1 /* 2 * Copyright 2014 IBM Corp. 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public License 6 * as published by the Free Software Foundation; either version 7 * 2 of the License, or (at your option) any later version. 8 */ 9 10 #include <linux/spinlock.h> 11 #include <linux/module.h> 12 #include <linux/export.h> 13 #include <linux/kernel.h> 14 #include <linux/bitmap.h> 15 #include <linux/sched/signal.h> 16 #include <linux/poll.h> 17 #include <linux/pid.h> 18 #include <linux/fs.h> 19 #include <linux/mm.h> 20 #include <linux/slab.h> 21 #include <asm/cputable.h> 22 #include <asm/current.h> 23 #include <asm/copro.h> 24 25 #include "cxl.h" 26 #include "trace.h" 27 28 #define CXL_NUM_MINORS 256 /* Total to reserve */ 29 30 #define CXL_AFU_MINOR_D(afu) (CXL_CARD_MINOR(afu->adapter) + 1 + (3 * afu->slice)) 31 #define CXL_AFU_MINOR_M(afu) (CXL_AFU_MINOR_D(afu) + 1) 32 #define CXL_AFU_MINOR_S(afu) (CXL_AFU_MINOR_D(afu) + 2) 33 #define CXL_AFU_MKDEV_D(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_D(afu)) 34 #define CXL_AFU_MKDEV_M(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_M(afu)) 35 #define CXL_AFU_MKDEV_S(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_S(afu)) 36 37 #define CXL_DEVT_AFU(dev) ((MINOR(dev) % CXL_DEV_MINORS - 1) / 3) 38 39 #define CXL_DEVT_IS_CARD(dev) (MINOR(dev) % CXL_DEV_MINORS == 0) 40 41 static dev_t cxl_dev; 42 43 static struct class *cxl_class; 44 45 static int __afu_open(struct inode *inode, struct file *file, bool master) 46 { 47 struct cxl *adapter; 48 struct cxl_afu *afu; 49 struct cxl_context *ctx; 50 int adapter_num = CXL_DEVT_ADAPTER(inode->i_rdev); 51 int slice = CXL_DEVT_AFU(inode->i_rdev); 52 int rc = -ENODEV; 53 54 pr_devel("afu_open afu%i.%i\n", slice, adapter_num); 55 56 if (!(adapter = get_cxl_adapter(adapter_num))) 57 return -ENODEV; 58 59 if (slice > adapter->slices) 60 goto err_put_adapter; 61 62 spin_lock(&adapter->afu_list_lock); 63 if (!(afu = adapter->afu[slice])) { 64 spin_unlock(&adapter->afu_list_lock); 65 goto err_put_adapter; 66 } 67 68 /* 69 * taking a ref to the afu so that it doesn't go away 70 * for rest of the function. This ref is released before 71 * we return. 72 */ 73 cxl_afu_get(afu); 74 spin_unlock(&adapter->afu_list_lock); 75 76 if (!afu->current_mode) 77 goto err_put_afu; 78 79 if (!cxl_ops->link_ok(adapter, afu)) { 80 rc = -EIO; 81 goto err_put_afu; 82 } 83 84 if (!(ctx = cxl_context_alloc())) { 85 rc = -ENOMEM; 86 goto err_put_afu; 87 } 88 89 rc = cxl_context_init(ctx, afu, master); 90 if (rc) 91 goto err_put_afu; 92 93 cxl_context_set_mapping(ctx, inode->i_mapping); 94 95 pr_devel("afu_open pe: %i\n", ctx->pe); 96 file->private_data = ctx; 97 cxl_ctx_get(); 98 99 /* indicate success */ 100 rc = 0; 101 102 err_put_afu: 103 /* release the ref taken earlier */ 104 cxl_afu_put(afu); 105 err_put_adapter: 106 put_device(&adapter->dev); 107 return rc; 108 } 109 110 int afu_open(struct inode *inode, struct file *file) 111 { 112 return __afu_open(inode, file, false); 113 } 114 115 static int afu_master_open(struct inode *inode, struct file *file) 116 { 117 return __afu_open(inode, file, true); 118 } 119 120 int afu_release(struct inode *inode, struct file *file) 121 { 122 struct cxl_context *ctx = file->private_data; 123 124 pr_devel("%s: closing cxl file descriptor. pe: %i\n", 125 __func__, ctx->pe); 126 cxl_context_detach(ctx); 127 128 129 /* 130 * Delete the context's mapping pointer, unless it's created by the 131 * kernel API, in which case leave it so it can be freed by reclaim_ctx() 132 */ 133 if (!ctx->kernelapi) { 134 mutex_lock(&ctx->mapping_lock); 135 ctx->mapping = NULL; 136 mutex_unlock(&ctx->mapping_lock); 137 } 138 139 /* 140 * At this this point all bottom halfs have finished and we should be 141 * getting no more IRQs from the hardware for this context. Once it's 142 * removed from the IDR (and RCU synchronised) it's safe to free the 143 * sstp and context. 144 */ 145 cxl_context_free(ctx); 146 147 return 0; 148 } 149 150 static long afu_ioctl_start_work(struct cxl_context *ctx, 151 struct cxl_ioctl_start_work __user *uwork) 152 { 153 struct cxl_ioctl_start_work work; 154 u64 amr = 0; 155 int rc; 156 157 pr_devel("%s: pe: %i\n", __func__, ctx->pe); 158 159 /* Do this outside the status_mutex to avoid a circular dependency with 160 * the locking in cxl_mmap_fault() */ 161 if (copy_from_user(&work, uwork, 162 sizeof(struct cxl_ioctl_start_work))) { 163 rc = -EFAULT; 164 goto out; 165 } 166 167 mutex_lock(&ctx->status_mutex); 168 if (ctx->status != OPENED) { 169 rc = -EIO; 170 goto out; 171 } 172 173 /* 174 * if any of the reserved fields are set or any of the unused 175 * flags are set it's invalid 176 */ 177 if (work.reserved1 || work.reserved2 || work.reserved3 || 178 work.reserved4 || work.reserved5 || work.reserved6 || 179 (work.flags & ~CXL_START_WORK_ALL)) { 180 rc = -EINVAL; 181 goto out; 182 } 183 184 if (!(work.flags & CXL_START_WORK_NUM_IRQS)) 185 work.num_interrupts = ctx->afu->pp_irqs; 186 else if ((work.num_interrupts < ctx->afu->pp_irqs) || 187 (work.num_interrupts > ctx->afu->irqs_max)) { 188 rc = -EINVAL; 189 goto out; 190 } 191 if ((rc = afu_register_irqs(ctx, work.num_interrupts))) 192 goto out; 193 194 if (work.flags & CXL_START_WORK_AMR) 195 amr = work.amr & mfspr(SPRN_UAMOR); 196 197 ctx->mmio_err_ff = !!(work.flags & CXL_START_WORK_ERR_FF); 198 199 /* 200 * Increment the mapped context count for adapter. This also checks 201 * if adapter_context_lock is taken. 202 */ 203 rc = cxl_adapter_context_get(ctx->afu->adapter); 204 if (rc) { 205 afu_release_irqs(ctx, ctx); 206 goto out; 207 } 208 209 /* 210 * We grab the PID here and not in the file open to allow for the case 211 * where a process (master, some daemon, etc) has opened the chardev on 212 * behalf of another process, so the AFU's mm gets bound to the process 213 * that performs this ioctl and not the process that opened the file. 214 * Also we grab the PID of the group leader so that if the task that 215 * has performed the attach operation exits the mm context of the 216 * process is still accessible. 217 */ 218 ctx->pid = get_task_pid(current, PIDTYPE_PID); 219 ctx->glpid = get_task_pid(current->group_leader, PIDTYPE_PID); 220 221 222 trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr); 223 224 if ((rc = cxl_ops->attach_process(ctx, false, work.work_element_descriptor, 225 amr))) { 226 afu_release_irqs(ctx, ctx); 227 cxl_adapter_context_put(ctx->afu->adapter); 228 put_pid(ctx->glpid); 229 put_pid(ctx->pid); 230 ctx->glpid = ctx->pid = NULL; 231 goto out; 232 } 233 234 ctx->status = STARTED; 235 rc = 0; 236 out: 237 mutex_unlock(&ctx->status_mutex); 238 return rc; 239 } 240 241 static long afu_ioctl_process_element(struct cxl_context *ctx, 242 int __user *upe) 243 { 244 pr_devel("%s: pe: %i\n", __func__, ctx->pe); 245 246 if (copy_to_user(upe, &ctx->external_pe, sizeof(__u32))) 247 return -EFAULT; 248 249 return 0; 250 } 251 252 static long afu_ioctl_get_afu_id(struct cxl_context *ctx, 253 struct cxl_afu_id __user *upafuid) 254 { 255 struct cxl_afu_id afuid = { 0 }; 256 257 afuid.card_id = ctx->afu->adapter->adapter_num; 258 afuid.afu_offset = ctx->afu->slice; 259 afuid.afu_mode = ctx->afu->current_mode; 260 261 /* set the flag bit in case the afu is a slave */ 262 if (ctx->afu->current_mode == CXL_MODE_DIRECTED && !ctx->master) 263 afuid.flags |= CXL_AFUID_FLAG_SLAVE; 264 265 if (copy_to_user(upafuid, &afuid, sizeof(afuid))) 266 return -EFAULT; 267 268 return 0; 269 } 270 271 long afu_ioctl(struct file *file, unsigned int cmd, unsigned long arg) 272 { 273 struct cxl_context *ctx = file->private_data; 274 275 if (ctx->status == CLOSED) 276 return -EIO; 277 278 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 279 return -EIO; 280 281 pr_devel("afu_ioctl\n"); 282 switch (cmd) { 283 case CXL_IOCTL_START_WORK: 284 return afu_ioctl_start_work(ctx, (struct cxl_ioctl_start_work __user *)arg); 285 case CXL_IOCTL_GET_PROCESS_ELEMENT: 286 return afu_ioctl_process_element(ctx, (__u32 __user *)arg); 287 case CXL_IOCTL_GET_AFU_ID: 288 return afu_ioctl_get_afu_id(ctx, (struct cxl_afu_id __user *) 289 arg); 290 } 291 return -EINVAL; 292 } 293 294 static long afu_compat_ioctl(struct file *file, unsigned int cmd, 295 unsigned long arg) 296 { 297 return afu_ioctl(file, cmd, arg); 298 } 299 300 int afu_mmap(struct file *file, struct vm_area_struct *vm) 301 { 302 struct cxl_context *ctx = file->private_data; 303 304 /* AFU must be started before we can MMIO */ 305 if (ctx->status != STARTED) 306 return -EIO; 307 308 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 309 return -EIO; 310 311 return cxl_context_iomap(ctx, vm); 312 } 313 314 static inline bool ctx_event_pending(struct cxl_context *ctx) 315 { 316 if (ctx->pending_irq || ctx->pending_fault || ctx->pending_afu_err) 317 return true; 318 319 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) 320 return true; 321 322 return false; 323 } 324 325 unsigned int afu_poll(struct file *file, struct poll_table_struct *poll) 326 { 327 struct cxl_context *ctx = file->private_data; 328 int mask = 0; 329 unsigned long flags; 330 331 332 poll_wait(file, &ctx->wq, poll); 333 334 pr_devel("afu_poll wait done pe: %i\n", ctx->pe); 335 336 spin_lock_irqsave(&ctx->lock, flags); 337 if (ctx_event_pending(ctx)) 338 mask |= POLLIN | POLLRDNORM; 339 else if (ctx->status == CLOSED) 340 /* Only error on closed when there are no futher events pending 341 */ 342 mask |= POLLERR; 343 spin_unlock_irqrestore(&ctx->lock, flags); 344 345 pr_devel("afu_poll pe: %i returning %#x\n", ctx->pe, mask); 346 347 return mask; 348 } 349 350 static ssize_t afu_driver_event_copy(struct cxl_context *ctx, 351 char __user *buf, 352 struct cxl_event *event, 353 struct cxl_event_afu_driver_reserved *pl) 354 { 355 /* Check event */ 356 if (!pl) { 357 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL); 358 return -EFAULT; 359 } 360 361 /* Check event size */ 362 event->header.size += pl->data_size; 363 if (event->header.size > CXL_READ_MIN_SIZE) { 364 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL); 365 return -EFAULT; 366 } 367 368 /* Copy event header */ 369 if (copy_to_user(buf, event, sizeof(struct cxl_event_header))) { 370 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT); 371 return -EFAULT; 372 } 373 374 /* Copy event data */ 375 buf += sizeof(struct cxl_event_header); 376 if (copy_to_user(buf, &pl->data, pl->data_size)) { 377 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT); 378 return -EFAULT; 379 } 380 381 ctx->afu_driver_ops->event_delivered(ctx, pl, 0); /* Success */ 382 return event->header.size; 383 } 384 385 ssize_t afu_read(struct file *file, char __user *buf, size_t count, 386 loff_t *off) 387 { 388 struct cxl_context *ctx = file->private_data; 389 struct cxl_event_afu_driver_reserved *pl = NULL; 390 struct cxl_event event; 391 unsigned long flags; 392 int rc; 393 DEFINE_WAIT(wait); 394 395 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 396 return -EIO; 397 398 if (count < CXL_READ_MIN_SIZE) 399 return -EINVAL; 400 401 spin_lock_irqsave(&ctx->lock, flags); 402 403 for (;;) { 404 prepare_to_wait(&ctx->wq, &wait, TASK_INTERRUPTIBLE); 405 if (ctx_event_pending(ctx) || (ctx->status == CLOSED)) 406 break; 407 408 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) { 409 rc = -EIO; 410 goto out; 411 } 412 413 if (file->f_flags & O_NONBLOCK) { 414 rc = -EAGAIN; 415 goto out; 416 } 417 418 if (signal_pending(current)) { 419 rc = -ERESTARTSYS; 420 goto out; 421 } 422 423 spin_unlock_irqrestore(&ctx->lock, flags); 424 pr_devel("afu_read going to sleep...\n"); 425 schedule(); 426 pr_devel("afu_read woken up\n"); 427 spin_lock_irqsave(&ctx->lock, flags); 428 } 429 430 finish_wait(&ctx->wq, &wait); 431 432 memset(&event, 0, sizeof(event)); 433 event.header.process_element = ctx->pe; 434 event.header.size = sizeof(struct cxl_event_header); 435 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) { 436 pr_devel("afu_read delivering AFU driver specific event\n"); 437 pl = ctx->afu_driver_ops->fetch_event(ctx); 438 atomic_dec(&ctx->afu_driver_events); 439 event.header.type = CXL_EVENT_AFU_DRIVER; 440 } else if (ctx->pending_irq) { 441 pr_devel("afu_read delivering AFU interrupt\n"); 442 event.header.size += sizeof(struct cxl_event_afu_interrupt); 443 event.header.type = CXL_EVENT_AFU_INTERRUPT; 444 event.irq.irq = find_first_bit(ctx->irq_bitmap, ctx->irq_count) + 1; 445 clear_bit(event.irq.irq - 1, ctx->irq_bitmap); 446 if (bitmap_empty(ctx->irq_bitmap, ctx->irq_count)) 447 ctx->pending_irq = false; 448 } else if (ctx->pending_fault) { 449 pr_devel("afu_read delivering data storage fault\n"); 450 event.header.size += sizeof(struct cxl_event_data_storage); 451 event.header.type = CXL_EVENT_DATA_STORAGE; 452 event.fault.addr = ctx->fault_addr; 453 event.fault.dsisr = ctx->fault_dsisr; 454 ctx->pending_fault = false; 455 } else if (ctx->pending_afu_err) { 456 pr_devel("afu_read delivering afu error\n"); 457 event.header.size += sizeof(struct cxl_event_afu_error); 458 event.header.type = CXL_EVENT_AFU_ERROR; 459 event.afu_error.error = ctx->afu_err; 460 ctx->pending_afu_err = false; 461 } else if (ctx->status == CLOSED) { 462 pr_devel("afu_read fatal error\n"); 463 spin_unlock_irqrestore(&ctx->lock, flags); 464 return -EIO; 465 } else 466 WARN(1, "afu_read must be buggy\n"); 467 468 spin_unlock_irqrestore(&ctx->lock, flags); 469 470 if (event.header.type == CXL_EVENT_AFU_DRIVER) 471 return afu_driver_event_copy(ctx, buf, &event, pl); 472 473 if (copy_to_user(buf, &event, event.header.size)) 474 return -EFAULT; 475 return event.header.size; 476 477 out: 478 finish_wait(&ctx->wq, &wait); 479 spin_unlock_irqrestore(&ctx->lock, flags); 480 return rc; 481 } 482 483 /* 484 * Note: if this is updated, we need to update api.c to patch the new ones in 485 * too 486 */ 487 const struct file_operations afu_fops = { 488 .owner = THIS_MODULE, 489 .open = afu_open, 490 .poll = afu_poll, 491 .read = afu_read, 492 .release = afu_release, 493 .unlocked_ioctl = afu_ioctl, 494 .compat_ioctl = afu_compat_ioctl, 495 .mmap = afu_mmap, 496 }; 497 498 static const struct file_operations afu_master_fops = { 499 .owner = THIS_MODULE, 500 .open = afu_master_open, 501 .poll = afu_poll, 502 .read = afu_read, 503 .release = afu_release, 504 .unlocked_ioctl = afu_ioctl, 505 .compat_ioctl = afu_compat_ioctl, 506 .mmap = afu_mmap, 507 }; 508 509 510 static char *cxl_devnode(struct device *dev, umode_t *mode) 511 { 512 if (cpu_has_feature(CPU_FTR_HVMODE) && 513 CXL_DEVT_IS_CARD(dev->devt)) { 514 /* 515 * These minor numbers will eventually be used to program the 516 * PSL and AFUs once we have dynamic reprogramming support 517 */ 518 return NULL; 519 } 520 return kasprintf(GFP_KERNEL, "cxl/%s", dev_name(dev)); 521 } 522 523 extern struct class *cxl_class; 524 525 static int cxl_add_chardev(struct cxl_afu *afu, dev_t devt, struct cdev *cdev, 526 struct device **chardev, char *postfix, char *desc, 527 const struct file_operations *fops) 528 { 529 struct device *dev; 530 int rc; 531 532 cdev_init(cdev, fops); 533 if ((rc = cdev_add(cdev, devt, 1))) { 534 dev_err(&afu->dev, "Unable to add %s chardev: %i\n", desc, rc); 535 return rc; 536 } 537 538 dev = device_create(cxl_class, &afu->dev, devt, afu, 539 "afu%i.%i%s", afu->adapter->adapter_num, afu->slice, postfix); 540 if (IS_ERR(dev)) { 541 dev_err(&afu->dev, "Unable to create %s chardev in sysfs: %i\n", desc, rc); 542 rc = PTR_ERR(dev); 543 goto err; 544 } 545 546 *chardev = dev; 547 548 return 0; 549 err: 550 cdev_del(cdev); 551 return rc; 552 } 553 554 int cxl_chardev_d_afu_add(struct cxl_afu *afu) 555 { 556 return cxl_add_chardev(afu, CXL_AFU_MKDEV_D(afu), &afu->afu_cdev_d, 557 &afu->chardev_d, "d", "dedicated", 558 &afu_master_fops); /* Uses master fops */ 559 } 560 561 int cxl_chardev_m_afu_add(struct cxl_afu *afu) 562 { 563 return cxl_add_chardev(afu, CXL_AFU_MKDEV_M(afu), &afu->afu_cdev_m, 564 &afu->chardev_m, "m", "master", 565 &afu_master_fops); 566 } 567 568 int cxl_chardev_s_afu_add(struct cxl_afu *afu) 569 { 570 return cxl_add_chardev(afu, CXL_AFU_MKDEV_S(afu), &afu->afu_cdev_s, 571 &afu->chardev_s, "s", "shared", 572 &afu_fops); 573 } 574 575 void cxl_chardev_afu_remove(struct cxl_afu *afu) 576 { 577 if (afu->chardev_d) { 578 cdev_del(&afu->afu_cdev_d); 579 device_unregister(afu->chardev_d); 580 afu->chardev_d = NULL; 581 } 582 if (afu->chardev_m) { 583 cdev_del(&afu->afu_cdev_m); 584 device_unregister(afu->chardev_m); 585 afu->chardev_m = NULL; 586 } 587 if (afu->chardev_s) { 588 cdev_del(&afu->afu_cdev_s); 589 device_unregister(afu->chardev_s); 590 afu->chardev_s = NULL; 591 } 592 } 593 594 int cxl_register_afu(struct cxl_afu *afu) 595 { 596 afu->dev.class = cxl_class; 597 598 return device_register(&afu->dev); 599 } 600 601 int cxl_register_adapter(struct cxl *adapter) 602 { 603 adapter->dev.class = cxl_class; 604 605 /* 606 * Future: When we support dynamically reprogramming the PSL & AFU we 607 * will expose the interface to do that via a chardev: 608 * adapter->dev.devt = CXL_CARD_MKDEV(adapter); 609 */ 610 611 return device_register(&adapter->dev); 612 } 613 614 dev_t cxl_get_dev(void) 615 { 616 return cxl_dev; 617 } 618 619 int __init cxl_file_init(void) 620 { 621 int rc; 622 623 /* 624 * If these change we really need to update API. Either change some 625 * flags or update API version number CXL_API_VERSION. 626 */ 627 BUILD_BUG_ON(CXL_API_VERSION != 3); 628 BUILD_BUG_ON(sizeof(struct cxl_ioctl_start_work) != 64); 629 BUILD_BUG_ON(sizeof(struct cxl_event_header) != 8); 630 BUILD_BUG_ON(sizeof(struct cxl_event_afu_interrupt) != 8); 631 BUILD_BUG_ON(sizeof(struct cxl_event_data_storage) != 32); 632 BUILD_BUG_ON(sizeof(struct cxl_event_afu_error) != 16); 633 634 if ((rc = alloc_chrdev_region(&cxl_dev, 0, CXL_NUM_MINORS, "cxl"))) { 635 pr_err("Unable to allocate CXL major number: %i\n", rc); 636 return rc; 637 } 638 639 pr_devel("CXL device allocated, MAJOR %i\n", MAJOR(cxl_dev)); 640 641 cxl_class = class_create(THIS_MODULE, "cxl"); 642 if (IS_ERR(cxl_class)) { 643 pr_err("Unable to create CXL class\n"); 644 rc = PTR_ERR(cxl_class); 645 goto err; 646 } 647 cxl_class->devnode = cxl_devnode; 648 649 return 0; 650 651 err: 652 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS); 653 return rc; 654 } 655 656 void cxl_file_exit(void) 657 { 658 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS); 659 class_destroy(cxl_class); 660 } 661