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 * We grab the PID here and not in the file open to allow for the case 198 * where a process (master, some daemon, etc) has opened the chardev on 199 * behalf of another process, so the AFU's mm gets bound to the process 200 * that performs this ioctl and not the process that opened the file. 201 * Also we grab the PID of the group leader so that if the task that 202 * has performed the attach operation exits the mm context of the 203 * process is still accessible. 204 */ 205 ctx->pid = get_task_pid(current, PIDTYPE_PID); 206 ctx->glpid = get_task_pid(current->group_leader, PIDTYPE_PID); 207 208 trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr); 209 210 if ((rc = cxl_ops->attach_process(ctx, false, work.work_element_descriptor, 211 amr))) { 212 afu_release_irqs(ctx, ctx); 213 goto out; 214 } 215 216 ctx->status = STARTED; 217 rc = 0; 218 out: 219 mutex_unlock(&ctx->status_mutex); 220 return rc; 221 } 222 223 static long afu_ioctl_process_element(struct cxl_context *ctx, 224 int __user *upe) 225 { 226 pr_devel("%s: pe: %i\n", __func__, ctx->pe); 227 228 if (copy_to_user(upe, &ctx->external_pe, sizeof(__u32))) 229 return -EFAULT; 230 231 return 0; 232 } 233 234 static long afu_ioctl_get_afu_id(struct cxl_context *ctx, 235 struct cxl_afu_id __user *upafuid) 236 { 237 struct cxl_afu_id afuid = { 0 }; 238 239 afuid.card_id = ctx->afu->adapter->adapter_num; 240 afuid.afu_offset = ctx->afu->slice; 241 afuid.afu_mode = ctx->afu->current_mode; 242 243 /* set the flag bit in case the afu is a slave */ 244 if (ctx->afu->current_mode == CXL_MODE_DIRECTED && !ctx->master) 245 afuid.flags |= CXL_AFUID_FLAG_SLAVE; 246 247 if (copy_to_user(upafuid, &afuid, sizeof(afuid))) 248 return -EFAULT; 249 250 return 0; 251 } 252 253 long afu_ioctl(struct file *file, unsigned int cmd, unsigned long arg) 254 { 255 struct cxl_context *ctx = file->private_data; 256 257 if (ctx->status == CLOSED) 258 return -EIO; 259 260 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 261 return -EIO; 262 263 pr_devel("afu_ioctl\n"); 264 switch (cmd) { 265 case CXL_IOCTL_START_WORK: 266 return afu_ioctl_start_work(ctx, (struct cxl_ioctl_start_work __user *)arg); 267 case CXL_IOCTL_GET_PROCESS_ELEMENT: 268 return afu_ioctl_process_element(ctx, (__u32 __user *)arg); 269 case CXL_IOCTL_GET_AFU_ID: 270 return afu_ioctl_get_afu_id(ctx, (struct cxl_afu_id __user *) 271 arg); 272 } 273 return -EINVAL; 274 } 275 276 static long afu_compat_ioctl(struct file *file, unsigned int cmd, 277 unsigned long arg) 278 { 279 return afu_ioctl(file, cmd, arg); 280 } 281 282 int afu_mmap(struct file *file, struct vm_area_struct *vm) 283 { 284 struct cxl_context *ctx = file->private_data; 285 286 /* AFU must be started before we can MMIO */ 287 if (ctx->status != STARTED) 288 return -EIO; 289 290 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 291 return -EIO; 292 293 return cxl_context_iomap(ctx, vm); 294 } 295 296 static inline bool ctx_event_pending(struct cxl_context *ctx) 297 { 298 if (ctx->pending_irq || ctx->pending_fault || ctx->pending_afu_err) 299 return true; 300 301 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) 302 return true; 303 304 return false; 305 } 306 307 unsigned int afu_poll(struct file *file, struct poll_table_struct *poll) 308 { 309 struct cxl_context *ctx = file->private_data; 310 int mask = 0; 311 unsigned long flags; 312 313 314 poll_wait(file, &ctx->wq, poll); 315 316 pr_devel("afu_poll wait done pe: %i\n", ctx->pe); 317 318 spin_lock_irqsave(&ctx->lock, flags); 319 if (ctx_event_pending(ctx)) 320 mask |= POLLIN | POLLRDNORM; 321 else if (ctx->status == CLOSED) 322 /* Only error on closed when there are no futher events pending 323 */ 324 mask |= POLLERR; 325 spin_unlock_irqrestore(&ctx->lock, flags); 326 327 pr_devel("afu_poll pe: %i returning %#x\n", ctx->pe, mask); 328 329 return mask; 330 } 331 332 static ssize_t afu_driver_event_copy(struct cxl_context *ctx, 333 char __user *buf, 334 struct cxl_event *event, 335 struct cxl_event_afu_driver_reserved *pl) 336 { 337 /* Check event */ 338 if (!pl) { 339 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL); 340 return -EFAULT; 341 } 342 343 /* Check event size */ 344 event->header.size += pl->data_size; 345 if (event->header.size > CXL_READ_MIN_SIZE) { 346 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL); 347 return -EFAULT; 348 } 349 350 /* Copy event header */ 351 if (copy_to_user(buf, event, sizeof(struct cxl_event_header))) { 352 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT); 353 return -EFAULT; 354 } 355 356 /* Copy event data */ 357 buf += sizeof(struct cxl_event_header); 358 if (copy_to_user(buf, &pl->data, pl->data_size)) { 359 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT); 360 return -EFAULT; 361 } 362 363 ctx->afu_driver_ops->event_delivered(ctx, pl, 0); /* Success */ 364 return event->header.size; 365 } 366 367 ssize_t afu_read(struct file *file, char __user *buf, size_t count, 368 loff_t *off) 369 { 370 struct cxl_context *ctx = file->private_data; 371 struct cxl_event_afu_driver_reserved *pl = NULL; 372 struct cxl_event event; 373 unsigned long flags; 374 int rc; 375 DEFINE_WAIT(wait); 376 377 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 378 return -EIO; 379 380 if (count < CXL_READ_MIN_SIZE) 381 return -EINVAL; 382 383 spin_lock_irqsave(&ctx->lock, flags); 384 385 for (;;) { 386 prepare_to_wait(&ctx->wq, &wait, TASK_INTERRUPTIBLE); 387 if (ctx_event_pending(ctx) || (ctx->status == CLOSED)) 388 break; 389 390 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) { 391 rc = -EIO; 392 goto out; 393 } 394 395 if (file->f_flags & O_NONBLOCK) { 396 rc = -EAGAIN; 397 goto out; 398 } 399 400 if (signal_pending(current)) { 401 rc = -ERESTARTSYS; 402 goto out; 403 } 404 405 spin_unlock_irqrestore(&ctx->lock, flags); 406 pr_devel("afu_read going to sleep...\n"); 407 schedule(); 408 pr_devel("afu_read woken up\n"); 409 spin_lock_irqsave(&ctx->lock, flags); 410 } 411 412 finish_wait(&ctx->wq, &wait); 413 414 memset(&event, 0, sizeof(event)); 415 event.header.process_element = ctx->pe; 416 event.header.size = sizeof(struct cxl_event_header); 417 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) { 418 pr_devel("afu_read delivering AFU driver specific event\n"); 419 pl = ctx->afu_driver_ops->fetch_event(ctx); 420 atomic_dec(&ctx->afu_driver_events); 421 event.header.type = CXL_EVENT_AFU_DRIVER; 422 } else if (ctx->pending_irq) { 423 pr_devel("afu_read delivering AFU interrupt\n"); 424 event.header.size += sizeof(struct cxl_event_afu_interrupt); 425 event.header.type = CXL_EVENT_AFU_INTERRUPT; 426 event.irq.irq = find_first_bit(ctx->irq_bitmap, ctx->irq_count) + 1; 427 clear_bit(event.irq.irq - 1, ctx->irq_bitmap); 428 if (bitmap_empty(ctx->irq_bitmap, ctx->irq_count)) 429 ctx->pending_irq = false; 430 } else if (ctx->pending_fault) { 431 pr_devel("afu_read delivering data storage fault\n"); 432 event.header.size += sizeof(struct cxl_event_data_storage); 433 event.header.type = CXL_EVENT_DATA_STORAGE; 434 event.fault.addr = ctx->fault_addr; 435 event.fault.dsisr = ctx->fault_dsisr; 436 ctx->pending_fault = false; 437 } else if (ctx->pending_afu_err) { 438 pr_devel("afu_read delivering afu error\n"); 439 event.header.size += sizeof(struct cxl_event_afu_error); 440 event.header.type = CXL_EVENT_AFU_ERROR; 441 event.afu_error.error = ctx->afu_err; 442 ctx->pending_afu_err = false; 443 } else if (ctx->status == CLOSED) { 444 pr_devel("afu_read fatal error\n"); 445 spin_unlock_irqrestore(&ctx->lock, flags); 446 return -EIO; 447 } else 448 WARN(1, "afu_read must be buggy\n"); 449 450 spin_unlock_irqrestore(&ctx->lock, flags); 451 452 if (event.header.type == CXL_EVENT_AFU_DRIVER) 453 return afu_driver_event_copy(ctx, buf, &event, pl); 454 455 if (copy_to_user(buf, &event, event.header.size)) 456 return -EFAULT; 457 return event.header.size; 458 459 out: 460 finish_wait(&ctx->wq, &wait); 461 spin_unlock_irqrestore(&ctx->lock, flags); 462 return rc; 463 } 464 465 /* 466 * Note: if this is updated, we need to update api.c to patch the new ones in 467 * too 468 */ 469 const struct file_operations afu_fops = { 470 .owner = THIS_MODULE, 471 .open = afu_open, 472 .poll = afu_poll, 473 .read = afu_read, 474 .release = afu_release, 475 .unlocked_ioctl = afu_ioctl, 476 .compat_ioctl = afu_compat_ioctl, 477 .mmap = afu_mmap, 478 }; 479 480 static const struct file_operations afu_master_fops = { 481 .owner = THIS_MODULE, 482 .open = afu_master_open, 483 .poll = afu_poll, 484 .read = afu_read, 485 .release = afu_release, 486 .unlocked_ioctl = afu_ioctl, 487 .compat_ioctl = afu_compat_ioctl, 488 .mmap = afu_mmap, 489 }; 490 491 492 static char *cxl_devnode(struct device *dev, umode_t *mode) 493 { 494 if (cpu_has_feature(CPU_FTR_HVMODE) && 495 CXL_DEVT_IS_CARD(dev->devt)) { 496 /* 497 * These minor numbers will eventually be used to program the 498 * PSL and AFUs once we have dynamic reprogramming support 499 */ 500 return NULL; 501 } 502 return kasprintf(GFP_KERNEL, "cxl/%s", dev_name(dev)); 503 } 504 505 extern struct class *cxl_class; 506 507 static int cxl_add_chardev(struct cxl_afu *afu, dev_t devt, struct cdev *cdev, 508 struct device **chardev, char *postfix, char *desc, 509 const struct file_operations *fops) 510 { 511 struct device *dev; 512 int rc; 513 514 cdev_init(cdev, fops); 515 if ((rc = cdev_add(cdev, devt, 1))) { 516 dev_err(&afu->dev, "Unable to add %s chardev: %i\n", desc, rc); 517 return rc; 518 } 519 520 dev = device_create(cxl_class, &afu->dev, devt, afu, 521 "afu%i.%i%s", afu->adapter->adapter_num, afu->slice, postfix); 522 if (IS_ERR(dev)) { 523 dev_err(&afu->dev, "Unable to create %s chardev in sysfs: %i\n", desc, rc); 524 rc = PTR_ERR(dev); 525 goto err; 526 } 527 528 *chardev = dev; 529 530 return 0; 531 err: 532 cdev_del(cdev); 533 return rc; 534 } 535 536 int cxl_chardev_d_afu_add(struct cxl_afu *afu) 537 { 538 return cxl_add_chardev(afu, CXL_AFU_MKDEV_D(afu), &afu->afu_cdev_d, 539 &afu->chardev_d, "d", "dedicated", 540 &afu_master_fops); /* Uses master fops */ 541 } 542 543 int cxl_chardev_m_afu_add(struct cxl_afu *afu) 544 { 545 return cxl_add_chardev(afu, CXL_AFU_MKDEV_M(afu), &afu->afu_cdev_m, 546 &afu->chardev_m, "m", "master", 547 &afu_master_fops); 548 } 549 550 int cxl_chardev_s_afu_add(struct cxl_afu *afu) 551 { 552 return cxl_add_chardev(afu, CXL_AFU_MKDEV_S(afu), &afu->afu_cdev_s, 553 &afu->chardev_s, "s", "shared", 554 &afu_fops); 555 } 556 557 void cxl_chardev_afu_remove(struct cxl_afu *afu) 558 { 559 if (afu->chardev_d) { 560 cdev_del(&afu->afu_cdev_d); 561 device_unregister(afu->chardev_d); 562 afu->chardev_d = NULL; 563 } 564 if (afu->chardev_m) { 565 cdev_del(&afu->afu_cdev_m); 566 device_unregister(afu->chardev_m); 567 afu->chardev_m = NULL; 568 } 569 if (afu->chardev_s) { 570 cdev_del(&afu->afu_cdev_s); 571 device_unregister(afu->chardev_s); 572 afu->chardev_s = NULL; 573 } 574 } 575 576 int cxl_register_afu(struct cxl_afu *afu) 577 { 578 afu->dev.class = cxl_class; 579 580 return device_register(&afu->dev); 581 } 582 583 int cxl_register_adapter(struct cxl *adapter) 584 { 585 adapter->dev.class = cxl_class; 586 587 /* 588 * Future: When we support dynamically reprogramming the PSL & AFU we 589 * will expose the interface to do that via a chardev: 590 * adapter->dev.devt = CXL_CARD_MKDEV(adapter); 591 */ 592 593 return device_register(&adapter->dev); 594 } 595 596 dev_t cxl_get_dev(void) 597 { 598 return cxl_dev; 599 } 600 601 int __init cxl_file_init(void) 602 { 603 int rc; 604 605 /* 606 * If these change we really need to update API. Either change some 607 * flags or update API version number CXL_API_VERSION. 608 */ 609 BUILD_BUG_ON(CXL_API_VERSION != 3); 610 BUILD_BUG_ON(sizeof(struct cxl_ioctl_start_work) != 64); 611 BUILD_BUG_ON(sizeof(struct cxl_event_header) != 8); 612 BUILD_BUG_ON(sizeof(struct cxl_event_afu_interrupt) != 8); 613 BUILD_BUG_ON(sizeof(struct cxl_event_data_storage) != 32); 614 BUILD_BUG_ON(sizeof(struct cxl_event_afu_error) != 16); 615 616 if ((rc = alloc_chrdev_region(&cxl_dev, 0, CXL_NUM_MINORS, "cxl"))) { 617 pr_err("Unable to allocate CXL major number: %i\n", rc); 618 return rc; 619 } 620 621 pr_devel("CXL device allocated, MAJOR %i\n", MAJOR(cxl_dev)); 622 623 cxl_class = class_create(THIS_MODULE, "cxl"); 624 if (IS_ERR(cxl_class)) { 625 pr_err("Unable to create CXL class\n"); 626 rc = PTR_ERR(cxl_class); 627 goto err; 628 } 629 cxl_class->devnode = cxl_devnode; 630 631 return 0; 632 633 err: 634 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS); 635 return rc; 636 } 637 638 void cxl_file_exit(void) 639 { 640 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS); 641 class_destroy(cxl_class); 642 } 643