |
| 110 | err_qp_alloc: |
| 111 | kfree(ep); |
| 112 | err_ep_alloc: |
| 113 | return NULL; |
| 114 | } |
| 115 | |
| 116 | scif_epd_t |
| 117 | scif_open(void) |
| 118 | { |
| 119 | struct endpt *ep; |
| 120 | ep = (struct endpt *)__scif_open(); |
| 121 | if (ep) |
| 122 | kref_init(&(ep->ref_count)); |
| 123 | return (scif_epd_t)ep; |
| 124 | } |
| 125 | EXPORT_SYMBOL(scif_open); |
| 126 | |
| 127 | /** |
| 128 | * scif_close() - Terminate a SCIF end point |
| 129 | * @epd: The end point address returned from scif_open() |
| 130 | * |
| 131 | * The function terminates a scif connection. It must ensure all traffic on |
| 132 | * the connection is finished before removing it. |
| 133 | * |
| 134 | * On Connection with memory mapped this become more difficult. Once normal |
| 135 | * DMA and message traffic has ended the end point must be placed in a zombie |
| 136 | * state and wait for the other side to also release it's memory references. |
| 137 | */ |
| 138 | int |
| 139 | __scif_close(scif_epd_t epd) |
| 140 | { |
| 141 | struct endpt *ep = (struct endpt *)epd; |
| 142 | struct endpt *tmpep; |
| 143 | struct list_head *pos, *tmpq; |
| 144 | unsigned long sflags; |
| 145 | enum endptstate oldstate; |
| 146 | int err; |
| 147 | bool flush_conn; |
| 148 | |
| 149 | pr_debug("SCIFAPI close: ep %p %s\n", ep, scif_ep_states[ep->state]); |
| 150 | |
| 151 | might_sleep(); |
| 152 | |
| 153 | spin_lock(&ep->lock); |
| 154 | flush_conn = (ep->conn_async_state == ASYNC_CONN_INPROGRESS); |
| 155 | spin_unlock(&ep->lock); |
| 156 | |
| 157 | if (flush_conn) |
| 158 | flush_workqueue(ms_info.mi_conn_wq); |
| 159 | |
| 160 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 161 | |
| 162 | spin_lock_irqsave(&ep->lock, sflags); |
| 163 | oldstate = ep->state; |
| 164 | |
| 165 | ep->state = SCIFEP_CLOSING; |
| 166 | |
| 167 | switch (oldstate) { |
| 168 | case SCIFEP_ZOMBIE: |
| 169 | BUG_ON(SCIFEP_ZOMBIE == oldstate); |
| 170 | case SCIFEP_CLOSED: |
| 171 | case SCIFEP_DISCONNECTED: |
| 172 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 173 | micscif_unregister_all_windows(epd); |
| 174 | // Remove from the disconnected list |
| 175 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 176 | list_for_each_safe(pos, tmpq, &ms_info.mi_disconnected) { |
| 177 | tmpep = list_entry(pos, struct endpt, list); |
| 178 | if (tmpep == ep) { |
| 179 | list_del(pos); |
| 180 | break; |
| 181 | } |
| 182 | } |
| 183 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 184 | break; |
| 185 | case SCIFEP_UNBOUND: |
| 186 | case SCIFEP_BOUND: |
| 187 | case SCIFEP_CONNECTING: |
| 188 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 189 | break; |
| 190 | case SCIFEP_MAPPING: |
| 191 | case SCIFEP_CONNECTED: |
| 192 | case SCIFEP_CLOSING: |
| 193 | { |
| 194 | struct nodemsg msg; |
| 195 | struct endpt *fep = NULL; |
| 196 | struct endpt *tmpep; |
| 197 | unsigned long ts = jiffies; |
| 198 | struct list_head *pos, *tmpq; |
| 199 | |
| 200 | // Very short time before mapping completes and state becomes connected |
| 201 | // and does a standard teardown. |
| 202 | ts = jiffies; |
| 203 | while (ep->state == SCIFEP_MAPPING) { |
| 204 | cpu_relax(); |
| 205 | if (time_after((unsigned long)jiffies,ts + NODE_ALIVE_TIMEOUT)) { |
| 206 | printk(KERN_ERR "%s %d ep->state %d\n", __func__, __LINE__, ep->state); |
| 207 | ep->state = SCIFEP_BOUND; |
| 208 | break; |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | init_waitqueue_head(&ep->disconwq); // Wait for connection queue |
| 213 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 214 | |
| 215 | micscif_unregister_all_windows(epd); |
| 216 | |
| 217 | // Remove from the connected list |
| 218 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 219 | list_for_each_safe(pos, tmpq, &ms_info.mi_connected) { |
| 220 | tmpep = list_entry(pos, struct endpt, list); |
| 221 | if (tmpep == ep) { |
| 222 | list_del(pos); |
| 223 | put_conn_count(ep->remote_dev); |
| 224 | fep = tmpep; |
| 225 | spin_lock(&ep->lock); |
| 226 | break; |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | if (fep == NULL) { |
| 231 | // The other side has completed the disconnect before |
| 232 | // the end point can be removed from the list. Therefore |
| 233 | // the ep lock is not locked, traverse the disconnected list |
| 234 | // to find the endpoint, release the conn lock and |
| 235 | // proceed to teardown the end point below. |
| 236 | list_for_each_safe(pos, tmpq, &ms_info.mi_disconnected) { |
| 237 | tmpep = list_entry(pos, struct endpt, list); |
| 238 | if (tmpep == ep) { |
| 239 | list_del(pos); |
| 240 | break; |
| 241 | } |
| 242 | } |
| 243 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 244 | break; |
| 245 | } |
| 246 | |
| 247 | spin_unlock(&ms_info.mi_connlock); |
| 248 | |
| 249 | // Now we are free to close out the connection |
| 250 | msg.uop = SCIF_DISCNCT; |
| 251 | msg.src = ep->port; |
| 252 | msg.dst = ep->peer; |
| 253 | msg.payload[0] = (uint64_t)ep; |
| 254 | msg.payload[1] = ep->remote_ep; |
| 255 | |
| 256 | err = micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 257 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 258 | |
| 259 | if (!err) |
| 260 | /* Now wait for the remote node to respond */ |
| 261 | wait_event_timeout(ep->disconwq, |
| 262 | (ep->state == SCIFEP_DISCONNECTED), NODE_ALIVE_TIMEOUT); |
| 263 | /* |
| 264 | * Grab and release the ep lock to synchronize with the |
| 265 | * thread waking us up. If we dont grab this lock, then |
| 266 | * the ep might be freed before the wakeup completes |
| 267 | * resulting in potential memory corruption. |
| 268 | */ |
| 269 | spin_lock_irqsave(&ep->lock, sflags); |
| 270 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 271 | break; |
| 272 | } |
| 273 | case SCIFEP_LISTENING: |
| 274 | case SCIFEP_CLLISTEN: |
| 275 | { |
| 276 | struct conreq *conreq; |
| 277 | struct nodemsg msg; |
| 278 | struct endpt *aep; |
| 279 | |
| 280 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 281 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); |
| 282 | |
| 283 | // remove from listen list |
| 284 | list_for_each_safe(pos, tmpq, &ms_info.mi_listen) { |
| 285 | tmpep = list_entry(pos, struct endpt, list); |
| 286 | if (tmpep == ep) { |
| 287 | list_del(pos); |
| 288 | } |
| 289 | } |
| 290 | // Remove any dangling accepts |
| 291 | while (ep->acceptcnt) { |
| 292 | aep = list_first_entry(&ep->li_accept, struct endpt, liacceptlist); |
| 293 | BUG_ON(!aep); |
| 294 | list_del(&aep->liacceptlist); |
| 295 | if (aep->port.port && !aep->accepted_ep) |
| 296 | put_scif_port(aep->port.port); |
| 297 | list_for_each_safe(pos, tmpq, &ms_info.mi_uaccept) { |
| 298 | tmpep = list_entry(pos, struct endpt, miacceptlist); |
| 299 | if (tmpep == aep) { |
| 300 | list_del(pos); |
| 301 | break; |
| 302 | } |
| 303 | } |
| 304 | spin_unlock_irqrestore(&ms_info.mi_eplock, sflags); |
| 305 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 306 | list_for_each_safe(pos, tmpq, &ms_info.mi_connected) { |
| 307 | tmpep = list_entry(pos, struct endpt, list); |
| 308 | if (tmpep == aep) { |
| 309 | list_del(pos); |
| 310 | put_conn_count(aep->remote_dev); |
| 311 | break; |
| 312 | } |
| 313 | } |
| 314 | list_for_each_safe(pos, tmpq, &ms_info.mi_disconnected) { |
| 315 | tmpep = list_entry(pos, struct endpt, list); |
| 316 | if (tmpep == aep) { |
| 317 | list_del(pos); |
| 318 | break; |
| 319 | } |
| 320 | } |
| 321 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 322 | micscif_teardown_ep(aep); |
| 323 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); |
| 324 | micscif_add_epd_to_zombie_list(aep, MI_EPLOCK_HELD); |
| 325 | ep->acceptcnt--; |
| 326 | } |
| 327 | |
| 328 | spin_lock(&ep->lock); |
| 329 | spin_unlock(&ms_info.mi_eplock); |
| 330 | |
| 331 | // Remove and reject any pending connection requests. |
| 332 | while (ep->conreqcnt) { |
| 333 | conreq = list_first_entry(&ep->conlist, struct conreq, list); |
| 334 | list_del(&conreq->list); |
| 335 | |
| 336 | msg.uop = SCIF_CNCT_REJ; |
| 337 | msg.dst.node = conreq->msg.src.node; |
| 338 | msg.dst.port = conreq->msg.src.port; |
| 339 | msg.payload[0] = conreq->msg.payload[0]; |
| 340 | msg.payload[1] = conreq->msg.payload[1]; |
| 341 | /* |
| 342 | * No Error Handling on purpose for micscif_nodeqp_send(). |
| 343 | * If the remote node is lost we still want free the connection |
| 344 | * requests on the self node. |
| 345 | */ |
| 346 | micscif_nodeqp_send(&scif_dev[conreq->msg.src.node], &msg, ep); |
| 347 | |
| 348 | ep->conreqcnt--; |
| 349 | kfree(conreq); |
| 350 | } |
| 351 | |
| 352 | // If a kSCIF accept is waiting wake it up |
| 353 | wake_up_interruptible(&ep->conwq); |
| 354 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 355 | break; |
| 356 | } |
| 357 | } |
| 358 | if (ep->port.port && !ep->accepted_ep) |
| 359 | put_scif_port(ep->port.port); |
| 360 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 361 | micscif_teardown_ep(ep); |
| 362 | micscif_add_epd_to_zombie_list(ep, !MI_EPLOCK_HELD); |
| 363 | return 0; |
| 364 | } |
| 365 | |
| 366 | void |
| 367 | scif_ref_rel(struct kref *kref_count) |
| 368 | { |
| 369 | struct endpt *epd; |
| 370 | epd = container_of(kref_count, struct endpt, ref_count); |
| 371 | __scif_close((scif_epd_t)epd); |
| 372 | } |
| 373 | |
| 374 | int |
| 375 | scif_close(scif_epd_t epd) |
| 376 | { |
| 377 | __scif_flush(epd); |
| 378 | put_kref_count(epd); |
| 379 | return 0; |
| 380 | } |
| 381 | EXPORT_SYMBOL(scif_close); |
| 382 | |
| 383 | /** |
| 384 | * scif_flush() - Flush the endpoint |
| 385 | * @epd: The end point address returned from scif_open() |
| 386 | * |
| 387 | */ |
| 388 | int |
| 389 | __scif_flush(scif_epd_t epd) |
| 390 | { |
| 391 | struct endpt *ep = (struct endpt *)epd; |
| 392 | struct endpt *tmpep; |
| 393 | struct list_head *pos, *tmpq; |
| 394 | unsigned long sflags; |
| 395 | int err; |
| 396 | |
| 397 | might_sleep(); |
| 398 | |
| 399 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 400 | |
| 401 | spin_lock_irqsave(&ep->lock, sflags); |
| 402 | |
| 403 | switch (ep->state) { |
| 404 | case SCIFEP_CONNECTED: |
| 405 | { |
| 406 | struct nodemsg msg; |
| 407 | struct endpt *fep = NULL; |
| 408 | |
| 409 | init_waitqueue_head(&ep->disconwq); // Wait for connection queue |
| 410 | WARN_ON(ep->files); // files should never be set while connected |
| 411 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 412 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 413 | |
| 414 | list_for_each_safe(pos, tmpq, &ms_info.mi_connected) { |
| 415 | tmpep = list_entry(pos, struct endpt, list); |
| 416 | if (tmpep == ep) { |
| 417 | list_del(pos); |
| 418 | put_conn_count(ep->remote_dev); |
| 419 | fep = tmpep; |
| 420 | spin_lock(&ep->lock); |
| 421 | break; |
| 422 | } |
| 423 | } |
| 424 | |
| 425 | if (fep == NULL) { |
| 426 | // The other side has completed the disconnect before |
| 427 | // the end point can be removed from the list. Therefore |
| 428 | // the ep lock is not locked, traverse the disconnected list |
| 429 | // to find the endpoint, release the conn lock. |
| 430 | list_for_each_safe(pos, tmpq, &ms_info.mi_disconnected) { |
| 431 | tmpep = list_entry(pos, struct endpt, list); |
| 432 | if (tmpep == ep) { |
| 433 | list_del(pos); |
| 434 | break; |
| 435 | } |
| 436 | } |
| 437 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 438 | break; |
| 439 | } |
| 440 | |
| 441 | spin_unlock(&ms_info.mi_connlock); |
| 442 | |
| 443 | msg.uop = SCIF_DISCNCT; |
| 444 | msg.src = ep->port; |
| 445 | msg.dst = ep->peer; |
| 446 | msg.payload[0] = (uint64_t)ep; |
| 447 | msg.payload[1] = ep->remote_ep; |
| 448 | |
| 449 | err = micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 450 | |
| 451 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 452 | if (!err) |
| 453 | /* Now wait for the remote node to respond */ |
| 454 | wait_event_timeout(ep->disconwq, |
| 455 | (ep->state == SCIFEP_DISCONNECTED), NODE_ALIVE_TIMEOUT); |
| 456 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 457 | spin_lock(&ep->lock); |
| 458 | list_add_tail(&ep->list, &ms_info.mi_disconnected); |
| 459 | ep->state = SCIFEP_DISCONNECTED; |
| 460 | spin_unlock(&ep->lock); |
| 461 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 462 | // Wake up threads blocked in send and recv |
| 463 | wake_up_interruptible(&ep->sendwq); |
| 464 | wake_up_interruptible(&ep->recvwq); |
| 465 | break; |
| 466 | } |
| 467 | case SCIFEP_LISTENING: |
| 468 | { |
| 469 | ep->state = SCIFEP_CLLISTEN; |
| 470 | |
| 471 | // If an accept is waiting wake it up |
| 472 | wake_up_interruptible(&ep->conwq); |
| 473 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 474 | break; |
| 475 | } |
| 476 | default: |
| 477 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 478 | break; |
| 479 | } |
| 480 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 481 | return 0; |
| 482 | } |
| 483 | |
| 484 | /** |
| 485 | * scif_bind() - Bind a SCIF end point to a port ID. |
| 486 | * @epd: The end point address returned from scif_open() |
| 487 | * @pn: Port ID (number) to bind to |
| 488 | * |
| 489 | * Set the port ID associated with the end point and place it in the bound state. |
| 490 | * If a port ID of zero is requested a non zero port ID is allocated for it. |
| 491 | * |
| 492 | * Upon successful compltion the port id (number) will be returned. |
| 493 | * |
| 494 | * If the end point is not in the unbound state then return -EISCONN. |
| 495 | * |
| 496 | * If port ID zero is specified and allocation of a port ID fails -ENOSPC |
| 497 | * will be returned. |
| 498 | */ |
| 499 | int |
| 500 | __scif_bind(scif_epd_t epd, uint16_t pn) |
| 501 | { |
| 502 | struct endpt *ep = (struct endpt *)epd; |
| 503 | unsigned long sflags; |
| 504 | int ret = 0; |
| 505 | int tmp; |
| 506 | |
| 507 | pr_debug("SCIFAPI bind: ep %p %s requested port number %d\n", |
| 508 | ep, scif_ep_states[ep->state], pn); |
| 509 | |
| 510 | might_sleep(); |
| 511 | |
| 512 | if (pn) { |
| 513 | /* |
| 514 | * Modeled on http://www.ietf.org/rfc/rfc1700.txt?number=1700 |
| 515 | * SCIF ports below SCIF_ADMIN_PORT_END can only be bound by |
| 516 | * system (or root) processes or by processes executed by |
| 517 | * privileged users. |
| 518 | */ |
| 519 | if ( pn < SCIF_ADMIN_PORT_END && !capable(CAP_SYS_ADMIN)) { |
| 520 | ret = -EACCES; |
| 521 | goto scif_bind_admin_exit; |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | spin_lock_irqsave(&ep->lock, sflags); |
| 526 | if (ep->state == SCIFEP_BOUND) { |
| 527 | ret = -EINVAL; |
| 528 | goto scif_bind_exit; |
| 529 | } else if (ep->state != SCIFEP_UNBOUND) { |
| 530 | ret = -EISCONN; |
| 531 | goto scif_bind_exit; |
| 532 | } |
| 533 | |
| 534 | if (pn) { |
| 535 | if ((tmp = rsrv_scif_port(pn)) != pn) { |
| 536 | ret = -EINVAL; |
| 537 | goto scif_bind_exit; |
| 538 | } |
| 539 | } else { |
| 540 | pn = get_scif_port(); |
| 541 | if (!pn) { |
| 542 | ret = -ENOSPC; |
| 543 | goto scif_bind_exit; |
| 544 | } |
| 545 | } |
| 546 | |
| 547 | ep->state = SCIFEP_BOUND; |
| 548 | ep->port.node = ms_info.mi_nodeid; |
| 549 | ep->port.port = pn; |
| 550 | ep->conn_async_state = ASYNC_CONN_IDLE; |
| 551 | ret = pn; |
| 552 | pr_debug("SCIFAPI bind: bound to port number %d\n", pn); |
| 553 | |
| 554 | scif_bind_exit: |
| 555 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 556 | scif_bind_admin_exit: |
| 557 | return ret; |
| 558 | } |
| 559 | |
| 560 | int |
| 561 | scif_bind(scif_epd_t epd, uint16_t pn) |
| 562 | { |
| 563 | int ret; |
| 564 | get_kref_count(epd); |
| 565 | ret = __scif_bind(epd, pn); |
| 566 | put_kref_count(epd); |
| 567 | return ret; |
| 568 | } |
| 569 | EXPORT_SYMBOL(scif_bind); |
| 570 | |
| 571 | /** |
| 572 | * scif_listen() - Place the end point in the listening state |
| 573 | * @epd: The end point address returned from scif_open() |
| 574 | * @backlog: Maximum number of pending connection requests. |
| 575 | * |
| 576 | * The end point is placed in the listening state ready to accept connection |
| 577 | * requests. The backlog paramter is saved to indicate the maximun number of |
| 578 | * connection requests from the remote node to save. The end point is |
| 579 | * placed on a list of listening end points to allow a connection request to |
| 580 | * find it. |
| 581 | * |
| 582 | * Upon successful completion a zero is returned. |
| 583 | * |
| 584 | * If the end point is not in the bound state -EINVAL or -EISCONN is returned. |
| 585 | * |
| 586 | */ |
| 587 | int |
| 588 | __scif_listen(scif_epd_t epd, int backlog) |
| 589 | { |
| 590 | struct endpt *ep = (struct endpt *)epd; |
| 591 | unsigned long sflags; |
| 592 | |
| 593 | pr_debug("SCIFAPI listen: ep %p %s\n", ep, scif_ep_states[ep->state]); |
| 594 | |
| 595 | might_sleep(); |
| 596 | spin_lock_irqsave(&ep->lock, sflags); |
| 597 | switch (ep->state) { |
| 598 | case SCIFEP_ZOMBIE: |
| 599 | BUG_ON(SCIFEP_ZOMBIE == ep->state); |
| 600 | case SCIFEP_CLOSED: |
| 601 | case SCIFEP_CLOSING: |
| 602 | case SCIFEP_CLLISTEN: |
| 603 | case SCIFEP_UNBOUND: |
| 604 | case SCIFEP_DISCONNECTED: |
| 605 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 606 | return -EINVAL; |
| 607 | case SCIFEP_LISTENING: |
| 608 | case SCIFEP_CONNECTED: |
| 609 | case SCIFEP_CONNECTING: |
| 610 | case SCIFEP_MAPPING: |
| 611 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 612 | return -EISCONN; |
| 613 | case SCIFEP_BOUND: |
| 614 | break; |
| 615 | } |
| 616 | |
| 617 | ep->state = SCIFEP_LISTENING; |
| 618 | ep->backlog = backlog; |
| 619 | |
| 620 | ep->conreqcnt = 0; |
| 621 | ep->acceptcnt = 0; |
| 622 | INIT_LIST_HEAD(&ep->conlist); // List of connection requests |
| 623 | init_waitqueue_head(&ep->conwq); // Wait for connection queue |
| 624 | INIT_LIST_HEAD(&ep->li_accept); // User ep list for ACCEPTREG calls |
| 625 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 626 | |
| 627 | // Listen status is complete so delete the qp information not needed |
| 628 | // on a listen before placing on the list of listening ep's |
| 629 | micscif_teardown_ep((void *)ep); |
| 630 | ep->qp_info.qp = NULL; |
| 631 | |
| 632 | spin_lock_irqsave(&ms_info.mi_eplock, sflags); |
| 633 | list_add_tail(&ep->list, &ms_info.mi_listen); |
| 634 | spin_unlock_irqrestore(&ms_info.mi_eplock, sflags); |
| 635 | return 0; |
| 636 | } |
| 637 | |
| 638 | int |
| 639 | scif_listen(scif_epd_t epd, int backlog) |
| 640 | { |
| 641 | int ret; |
| 642 | get_kref_count(epd); |
| 643 | ret = __scif_listen(epd, backlog); |
| 644 | put_kref_count(epd); |
| 645 | return ret; |
| 646 | } |
| 647 | EXPORT_SYMBOL(scif_listen); |
| 648 | |
| 649 | #ifdef _MIC_SCIF_ |
| 650 | /* |
| 651 | * scif_p2p_connect: |
| 652 | * @node: destination node id |
| 653 | * |
| 654 | * Try to setup a p2p connection between the current |
| 655 | * node and the desitination node. We need host to |
| 656 | * setup the initial p2p connections. So we send |
| 657 | * this message to the host which acts like proxy |
| 658 | * in setting up p2p connection. |
| 659 | */ |
| 660 | static int scif_p2p_connect(int node) |
| 661 | { |
| 662 | struct micscif_dev *remote_dev = &scif_dev[node]; |
| 663 | struct nodemsg msg; |
| 664 | int err; |
| 665 | |
| 666 | pr_debug("%s:%d SCIF_NODE_CONNECT to host\n", __func__, __LINE__); |
| 667 | micscif_inc_node_refcnt(&scif_dev[SCIF_HOST_NODE], 1); |
| 668 | |
| 669 | msg.dst.node = SCIF_HOST_NODE; |
| 670 | msg.payload[0] = node; |
| 671 | msg.uop = SCIF_NODE_CONNECT; |
| 672 | |
| 673 | if ((err = micscif_nodeqp_send(&scif_dev[SCIF_HOST_NODE], |
| 674 | &msg, NULL))) { |
| 675 | printk(KERN_ERR "%s:%d error while sending SCIF_NODE_CONNECT to" |
| 676 | " node %d\n", __func__, __LINE__, node); |
| 677 | micscif_dec_node_refcnt(&scif_dev[SCIF_HOST_NODE], 1); |
| 678 | goto error; |
| 679 | } |
| 680 | |
| 681 | wait_event_interruptible_timeout(remote_dev->sd_p2p_wq, |
| 682 | (remote_dev->sd_state == SCIFDEV_RUNNING) || |
| 683 | (remote_dev->sd_state == SCIFDEV_NOTPRESENT), NODE_ALIVE_TIMEOUT); |
| 684 | |
| 685 | pr_debug("%s:%d SCIF_NODE_CONNECT state:%d\n", __func__, __LINE__, |
| 686 | remote_dev->sd_state); |
| 687 | micscif_dec_node_refcnt(&scif_dev[SCIF_HOST_NODE], 1); |
| 688 | error: |
| 689 | return err; |
| 690 | } |
| 691 | #endif |
| 692 | |
| 693 | static int scif_conn_func(struct endpt *ep) |
| 694 | { |
| 695 | int err = 0; |
| 696 | struct nodemsg msg; |
| 697 | unsigned long sflags; |
| 698 | int term_sent = 0; |
| 699 | |
| 700 | if ((err = micscif_reserve_dma_chan(ep))) { |
| 701 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 702 | ep->state = SCIFEP_BOUND; |
| 703 | goto connect_error_simple; |
| 704 | } |
| 705 | // Initiate the first part of the endpoint QP setup |
| 706 | err = micscif_setup_qp_connect(ep->qp_info.qp, &ep->qp_info.qp_offset, |
| 707 | ENDPT_QP_SIZE, ep->remote_dev); |
| 708 | if (err) { |
| 709 | printk(KERN_ERR "%s err %d qp_offset 0x%llx\n", |
| 710 | __func__, err, ep->qp_info.qp_offset); |
| 711 | ep->state = SCIFEP_BOUND; |
| 712 | goto connect_error_simple; |
| 713 | } |
| 714 | |
| 715 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 716 | |
| 717 | // Format connect message and send it |
| 718 | msg.src = ep->port; |
| 719 | msg.dst = ep->conn_port; |
| 720 | msg.uop = SCIF_CNCT_REQ; |
| 721 | msg.payload[0] = (uint64_t)ep; |
| 722 | msg.payload[1] = ep->qp_info.qp_offset; |
| 723 | if ((err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) { |
| 724 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 725 | goto connect_error_simple; |
| 726 | } |
| 727 | // Wait for request to be processed. |
| 728 | while ((err = wait_event_interruptible_timeout(ep->conwq, |
| 729 | (ep->state != SCIFEP_CONNECTING), NODE_ALIVE_TIMEOUT)) <= 0) { |
| 730 | if (!err) |
| 731 | err = -ENODEV; |
| 732 | |
| 733 | pr_debug("SCIFAPI connect: ep %p ^C detected\n", ep); |
| 734 | // interrupted out of the wait |
| 735 | if (!term_sent++) { |
| 736 | int bak_err = err; |
| 737 | msg.uop = SCIF_CNCT_TERM; |
| 738 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) { |
| 739 | retry: |
| 740 | err = wait_event_timeout(ep->diswq, |
| 741 | (ep->state != SCIFEP_CONNECTING), NODE_ALIVE_TIMEOUT); |
| 742 | if (!err && scifdev_alive(ep)) |
| 743 | goto retry; |
| 744 | if (!err) |
| 745 | err = -ENODEV; |
| 746 | if (err > 0) |
| 747 | err = 0; |
| 748 | } |
| 749 | if (ep->state == SCIFEP_MAPPING) { |
| 750 | micscif_setup_qp_connect_response(ep->remote_dev, |
| 751 | ep->qp_info.qp, ep->qp_info.cnct_gnt_payload); |
| 752 | // Send grant nack |
| 753 | msg.uop = SCIF_CNCT_GNTNACK; |
| 754 | msg.payload[0] = ep->remote_ep; |
| 755 | /* No error handling for Notification messages */ |
| 756 | micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 757 | } |
| 758 | // Ensure after that even after a timeout the state of the end point is bound |
| 759 | ep->state = SCIFEP_BOUND; |
| 760 | if (bak_err) |
| 761 | err = bak_err; |
| 762 | break; |
| 763 | } |
| 764 | } |
| 765 | |
| 766 | if (err > 0) |
| 767 | err = 0; |
| 768 | |
| 769 | if (term_sent || err) { |
| 770 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 771 | goto connect_error_simple; |
| 772 | } |
| 773 | |
| 774 | if (ep->state == SCIFEP_MAPPING) { |
| 775 | err = micscif_setup_qp_connect_response(ep->remote_dev, |
| 776 | ep->qp_info.qp, ep->qp_info.cnct_gnt_payload); |
| 777 | |
| 778 | // If the resource to map the queue are not available then we need |
| 779 | // to tell the other side to terminate the accept |
| 780 | if (err) { |
| 781 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 782 | |
| 783 | // Send grant nack |
| 784 | msg.uop = SCIF_CNCT_GNTNACK; |
| 785 | msg.payload[0] = ep->remote_ep; |
| 786 | /* No error handling for Notification messages */ |
| 787 | micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 788 | |
| 789 | ep->state = SCIFEP_BOUND; |
| 790 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 791 | goto connect_error_simple; |
| 792 | } |
| 793 | |
| 794 | // Send a grant ack to inform the accept we are done mapping its resources. |
| 795 | msg.uop = SCIF_CNCT_GNTACK; |
| 796 | msg.payload[0] = ep->remote_ep; |
| 797 | if (!(err = micscif_nodeqp_send(ep->remote_dev, &msg, ep))) { |
| 798 | ep->state = SCIFEP_CONNECTED; |
| 799 | spin_lock_irqsave(&ms_info.mi_connlock, sflags); |
| 800 | list_add_tail(&ep->list, &ms_info.mi_connected); |
| 801 | get_conn_count(ep->remote_dev); |
| 802 | spin_unlock_irqrestore(&ms_info.mi_connlock, sflags); |
| 803 | pr_debug("SCIFAPI connect: ep %p connected\n", ep); |
| 804 | } else |
| 805 | ep->state = SCIFEP_BOUND; |
| 806 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 807 | goto connect_error_simple; |
| 808 | |
| 809 | } else if (ep->state == SCIFEP_BOUND) { |
| 810 | pr_debug("SCIFAPI connect: ep %p connection refused\n", ep); |
| 811 | err = -ECONNREFUSED; |
| 812 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 813 | goto connect_error_simple; |
| 814 | |
| 815 | } else { |
| 816 | pr_debug("SCIFAPI connect: ep %p connection interrupted\n", ep); |
| 817 | err = -EINTR; |
| 818 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 819 | goto connect_error_simple; |
| 820 | } |
| 821 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 822 | connect_error_simple: |
| 823 | return err; |
| 824 | } |
| 825 | |
| 826 | /* |
| 827 | * micscif_conn_handler: |
| 828 | * |
| 829 | * Workqueue handler for servicing non-blocking SCIF connect |
| 830 | * |
| 831 | */ |
| 832 | void micscif_conn_handler(struct work_struct *work) |
| 833 | { |
| 834 | struct endpt *ep; |
| 835 | |
| 836 | do { |
| 837 | ep = NULL; |
| 838 | spin_lock(&ms_info.mi_nb_connect_lock); |
| 839 | if (!list_empty(&ms_info.mi_nb_connect_list)) { |
| 840 | ep = list_first_entry(&ms_info.mi_nb_connect_list, |
| 841 | struct endpt, conn_list); |
| 842 | list_del(&ep->conn_list); |
| 843 | } |
| 844 | spin_unlock(&ms_info.mi_nb_connect_lock); |
| 845 | if (ep) { |
| 846 | ep->conn_err = scif_conn_func(ep); |
| 847 | wake_up_interruptible(&ep->conn_pend_wq); |
| 848 | } |
| 849 | } while (ep); |
| 850 | } |
| 851 | |
| 852 | /** |
| 853 | * scif_connect() - Request a connection to a remote node |
| 854 | * @epd: The end point address returned from scif_open() |
| 855 | * @dst: Remote note address informtion |
| 856 | * |
| 857 | * The function requests a scif connection to the remote node |
| 858 | * identified by the dst parameter. "dst" contains the remote node and |
| 859 | * port ids. |
| 860 | * |
| 861 | * Upon successful complete a zero will be returned. |
| 862 | * |
| 863 | * If the end point is not in the bound state -EINVAL will be returned. |
| 864 | * |
| 865 | * If during the connection sequence resource allocation fails the -ENOMEM |
| 866 | * will be returned. |
| 867 | * |
| 868 | * If the remote side is not responding to connection requests the caller may |
| 869 | * terminate this funciton with a signal. If so a -EINTR will be returned. |
| 870 | */ |
| 871 | int |
| 872 | __scif_connect(scif_epd_t epd, struct scif_portID *dst, bool non_block) |
| 873 | { |
| 874 | struct endpt *ep = (struct endpt *)epd; |
| 875 | unsigned long sflags; |
| 876 | int err = 0; |
| 877 | #ifdef _MIC_SCIF_ |
| 878 | struct micscif_dev *remote_dev; |
| 879 | #endif |
| 880 | |
| 881 | pr_debug("SCIFAPI connect: ep %p %s\n", ep, |
| 882 | scif_ep_states[ep->state]); |
| 883 | |
| 884 | if (dst->node > MAX_BOARD_SUPPORTED) |
| 885 | return -ENODEV; |
| 886 | |
| 887 | might_sleep(); |
| 888 | |
| 889 | #ifdef _MIC_SCIF_ |
| 890 | remote_dev = &scif_dev[dst->node]; |
| 891 | if ((SCIFDEV_INIT == remote_dev->sd_state || |
| 892 | SCIFDEV_STOPPED == remote_dev->sd_state) && mic_p2p_enable) |
| 893 | if ((err = scif_p2p_connect(dst->node))) |
| 894 | return err; |
| 895 | #endif |
| 896 | |
| 897 | if (SCIFDEV_RUNNING != scif_dev[dst->node].sd_state && |
| 898 | SCIFDEV_SLEEPING != scif_dev[dst->node].sd_state) |
| 899 | return -ENODEV; |
| 900 | |
| 901 | spin_lock_irqsave(&ep->lock, sflags); |
| 902 | switch (ep->state) { |
| 903 | case SCIFEP_ZOMBIE: |
| 904 | BUG_ON(SCIFEP_ZOMBIE == ep->state); |
| 905 | |
| 906 | case SCIFEP_CLOSED: |
| 907 | case SCIFEP_CLOSING: |
| 908 | err = -EINVAL; |
| 909 | break; |
| 910 | |
| 911 | case SCIFEP_DISCONNECTED: |
| 912 | if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) |
| 913 | ep->conn_async_state = ASYNC_CONN_FLUSH_WORK; |
| 914 | else |
| 915 | err = -EINVAL; |
| 916 | break; |
| 917 | |
| 918 | case SCIFEP_LISTENING: |
| 919 | case SCIFEP_CLLISTEN: |
| 920 | err = -EOPNOTSUPP; |
| 921 | break; |
| 922 | |
| 923 | case SCIFEP_CONNECTING: |
| 924 | case SCIFEP_MAPPING: |
| 925 | if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) |
| 926 | err = -EINPROGRESS; |
| 927 | else |
| 928 | err = -EISCONN; |
| 929 | break; |
| 930 | |
| 931 | case SCIFEP_CONNECTED: |
| 932 | if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) |
| 933 | ep->conn_async_state = ASYNC_CONN_FLUSH_WORK; |
| 934 | else |
| 935 | err = -EISCONN; |
| 936 | break; |
| 937 | |
| 938 | case SCIFEP_UNBOUND: |
| 939 | if ((ep->port.port = get_scif_port()) == 0) |
| 940 | err = -ENOSPC; |
| 941 | else { |
| 942 | ep->port.node = ms_info.mi_nodeid; |
| 943 | ep->conn_async_state = ASYNC_CONN_IDLE; |
| 944 | } |
| 945 | /* Fall through */ |
| 946 | case SCIFEP_BOUND: |
| 947 | /* |
| 948 | * If a non-blocking connect has been already initiated (conn_async_state |
| 949 | * is either ASYNC_CONN_INPROGRESS or ASYNC_CONN_FLUSH_WORK), the end point |
| 950 | * could end up in SCIF_BOUND due an error in the connection |
| 951 | * process (e.g., connnection refused) |
| 952 | * If conn_async_state is ASYNC_CONN_INPROGRESS - transition to |
| 953 | * ASYNC_CONN_FLUSH_WORK so that the error status can be collected. |
| 954 | * If the state is already ASYNC_CONN_FLUSH_WORK - then set the error |
| 955 | * to EINPROGRESS since some other thread is waiting to collect error status. |
| 956 | */ |
| 957 | if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) |
| 958 | ep->conn_async_state = ASYNC_CONN_FLUSH_WORK; |
| 959 | else if (ep->conn_async_state == ASYNC_CONN_FLUSH_WORK) |
| 960 | err = -EINPROGRESS; |
| 961 | else { |
| 962 | ep->conn_port = *dst; |
| 963 | init_waitqueue_head(&ep->sendwq); |
| 964 | init_waitqueue_head(&ep->recvwq); |
| 965 | init_waitqueue_head(&ep->conwq); |
| 966 | init_waitqueue_head(&ep->diswq); |
| 967 | ep->conn_async_state = 0; |
| 968 | |
| 969 | if (unlikely(non_block)) |
| 970 | ep->conn_async_state = ASYNC_CONN_INPROGRESS; |
| 971 | } |
| 972 | break; |
| 973 | } |
| 974 | |
| 975 | if (err || ep->conn_async_state == ASYNC_CONN_FLUSH_WORK) |
| 976 | goto connect_simple_unlock1; |
| 977 | |
| 978 | ep->state = SCIFEP_CONNECTING; |
| 979 | ep->remote_dev = &scif_dev[dst->node]; |
| 980 | ep->sd_state = SCIFDEV_RUNNING; |
| 981 | ep->qp_info.qp->magic = SCIFEP_MAGIC; |
| 982 | ep->qp_info.qp->ep = (uint64_t)ep; |
| 983 | if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) { |
| 984 | init_waitqueue_head(&ep->conn_pend_wq); |
| 985 | spin_lock(&ms_info.mi_nb_connect_lock); |
| 986 | list_add_tail(&ep->conn_list, |
| 987 | &ms_info.mi_nb_connect_list); |
| 988 | spin_unlock(&ms_info.mi_nb_connect_lock); |
| 989 | err = -EINPROGRESS; |
| 990 | queue_work(ms_info.mi_conn_wq, &ms_info.mi_conn_work); |
| 991 | } |
| 992 | connect_simple_unlock1: |
| 993 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 994 | |
| 995 | if (err) |
| 996 | return err; |
| 997 | else if (ep->conn_async_state == ASYNC_CONN_FLUSH_WORK) { |
| 998 | flush_workqueue(ms_info.mi_conn_wq); |
| 999 | err = ep->conn_err; |
| 1000 | spin_lock_irqsave(&ep->lock, sflags); |
| 1001 | ep->conn_async_state = ASYNC_CONN_IDLE; |
| 1002 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 1003 | } else { |
| 1004 | err = scif_conn_func(ep); |
| 1005 | } |
| 1006 | return err; |
| 1007 | } |
| 1008 | |
| 1009 | int |
| 1010 | scif_connect(scif_epd_t epd, struct scif_portID *dst) |
| 1011 | { |
| 1012 | int ret; |
| 1013 | get_kref_count(epd); |
| 1014 | ret = __scif_connect(epd, dst, false); |
| 1015 | put_kref_count(epd); |
| 1016 | return ret; |
| 1017 | } |
| 1018 | EXPORT_SYMBOL(scif_connect); |
| 1019 | |
| 1020 | /** |
| 1021 | * scif_accept() - Accept a connection request from the remote node |
| 1022 | * @epd: The end point address returned from scif_open() |
| 1023 | * @peer: Filled in with pear node and port information |
| 1024 | * @newepd: New end point created for connection |
| 1025 | * @flags: Indicates sychronous or asynchronous mode |
| 1026 | * |
| 1027 | * The function accepts a connection request from the remote node. Successful |
| 1028 | * complete is indicate by a new end point being created and passed back |
| 1029 | * to the caller for future reference. |
| 1030 | * |
| 1031 | * Upon successful complete a zero will be returned and the peer information |
| 1032 | * will be filled in. |
| 1033 | * |
| 1034 | * If the end point is not in the listening state -EINVAL will be returned. |
| 1035 | * |
| 1036 | * If during the connection sequence resource allocation fails the -ENOMEM |
| 1037 | * will be returned. |
| 1038 | * |
| 1039 | * If the function is called asynchronously and not connection request are |
| 1040 | * pending it will return -EAGAIN. |
| 1041 | * |
| 1042 | * If the remote side is not sending any connection requests the caller may |
| 1043 | * terminate this funciton with a signal. If so a -EINTR will be returned. |
| 1044 | */ |
| 1045 | int |
| 1046 | __scif_accept(scif_epd_t epd, struct scif_portID *peer, scif_epd_t *newepd, int flags) |
| 1047 | { |
| 1048 | struct endpt *lep = (struct endpt *)epd; |
| 1049 | struct endpt *cep; |
| 1050 | struct conreq *conreq; |
| 1051 | struct nodemsg msg; |
| 1052 | unsigned long sflags; |
| 1053 | int err; |
| 1054 | |
| 1055 | pr_debug("SCIFAPI accept: ep %p %s\n", lep, scif_ep_states[lep->state]); |
| 1056 | |
| 1057 | // Error if flags other than SCIF_ACCEPT_SYNC are set |
| 1058 | if (flags & ~SCIF_ACCEPT_SYNC) { |
| 1059 | pr_debug("SCIFAPI accept: ep %p invalid flags %x\n", lep, flags & ~SCIF_ACCEPT_SYNC); |
| 1060 | return -EINVAL; |
| 1061 | } |
| 1062 | |
| 1063 | if (!peer || !newepd) { |
| 1064 | pr_debug("SCIFAPI accept: ep %p peer %p or newepd %p NULL\n", |
| 1065 | lep, peer, newepd); |
| 1066 | return -EINVAL; |
| 1067 | } |
| 1068 | |
| 1069 | might_sleep(); |
| 1070 | spin_lock_irqsave(&lep->lock, sflags); |
| 1071 | if (lep->state != SCIFEP_LISTENING) { |
| 1072 | pr_debug("SCIFAPI accept: ep %p not listending\n", lep); |
| 1073 | spin_unlock_irqrestore(&lep->lock, sflags); |
| 1074 | return -EINVAL; |
| 1075 | } |
| 1076 | |
| 1077 | if (!lep->conreqcnt && !(flags & SCIF_ACCEPT_SYNC)) { |
| 1078 | // No connection request present and we do not want to wait |
| 1079 | pr_debug("SCIFAPI accept: ep %p async request with nothing pending\n", lep); |
| 1080 | spin_unlock_irqrestore(&lep->lock, sflags); |
| 1081 | return -EAGAIN; |
| 1082 | } |
| 1083 | |
| 1084 | retry_connection: |
| 1085 | spin_unlock_irqrestore(&lep->lock, sflags); |
| 1086 | lep->files = current ? current->files : NULL; |
| 1087 | if ((err = wait_event_interruptible(lep->conwq, |
| 1088 | (lep->conreqcnt || (lep->state != SCIFEP_LISTENING)))) != 0) { |
| 1089 | // wait was interrupted |
| 1090 | pr_debug("SCIFAPI accept: ep %p ^C detected\n", lep); |
| 1091 | return err; // -ERESTARTSYS |
| 1092 | } |
| 1093 | |
| 1094 | if (lep->state != SCIFEP_LISTENING) { |
| 1095 | return -EINTR; |
| 1096 | } |
| 1097 | |
| 1098 | spin_lock_irqsave(&lep->lock, sflags); |
| 1099 | |
| 1100 | if (!lep->conreqcnt) { |
| 1101 | goto retry_connection; |
| 1102 | } |
| 1103 | |
| 1104 | // Get the first connect request off the list |
| 1105 | conreq = list_first_entry(&lep->conlist, struct conreq, list); |
| 1106 | list_del(&conreq->list); |
| 1107 | lep->conreqcnt--; |
| 1108 | spin_unlock_irqrestore(&lep->lock, sflags); |
| 1109 | |
| 1110 | // Fill in the peer information |
| 1111 | peer->node = conreq->msg.src.node; |
| 1112 | peer->port = conreq->msg.src.port; |
| 1113 | |
| 1114 | // Create the connection endpoint |
| 1115 | cep = (struct endpt *)kzalloc(sizeof(struct endpt), GFP_KERNEL); |
| 1116 | if (!cep) { |
| 1117 | pr_debug("SCIFAPI accept: ep %p new end point allocation failed\n", lep); |
| 1118 | err = -ENOMEM; |
| 1119 | goto scif_accept_error_epalloc; |
| 1120 | } |
| 1121 | spin_lock_init(&cep->lock); |
| 1122 | mutex_init (&cep->sendlock); |
| 1123 | mutex_init (&cep->recvlock); |
| 1124 | cep->state = SCIFEP_CONNECTING; |
| 1125 | cep->remote_dev = &scif_dev[peer->node]; |
| 1126 | cep->remote_ep = conreq->msg.payload[0]; |
| 1127 | cep->sd_state = SCIFDEV_RUNNING; |
| 1128 | |
| 1129 | if (!scifdev_alive(cep)) { |
| 1130 | err = -ENODEV; |
| 1131 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 1132 | goto scif_accept_error_qpalloc; |
| 1133 | } |
| 1134 | |
| 1135 | if (micscif_rma_ep_init(cep) < 0) { |
| 1136 | pr_debug("SCIFAPI accept: ep %p new %p RMA EP init failed\n", lep, cep); |
| 1137 | err = -ENOMEM; |
| 1138 | goto scif_accept_error_qpalloc; |
| 1139 | } |
| 1140 | |
| 1141 | if ((err = micscif_reserve_dma_chan(cep))) { |
| 1142 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 1143 | goto scif_accept_error_qpalloc; |
| 1144 | } |
| 1145 | |
| 1146 | cep->qp_info.qp = (struct micscif_qp *)kzalloc(sizeof(struct micscif_qp), GFP_KERNEL); |
| 1147 | if (!cep->qp_info.qp) { |
| 1148 | printk(KERN_ERR "Port Qp Allocation Failed\n"); |
| 1149 | err = -ENOMEM; |
| 1150 | goto scif_accept_error_qpalloc; |
| 1151 | } |
| 1152 | |
| 1153 | cep->qp_info.qp->magic = SCIFEP_MAGIC; |
| 1154 | cep->qp_info.qp->ep = (uint64_t)cep; |
| 1155 | micscif_inc_node_refcnt(cep->remote_dev, 1); |
| 1156 | err = micscif_setup_qp_accept(cep->qp_info.qp, &cep->qp_info.qp_offset, |
| 1157 | conreq->msg.payload[1], ENDPT_QP_SIZE, cep->remote_dev); |
| 1158 | if (err) { |
| 1159 | pr_debug("SCIFAPI accept: ep %p new %p micscif_setup_qp_accept %d qp_offset 0x%llx\n", |
| 1160 | lep, cep, err, cep->qp_info.qp_offset); |
| 1161 | micscif_dec_node_refcnt(cep->remote_dev, 1); |
| 1162 | goto scif_accept_error_map; |
| 1163 | } |
| 1164 | |
| 1165 | cep->port.node = lep->port.node; |
| 1166 | cep->port.port = lep->port.port; |
| 1167 | cep->peer.node = peer->node; |
| 1168 | cep->peer.port = peer->port; |
| 1169 | cep->accepted_ep = true; |
| 1170 | init_waitqueue_head(&cep->sendwq); // Wait for data to be consumed |
| 1171 | init_waitqueue_head(&cep->recvwq); // Wait for data to be produced |
| 1172 | init_waitqueue_head(&cep->conwq); // Wait for connection request |
| 1173 | |
| 1174 | // Return the grant message |
| 1175 | msg.uop = SCIF_CNCT_GNT; |
| 1176 | msg.src = cep->port; |
| 1177 | msg.payload[0] = cep->remote_ep; |
| 1178 | msg.payload[1] = cep->qp_info.qp_offset; |
| 1179 | msg.payload[2] = (uint64_t)cep; |
| 1180 | |
| 1181 | err = micscif_nodeqp_send(cep->remote_dev, &msg, cep); |
| 1182 | |
| 1183 | micscif_dec_node_refcnt(cep->remote_dev, 1); |
| 1184 | if (err) |
| 1185 | goto scif_accept_error_map; |
| 1186 | retry: |
| 1187 | err = wait_event_timeout(cep->conwq, |
| 1188 | (cep->state != SCIFEP_CONNECTING), NODE_ACCEPT_TIMEOUT); |
| 1189 | if (!err && scifdev_alive(cep)) |
| 1190 | goto retry; |
| 1191 | |
| 1192 | if (!err) { |
| 1193 | err = -ENODEV; |
| 1194 | goto scif_accept_error_map; |
| 1195 | } |
| 1196 | |
| 1197 | if (err > 0) |
| 1198 | err = 0; |
| 1199 | |
| 1200 | kfree(conreq); |
| 1201 | |
| 1202 | spin_lock_irqsave(&cep->lock, sflags); |
| 1203 | |
| 1204 | if (cep->state == SCIFEP_CONNECTED) { |
| 1205 | // Connect sequence complete return new endpoint information |
| 1206 | *newepd = (scif_epd_t)cep; |
| 1207 | spin_unlock_irqrestore(&cep->lock, sflags); |
| 1208 | pr_debug("SCIFAPI accept: ep %p new %p returning new epnd point\n", lep, cep); |
| 1209 | return 0; |
| 1210 | } |
| 1211 | |
| 1212 | if (cep->state == SCIFEP_CLOSING) { |
| 1213 | // Remote failed to allocate resources and NAKed the grant. |
| 1214 | // There is at this point nothing referencing the new end point. |
| 1215 | spin_unlock_irqrestore(&cep->lock, sflags); |
| 1216 | micscif_teardown_ep((void *)cep); |
| 1217 | kfree(cep); |
| 1218 | |
| 1219 | // If call with sync flag then go back and wait. |
| 1220 | if (flags & SCIF_ACCEPT_SYNC) { |
| 1221 | spin_lock_irqsave(&lep->lock, sflags); |
| 1222 | goto retry_connection; |
| 1223 | } |
| 1224 | |
| 1225 | pr_debug("SCIFAPI accept: ep %p new %p remote failed to allocate resources\n", lep, cep); |
| 1226 | return -EAGAIN; |
| 1227 | } |
| 1228 | |
| 1229 | // While connect was in progress the other side closed and sent a disconnect |
| 1230 | // so set the end point status to closed but return anyway. This will allow |
| 1231 | // the caller to drain anything the other side may have put in the message queue. |
| 1232 | *newepd = (scif_epd_t)cep; |
| 1233 | spin_unlock_irqrestore(&cep->lock, sflags); |
| 1234 | return 0; |
| 1235 | |
| 1236 | // Error allocating or mapping resources |
| 1237 | scif_accept_error_map: |
| 1238 | kfree(cep->qp_info.qp); |
| 1239 | |
| 1240 | scif_accept_error_qpalloc: |
| 1241 | kfree(cep); |
| 1242 | |
| 1243 | scif_accept_error_epalloc: |
| 1244 | micscif_inc_node_refcnt(&scif_dev[conreq->msg.src.node], 1); |
| 1245 | // New reject the connection request due to lack of resources |
| 1246 | msg.uop = SCIF_CNCT_REJ; |
| 1247 | msg.dst.node = conreq->msg.src.node; |
| 1248 | msg.dst.port = conreq->msg.src.port; |
| 1249 | msg.payload[0] = conreq->msg.payload[0]; |
| 1250 | msg.payload[1] = conreq->msg.payload[1]; |
| 1251 | /* No error handling for Notification messages */ |
| 1252 | micscif_nodeqp_send(&scif_dev[conreq->msg.src.node], &msg, NULL); |
| 1253 | micscif_dec_node_refcnt(&scif_dev[conreq->msg.src.node], 1); |
| 1254 | |
| 1255 | kfree(conreq); |
| 1256 | return err; |
| 1257 | } |
| 1258 | |
| 1259 | int |
| 1260 | scif_accept(scif_epd_t epd, struct scif_portID *peer, scif_epd_t *newepd, int flags) |
| 1261 | { |
| 1262 | int ret; |
| 1263 | get_kref_count(epd); |
| 1264 | ret = __scif_accept(epd, peer, newepd, flags); |
| 1265 | if (ret == 0) { |
| 1266 | kref_init(&((*newepd)->ref_count)); |
| 1267 | } |
| 1268 | put_kref_count(epd); |
| 1269 | return ret; |
| 1270 | } |
| 1271 | EXPORT_SYMBOL(scif_accept); |
| 1272 | |
| 1273 | /* |
| 1274 | * scif_msg_param_check: |
| 1275 | * @epd: The end point address returned from scif_open() |
| 1276 | * @len: Length to receive |
| 1277 | * @flags: Syncronous or asynchronous access |
| 1278 | * |
| 1279 | * Validate parameters for messaging APIs scif_send(..)/scif_recv(..). |
| 1280 | */ |
| 1281 | static inline int |
| 1282 | scif_msg_param_check(scif_epd_t epd, int len, int flags) |
| 1283 | { |
| 1284 | int ret = -EINVAL; |
| 1285 | |
| 1286 | if (len < 0) |
| 1287 | goto err_ret; |
| 1288 | |
| 1289 | if (flags && (!(flags & SCIF_RECV_BLOCK))) |
| 1290 | goto err_ret; |
| 1291 | |
| 1292 | ret = 0; |
| 1293 | |
| 1294 | err_ret: |
| 1295 | return ret; |
| 1296 | } |
| 1297 | |
| 1298 | #define SCIF_BLAST (1 << 1) /* Use bit 1 of flags field */ |
| 1299 | |
| 1300 | #ifdef SCIF_BLAST |
| 1301 | /* |
| 1302 | * Added a temporary implementation of the exception path. |
| 1303 | * The cost to the normal path is 1 local variable (set once and |
| 1304 | * tested once) plus 2 tests for the 'blast' flag. |
| 1305 | * This only apply to the card side kernel API. |
| 1306 | */ |
| 1307 | #ifndef _MIC_SCIF_ |
| 1308 | #undef SCIF_BLAST |
| 1309 | #endif |
| 1310 | #endif |
| 1311 | |
| 1312 | /** |
| 1313 | * _scif_send() - Send data to connection queue |
| 1314 | * @epd: The end point address returned from scif_open() |
| 1315 | * @msg: Address to place data |
| 1316 | * @len: Length to receive |
| 1317 | * @flags: Syncronous or asynchronous access |
| 1318 | * |
| 1319 | * This function sends a packet of data to the queue * created by the |
| 1320 | * connection establishment sequence. It returns when the packet has |
| 1321 | * been completely sent. |
| 1322 | * |
| 1323 | * Successful completion returns the number of bytes sent. |
| 1324 | * |
| 1325 | * If the end point is not in the connect state returns -ENOTCONN; |
| 1326 | * |
| 1327 | * This function may be interrupted by a signal and will return -EINTR. |
| 1328 | */ |
| 1329 | int |
| 1330 | _scif_send(scif_epd_t epd, void *msg, int len, int flags) |
| 1331 | { |
| 1332 | struct endpt *ep = (struct endpt *)epd; |
| 1333 | struct nodemsg notif_msg; |
| 1334 | unsigned long sflags; |
| 1335 | size_t curr_xfer_len = 0; |
| 1336 | size_t sent_len = 0; |
| 1337 | size_t write_count; |
| 1338 | int ret; |
| 1339 | #ifdef SCIF_BLAST |
| 1340 | int tl; |
| 1341 | #endif |
| 1342 | |
| 1343 | if (flags & SCIF_SEND_BLOCK) |
| 1344 | might_sleep(); |
| 1345 | |
| 1346 | #ifdef SCIF_BLAST |
| 1347 | if (flags & SCIF_BLAST) { |
| 1348 | /* |
| 1349 | * Do a decent try to acquire lock (~100 uSec) |
| 1350 | */ |
| 1351 | for (ret = tl = 0; ret < 100 && !tl; ret++) { |
| 1352 | tl = spin_trylock_irqsave(&ep->lock, sflags); |
| 1353 | cpu_relax(); |
| 1354 | } |
| 1355 | } else { |
| 1356 | tl = 1; |
| 1357 | spin_lock_irqsave(&ep->lock, sflags); |
| 1358 | } |
| 1359 | #else |
| 1360 | spin_lock_irqsave(&ep->lock, sflags); |
| 1361 | #endif |
| 1362 | |
| 1363 | while (sent_len != len) { |
| 1364 | if (ep->state == SCIFEP_DISCONNECTED) { |
| 1365 | ret = (int)(sent_len ? sent_len : -ECONNRESET); |
| 1366 | goto unlock_dec_return; |
| 1367 | } |
| 1368 | if (ep->state != SCIFEP_CONNECTED) { |
| 1369 | ret = (int)(sent_len ? sent_len : -ENOTCONN); |
| 1370 | goto unlock_dec_return; |
| 1371 | } |
| 1372 | if (!scifdev_alive(ep)) { |
| 1373 | ret = (int) (sent_len ? sent_len : -ENODEV); |
| 1374 | goto unlock_dec_return; |
| 1375 | } |
| 1376 | write_count = micscif_rb_space(&ep->qp_info.qp->outbound_q); |
| 1377 | if (write_count) { |
| 1378 | /* |
| 1379 | * Best effort to send as much data as there |
| 1380 | * is space in the RB particularly important for the |
| 1381 | * Non Blocking case. |
| 1382 | */ |
| 1383 | curr_xfer_len = min(len - sent_len, write_count); |
| 1384 | ret = micscif_rb_write(&ep->qp_info.qp->outbound_q, msg, |
| 1385 | (uint32_t)curr_xfer_len); |
| 1386 | if (ret < 0) { |
| 1387 | ret = -EFAULT; |
| 1388 | goto unlock_dec_return; |
| 1389 | } |
| 1390 | if (ret) { |
| 1391 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 1392 | /* |
| 1393 | * If there is space in the RB and we have the |
| 1394 | * EP lock held then writing to the RB should |
| 1395 | * succeed. Releasing spin lock before asserting |
| 1396 | * to avoid deadlocking the system. |
| 1397 | */ |
| 1398 | BUG_ON(ret); |
| 1399 | } |
| 1400 | /* |
| 1401 | * Success. Update write pointer. |
| 1402 | */ |
| 1403 | micscif_rb_commit(&ep->qp_info.qp->outbound_q); |
| 1404 | #ifdef SCIF_BLAST |
| 1405 | if (flags & SCIF_BLAST) { |
| 1406 | /* |
| 1407 | * Bypass-path; set flag int the host side node_qp |
| 1408 | * and ring the doorbell. Host will wake-up all |
| 1409 | * listeners, such that the message will be seen. |
| 1410 | * Need micscif_send_host_intr() to be non-static. |
| 1411 | */ |
| 1412 | extern int micscif_send_host_intr(struct micscif_dev *, uint32_t); |
| 1413 | ep->remote_dev->qpairs->remote_qp->blast = 1; |
| 1414 | smp_wmb(); /* Sufficient or need sfence? */ |
| 1415 | micscif_send_host_intr(ep->remote_dev, 0); |
| 1416 | } else { |
| 1417 | /* |
| 1418 | * Normal path: send notification on the |
| 1419 | * node_qp ring buffer and ring the doorbell. |
| 1420 | */ |
| 1421 | notif_msg.src = ep->port; |
| 1422 | notif_msg.uop = SCIF_CLIENT_SENT; |
| 1423 | notif_msg.payload[0] = ep->remote_ep; |
| 1424 | if ((ret = micscif_nodeqp_send(ep->remote_dev, ¬if_msg, ep))) { |
| 1425 | ret = sent_len ? sent_len : ret; |
| 1426 | goto unlock_dec_return; |
| 1427 | } |
| 1428 | } |
| 1429 | #else |
| 1430 | /* |
| 1431 | * Send a notification to the peer about the |
| 1432 | * produced data message. |
| 1433 | */ |
| 1434 | notif_msg.src = ep->port; |
| 1435 | notif_msg.uop = SCIF_CLIENT_SENT; |
| 1436 | notif_msg.payload[0] = ep->remote_ep; |
| 1437 | if ((ret = micscif_nodeqp_send(ep->remote_dev, ¬if_msg, ep))) { |
| 1438 | ret = (int)(sent_len ? sent_len : ret); |
| 1439 | goto unlock_dec_return; |
| 1440 | } |
| 1441 | #endif |
| 1442 | sent_len += curr_xfer_len; |
| 1443 | msg = (char *)msg + curr_xfer_len; |
| 1444 | continue; |
| 1445 | } |
| 1446 | curr_xfer_len = min(len - sent_len, (size_t)(ENDPT_QP_SIZE - 1)); |
| 1447 | /* |
| 1448 | * Not enough space in the RB. Return in the Non Blocking case. |
| 1449 | */ |
| 1450 | if (!(flags & SCIF_SEND_BLOCK)) { |
| 1451 | ret = (int)sent_len; |
| 1452 | goto unlock_dec_return; |
| 1453 | } |
| 1454 | #ifdef SCIF_BLAST |
| 1455 | /* |
| 1456 | * Flags SCIF_BLAST and SCIF_SEND_BLOCK are mutually |
| 1457 | * exclusive, so if we get here we know that SCIF_BLAST |
| 1458 | * was not set and thus we _do_ have the spinlock. |
| 1459 | * No need to check variable tl here |
| 1460 | */ |
| 1461 | #endif |
| 1462 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 1463 | /* |
| 1464 | * Wait for a message now in the Blocking case. |
| 1465 | */ |
| 1466 | if ((ret = wait_event_interruptible(ep->sendwq, |
| 1467 | (SCIFEP_CONNECTED != ep->state) || |
| 1468 | (micscif_rb_space(&ep->qp_info.qp->outbound_q) |
| 1469 | >= curr_xfer_len) || (!scifdev_alive(ep))))) { |
| 1470 | ret = (int) (sent_len ? sent_len : ret); |
| 1471 | goto dec_return; |
| 1472 | } |
| 1473 | spin_lock_irqsave(&ep->lock, sflags); |
| 1474 | } |
| 1475 | ret = len; |
| 1476 | unlock_dec_return: |
| 1477 | #ifdef SCIF_BLAST |
| 1478 | if (tl) |
| 1479 | #endif |
| 1480 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 1481 | dec_return: |
| 1482 | return ret; |
| 1483 | } |
| 1484 | |
| 1485 | /** |
| 1486 | * _scif_recv() - Recieve data from connection queue |
| 1487 | * @epd: The end point address returned from scif_open() |
| 1488 | * @msg: Address to place data |
| 1489 | * @len: Length to receive |
| 1490 | * @flags: Syncronous or asynchronous access |
| 1491 | * @touser: package send to user buffer or kernel |
| 1492 | * |
| 1493 | * This function requests to receive a packet of data from the queue |
| 1494 | * created by the connection establishment sequence. It reads the amount |
| 1495 | * of data requested before returning. |
| 1496 | * |
| 1497 | * This function differs from the scif_send() by also returning data if the |
| 1498 | * end point is in the disconnected state and data is present. |
| 1499 | * |
| 1500 | * Successful completion returns the number of bytes read. |
| 1501 | * |
| 1502 | * If the end point is not in the connect state or in the disconnected state |
| 1503 | * with data prosent it returns -ENOTCONN; |
| 1504 | * |
| 1505 | * This function may be interrupted by a signal and will return -EINTR. |
| 1506 | */ |
| 1507 | int |
| 1508 | _scif_recv(scif_epd_t epd, void *msg, int len, int flags) |
| 1509 | { |
| 1510 | int read_size; |
| 1511 | struct endpt *ep = (struct endpt *)epd; |
| 1512 | unsigned long sflags; |
| 1513 | struct nodemsg notif_msg; |
| 1514 | size_t curr_recv_len = 0; |
| 1515 | size_t remaining_len = len; |
| 1516 | size_t read_count; |
| 1517 | int ret; |
| 1518 | |
| 1519 | if (flags & SCIF_RECV_BLOCK) |
| 1520 | might_sleep(); |
| 1521 | |
| 1522 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 1523 | spin_lock_irqsave(&ep->lock, sflags); |
| 1524 | while (remaining_len) { |
| 1525 | if (ep->state != SCIFEP_CONNECTED && |
| 1526 | ep->state != SCIFEP_DISCONNECTED) { |
| 1527 | ret = (int) (len - remaining_len) ? |
| 1528 | (int) (len - remaining_len) : -ENOTCONN; |
| 1529 | goto unlock_dec_return; |
| 1530 | } |
| 1531 | read_count = micscif_rb_count(&ep->qp_info.qp->inbound_q, |
| 1532 | (int) remaining_len); |
| 1533 | if (read_count) { |
| 1534 | /* |
| 1535 | * Best effort to recv as much data as there |
| 1536 | * are bytes to read in the RB particularly |
| 1537 | * important for the Non Blocking case. |
| 1538 | */ |
| 1539 | curr_recv_len = min(remaining_len, read_count); |
| 1540 | read_size = micscif_rb_get_next( |
| 1541 | &ep->qp_info.qp->inbound_q, |
| 1542 | msg, (int) curr_recv_len); |
| 1543 | if (read_size < 0){ |
| 1544 | /* only could happen when copy to USER buffer |
| 1545 | */ |
| 1546 | ret = -EFAULT; |
| 1547 | goto unlock_dec_return; |
| 1548 | } |
| 1549 | if (read_size != curr_recv_len) { |
| 1550 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 1551 | /* |
| 1552 | * If there are bytes to be read from the RB and |
| 1553 | * we have the EP lock held then reading from |
| 1554 | * RB should succeed. Releasing spin lock before |
| 1555 | * asserting to avoid deadlocking the system. |
| 1556 | */ |
| 1557 | BUG_ON(read_size != curr_recv_len); |
| 1558 | } |
| 1559 | if (ep->state == SCIFEP_CONNECTED) { |
| 1560 | /* |
| 1561 | * Update the read pointer only if the endpoint is |
| 1562 | * still connected else the read pointer might no |
| 1563 | * longer exist since the peer has freed resources! |
| 1564 | */ |
| 1565 | micscif_rb_update_read_ptr(&ep->qp_info.qp->inbound_q); |
| 1566 | /* |
| 1567 | * Send a notification to the peer about the |
| 1568 | * consumed data message only if the EP is in |
| 1569 | * SCIFEP_CONNECTED state. |
| 1570 | */ |
| 1571 | notif_msg.src = ep->port; |
| 1572 | notif_msg.uop = SCIF_CLIENT_RCVD; |
| 1573 | notif_msg.payload[0] = ep->remote_ep; |
| 1574 | if ((ret = micscif_nodeqp_send(ep->remote_dev, ¬if_msg, ep))) { |
| 1575 | ret = (len - (int)remaining_len) ? |
| 1576 | (len - (int)remaining_len) : ret; |
| 1577 | goto unlock_dec_return; |
| 1578 | } |
| 1579 | } |
| 1580 | remaining_len -= curr_recv_len; |
| 1581 | msg = (char *)msg + curr_recv_len; |
| 1582 | continue; |
| 1583 | } |
| 1584 | curr_recv_len = min(remaining_len, (size_t)(ENDPT_QP_SIZE - 1)); |
| 1585 | /* |
| 1586 | * Bail out now if the EP is in SCIFEP_DISCONNECTED state else |
| 1587 | * we will keep looping forever. |
| 1588 | */ |
| 1589 | if (ep->state == SCIFEP_DISCONNECTED) { |
| 1590 | ret = (len - (int)remaining_len) ? |
| 1591 | (len - (int)remaining_len) : -ECONNRESET; |
| 1592 | goto unlock_dec_return; |
| 1593 | } |
| 1594 | /* |
| 1595 | * Return in the Non Blocking case if there is no data |
| 1596 | * to read in this iteration. |
| 1597 | */ |
| 1598 | if (!(flags & SCIF_RECV_BLOCK)) { |
| 1599 | ret = len - (int)remaining_len; |
| 1600 | goto unlock_dec_return; |
| 1601 | } |
| 1602 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 1603 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 1604 | /* |
| 1605 | * Wait for a message now in the Blocking case. |
| 1606 | * or until other side disconnects. |
| 1607 | */ |
| 1608 | if ((ret = wait_event_interruptible(ep->recvwq, |
| 1609 | (SCIFEP_CONNECTED != ep->state) || |
| 1610 | (micscif_rb_count(&ep->qp_info.qp->inbound_q, |
| 1611 | curr_recv_len) >= curr_recv_len) || (!scifdev_alive(ep))))) { |
| 1612 | ret = (len - remaining_len) ? |
| 1613 | (len - (int)remaining_len) : ret; |
| 1614 | goto dec_return; |
| 1615 | } |
| 1616 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 1617 | spin_lock_irqsave(&ep->lock, sflags); |
| 1618 | } |
| 1619 | ret = len; |
| 1620 | unlock_dec_return: |
| 1621 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 1622 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 1623 | dec_return: |
| 1624 | return ret; |
| 1625 | } |
| 1626 | |
| 1627 | |
| 1628 | /** |
| 1629 | * scif_user_send() - Send data to connection queue |
| 1630 | * @epd: The end point address returned from scif_open() |
| 1631 | * @msg: Address to place data |
| 1632 | * @len: Length to receive |
| 1633 | * @flags: Syncronous or asynchronous access |
| 1634 | * |
| 1635 | * This function is called from the driver IOCTL entry point |
| 1636 | * only and is a wrapper for _scif_send(). |
| 1637 | */ |
| 1638 | int |
| 1639 | scif_user_send(scif_epd_t epd, void *msg, int len, int flags) |
| 1640 | { |
| 1641 | struct endpt *ep = (struct endpt *)epd; |
| 1642 | int err = 0; |
| 1643 | int sent_len = 0; |
| 1644 | char *tmp; |
| 1645 | int loop_len; |
| 1646 | int chunk_len = min(len, (1 << (MAX_ORDER + PAGE_SHIFT - 1)));; |
| 1647 | pr_debug("SCIFAPI send (U): ep %p %s\n", ep, scif_ep_states[ep->state]); |
| 1648 | |
| 1649 | if (!len) |
| 1650 | return 0; |
| 1651 | |
| 1652 | if ((err = scif_msg_param_check(epd, len, flags))) |
| 1653 | goto send_err; |
| 1654 | |
| 1655 | if (!(tmp = kmalloc(chunk_len, GFP_KERNEL))) { |
| 1656 | err = -ENOMEM; |
| 1657 | goto send_err; |
| 1658 | } |
| 1659 | err = 0; |
| 1660 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 1661 | /* |
| 1662 | * Grabbing the lock before breaking up the transfer in |
| 1663 | * multiple chunks is required to ensure that messages do |
| 1664 | * not get fragmented and reordered. |
| 1665 | */ |
| 1666 | mutex_lock(&ep->sendlock); |
| 1667 | |
| 1668 | while (sent_len != len) { |
| 1669 | msg = (void *)((char *)msg + err); |
| 1670 | loop_len = len - sent_len; |
| 1671 | loop_len = min(chunk_len, loop_len); |
| 1672 | if (copy_from_user(tmp, msg, loop_len)) { |
| 1673 | err = -EFAULT; |
| 1674 | goto send_free_err; |
| 1675 | } |
| 1676 | err = _scif_send(epd, (void *)tmp, loop_len, flags); |
| 1677 | if (err < 0) { |
| 1678 | goto send_free_err; |
| 1679 | } |
| 1680 | sent_len += err; |
| 1681 | if (err !=loop_len) { |
| 1682 | goto send_free_err; |
| 1683 | } |
| 1684 | } |
| 1685 | send_free_err: |
| 1686 | mutex_unlock(&ep->sendlock); |
| 1687 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 1688 | kfree(tmp); |
| 1689 | send_err: |
| 1690 | return err < 0 ? err : sent_len; |
| 1691 | } |
| 1692 | |
| 1693 | /** |
| 1694 | * scif_user_recv() - Recieve data from connection queue |
| 1695 | * @epd: The end point address returned from scif_open() |
| 1696 | * @msg: Address to place data |
| 1697 | * @len: Length to receive |
| 1698 | * @flags: Syncronous or asynchronous access |
| 1699 | * |
| 1700 | * This function is called from the driver IOCTL entry point |
| 1701 | * only and is a wrapper for _scif_recv(). |
| 1702 | */ |
| 1703 | int |
| 1704 | scif_user_recv(scif_epd_t epd, void *msg, int len, int flags) |
| 1705 | { |
| 1706 | struct endpt *ep = (struct endpt *)epd; |
| 1707 | int err = 0; |
| 1708 | int recv_len = 0; |
| 1709 | char *tmp; |
| 1710 | int loop_len; |
| 1711 | int chunk_len = min(len, (1 << (MAX_ORDER + PAGE_SHIFT - 1)));; |
| 1712 | pr_debug("SCIFAPI recv (U): ep %p %s\n", ep, scif_ep_states[ep->state]); |
| 1713 | |
| 1714 | if (!len) |
| 1715 | return 0; |
| 1716 | |
| 1717 | if ((err = scif_msg_param_check(epd, len, flags))) |
| 1718 | goto recv_err; |
| 1719 | |
| 1720 | if (!(tmp = kmalloc(chunk_len, GFP_KERNEL))) { |
| 1721 | err = -ENOMEM; |
| 1722 | goto recv_err; |
| 1723 | } |
| 1724 | err = 0; |
| 1725 | /* |
| 1726 | * Grabbing the lock before breaking up the transfer in |
| 1727 | * multiple chunks is required to ensure that messages do |
| 1728 | * not get fragmented and reordered. |
| 1729 | */ |
| 1730 | mutex_lock(&ep->recvlock); |
| 1731 | |
| 1732 | while (recv_len != len) { |
| 1733 | msg = (void *)((char *)msg + err); |
| 1734 | loop_len = len - recv_len; |
| 1735 | loop_len = min(chunk_len, loop_len); |
| 1736 | if ((err = _scif_recv(epd, tmp, loop_len, flags)) < 0) |
| 1737 | goto recv_free_err; |
| 1738 | if (copy_to_user(msg, tmp, err)) { |
| 1739 | err = -EFAULT; |
| 1740 | goto recv_free_err; |
| 1741 | } |
| 1742 | recv_len += err; |
| 1743 | if (err !=loop_len) { |
| 1744 | goto recv_free_err; |
| 1745 | } |
| 1746 | } |
| 1747 | recv_free_err: |
| 1748 | mutex_unlock(&ep->recvlock); |
| 1749 | kfree(tmp); |
| 1750 | recv_err: |
| 1751 | return err < 0 ? err : recv_len; |
| 1752 | } |
| 1753 | |
| 1754 | #ifdef SCIF_BLAST |
| 1755 | /* |
| 1756 | * Added a temporary implementation of the exception path. |
| 1757 | * The cost to the normal path testing of 2 flag bits instead |
| 1758 | * of just one and a change to condition for node-wakeup. |
| 1759 | */ |
| 1760 | #endif |
| 1761 | |
| 1762 | /** |
| 1763 | * scif_send() - Send data to connection queue |
| 1764 | * @epd: The end point address returned from scif_open() |
| 1765 | * @msg: Address to place data |
| 1766 | * @len: Length to receive |
| 1767 | * @flags: Syncronous or asynchronous access |
| 1768 | * |
| 1769 | * This function is called from the kernel mode only and is |
| 1770 | * a wrapper for _scif_send(). |
| 1771 | */ |
| 1772 | int |
| 1773 | __scif_send(scif_epd_t epd, void *msg, int len, int flags) |
| 1774 | { |
| 1775 | struct endpt *ep = (struct endpt *)epd; |
| 1776 | int ret; |
| 1777 | |
| 1778 | pr_debug("SCIFAPI send (K): ep %p %s\n", ep, scif_ep_states[ep->state]); |
| 1779 | if (!len) |
| 1780 | return 0; |
| 1781 | |
| 1782 | #ifdef SCIF_BLAST |
| 1783 | /* |
| 1784 | * KAA: this is same code as scif_msg_param_check(), |
| 1785 | * but since that routine is shared with scif_recv |
| 1786 | * I thought is safer to replicate code here. |
| 1787 | */ |
| 1788 | if (len < 0) |
| 1789 | return -EINVAL; |
| 1790 | |
| 1791 | if (flags && !(flags & (SCIF_SEND_BLOCK | SCIF_BLAST))) |
| 1792 | return -EINVAL; |
| 1793 | |
| 1794 | if ((flags & (SCIF_SEND_BLOCK | SCIF_BLAST)) == |
| 1795 | (SCIF_SEND_BLOCK | SCIF_BLAST)) |
| 1796 | return -EINVAL; |
| 1797 | #else |
| 1798 | if ((ret = scif_msg_param_check(epd, len, flags))) |
| 1799 | return ret; |
| 1800 | #endif |
| 1801 | /* |
| 1802 | * Cannot block while waiting for node to wake up |
| 1803 | * if non blocking messaging mode is requested. Return |
| 1804 | * ENODEV if the remote node is idle. |
| 1805 | */ |
| 1806 | if (!(flags & SCIF_SEND_BLOCK) && ep->remote_dev && |
| 1807 | SCIF_NODE_IDLE == atomic_long_read( |
| 1808 | &ep->remote_dev->scif_ref_cnt)) |
| 1809 | return -ENODEV; |
| 1810 | |
| 1811 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 1812 | |
| 1813 | /* |
| 1814 | * Grab the mutex lock in the blocking case only |
| 1815 | * to ensure messages do not get fragmented/reordered. |
| 1816 | * The non blocking mode is protected using spin locks |
| 1817 | * in _scif_send(). |
| 1818 | */ |
| 1819 | if (flags & SCIF_SEND_BLOCK) |
| 1820 | mutex_lock(&ep->sendlock); |
| 1821 | |
| 1822 | ret = _scif_send(epd, msg, len, flags); |
| 1823 | |
| 1824 | if (flags & SCIF_SEND_BLOCK) |
| 1825 | mutex_unlock(&ep->sendlock); |
| 1826 | |
| 1827 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 1828 | return ret; |
| 1829 | } |
| 1830 | |
| 1831 | int |
| 1832 | scif_send(scif_epd_t epd, void *msg, int len, int flags) |
| 1833 | { |
| 1834 | int ret; |
| 1835 | get_kref_count(epd); |
| 1836 | ret = __scif_send(epd, msg, len, flags); |
| 1837 | put_kref_count(epd); |
| 1838 | return ret; |
| 1839 | } |
| 1840 | EXPORT_SYMBOL(scif_send); |
| 1841 | |
| 1842 | /** |
| 1843 | * scif_recv() - Recieve data from connection queue |
| 1844 | * @epd: The end point address returned from scif_open() |
| 1845 | * @msg: Address to place data |
| 1846 | * @len: Length to receive |
| 1847 | * @flags: Syncronous or asynchronous access |
| 1848 | * |
| 1849 | * This function is called from the kernel mode only and is |
| 1850 | * a wrapper for _scif_recv(). |
| 1851 | */ |
| 1852 | int |
| 1853 | __scif_recv(scif_epd_t epd, void *msg, int len, int flags) |
| 1854 | { |
| 1855 | struct endpt *ep = (struct endpt *)epd; |
| 1856 | int ret; |
| 1857 | |
| 1858 | pr_debug("SCIFAPI recv (K): ep %p %s\n", ep, scif_ep_states[ep->state]); |
| 1859 | |
| 1860 | if (!len) |
| 1861 | return 0; |
| 1862 | |
| 1863 | if ((ret = scif_msg_param_check(epd, len, flags))) |
| 1864 | return ret; |
| 1865 | |
| 1866 | /* |
| 1867 | * Cannot block while waiting for node to wake up |
| 1868 | * if non blocking messaging mode is requested. Return |
| 1869 | * ENODEV if the remote node is idle. |
| 1870 | */ |
| 1871 | if (!flags && ep->remote_dev && |
| 1872 | SCIF_NODE_IDLE == atomic_long_read( |
| 1873 | &ep->remote_dev->scif_ref_cnt)) |
| 1874 | return -ENODEV; |
| 1875 | |
| 1876 | /* |
| 1877 | * Grab the mutex lock in the blocking case only |
| 1878 | * to ensure messages do not get fragmented/reordered. |
| 1879 | * The non blocking mode is protected using spin locks |
| 1880 | * in _scif_send(). |
| 1881 | */ |
| 1882 | if (flags & SCIF_RECV_BLOCK) |
| 1883 | mutex_lock(&ep->recvlock); |
| 1884 | |
| 1885 | ret = _scif_recv(epd, msg, len, flags); |
| 1886 | |
| 1887 | if (flags & SCIF_RECV_BLOCK) |
| 1888 | mutex_unlock(&ep->recvlock); |
| 1889 | |
| 1890 | return ret; |
| 1891 | } |
| 1892 | |
| 1893 | int |
| 1894 | scif_recv(scif_epd_t epd, void *msg, int len, int flags) |
| 1895 | { |
| 1896 | int ret; |
| 1897 | get_kref_count(epd); |
| 1898 | ret = __scif_recv(epd, msg, len, flags); |
| 1899 | put_kref_count(epd); |
| 1900 | return ret; |
| 1901 | } |
| 1902 | EXPORT_SYMBOL(scif_recv); |
| 1903 | |
| 1904 | /** |
| 1905 | * __scif_pin_pages - __scif_pin_pages() pins the physical pages which back |
| 1906 | * the range of virtual address pages starting at addr and continuing for |
| 1907 | * len bytes. addr and len are constrained to be multiples of the page size. |
| 1908 | * A successful scif_register() call returns an opaque pointer value |
| 1909 | * which may be used in subsequent calls to scif_register_pinned_pages(). |
| 1910 | * |
| 1911 | * Return Values |
| 1912 | * Upon successful completion, __scif_pin_pages() returns a |
| 1913 | * scif_pinned_pages_t value else an apt error is returned as documented |
| 1914 | * in scif.h. Protections of the set of pinned pages are also returned by |
| 1915 | * reference via out_prot. |
| 1916 | */ |
| 1917 | int |
| 1918 | __scif_pin_pages(void *addr, size_t len, int *out_prot, |
| 1919 | int map_flags, scif_pinned_pages_t *pages) |
| 1920 | { |
| 1921 | struct scif_pinned_pages *pinned_pages; |
| 1922 | int nr_pages, err = 0, i; |
| 1923 | bool vmalloc_addr = false; |
| 1924 | bool try_upgrade = false; |
| 1925 | int prot = *out_prot; |
| 1926 | int ulimit = 0; |
| 1927 | struct mm_struct *mm = NULL; |
| 1928 | |
| 1929 | /* Unsupported flags */ |
| 1930 | if (map_flags & ~(SCIF_MAP_KERNEL | SCIF_MAP_ULIMIT)) |
| 1931 | return -EINVAL; |
| 1932 | ulimit = !!(map_flags & SCIF_MAP_ULIMIT); |
| 1933 | |
| 1934 | /* Unsupported protection requested */ |
| 1935 | if (prot & ~(SCIF_PROT_READ | SCIF_PROT_WRITE)) |
| 1936 | return -EINVAL; |
| 1937 | |
| 1938 | /* addr/len must be page aligned. len should be non zero */ |
| 1939 | if ((!len) || |
| 1940 | (align_low((uint64_t)addr, PAGE_SIZE) != (uint64_t)addr) || |
| 1941 | (align_low((uint64_t)len, PAGE_SIZE) != (uint64_t)len)) |
| 1942 | return -EINVAL; |
| 1943 | |
| 1944 | might_sleep(); |
| 1945 | |
| 1946 | nr_pages = (int)(len >> PAGE_SHIFT); |
| 1947 | |
| 1948 | /* Allocate a set of pinned pages */ |
| 1949 | if (!(pinned_pages = micscif_create_pinned_pages(nr_pages, prot))) |
| 1950 | return -ENOMEM; |
| 1951 | |
| 1952 | if (unlikely(map_flags & SCIF_MAP_KERNEL)) { |
| 1953 | if (is_vmalloc_addr(addr)) |
| 1954 | vmalloc_addr = true; |
| 1955 | |
| 1956 | for (i = 0; i < nr_pages; i++) { |
| 1957 | if (unlikely(vmalloc_addr)) |
| 1958 | pinned_pages->pages[i] = |
| 1959 | vmalloc_to_page((char *)addr + (i * PAGE_SIZE) ); |
| 1960 | else |
| 1961 | pinned_pages->pages[i] = |
| 1962 | virt_to_page((char *)addr + (i * PAGE_SIZE) ); |
| 1963 | pinned_pages->num_pages[i] = 1; |
| 1964 | pinned_pages->nr_contig_chunks++; |
| 1965 | } |
| 1966 | pinned_pages->nr_pages = nr_pages; |
| 1967 | pinned_pages->map_flags = SCIF_MAP_KERNEL; |
| 1968 | } else { |
| 1969 | if (prot == SCIF_PROT_READ) |
| 1970 | try_upgrade = true; |
| 1971 | prot |= SCIF_PROT_WRITE; |
| 1972 | retry: |
| 1973 | mm = current->mm; |
| 1974 | down_write(&mm->mmap_sem); |
| 1975 | if (ulimit) { |
| 1976 | err = __scif_check_inc_pinned_vm(mm, nr_pages); |
| 1977 | if (err) { |
| 1978 | up_write(&mm->mmap_sem); |
| 1979 | pinned_pages->nr_pages = 0; |
| 1980 | goto error_unmap; |
| 1981 | } |
| 1982 | } |
| 1983 | |
| 1984 | pinned_pages->nr_pages = get_user_pages( |
| 1985 | (uint64_t)addr, |
| 1986 | nr_pages, |
| 1987 | prot & SCIF_PROT_WRITE ? FOLL_WRITE : 0, |
| 1988 | pinned_pages->pages, |
| 1989 | pinned_pages->vma); |
| 1990 | up_write(&mm->mmap_sem); |
| 1991 | if (nr_pages == pinned_pages->nr_pages) { |
| 1992 | #ifdef RMA_DEBUG |
| 1993 | atomic_long_add_return(nr_pages, &ms_info.rma_pin_cnt); |
| 1994 | #endif |
| 1995 | micscif_detect_large_page(pinned_pages, addr); |
| 1996 | } else { |
| 1997 | if (try_upgrade) { |
| 1998 | if (ulimit) |
| 1999 | __scif_dec_pinned_vm_lock(mm, nr_pages, 0); |
| 2000 | #ifdef RMA_DEBUG |
| 2001 | WARN_ON(atomic_long_sub_return(1, |
| 2002 | &ms_info.rma_mm_cnt) < 0); |
| 2003 | #endif |
| 2004 | /* Roll back any pinned pages */ |
| 2005 | for (i = 0; i < pinned_pages->nr_pages; i++) { |
| 2006 | if (pinned_pages->pages[i]) |
| 2007 | put_page(pinned_pages->pages[i]); |
| 2008 | } |
| 2009 | prot &= ~SCIF_PROT_WRITE; |
| 2010 | try_upgrade = false; |
| 2011 | goto retry; |
| 2012 | } |
| 2013 | } |
| 2014 | pinned_pages->map_flags = 0; |
| 2015 | } |
| 2016 | |
| 2017 | if (pinned_pages->nr_pages < nr_pages) { |
| 2018 | err = -EFAULT; |
| 2019 | pinned_pages->nr_pages = nr_pages; |
| 2020 | goto dec_pinned; |
| 2021 | } |
| 2022 | |
| 2023 | *out_prot = prot; |
| 2024 | atomic_set(&pinned_pages->ref_count, nr_pages); |
| 2025 | *pages = pinned_pages; |
| 2026 | return err; |
| 2027 | dec_pinned: |
| 2028 | if (ulimit) |
| 2029 | __scif_dec_pinned_vm_lock(mm, nr_pages, 0); |
| 2030 | /* Something went wrong! Rollback */ |
| 2031 | error_unmap: |
| 2032 | pinned_pages->nr_pages = nr_pages; |
| 2033 | micscif_destroy_pinned_pages(pinned_pages); |
| 2034 | *pages = NULL; |
| 2035 | pr_debug("%s %d err %d len 0x%lx\n", __func__, __LINE__, err, len); |
| 2036 | return err; |
| 2037 | |
| 2038 | } |
| 2039 | |
| 2040 | /** |
| 2041 | * scif_pin_pages - scif_pin_pages() pins the physical pages which back |
| 2042 | * the range of virtual address pages starting at addr and continuing for |
| 2043 | * len bytes. addr and len are constrained to be multiples of the page size. |
| 2044 | * A successful scif_register() call returns an opaque pointer value |
| 2045 | * which may be used in subsequent calls to scif_register_pinned_pages(). |
| 2046 | * |
| 2047 | * Return Values |
| 2048 | * Upon successful completion, scif_register() returns a |
| 2049 | * scif_pinned_pages_t value else an apt error is returned as documented |
| 2050 | * in scif.h |
| 2051 | */ |
| 2052 | int |
| 2053 | scif_pin_pages(void *addr, size_t len, int prot, |
| 2054 | int map_flags, scif_pinned_pages_t *pages) |
| 2055 | { |
| 2056 | return __scif_pin_pages(addr, len, &prot, map_flags, pages); |
| 2057 | } |
| 2058 | EXPORT_SYMBOL(scif_pin_pages); |
| 2059 | |
| 2060 | /** |
| 2061 | * scif_unpin_pages: Unpin a set of pages |
| 2062 | * |
| 2063 | * Return Values: |
| 2064 | * Upon successful completion, scif_unpin_pages() returns 0; |
| 2065 | * else an apt error is returned as documented in scif.h |
| 2066 | */ |
| 2067 | int |
| 2068 | scif_unpin_pages(scif_pinned_pages_t pinned_pages) |
| 2069 | { |
| 2070 | int err = 0, ret; |
| 2071 | |
| 2072 | if (!pinned_pages || SCIFEP_MAGIC != pinned_pages->magic) |
| 2073 | return -EINVAL; |
| 2074 | |
| 2075 | ret = atomic_sub_return((int32_t)pinned_pages->nr_pages, |
| 2076 | &pinned_pages->ref_count); |
| 2077 | BUG_ON(ret < 0); |
| 2078 | |
| 2079 | /* |
| 2080 | * Destroy the window if the ref count for this set of pinned |
| 2081 | * pages has dropped to zero. If it is positive then there is |
| 2082 | * a valid registered window which is backed by these pages and |
| 2083 | * it will be destroyed once all such windows are unregistered. |
| 2084 | */ |
| 2085 | if (!ret) |
| 2086 | err = micscif_destroy_pinned_pages(pinned_pages); |
| 2087 | |
| 2088 | return err; |
| 2089 | } |
| 2090 | EXPORT_SYMBOL(scif_unpin_pages); |
| 2091 | |
| 2092 | /** |
| 2093 | * scif_register_pinned_pages: Mark a memory region for remote access. |
| 2094 | * |
| 2095 | * The scif_register_pinned_pages() function opens a window, a range |
| 2096 | * of whole pages of the registered address space of the endpoint epd, |
| 2097 | * starting at offset po. The value of po, further described below, is |
| 2098 | * a function of the parameters offset and pinned_pages, and the value |
| 2099 | * of map_flags. Each page of the window represents a corresponding |
| 2100 | * physical memory page of pinned_pages; the length of the window is |
| 2101 | * the same as the length of pinned_pages. A successful scif_register() |
| 2102 | * call returns po as the return value. |
| 2103 | * |
| 2104 | * Return Values |
| 2105 | * Upon successful completion, scif_register_pinned_pages() returns |
| 2106 | * the offset at which the mapping was placed (po); |
| 2107 | * else an apt error is returned as documented in scif.h |
| 2108 | */ |
| 2109 | off_t |
| 2110 | __scif_register_pinned_pages(scif_epd_t epd, |
| 2111 | scif_pinned_pages_t pinned_pages, off_t offset, int map_flags) |
| 2112 | { |
| 2113 | struct endpt *ep = (struct endpt *)epd; |
| 2114 | uint64_t computed_offset; |
| 2115 | struct reg_range_t *window; |
| 2116 | int err; |
| 2117 | size_t len; |
| 2118 | |
| 2119 | #ifdef DEBUG |
| 2120 | /* Bad EP */ |
| 2121 | if (!ep || !pinned_pages || pinned_pages->magic != SCIFEP_MAGIC) |
| 2122 | return -EINVAL; |
| 2123 | #endif |
| 2124 | /* Unsupported flags */ |
| 2125 | if (map_flags & ~SCIF_MAP_FIXED) |
| 2126 | return -EINVAL; |
| 2127 | |
| 2128 | len = pinned_pages->nr_pages << PAGE_SHIFT; |
| 2129 | |
| 2130 | /* |
| 2131 | * Offset is not page aligned/negative or offset+len |
| 2132 | * wraps around with SCIF_MAP_FIXED. |
| 2133 | */ |
| 2134 | if ((map_flags & SCIF_MAP_FIXED) && |
| 2135 | ((align_low(offset, PAGE_SIZE) != offset) || |
| 2136 | (offset < 0) || |
| 2137 | (offset + (off_t)len < offset))) |
| 2138 | return -EINVAL; |
| 2139 | |
| 2140 | might_sleep(); |
| 2141 | |
| 2142 | if ((err = verify_epd(ep))) |
| 2143 | return err; |
| 2144 | |
| 2145 | /* Compute the offset for this registration */ |
| 2146 | if ((err = micscif_get_window_offset(ep, map_flags, offset, |
| 2147 | len, &computed_offset))) |
| 2148 | return err; |
| 2149 | |
| 2150 | /* Allocate and prepare self registration window */ |
| 2151 | if (!(window = micscif_create_window(ep, pinned_pages->nr_pages, |
| 2152 | computed_offset, false))) { |
| 2153 | micscif_free_window_offset(ep, computed_offset, len); |
| 2154 | return -ENOMEM; |
| 2155 | } |
| 2156 | |
| 2157 | window->pinned_pages = pinned_pages; |
| 2158 | window->nr_pages = pinned_pages->nr_pages; |
| 2159 | window->nr_contig_chunks = pinned_pages->nr_contig_chunks; |
| 2160 | window->prot = pinned_pages->prot; |
| 2161 | |
| 2162 | /* |
| 2163 | * This set of pinned pages now belongs to this window as well. |
| 2164 | * Assert if the ref count is zero since it is an error to |
| 2165 | * pass pinned_pages to scif_register_pinned_pages() after |
| 2166 | * calling scif_unpin_pages(). |
| 2167 | */ |
| 2168 | if (!atomic_add_unless(&pinned_pages->ref_count, |
| 2169 | (int32_t)pinned_pages->nr_pages, 0)) |
| 2170 | BUG_ON(1); |
| 2171 | |
| 2172 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2173 | |
| 2174 | if ((err = micscif_send_alloc_request(ep, window))) { |
| 2175 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2176 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2177 | goto error_unmap; |
| 2178 | } |
| 2179 | |
| 2180 | /* Prepare the remote registration window */ |
| 2181 | if ((err = micscif_prep_remote_window(ep, window))) { |
| 2182 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2183 | micscif_set_nr_pages(ep->remote_dev, window); |
| 2184 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2185 | goto error_unmap; |
| 2186 | } |
| 2187 | |
| 2188 | /* Tell the peer about the new window */ |
| 2189 | if ((err = micscif_send_scif_register(ep, window))) { |
| 2190 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2191 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2192 | goto error_unmap; |
| 2193 | } |
| 2194 | |
| 2195 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2196 | |
| 2197 | /* No further failures expected. Insert new window */ |
| 2198 | mutex_lock(&ep->rma_info.rma_lock); |
| 2199 | set_window_ref_count(window, pinned_pages->nr_pages); |
| 2200 | micscif_insert_window(window, &ep->rma_info.reg_list); |
| 2201 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2202 | |
| 2203 | return computed_offset; |
| 2204 | error_unmap: |
| 2205 | micscif_destroy_window(ep, window); |
| 2206 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2207 | return err; |
| 2208 | } |
| 2209 | |
| 2210 | off_t |
| 2211 | scif_register_pinned_pages(scif_epd_t epd, |
| 2212 | scif_pinned_pages_t pinned_pages, off_t offset, int map_flags) |
| 2213 | { |
| 2214 | off_t ret; |
| 2215 | get_kref_count(epd); |
| 2216 | ret = __scif_register_pinned_pages(epd, pinned_pages, offset, map_flags); |
| 2217 | put_kref_count(epd); |
| 2218 | return ret; |
| 2219 | } |
| 2220 | EXPORT_SYMBOL(scif_register_pinned_pages); |
| 2221 | |
| 2222 | /** |
| 2223 | * scif_get_pages - Add references to remote registered pages |
| 2224 | * |
| 2225 | * scif_get_pages() returns the addresses of the physical pages represented |
| 2226 | * by those pages of the registered address space of the peer of epd, starting |
| 2227 | * at offset offset and continuing for len bytes. offset and len are constrained |
| 2228 | * to be multiples of the page size. |
| 2229 | * |
| 2230 | * Return Values |
| 2231 | * Upon successful completion, scif_get_pages() returns 0; |
| 2232 | * else an apt error is returned as documented in scif.h. |
| 2233 | */ |
| 2234 | int |
| 2235 | __scif_get_pages(scif_epd_t epd, off_t offset, size_t len, struct scif_range **pages) |
| 2236 | { |
| 2237 | struct endpt *ep = (struct endpt *)epd; |
| 2238 | struct micscif_rma_req req; |
| 2239 | struct reg_range_t *window = NULL; |
| 2240 | int nr_pages, err, i; |
| 2241 | |
| 2242 | pr_debug("SCIFAPI get_pinned_pages: ep %p %s offset 0x%lx len 0x%lx\n", |
| 2243 | ep, scif_ep_states[ep->state], offset, len); |
| 2244 | |
| 2245 | if ((err = verify_epd(ep))) |
| 2246 | return err; |
| 2247 | |
| 2248 | if ((!len) || |
| 2249 | (offset < 0) || |
| 2250 | (offset + len < offset) || |
| 2251 | (align_low((uint64_t)offset, PAGE_SIZE) != (uint64_t)offset) || |
| 2252 | (align_low((uint64_t)len, PAGE_SIZE) != (uint64_t)len)) |
| 2253 | return -EINVAL; |
| 2254 | |
| 2255 | nr_pages = len >> PAGE_SHIFT; |
| 2256 | |
| 2257 | req.out_window = &window; |
| 2258 | req.offset = offset; |
| 2259 | req.prot = 0; |
| 2260 | req.nr_bytes = len; |
| 2261 | req.type = WINDOW_SINGLE; |
| 2262 | req.head = &ep->rma_info.remote_reg_list; |
| 2263 | |
| 2264 | mutex_lock(&ep->rma_info.rma_lock); |
| 2265 | /* Does a valid window exist? */ |
| 2266 | if ((err = micscif_query_window(&req))) { |
| 2267 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2268 | goto error; |
| 2269 | } |
| 2270 | RMA_MAGIC(window); |
| 2271 | |
| 2272 | /* Allocate scif_range */ |
| 2273 | if (!(*pages = kzalloc(sizeof(struct scif_range), GFP_KERNEL))) { |
| 2274 | err = -ENOMEM; |
| 2275 | goto error; |
| 2276 | } |
| 2277 | |
| 2278 | /* Allocate phys addr array */ |
| 2279 | if (!((*pages)->phys_addr = scif_zalloc(nr_pages * sizeof(dma_addr_t)))) { |
| 2280 | err = -ENOMEM; |
| 2281 | goto error; |
| 2282 | } |
| 2283 | |
| 2284 | #ifndef _MIC_SCIF_ |
| 2285 | /* Allocate virtual address array */ |
| 2286 | if (!((*pages)->va = scif_zalloc(nr_pages * sizeof(void *)))) { |
| 2287 | err = -ENOMEM; |
| 2288 | goto error; |
| 2289 | } |
| 2290 | #endif |
| 2291 | /* Populate the values */ |
| 2292 | (*pages)->cookie = window; |
| 2293 | (*pages)->nr_pages = nr_pages; |
| 2294 | (*pages)->prot_flags = window->prot; |
| 2295 | |
| 2296 | for (i = 0; i < nr_pages; i++) { |
| 2297 | (*pages)->phys_addr[i] = |
| 2298 | #if !defined(_MIC_SCIF_) && defined(CONFIG_ML1OM) |
| 2299 | is_self_scifdev(ep->remote_dev) ? |
| 2300 | micscif_get_dma_addr(window, offset + (i * PAGE_SIZE), |
| 2301 | NULL, NULL, NULL) : window->phys_addr[i]; |
| 2302 | #else |
| 2303 | get_phys_addr(micscif_get_dma_addr(window, offset + (i * PAGE_SIZE), |
| 2304 | NULL, NULL, NULL), ep->remote_dev); |
| 2305 | #endif |
| 2306 | #ifndef _MIC_SCIF_ |
| 2307 | if (!is_self_scifdev(ep->remote_dev)) |
| 2308 | (*pages)->va[i] = |
| 2309 | get_per_dev_ctx(ep->remote_dev->sd_node - 1)->aper.va + |
| 2310 | (*pages)->phys_addr[i] - |
| 2311 | get_per_dev_ctx(ep->remote_dev->sd_node - 1)->aper.pa; |
| 2312 | #endif |
| 2313 | } |
| 2314 | |
| 2315 | window->get_put_ref_count += nr_pages; |
| 2316 | get_window_ref_count(window, nr_pages); |
| 2317 | error: |
| 2318 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2319 | if (err) { |
| 2320 | if (*pages) { |
| 2321 | if ((*pages)->phys_addr) |
| 2322 | scif_free((*pages)->phys_addr, nr_pages * sizeof(dma_addr_t)); |
| 2323 | #ifndef _MIC_SCIF_ |
| 2324 | if ((*pages)->va) |
| 2325 | scif_free((*pages)->va, nr_pages * sizeof(void *)); |
| 2326 | #endif |
| 2327 | kfree(*pages); |
| 2328 | *pages = NULL; |
| 2329 | } |
| 2330 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2331 | } else { |
| 2332 | micscif_create_node_dep(ep->remote_dev, nr_pages); |
| 2333 | } |
| 2334 | return err; |
| 2335 | } |
| 2336 | |
| 2337 | int |
| 2338 | scif_get_pages(scif_epd_t epd, off_t offset, size_t len, struct scif_range **pages) |
| 2339 | { |
| 2340 | int ret; |
| 2341 | get_kref_count(epd); |
| 2342 | ret = __scif_get_pages(epd, offset, len, pages); |
| 2343 | put_kref_count(epd); |
| 2344 | return ret; |
| 2345 | } |
| 2346 | EXPORT_SYMBOL(scif_get_pages); |
| 2347 | |
| 2348 | /** |
| 2349 | * scif_put_pages - Remove references from remote registered pages |
| 2350 | * |
| 2351 | * scif_put_pages() returns a scif_range structure previously obtained by |
| 2352 | * calling scif_get_pages(). When control returns, the physical pages may |
| 2353 | * become available for reuse if and when the window which represented |
| 2354 | * those pages is unregistered. Therefore, those pages must never be accessed. |
| 2355 | * |
| 2356 | * Return Values |
| 2357 | * Upon success, zero is returned. |
| 2358 | * else an apt error is returned as documented in scif.h. |
| 2359 | */ |
| 2360 | int |
| 2361 | __scif_put_pages(struct scif_range *pages) |
| 2362 | { |
| 2363 | struct endpt *ep; |
| 2364 | struct reg_range_t *window; |
| 2365 | struct nodemsg msg; |
| 2366 | |
| 2367 | if (!pages || !pages->cookie) |
| 2368 | return -EINVAL; |
| 2369 | |
| 2370 | window = pages->cookie; |
| 2371 | |
| 2372 | if (!window || window->magic != SCIFEP_MAGIC || |
| 2373 | !window->get_put_ref_count) |
| 2374 | return -EINVAL; |
| 2375 | |
| 2376 | ep = (struct endpt *)window->ep; |
| 2377 | |
| 2378 | /* |
| 2379 | * If the state is SCIFEP_CONNECTED or SCIFEP_DISCONNECTED then the |
| 2380 | * callee should be allowed to release references to the pages, |
| 2381 | * else the endpoint was not connected in the first place, |
| 2382 | * hence the ENOTCONN. |
| 2383 | */ |
| 2384 | if (ep->state != SCIFEP_CONNECTED && ep->state != SCIFEP_DISCONNECTED) |
| 2385 | return -ENOTCONN; |
| 2386 | |
| 2387 | /* |
| 2388 | * TODO: Re-enable this check once ref counts for kernel mode APIs |
| 2389 | * have been implemented and node remove call backs are called before |
| 2390 | * the node is removed. This check results in kernel mode APIs not |
| 2391 | * being able to release pages correctly since node remove callbacks |
| 2392 | * are called after the node is removed currently. |
| 2393 | * if (!scifdev_alive(ep)) |
| 2394 | * return -ENODEV; |
| 2395 | */ |
| 2396 | |
| 2397 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2398 | mutex_lock(&ep->rma_info.rma_lock); |
| 2399 | |
| 2400 | /* Decrement the ref counts and check for errors */ |
| 2401 | window->get_put_ref_count -= pages->nr_pages; |
| 2402 | BUG_ON(window->get_put_ref_count < 0); |
| 2403 | put_window_ref_count(window, pages->nr_pages); |
| 2404 | |
| 2405 | /* Initiate window destruction if ref count is zero */ |
| 2406 | if (!window->ref_count) { |
| 2407 | drain_dma_intr(ep->rma_info.dma_chan); |
| 2408 | /* Inform the peer about this window being destroyed. */ |
| 2409 | msg.uop = SCIF_MUNMAP; |
| 2410 | msg.src = ep->port; |
| 2411 | msg.payload[0] = window->peer_window; |
| 2412 | /* No error handling for notification messages */ |
| 2413 | micscif_nodeqp_send(ep->remote_dev, &msg, ep); |
| 2414 | list_del(&window->list_member); |
| 2415 | /* Destroy this window from the peer's registered AS */ |
| 2416 | micscif_destroy_remote_window(ep, window); |
| 2417 | } |
| 2418 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2419 | |
| 2420 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2421 | micscif_destroy_node_dep(ep->remote_dev, pages->nr_pages); |
| 2422 | scif_free(pages->phys_addr, pages->nr_pages * sizeof(dma_addr_t)); |
| 2423 | #ifndef _MIC_SCIF_ |
| 2424 | scif_free(pages->va, pages->nr_pages * sizeof(void*)); |
| 2425 | #endif |
| 2426 | kfree(pages); |
| 2427 | return 0; |
| 2428 | } |
| 2429 | |
| 2430 | int |
| 2431 | scif_put_pages(struct scif_range *pages) |
| 2432 | { |
| 2433 | int ret; |
| 2434 | struct reg_range_t *window = pages->cookie; |
| 2435 | struct endpt *ep = (struct endpt *)window->ep; |
| 2436 | if (atomic_read(&(&(&(ep->ref_count))->refcount)->refs) > 0) { |
| 2437 | kref_get(&(ep->ref_count)); |
| 2438 | } else { |
| 2439 | WARN_ON(1); |
| 2440 | } |
| 2441 | ret = __scif_put_pages(pages); |
| 2442 | if (atomic_read(&(&(&(ep->ref_count))->refcount)->refs) > 0) { |
| 2443 | kref_put(&(ep->ref_count), scif_ref_rel); |
| 2444 | } else { |
| 2445 | //WARN_ON(1); |
| 2446 | } |
| 2447 | return ret; |
| 2448 | } |
| 2449 | EXPORT_SYMBOL(scif_put_pages); |
| 2450 | |
| 2451 | int scif_event_register(scif_callback_t handler) |
| 2452 | { |
| 2453 | /* Add to the list of event handlers */ |
| 2454 | struct scif_callback *cb = kmalloc(sizeof(*cb), GFP_KERNEL); |
| 2455 | if (!cb) |
| 2456 | return -ENOMEM; |
| 2457 | mutex_lock(&ms_info.mi_event_cblock); |
| 2458 | cb->callback_handler = handler; |
| 2459 | list_add_tail(&cb->list_member, &ms_info.mi_event_cb); |
| 2460 | mutex_unlock(&ms_info.mi_event_cblock); |
| 2461 | return 0; |
| 2462 | } |
| 2463 | EXPORT_SYMBOL(scif_event_register); |
| 2464 | |
| 2465 | int scif_event_unregister(scif_callback_t handler) |
| 2466 | { |
| 2467 | struct list_head *pos, *unused; |
| 2468 | struct scif_callback *temp; |
| 2469 | int err = -EINVAL; |
| 2470 | |
| 2471 | mutex_lock(&ms_info.mi_event_cblock); |
| 2472 | list_for_each_safe(pos, unused, &ms_info.mi_event_cb) { |
| 2473 | temp = list_entry(pos, struct scif_callback, list_member); |
| 2474 | if (temp->callback_handler == handler) { |
| 2475 | err = 0; |
| 2476 | list_del(pos); |
| 2477 | kfree(temp); |
| 2478 | break; |
| 2479 | } |
| 2480 | } |
| 2481 | |
| 2482 | mutex_unlock(&ms_info.mi_event_cblock); |
| 2483 | return err; |
| 2484 | } |
| 2485 | EXPORT_SYMBOL(scif_event_unregister); |
| 2486 | |
| 2487 | /** |
| 2488 | * scif_register - Mark a memory region for remote access. |
| 2489 | * @epd: endpoint descriptor |
| 2490 | * @addr: starting virtual address |
| 2491 | * @len: length of range |
| 2492 | * @offset: offset of window |
| 2493 | * @prot: read/write protection |
| 2494 | * @map_flags: flags |
| 2495 | * |
| 2496 | * Return Values |
| 2497 | * Upon successful completion, scif_register() returns the offset |
| 2498 | * at which the mapping was placed else an apt error is returned |
| 2499 | * as documented in scif.h. |
| 2500 | */ |
| 2501 | off_t |
| 2502 | __scif_register(scif_epd_t epd, void *addr, size_t len, off_t offset, |
| 2503 | int prot, int map_flags) |
| 2504 | { |
| 2505 | scif_pinned_pages_t pinned_pages; |
| 2506 | off_t err; |
| 2507 | struct endpt *ep = (struct endpt *)epd; |
| 2508 | uint64_t computed_offset; |
| 2509 | struct reg_range_t *window; |
| 2510 | struct mm_struct *mm = NULL; |
| 2511 | |
| 2512 | pr_debug("SCIFAPI register: ep %p %s addr %p len 0x%lx" |
| 2513 | " offset 0x%lx prot 0x%x map_flags 0x%x\n", |
| 2514 | epd, scif_ep_states[epd->state], addr, len, offset, prot, map_flags); |
| 2515 | |
| 2516 | /* Unsupported flags */ |
| 2517 | if (map_flags & ~(SCIF_MAP_FIXED | SCIF_MAP_KERNEL)) |
| 2518 | return -EINVAL; |
| 2519 | |
| 2520 | /* Unsupported protection requested */ |
| 2521 | if (prot & ~(SCIF_PROT_READ | SCIF_PROT_WRITE)) |
| 2522 | return -EINVAL; |
| 2523 | |
| 2524 | /* addr/len must be page aligned. len should be non zero */ |
| 2525 | if ((!len) || |
| 2526 | (align_low((uint64_t)addr, PAGE_SIZE) != (uint64_t)addr) || |
| 2527 | (align_low((uint64_t)len, PAGE_SIZE) != (uint64_t)len)) |
| 2528 | return -EINVAL; |
| 2529 | |
| 2530 | /* |
| 2531 | * Offset is not page aligned/negative or offset+len |
| 2532 | * wraps around with SCIF_MAP_FIXED. |
| 2533 | */ |
| 2534 | if ((map_flags & SCIF_MAP_FIXED) && |
| 2535 | ((align_low(offset, PAGE_SIZE) != offset) || |
| 2536 | (offset < 0) || |
| 2537 | (offset + (off_t)len < offset))) |
| 2538 | return -EINVAL; |
| 2539 | |
| 2540 | |
| 2541 | might_sleep(); |
| 2542 | |
| 2543 | #ifdef DEBUG |
| 2544 | /* Bad EP */ |
| 2545 | if (!ep) |
| 2546 | return -EINVAL; |
| 2547 | #endif |
| 2548 | |
| 2549 | if ((err = verify_epd(ep))) |
| 2550 | return err; |
| 2551 | |
| 2552 | /* Compute the offset for this registration */ |
| 2553 | if ((err = micscif_get_window_offset(ep, map_flags, offset, |
| 2554 | len, &computed_offset))) |
| 2555 | return err; |
| 2556 | |
| 2557 | /* Allocate and prepare self registration window */ |
| 2558 | if (!(window = micscif_create_window(ep, len >> PAGE_SHIFT, |
| 2559 | computed_offset, false))) { |
| 2560 | micscif_free_window_offset(ep, computed_offset, len); |
| 2561 | return -ENOMEM; |
| 2562 | } |
| 2563 | |
| 2564 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2565 | |
| 2566 | window->nr_pages = len >> PAGE_SHIFT; |
| 2567 | |
| 2568 | if ((err = micscif_send_alloc_request(ep, window))) { |
| 2569 | micscif_destroy_incomplete_window(ep, window); |
| 2570 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2571 | return err; |
| 2572 | } |
| 2573 | |
| 2574 | if (!(map_flags & SCIF_MAP_KERNEL)) { |
| 2575 | mm = __scif_acquire_mm(); |
| 2576 | map_flags |= SCIF_MAP_ULIMIT; |
| 2577 | } |
| 2578 | /* Pin down the pages */ |
| 2579 | if ((err = scif_pin_pages(addr, len, prot, |
| 2580 | map_flags & (SCIF_MAP_KERNEL | SCIF_MAP_ULIMIT), |
| 2581 | &pinned_pages))) { |
| 2582 | micscif_destroy_incomplete_window(ep, window); |
| 2583 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2584 | __scif_release_mm(mm); |
| 2585 | goto error; |
| 2586 | } |
| 2587 | |
| 2588 | window->pinned_pages = pinned_pages; |
| 2589 | window->nr_contig_chunks = pinned_pages->nr_contig_chunks; |
| 2590 | window->prot = pinned_pages->prot; |
| 2591 | window->mm = mm; |
| 2592 | |
| 2593 | /* Prepare the remote registration window */ |
| 2594 | if ((err = micscif_prep_remote_window(ep, window))) { |
| 2595 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2596 | micscif_set_nr_pages(ep->remote_dev, window); |
| 2597 | printk(KERN_ERR "%s %d err %ld\n", __func__, __LINE__, err); |
| 2598 | goto error_unmap; |
| 2599 | } |
| 2600 | |
| 2601 | /* Tell the peer about the new window */ |
| 2602 | if ((err = micscif_send_scif_register(ep, window))) { |
| 2603 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2604 | printk(KERN_ERR "%s %d err %ld\n", __func__, __LINE__, err); |
| 2605 | goto error_unmap; |
| 2606 | } |
| 2607 | |
| 2608 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2609 | |
| 2610 | /* No further failures expected. Insert new window */ |
| 2611 | mutex_lock(&ep->rma_info.rma_lock); |
| 2612 | set_window_ref_count(window, pinned_pages->nr_pages); |
| 2613 | micscif_insert_window(window, &ep->rma_info.reg_list); |
| 2614 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2615 | |
| 2616 | pr_debug("SCIFAPI register: ep %p %s addr %p" |
| 2617 | " len 0x%lx computed_offset 0x%llx\n", |
| 2618 | epd, scif_ep_states[epd->state], addr, len, computed_offset); |
| 2619 | return computed_offset; |
| 2620 | error_unmap: |
| 2621 | micscif_destroy_window(ep, window); |
| 2622 | error: |
| 2623 | printk(KERN_ERR "%s %d err %ld\n", __func__, __LINE__, err); |
| 2624 | return err; |
| 2625 | } |
| 2626 | |
| 2627 | off_t |
| 2628 | scif_register(scif_epd_t epd, void *addr, size_t len, off_t offset, |
| 2629 | int prot, int map_flags) |
| 2630 | { |
| 2631 | off_t ret; |
| 2632 | get_kref_count(epd); |
| 2633 | ret = __scif_register(epd, addr, len, offset, prot, map_flags); |
| 2634 | put_kref_count(epd); |
| 2635 | return ret; |
| 2636 | } |
| 2637 | EXPORT_SYMBOL(scif_register); |
| 2638 | |
| 2639 | /** |
| 2640 | * scif_unregister - Release a memory region registered for remote access. |
| 2641 | * @epd: endpoint descriptor |
| 2642 | * @offset: start of range to unregister |
| 2643 | * @len: length of range to unregister |
| 2644 | * |
| 2645 | * Return Values |
| 2646 | * Upon successful completion, scif_unegister() returns zero |
| 2647 | * else an apt error is returned as documented in scif.h. |
| 2648 | */ |
| 2649 | int |
| 2650 | __scif_unregister(scif_epd_t epd, off_t offset, size_t len) |
| 2651 | { |
| 2652 | struct endpt *ep = (struct endpt *)epd; |
| 2653 | struct reg_range_t *window = NULL; |
| 2654 | struct micscif_rma_req req; |
| 2655 | int nr_pages, err; |
| 2656 | |
| 2657 | pr_debug("SCIFAPI unregister: ep %p %s offset 0x%lx len 0x%lx\n", |
| 2658 | ep, scif_ep_states[ep->state], offset, len); |
| 2659 | |
| 2660 | /* len must be page aligned. len should be non zero */ |
| 2661 | if ((!len) || |
| 2662 | (align_low((uint64_t)len, PAGE_SIZE) != (uint64_t)len)) |
| 2663 | return -EINVAL; |
| 2664 | |
| 2665 | /* Offset is not page aligned or offset+len wraps around */ |
| 2666 | if ((align_low(offset, PAGE_SIZE) != offset) || |
| 2667 | (offset + (off_t)len < offset)) |
| 2668 | return -EINVAL; |
| 2669 | |
| 2670 | if ((err = verify_epd(ep))) |
| 2671 | return err; |
| 2672 | |
| 2673 | might_sleep(); |
| 2674 | nr_pages = (int)(len >> PAGE_SHIFT); |
| 2675 | |
| 2676 | req.out_window = &window; |
| 2677 | req.offset = offset; |
| 2678 | req.prot = 0; |
| 2679 | req.nr_bytes = len; |
| 2680 | req.type = WINDOW_FULL; |
| 2681 | req.head = &ep->rma_info.reg_list; |
| 2682 | |
| 2683 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2684 | mutex_lock(&ep->rma_info.rma_lock); |
| 2685 | /* Does a valid window exist? */ |
| 2686 | if ((err = micscif_query_window(&req))) { |
| 2687 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2688 | goto error; |
| 2689 | } |
| 2690 | /* Unregister all the windows in this range */ |
| 2691 | if ((err = micscif_rma_list_unregister(window, offset, nr_pages))) |
| 2692 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2693 | error: |
| 2694 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2695 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2696 | return err; |
| 2697 | } |
| 2698 | |
| 2699 | int |
| 2700 | scif_unregister(scif_epd_t epd, off_t offset, size_t len) |
| 2701 | { |
| 2702 | int ret; |
| 2703 | get_kref_count(epd); |
| 2704 | ret = __scif_unregister(epd, offset, len); |
| 2705 | put_kref_count(epd); |
| 2706 | return ret; |
| 2707 | } |
| 2708 | EXPORT_SYMBOL(scif_unregister); |
| 2709 | |
| 2710 | unsigned int scif_pollfd(struct file *f, poll_table *wait, scif_epd_t epd) |
| 2711 | { |
| 2712 | unsigned int ret; |
| 2713 | get_kref_count(epd); |
| 2714 | ret = __scif_pollfd(f, wait, (struct endpt *)epd); |
| 2715 | put_kref_count(epd); |
| 2716 | return ret; |
| 2717 | } |
| 2718 | |
| 2719 | unsigned int __scif_pollfd(struct file *f, poll_table *wait, struct endpt *ep) |
| 2720 | { |
| 2721 | unsigned int mask = 0; |
| 2722 | unsigned long sflags; |
| 2723 | |
| 2724 | pr_debug("SCIFAPI pollfd: ep %p %s\n", ep, scif_ep_states[ep->state]); |
| 2725 | |
| 2726 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2727 | spin_lock_irqsave(&ep->lock, sflags); |
| 2728 | |
| 2729 | if (ep->conn_async_state == ASYNC_CONN_INPROGRESS) { |
| 2730 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 2731 | if (!wait || poll_requested_events(wait) & SCIF_POLLOUT) { |
| 2732 | #else |
| 2733 | if (!wait || wait->key & SCIF_POLLOUT) { |
| 2734 | #endif |
| 2735 | poll_wait(f, &ep->conn_pend_wq, wait); |
| 2736 | if (ep->state == SCIFEP_CONNECTED || |
| 2737 | ep->state == SCIFEP_DISCONNECTED || |
| 2738 | ep->conn_err) { |
| 2739 | mask |= SCIF_POLLOUT; |
| 2740 | } |
| 2741 | goto return_scif_poll; |
| 2742 | } |
| 2743 | } |
| 2744 | |
| 2745 | /* Is it OK to use wait->key?? */ |
| 2746 | if (ep->state == SCIFEP_LISTENING) { |
| 2747 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 2748 | if (!wait || poll_requested_events(wait) & SCIF_POLLIN) { |
| 2749 | #else |
| 2750 | if (!wait || wait->key & SCIF_POLLIN) { |
| 2751 | #endif |
| 2752 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 2753 | poll_wait(f, &ep->conwq, wait); |
| 2754 | spin_lock_irqsave(&ep->lock, sflags); |
| 2755 | if (ep->conreqcnt) |
| 2756 | mask |= SCIF_POLLIN; |
| 2757 | } else { |
| 2758 | mask |= SCIF_POLLERR; |
| 2759 | } |
| 2760 | goto return_scif_poll; |
| 2761 | } |
| 2762 | |
| 2763 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 2764 | if (!wait || poll_requested_events(wait) & SCIF_POLLIN) { |
| 2765 | #else |
| 2766 | if (!wait || wait->key & SCIF_POLLIN) { |
| 2767 | #endif |
| 2768 | if (ep->state != SCIFEP_CONNECTED && |
| 2769 | ep->state != SCIFEP_LISTENING && |
| 2770 | ep->state != SCIFEP_DISCONNECTED) { |
| 2771 | mask |= SCIF_POLLERR; |
| 2772 | goto return_scif_poll; |
| 2773 | } |
| 2774 | |
| 2775 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 2776 | poll_wait(f, &ep->recvwq, wait); |
| 2777 | spin_lock_irqsave(&ep->lock, sflags); |
| 2778 | if (micscif_rb_count(&ep->qp_info.qp->inbound_q, 1)) |
| 2779 | mask |= SCIF_POLLIN; |
| 2780 | } |
| 2781 | |
| 2782 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 2783 | if (!wait || poll_requested_events(wait) & SCIF_POLLOUT) { |
| 2784 | #else |
| 2785 | if (!wait || wait->key & SCIF_POLLOUT) { |
| 2786 | #endif |
| 2787 | if (ep->state != SCIFEP_CONNECTED && |
| 2788 | ep->state != SCIFEP_LISTENING) { |
| 2789 | mask |= SCIF_POLLERR; |
| 2790 | goto return_scif_poll; |
| 2791 | } |
| 2792 | |
| 2793 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 2794 | poll_wait(f, &ep->sendwq, wait); |
| 2795 | spin_lock_irqsave(&ep->lock, sflags); |
| 2796 | if (micscif_rb_space(&ep->qp_info.qp->outbound_q)) |
| 2797 | mask |= SCIF_POLLOUT; |
| 2798 | } |
| 2799 | |
| 2800 | return_scif_poll: |
| 2801 | /* If the endpoint is in the diconnected state then return hangup instead of error */ |
| 2802 | if (ep->state == SCIFEP_DISCONNECTED) { |
| 2803 | mask &= ~SCIF_POLLERR; |
| 2804 | mask |= SCIF_POLLHUP; |
| 2805 | } |
| 2806 | |
| 2807 | spin_unlock_irqrestore(&ep->lock, sflags); |
| 2808 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2809 | return mask; |
| 2810 | } |
| 2811 | |
| 2812 | /* |
| 2813 | * The private data field of each VMA used to mmap a remote window |
| 2814 | * points to an instance of struct vma_pvt |
| 2815 | */ |
| 2816 | struct vma_pvt { |
| 2817 | struct endpt *ep; /* End point for remote window */ |
| 2818 | uint64_t offset; /* offset within remote window */ |
| 2819 | bool valid_offset; /* offset is valid only if the original |
| 2820 | * mmap request was for a single page |
| 2821 | * else the offset within the vma is |
| 2822 | * the correct offset |
| 2823 | */ |
| 2824 | struct kref ref; |
| 2825 | }; |
| 2826 | |
| 2827 | static void vma_pvt_release(struct kref *ref) |
| 2828 | { |
| 2829 | struct vma_pvt *vmapvt = container_of(ref, struct vma_pvt, ref); |
| 2830 | kfree(vmapvt); |
| 2831 | } |
| 2832 | |
| 2833 | /** |
| 2834 | * scif_vma_open - VMA open driver callback |
| 2835 | * @vma: VMM memory area. |
| 2836 | * The open method is called by the kernel to allow the subsystem implementing |
| 2837 | * the VMA to initialize the area. This method is invoked any time a new |
| 2838 | * reference to the VMA is made (when a process forks, for example). |
| 2839 | * The one exception happens when the VMA is first created by mmap; |
| 2840 | * in this case, the driver's mmap method is called instead. |
| 2841 | * This function is also invoked when an existing VMA is split by the kernel |
| 2842 | * due to a call to munmap on a subset of the VMA resulting in two VMAs. |
| 2843 | * The kernel invokes this function only on one of the two VMAs. |
| 2844 | * |
| 2845 | * Return Values: None. |
| 2846 | */ |
| 2847 | static void scif_vma_open(struct vm_area_struct *vma) |
| 2848 | { |
| 2849 | struct vma_pvt *vmapvt = ((vma)->vm_private_data); |
| 2850 | pr_debug("SCIFAPI vma open: vma_start 0x%lx vma_end 0x%lx\n", |
| 2851 | ((vma)->vm_start), ((vma)->vm_end)); |
| 2852 | kref_get(&vmapvt->ref); |
| 2853 | } |
| 2854 | |
| 2855 | /** |
| 2856 | * scif_munmap - VMA close driver callback. |
| 2857 | * @vma: VMM memory area. |
| 2858 | * When an area is destroyed, the kernel calls its close operation. |
| 2859 | * Note that there's no usage count associated with VMA's; the area |
| 2860 | * is opened and closed exactly once by each process that uses it. |
| 2861 | * |
| 2862 | * Return Values: None. |
| 2863 | */ |
| 2864 | void scif_munmap(struct vm_area_struct *vma) |
| 2865 | { |
| 2866 | struct endpt *ep; |
| 2867 | struct vma_pvt *vmapvt = ((vma)->vm_private_data); |
| 2868 | int nr_pages = (int)( (((vma)->vm_end) - ((vma)->vm_start)) >> PAGE_SHIFT ); |
| 2869 | uint64_t offset; |
| 2870 | struct micscif_rma_req req; |
| 2871 | struct reg_range_t *window = NULL; |
| 2872 | int err; |
| 2873 | |
| 2874 | might_sleep(); |
| 2875 | pr_debug("SCIFAPI munmap: vma_start 0x%lx vma_end 0x%lx\n", |
| 2876 | ((vma)->vm_start), ((vma)->vm_end)); |
| 2877 | /* used to be a BUG_ON(), prefer keeping the kernel alive */ |
| 2878 | if (!vmapvt) { |
| 2879 | WARN_ON(1); |
| 2880 | printk(KERN_ERR "SCIFAPI munmap: vma_start 0x%lx vma_end 0x%lx\n", |
| 2881 | ((vma)->vm_start), ((vma)->vm_end)); |
| 2882 | return; |
| 2883 | } |
| 2884 | |
| 2885 | ep = vmapvt->ep; |
| 2886 | offset = vmapvt->valid_offset ? vmapvt->offset : |
| 2887 | ((vma)->vm_pgoff) << PAGE_SHIFT; |
| 2888 | pr_debug("SCIFAPI munmap: ep %p %s nr_pages 0x%x offset 0x%llx\n", |
| 2889 | ep, scif_ep_states[ep->state], nr_pages, offset); |
| 2890 | |
| 2891 | req.out_window = &window; |
| 2892 | req.offset = offset; |
| 2893 | req.nr_bytes = ((vma)->vm_end) - ((vma)->vm_start); |
| 2894 | req.prot = ((vma)->vm_flags) & (VM_READ | VM_WRITE); |
| 2895 | req.type = WINDOW_PARTIAL; |
| 2896 | req.head = &ep->rma_info.remote_reg_list; |
| 2897 | |
| 2898 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2899 | mutex_lock(&ep->rma_info.rma_lock); |
| 2900 | |
| 2901 | if ((err = micscif_query_window(&req))) |
| 2902 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2903 | else |
| 2904 | micscif_rma_list_munmap(window, offset, nr_pages); |
| 2905 | |
| 2906 | mutex_unlock(&ep->rma_info.rma_lock); |
| 2907 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 2908 | |
| 2909 | micscif_destroy_node_dep(ep->remote_dev, nr_pages); |
| 2910 | |
| 2911 | /* |
| 2912 | * The kernel probably zeroes these out but we still want |
| 2913 | * to clean up our own mess just in case. |
| 2914 | */ |
| 2915 | vma->vm_ops = NULL; |
| 2916 | ((vma)->vm_private_data) = NULL; |
| 2917 | kref_put(&vmapvt->ref, vma_pvt_release); |
| 2918 | micscif_rma_put_task(ep, nr_pages); |
| 2919 | } |
| 2920 | |
| 2921 | static const struct vm_operations_struct micscif_vm_ops = { |
| 2922 | .open = scif_vma_open, |
| 2923 | .close = scif_munmap, |
| 2924 | }; |
| 2925 | |
| 2926 | /** |
| 2927 | * scif_mmap - Map pages in virtual address space to a remote window. |
| 2928 | * @vma: VMM memory area. |
| 2929 | * @epd: endpoint descriptor |
| 2930 | * |
| 2931 | * Return Values |
| 2932 | * Upon successful completion, scif_mmap() returns zero |
| 2933 | * else an apt error is returned as documented in scif.h. |
| 2934 | */ |
| 2935 | int |
| 2936 | scif_mmap(struct vm_area_struct *vma, scif_epd_t epd) |
| 2937 | { |
| 2938 | struct micscif_rma_req req; |
| 2939 | struct reg_range_t *window = NULL; |
| 2940 | struct endpt *ep = (struct endpt *)epd; |
| 2941 | uint64_t start_offset = ((vma)->vm_pgoff) << PAGE_SHIFT; |
| 2942 | int nr_pages = (int)( (((vma)->vm_end) - ((vma)->vm_start)) >> PAGE_SHIFT); |
| 2943 | int err; |
| 2944 | struct vma_pvt *vmapvt; |
| 2945 | |
| 2946 | pr_debug("SCIFAPI mmap: ep %p %s start_offset 0x%llx nr_pages 0x%x\n", |
| 2947 | ep, scif_ep_states[ep->state], start_offset, nr_pages); |
| 2948 | |
| 2949 | if ((err = verify_epd(ep))) |
| 2950 | return err; |
| 2951 | |
| 2952 | might_sleep(); |
| 2953 | |
| 2954 | if ((err = micscif_rma_get_task(ep, nr_pages))) |
| 2955 | return err; |
| 2956 | |
| 2957 | if (!(vmapvt = kzalloc(sizeof(*vmapvt), GFP_KERNEL))) { |
| 2958 | micscif_rma_put_task(ep, nr_pages); |
| 2959 | return -ENOMEM; |
| 2960 | } |
| 2961 | |
| 2962 | vmapvt->ep = ep; |
| 2963 | kref_init(&vmapvt->ref); |
| 2964 | |
| 2965 | micscif_create_node_dep(ep->remote_dev, nr_pages); |
| 2966 | |
| 2967 | req.out_window = &window; |
| 2968 | req.offset = start_offset; |
| 2969 | req.nr_bytes = ((vma)->vm_end) - ((vma)->vm_start); |
| 2970 | req.prot = ((vma)->vm_flags) & (VM_READ | VM_WRITE); |
| 2971 | req.type = WINDOW_PARTIAL; |
| 2972 | req.head = &ep->rma_info.remote_reg_list; |
| 2973 | |
| 2974 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 2975 | mutex_lock(&ep->rma_info.rma_lock); |
| 2976 | /* Does a valid window exist? */ |
| 2977 | if ((err = micscif_query_window(&req))) { |
| 2978 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 2979 | goto error; |
| 2980 | } |
| 2981 | RMA_MAGIC(window); |
| 2982 | |
| 2983 | /* Default prot for loopback */ |
| 2984 | if (!is_self_scifdev(ep->remote_dev)) { |
| 2985 | #ifdef _MIC_SCIF_ |
| 2986 | vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); |
| 2987 | #else |
| 2988 | vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot); |
| 2989 | #endif |
| 2990 | } |
| 2991 | |
| 2992 | /* |
| 2993 | * VM_DONTCOPY - Do not copy this vma on fork |
| 2994 | * VM_DONTEXPAND - Cannot expand with mremap() |
| 2995 | * VM_RESERVED - Count as reserved_vm like IO |
| 2996 | * VM_PFNMAP - Page-ranges managed without "struct page" |
| 2997 | * VM_IO - Memory mapped I/O or similar |
| 2998 | * |
| 2999 | * We do not want to copy this VMA automatically on a fork(), |
| 3000 | * expand this VMA due to mremap() or swap out these pages since |
| 3001 | * the VMA is actually backed by physical pages in the remote |
| 3002 | * node's physical memory and not via a struct page. |
| 3003 | */ |
| 3004 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 3005 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP | VM_PFNMAP; |
| 3006 | #else |
| 3007 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP; |
| 3008 | #endif |
| 3009 | |
| 3010 | if (!is_self_scifdev(ep->remote_dev)) |
| 3011 | ((vma)->vm_flags) |= VM_IO; |
| 3012 | |
| 3013 | /* Map this range of windows */ |
| 3014 | if ((err = micscif_rma_list_mmap(window, |
| 3015 | start_offset, nr_pages, vma))) { |
| 3016 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 3017 | goto error; |
| 3018 | } |
| 3019 | /* Set up the driver call back */ |
| 3020 | vma->vm_ops = &micscif_vm_ops; |
| 3021 | ((vma)->vm_private_data) = vmapvt; |
| 3022 | /* |
| 3023 | * For 1 page sized VMAs the kernel (remap_pfn_range) replaces the |
| 3024 | * offset in the VMA with the pfn, so in that case save off the |
| 3025 | * original offset, since the page sized VMA can't be split into |
| 3026 | * smaller VMAs the offset is not going to change. |
| 3027 | */ |
| 3028 | if (nr_pages == 1) { |
| 3029 | vmapvt->offset = start_offset; |
| 3030 | vmapvt->valid_offset = true; |
| 3031 | } |
| 3032 | err = 0; |
| 3033 | error: |
| 3034 | mutex_unlock(&ep->rma_info.rma_lock); |
| 3035 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 3036 | if (err) { |
| 3037 | micscif_destroy_node_dep(ep->remote_dev, nr_pages); |
| 3038 | kfree(vmapvt); |
| 3039 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 3040 | micscif_rma_put_task(ep, nr_pages); |
| 3041 | } |
| 3042 | return err; |
| 3043 | } |
| 3044 | |
| 3045 | /** |
| 3046 | * scif_readfrom() - Read SCIF offset data from remote connection |
| 3047 | * @epd: endpoint descriptor |
| 3048 | * @loffset: offset in local registered address space to which to copy |
| 3049 | * @len: length of range to copy |
| 3050 | * @roffset: offset in remote registered address space from which to copy |
| 3051 | * @flags: flags |
| 3052 | * |
| 3053 | * Return Values |
| 3054 | * Upon successful completion, scif_readfrom() returns zero |
| 3055 | * else an apt error is returned as documented in scif.h. |
| 3056 | */ |
| 3057 | int |
| 3058 | scif_readfrom(scif_epd_t epd, off_t loffset, size_t len, |
| 3059 | off_t roffset, int flags) |
| 3060 | { |
| 3061 | int ret; |
| 3062 | get_kref_count(epd); |
| 3063 | ret = __scif_readfrom(epd, loffset, len, roffset, flags); |
| 3064 | put_kref_count(epd); |
| 3065 | return ret; |
| 3066 | } |
| 3067 | EXPORT_SYMBOL(scif_readfrom); |
| 3068 | |
| 3069 | /** |
| 3070 | * scif_writeto() - Send SCIF offset data to remote connection |
| 3071 | * @epd: endpoint descriptor |
| 3072 | * @loffset: offset in local registered address space from which to copy |
| 3073 | * @len: length of range to copy |
| 3074 | * @roffset: offset in remote registered address space to which to copy |
| 3075 | * @flags: flags |
| 3076 | * |
| 3077 | * Return Values |
| 3078 | * Upon successful completion, scif_writeto() returns zero |
| 3079 | * else an apt error is returned as documented in scif.h. |
| 3080 | * |
| 3081 | */ |
| 3082 | int scif_writeto(scif_epd_t epd, off_t loffset, size_t len, |
| 3083 | off_t roffset, int flags) |
| 3084 | { |
| 3085 | int ret; |
| 3086 | get_kref_count(epd); |
| 3087 | ret = __scif_writeto(epd, loffset, len, roffset, flags); |
| 3088 | put_kref_count(epd); |
| 3089 | return ret; |
| 3090 | } |
| 3091 | EXPORT_SYMBOL(scif_writeto); |
| 3092 | |
| 3093 | #define HOST_LOOPB_MAGIC_MARK 0xdead |
| 3094 | |
| 3095 | /** |
| 3096 | * scif_fence_mark: |
| 3097 | * @epd: endpoint descriptor |
| 3098 | * @flags: control flags |
| 3099 | * @mark: marked handle returned as output. |
| 3100 | * |
| 3101 | * scif_fence_mark() returns after marking the current set of all uncompleted |
| 3102 | * RMAs initiated through the endpoint epd or marking the current set of all |
| 3103 | * uncompleted RMAs initiated through the peer of endpoint epd. The RMAs are |
| 3104 | * marked with a value returned in mark. The application may subsequently |
| 3105 | * await completion of all RMAs so marked. |
| 3106 | * |
| 3107 | * Return Values |
| 3108 | * Upon successful completion, scif_fence_mark() returns 0; |
| 3109 | * else an apt error is returned as documented in scif.h. |
| 3110 | */ |
| 3111 | int __scif_fence_mark(scif_epd_t epd, int flags, int *mark) |
| 3112 | { |
| 3113 | struct endpt *ep = (struct endpt *)epd; |
| 3114 | int err = 0; |
| 3115 | |
| 3116 | pr_debug("SCIFAPI fence_mark: ep %p %s flags 0x%x mark 0x%x\n", |
| 3117 | ep, scif_ep_states[ep->state], flags, *mark); |
| 3118 | |
| 3119 | if ((err = verify_epd(ep))) |
| 3120 | return err; |
| 3121 | |
| 3122 | /* Invalid flags? */ |
| 3123 | if (flags & ~(SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER)) |
| 3124 | return -EINVAL; |
| 3125 | |
| 3126 | /* At least one of init self or peer RMA should be set */ |
| 3127 | if (!(flags & (SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER))) |
| 3128 | return -EINVAL; |
| 3129 | |
| 3130 | /* Exactly one of init self or peer RMA should be set but not both */ |
| 3131 | if ((flags & SCIF_FENCE_INIT_SELF) && (flags & SCIF_FENCE_INIT_PEER)) |
| 3132 | return -EINVAL; |
| 3133 | |
| 3134 | #ifndef _MIC_SCIF_ |
| 3135 | /* |
| 3136 | * Host Loopback does not need to use DMA. |
| 3137 | * Return a valid mark to be symmetric. |
| 3138 | */ |
| 3139 | if (is_self_scifdev(ep->remote_dev)) { |
| 3140 | *mark = HOST_LOOPB_MAGIC_MARK; |
| 3141 | return 0; |
| 3142 | } |
| 3143 | #endif |
| 3144 | |
| 3145 | if (flags & SCIF_FENCE_INIT_SELF) { |
| 3146 | if ((*mark = micscif_fence_mark(epd)) < 0) |
| 3147 | err = *mark; |
| 3148 | } else { |
| 3149 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 3150 | err = micscif_send_fence_mark(ep, mark); |
| 3151 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 3152 | } |
| 3153 | if (err) |
| 3154 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 3155 | |
| 3156 | pr_debug("SCIFAPI fence_mark: ep %p %s flags 0x%x mark 0x%x err %d\n", |
| 3157 | ep, scif_ep_states[ep->state], flags, *mark, err); |
| 3158 | return err; |
| 3159 | } |
| 3160 | |
| 3161 | int scif_fence_mark(scif_epd_t epd, int flags, int *mark) |
| 3162 | { |
| 3163 | int ret; |
| 3164 | get_kref_count(epd); |
| 3165 | ret = __scif_fence_mark(epd, flags, mark); |
| 3166 | put_kref_count(epd); |
| 3167 | return ret; |
| 3168 | } |
| 3169 | EXPORT_SYMBOL(scif_fence_mark); |
| 3170 | |
| 3171 | /** |
| 3172 | * scif_fence_wait: |
| 3173 | * @epd: endpoint descriptor |
| 3174 | * @mark: mark request. |
| 3175 | * |
| 3176 | * scif_fence_wait() returns after all RMAs marked with mark have completed. |
| 3177 | * |
| 3178 | * Return Values |
| 3179 | * Upon successful completion, scif_fence_wait() returns 0; |
| 3180 | * else an apt error is returned as documented in scif.h. |
| 3181 | */ |
| 3182 | int __scif_fence_wait(scif_epd_t epd, int mark) |
| 3183 | { |
| 3184 | struct endpt *ep = (struct endpt *)epd; |
| 3185 | int err = 0; |
| 3186 | |
| 3187 | pr_debug("SCIFAPI fence_wait: ep %p %s mark 0x%x\n", |
| 3188 | ep, scif_ep_states[ep->state], mark); |
| 3189 | |
| 3190 | if ((err = verify_epd(ep))) |
| 3191 | return err; |
| 3192 | |
| 3193 | #ifndef _MIC_SCIF_ |
| 3194 | /* |
| 3195 | * Host Loopback does not need to use DMA. |
| 3196 | * The only valid mark provided is 0 so simply |
| 3197 | * return success if the mark is valid. |
| 3198 | */ |
| 3199 | if (is_self_scifdev(ep->remote_dev)) { |
| 3200 | if (HOST_LOOPB_MAGIC_MARK == mark) |
| 3201 | return 0; |
| 3202 | else |
| 3203 | return -EINVAL; |
| 3204 | } |
| 3205 | #endif |
| 3206 | if (mark & SCIF_REMOTE_FENCE) { |
| 3207 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 3208 | err = micscif_send_fence_wait(epd, mark); |
| 3209 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 3210 | } else { |
| 3211 | err = dma_mark_wait(epd->rma_info.dma_chan, mark, true); |
| 3212 | if (!err && atomic_read(&ep->rma_info.tw_refcount)) |
| 3213 | queue_work(ms_info.mi_misc_wq, &ms_info.mi_misc_work); |
| 3214 | } |
| 3215 | |
| 3216 | if (err < 0) |
| 3217 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 3218 | return err; |
| 3219 | } |
| 3220 | |
| 3221 | int scif_fence_wait(scif_epd_t epd, int mark) |
| 3222 | { |
| 3223 | int ret; |
| 3224 | get_kref_count(epd); |
| 3225 | ret = __scif_fence_wait(epd, mark); |
| 3226 | put_kref_count(epd); |
| 3227 | return ret; |
| 3228 | } |
| 3229 | EXPORT_SYMBOL(scif_fence_wait); |
| 3230 | |
| 3231 | /* |
| 3232 | * scif_fence_signal: |
| 3233 | * @loff: local offset |
| 3234 | * @lval: local value to write to loffset |
| 3235 | * @roff: remote offset |
| 3236 | * @rval: remote value to write to roffset |
| 3237 | * @flags: flags |
| 3238 | * |
| 3239 | * scif_fence_signal() returns after marking the current set of all |
| 3240 | * uncompleted RMAs initiated through the endpoint epd or marking |
| 3241 | * the current set of all uncompleted RMAs initiated through the peer |
| 3242 | * of endpoint epd. |
| 3243 | * |
| 3244 | * Return Values |
| 3245 | * Upon successful completion, scif_fence_signal() returns 0; |
| 3246 | * else an apt error is returned as documented in scif.h. |
| 3247 | */ |
| 3248 | int __scif_fence_signal(scif_epd_t epd, off_t loff, uint64_t lval, |
| 3249 | off_t roff, uint64_t rval, int flags) |
| 3250 | { |
| 3251 | struct endpt *ep = (struct endpt *)epd; |
| 3252 | int err = 0; |
| 3253 | |
| 3254 | pr_debug("SCIFAPI fence_signal: ep %p %s loff 0x%lx lval 0x%llx " |
| 3255 | "roff 0x%lx rval 0x%llx flags 0x%x\n", |
| 3256 | ep, scif_ep_states[ep->state], loff, lval, roff, rval, flags); |
| 3257 | |
| 3258 | if ((err = verify_epd(ep))) |
| 3259 | return err; |
| 3260 | |
| 3261 | /* Invalid flags? */ |
| 3262 | if (flags & ~(SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER | |
| 3263 | SCIF_SIGNAL_LOCAL | SCIF_SIGNAL_REMOTE)) |
| 3264 | return -EINVAL; |
| 3265 | |
| 3266 | /* At least one of init self or peer RMA should be set */ |
| 3267 | if (!(flags & (SCIF_FENCE_INIT_SELF | SCIF_FENCE_INIT_PEER))) |
| 3268 | return -EINVAL; |
| 3269 | |
| 3270 | /* Exactly one of init self or peer RMA should be set but not both */ |
| 3271 | if ((flags & SCIF_FENCE_INIT_SELF) && (flags & SCIF_FENCE_INIT_PEER)) |
| 3272 | return -EINVAL; |
| 3273 | |
| 3274 | /* At least one of SCIF_SIGNAL_LOCAL or SCIF_SIGNAL_REMOTE required */ |
| 3275 | if (!(flags & (SCIF_SIGNAL_LOCAL | SCIF_SIGNAL_REMOTE))) |
| 3276 | return -EINVAL; |
| 3277 | |
| 3278 | /* Only Dword offsets allowed */ |
| 3279 | if ((flags & SCIF_SIGNAL_LOCAL) && (loff & (sizeof(uint32_t) - 1))) |
| 3280 | return -EINVAL; |
| 3281 | |
| 3282 | /* Only Dword aligned offsets allowed */ |
| 3283 | if ((flags & SCIF_SIGNAL_REMOTE) && (roff & (sizeof(uint32_t) - 1))) |
| 3284 | return -EINVAL; |
| 3285 | |
| 3286 | if (flags & SCIF_FENCE_INIT_PEER) { |
| 3287 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 3288 | err = micscif_send_fence_signal(epd, roff, |
| 3289 | rval, loff, lval, flags); |
| 3290 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 3291 | } else { |
| 3292 | /* Local Signal in Local RAS */ |
| 3293 | if (flags & SCIF_SIGNAL_LOCAL) |
| 3294 | if ((err = micscif_prog_signal(epd, loff, |
| 3295 | lval, RMA_WINDOW_SELF))) |
| 3296 | goto error_ret; |
| 3297 | |
| 3298 | /* Signal in Remote RAS */ |
| 3299 | if (flags & SCIF_SIGNAL_REMOTE) { |
| 3300 | micscif_inc_node_refcnt(ep->remote_dev, 1); |
| 3301 | err = micscif_prog_signal(epd, roff, |
| 3302 | rval, RMA_WINDOW_PEER); |
| 3303 | micscif_dec_node_refcnt(ep->remote_dev, 1); |
| 3304 | } |
| 3305 | } |
| 3306 | error_ret: |
| 3307 | if (err) |
| 3308 | printk(KERN_ERR "%s %d err %d\n", __func__, __LINE__, err); |
| 3309 | else if (atomic_read(&ep->rma_info.tw_refcount)) |
| 3310 | queue_work(ms_info.mi_misc_wq, &ms_info.mi_misc_work); |
| 3311 | return err; |
| 3312 | } |
| 3313 | |
| 3314 | int scif_fence_signal(scif_epd_t epd, off_t loff, uint64_t lval, |
| 3315 | off_t roff, uint64_t rval, int flags) |
| 3316 | { |
| 3317 | int ret; |
| 3318 | get_kref_count(epd); |
| 3319 | ret = __scif_fence_signal(epd, loff, lval, roff, rval, flags); |
| 3320 | put_kref_count(epd); |
| 3321 | return ret; |
| 3322 | } |
| 3323 | EXPORT_SYMBOL(scif_fence_signal); |
| 3324 | |
| 3325 | /** |
| 3326 | * scif_get_nodeIDs - Return information about online nodes |
| 3327 | * @nodes: array space reserved for returning online node IDs |
| 3328 | * @len: number of entries on the nodes array |
| 3329 | * @self: address to place the node ID of this system |
| 3330 | * |
| 3331 | * Return Values |
| 3332 | * scif_get_nodeIDs() returns the total number of scif nodes |
| 3333 | * (including host) in the system |
| 3334 | */ |
| 3335 | int |
| 3336 | scif_get_nodeIDs(uint16_t *nodes, int len, uint16_t *self) |
| 3337 | { |
| 3338 | int online = 0; |
| 3339 | int offset = 0; |
| 3340 | int node; |
| 3341 | #ifdef _MIC_SCIF_ |
| 3342 | micscif_get_node_info(); |
| 3343 | #endif |
| 3344 | |
| 3345 | *self = ms_info.mi_nodeid; |
| 3346 | mutex_lock(&ms_info.mi_conflock); |
| 3347 | len = SCIF_MIN(len, (int32_t)ms_info.mi_total); |
| 3348 | for (node = 0; node <=(int32_t)ms_info.mi_maxid; node++) { |
| 3349 | if (ms_info.mi_mask & (1UL << node)) { |
| 3350 | online++; |
| 3351 | if (offset < len) |
| 3352 | nodes[offset++] = node; |
| 3353 | } |
| 3354 | } |
| 3355 | pr_debug("SCIFAPI get_nodeIDs total %d online %d filled in %d nodes\n", |
| 3356 | ms_info.mi_total, online, len); |
| 3357 | mutex_unlock(&ms_info.mi_conflock); |
| 3358 | |
| 3359 | return online; |
| 3360 | } |
| 3361 | |
| 3362 | EXPORT_SYMBOL(scif_get_nodeIDs); |
| 3363 | |
| 3364 | /** |
| 3365 | * micscif_pci_dev: |
| 3366 | * @node: node ID |
| 3367 | * |
| 3368 | * Return the pci_dev associated with a node. |
| 3369 | */ |
| 3370 | int micscif_pci_dev(uint16_t node, struct pci_dev **pdev) |
| 3371 | { |
| 3372 | #ifdef _MIC_SCIF_ |
| 3373 | /* This *is* a PCI device, therefore no pdev to return. */ |
| 3374 | return -ENODEV; |
| 3375 | #else |
| 3376 | mic_ctx_t *mic_ctx = get_per_dev_ctx(node - 1); |
| 3377 | *pdev = mic_ctx->bi_pdev; |
| 3378 | return 0; |
| 3379 | #endif |
| 3380 | } |
| 3381 | |
| 3382 | #ifndef _MIC_SCIF_ |
| 3383 | /** |
| 3384 | * micscif_pci_info: |
| 3385 | * @node: node ID |
| 3386 | * |
| 3387 | * Populate the pci device info pointer associated with a node. |
| 3388 | */ |
| 3389 | int micscif_pci_info(uint16_t node, struct scif_pci_info *dev) |
| 3390 | { |
| 3391 | int i; |
| 3392 | mic_ctx_t *mic_ctx = get_per_dev_ctx(node - 1); |
| 3393 | struct pci_dev *pdev; |
| 3394 | |
| 3395 | if (!mic_ctx) |
| 3396 | return -ENODEV; |
| 3397 | |
| 3398 | dev->pdev = pdev = mic_ctx->bi_pdev; |
| 3399 | for (i = 0; i < PCI_NUM_RESOURCES; i++) { |
| 3400 | if (!pci_resource_start(pdev, i)) { |
| 3401 | dev->va[i] = NULL; |
| 3402 | continue; |
| 3403 | } |
| 3404 | if (pci_resource_flags(pdev, i) & IORESOURCE_PREFETCH) { |
| 3405 | /* TODO: Change comparison check for KNL. */ |
| 3406 | if (pci_resource_start(pdev, i) == mic_ctx->aper.pa) |
| 3407 | dev->va[i] = mic_ctx->aper.va; |
| 3408 | else |
| 3409 | dev->va[i] = NULL; |
| 3410 | } else { |
| 3411 | dev->va[i] = mic_ctx->mmio.va; |
| 3412 | } |
| 3413 | } |
| 3414 | return 0; |
| 3415 | } |
| 3416 | #endif |
| 3417 | |
| 3418 | /** |
| 3419 | * scif_pci_info - Populate the pci device info pointer associated with a node |
| 3420 | * @node: the node to query |
| 3421 | * @scif_pdev: The scif_pci_info structure to populate. |
| 3422 | * |
| 3423 | * scif_pci_info() populates the provided scif_pci_info structure |
| 3424 | * associated with a node. The requested node ID cannot be the same as |
| 3425 | * the current node. This routine may only return success when called from |
| 3426 | * the host. |
| 3427 | * |
| 3428 | * Return Values |
| 3429 | * Upon successful completion, scif_pci_info() returns 0; otherwise the |
| 3430 | * an appropriate error is returned as documented in scif.h. |
| 3431 | */ |
| 3432 | int scif_pci_info(uint16_t node, struct scif_pci_info *dev) |
| 3433 | { |
| 3434 | #ifdef _MIC_SCIF_ |
| 3435 | return -EINVAL; |
| 3436 | #else |
| 3437 | if (node > ms_info.mi_maxid) |
| 3438 | return -EINVAL; |
| 3439 | |
| 3440 | if ((scif_dev[node].sd_state == SCIFDEV_NOTPRESENT) || |
| 3441 | is_self_scifdev(&scif_dev[node])) |
| 3442 | return -ENODEV; |
| 3443 | |
| 3444 | return micscif_pci_info(node, dev); |
| 3445 | #endif |
| 3446 | } |
| 3447 | EXPORT_SYMBOL(scif_pci_info); |
| 3448 | |
| 3449 | /* |
| 3450 | * DEBUG helper functions |
| 3451 | */ |
| 3452 | void |
| 3453 | print_ep_state(struct endpt *ep, char *label) |
| 3454 | { |
| 3455 | if (ep) |
| 3456 | printk("%s: EP %p state %s\n", |
| 3457 | label, ep, scif_ep_states[ep->state]); |
| 3458 | else |
| 3459 | printk("%s: EP %p\n state ?\n", label, ep); |
| 3460 | } |
| 3461 | |