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