Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright 2014 IBM Corp.
4 */
5
6#include <linux/spinlock.h>
7#include <linux/module.h>
8#include <linux/export.h>
9#include <linux/kernel.h>
10#include <linux/bitmap.h>
11#include <linux/sched/signal.h>
12#include <linux/poll.h>
13#include <linux/pid.h>
14#include <linux/fs.h>
15#include <linux/mm.h>
16#include <linux/slab.h>
17#include <linux/sched/mm.h>
18#include <linux/mmu_context.h>
19#include <asm/cputable.h>
20#include <asm/current.h>
21#include <asm/copro.h>
22
23#include "cxl.h"
24#include "trace.h"
25
26#define CXL_NUM_MINORS 256 /* Total to reserve */
27
28#define CXL_AFU_MINOR_D(afu) (CXL_CARD_MINOR(afu->adapter) + 1 + (3 * afu->slice))
29#define CXL_AFU_MINOR_M(afu) (CXL_AFU_MINOR_D(afu) + 1)
30#define CXL_AFU_MINOR_S(afu) (CXL_AFU_MINOR_D(afu) + 2)
31#define CXL_AFU_MKDEV_D(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_D(afu))
32#define CXL_AFU_MKDEV_M(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_M(afu))
33#define CXL_AFU_MKDEV_S(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_S(afu))
34
35#define CXL_DEVT_AFU(dev) ((MINOR(dev) % CXL_DEV_MINORS - 1) / 3)
36
37#define CXL_DEVT_IS_CARD(dev) (MINOR(dev) % CXL_DEV_MINORS == 0)
38
39static dev_t cxl_dev;
40
41static int __afu_open(struct inode *inode, struct file *file, bool master)
42{
43 struct cxl *adapter;
44 struct cxl_afu *afu;
45 struct cxl_context *ctx;
46 int adapter_num = CXL_DEVT_ADAPTER(inode->i_rdev);
47 int slice = CXL_DEVT_AFU(inode->i_rdev);
48 int rc = -ENODEV;
49
50 pr_devel("afu_open afu%i.%i\n", slice, adapter_num);
51
52 if (!(adapter = get_cxl_adapter(adapter_num)))
53 return -ENODEV;
54
55 if (slice > adapter->slices)
56 goto err_put_adapter;
57
58 spin_lock(&adapter->afu_list_lock);
59 if (!(afu = adapter->afu[slice])) {
60 spin_unlock(&adapter->afu_list_lock);
61 goto err_put_adapter;
62 }
63
64 /*
65 * taking a ref to the afu so that it doesn't go away
66 * for rest of the function. This ref is released before
67 * we return.
68 */
69 cxl_afu_get(afu);
70 spin_unlock(&adapter->afu_list_lock);
71
72 if (!afu->current_mode)
73 goto err_put_afu;
74
75 if (!cxl_ops->link_ok(adapter, afu)) {
76 rc = -EIO;
77 goto err_put_afu;
78 }
79
80 if (!(ctx = cxl_context_alloc())) {
81 rc = -ENOMEM;
82 goto err_put_afu;
83 }
84
85 rc = cxl_context_init(ctx, afu, master);
86 if (rc)
87 goto err_put_afu;
88
89 cxl_context_set_mapping(ctx, inode->i_mapping);
90
91 pr_devel("afu_open pe: %i\n", ctx->pe);
92 file->private_data = ctx;
93
94 /* indicate success */
95 rc = 0;
96
97err_put_afu:
98 /* release the ref taken earlier */
99 cxl_afu_put(afu);
100err_put_adapter:
101 put_device(&adapter->dev);
102 return rc;
103}
104
105int afu_open(struct inode *inode, struct file *file)
106{
107 return __afu_open(inode, file, false);
108}
109
110static int afu_master_open(struct inode *inode, struct file *file)
111{
112 return __afu_open(inode, file, true);
113}
114
115int afu_release(struct inode *inode, struct file *file)
116{
117 struct cxl_context *ctx = file->private_data;
118
119 pr_devel("%s: closing cxl file descriptor. pe: %i\n",
120 __func__, ctx->pe);
121 cxl_context_detach(ctx);
122
123
124 /*
125 * Delete the context's mapping pointer, unless it's created by the
126 * kernel API, in which case leave it so it can be freed by reclaim_ctx()
127 */
128 if (!ctx->kernelapi) {
129 mutex_lock(&ctx->mapping_lock);
130 ctx->mapping = NULL;
131 mutex_unlock(&ctx->mapping_lock);
132 }
133
134 /*
135 * At this this point all bottom halfs have finished and we should be
136 * getting no more IRQs from the hardware for this context. Once it's
137 * removed from the IDR (and RCU synchronised) it's safe to free the
138 * sstp and context.
139 */
140 cxl_context_free(ctx);
141
142 return 0;
143}
144
145static long afu_ioctl_start_work(struct cxl_context *ctx,
146 struct cxl_ioctl_start_work __user *uwork)
147{
148 struct cxl_ioctl_start_work work;
149 u64 amr = 0;
150 int rc;
151
152 pr_devel("%s: pe: %i\n", __func__, ctx->pe);
153
154 /* Do this outside the status_mutex to avoid a circular dependency with
155 * the locking in cxl_mmap_fault() */
156 if (copy_from_user(&work, uwork, sizeof(work)))
157 return -EFAULT;
158
159 mutex_lock(&ctx->status_mutex);
160 if (ctx->status != OPENED) {
161 rc = -EIO;
162 goto out;
163 }
164
165 /*
166 * if any of the reserved fields are set or any of the unused
167 * flags are set it's invalid
168 */
169 if (work.reserved1 || work.reserved2 || work.reserved3 ||
170 work.reserved4 || work.reserved5 ||
171 (work.flags & ~CXL_START_WORK_ALL)) {
172 rc = -EINVAL;
173 goto out;
174 }
175
176 if (!(work.flags & CXL_START_WORK_NUM_IRQS))
177 work.num_interrupts = ctx->afu->pp_irqs;
178 else if ((work.num_interrupts < ctx->afu->pp_irqs) ||
179 (work.num_interrupts > ctx->afu->irqs_max)) {
180 rc = -EINVAL;
181 goto out;
182 }
183
184 if ((rc = afu_register_irqs(ctx, work.num_interrupts)))
185 goto out;
186
187 if (work.flags & CXL_START_WORK_AMR)
188 amr = work.amr & mfspr(SPRN_UAMOR);
189
190 if (work.flags & CXL_START_WORK_TID)
191 ctx->assign_tidr = true;
192
193 ctx->mmio_err_ff = !!(work.flags & CXL_START_WORK_ERR_FF);
194
195 /*
196 * Increment the mapped context count for adapter. This also checks
197 * if adapter_context_lock is taken.
198 */
199 rc = cxl_adapter_context_get(ctx->afu->adapter);
200 if (rc) {
201 afu_release_irqs(ctx, ctx);
202 goto out;
203 }
204
205 /*
206 * We grab the PID here and not in the file open to allow for the case
207 * where a process (master, some daemon, etc) has opened the chardev on
208 * behalf of another process, so the AFU's mm gets bound to the process
209 * that performs this ioctl and not the process that opened the file.
210 * Also we grab the PID of the group leader so that if the task that
211 * has performed the attach operation exits the mm context of the
212 * process is still accessible.
213 */
214 ctx->pid = get_task_pid(current, PIDTYPE_PID);
215
216 /* acquire a reference to the task's mm */
217 ctx->mm = get_task_mm(current);
218
219 /* ensure this mm_struct can't be freed */
220 cxl_context_mm_count_get(ctx);
221
222 if (ctx->mm) {
223 /* decrement the use count from above */
224 mmput(ctx->mm);
225 /* make TLBIs for this context global */
226 mm_context_add_copro(ctx->mm);
227 }
228
229 /*
230 * Increment driver use count. Enables global TLBIs for hash
231 * and callbacks to handle the segment table
232 */
233 cxl_ctx_get();
234
235 /*
236 * A barrier is needed to make sure all TLBIs are global
237 * before we attach and the context starts being used by the
238 * adapter.
239 *
240 * Needed after mm_context_add_copro() for radix and
241 * cxl_ctx_get() for hash/p8.
242 *
243 * The barrier should really be mb(), since it involves a
244 * device. However, it's only useful when we have local
245 * vs. global TLBIs, i.e SMP=y. So keep smp_mb().
246 */
247 smp_mb();
248
249 trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr);
250
251 if ((rc = cxl_ops->attach_process(ctx, false, work.work_element_descriptor,
252 amr))) {
253 afu_release_irqs(ctx, ctx);
254 cxl_adapter_context_put(ctx->afu->adapter);
255 put_pid(ctx->pid);
256 ctx->pid = NULL;
257 cxl_ctx_put();
258 cxl_context_mm_count_put(ctx);
259 if (ctx->mm)
260 mm_context_remove_copro(ctx->mm);
261 goto out;
262 }
263
264 rc = 0;
265 if (work.flags & CXL_START_WORK_TID) {
266 work.tid = ctx->tidr;
267 if (copy_to_user(uwork, &work, sizeof(work)))
268 rc = -EFAULT;
269 }
270
271 ctx->status = STARTED;
272
273out:
274 mutex_unlock(&ctx->status_mutex);
275 return rc;
276}
277
278static long afu_ioctl_process_element(struct cxl_context *ctx,
279 int __user *upe)
280{
281 pr_devel("%s: pe: %i\n", __func__, ctx->pe);
282
283 if (copy_to_user(upe, &ctx->external_pe, sizeof(__u32)))
284 return -EFAULT;
285
286 return 0;
287}
288
289static long afu_ioctl_get_afu_id(struct cxl_context *ctx,
290 struct cxl_afu_id __user *upafuid)
291{
292 struct cxl_afu_id afuid = { 0 };
293
294 afuid.card_id = ctx->afu->adapter->adapter_num;
295 afuid.afu_offset = ctx->afu->slice;
296 afuid.afu_mode = ctx->afu->current_mode;
297
298 /* set the flag bit in case the afu is a slave */
299 if (ctx->afu->current_mode == CXL_MODE_DIRECTED && !ctx->master)
300 afuid.flags |= CXL_AFUID_FLAG_SLAVE;
301
302 if (copy_to_user(upafuid, &afuid, sizeof(afuid)))
303 return -EFAULT;
304
305 return 0;
306}
307
308long afu_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
309{
310 struct cxl_context *ctx = file->private_data;
311
312 if (ctx->status == CLOSED)
313 return -EIO;
314
315 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
316 return -EIO;
317
318 pr_devel("afu_ioctl\n");
319 switch (cmd) {
320 case CXL_IOCTL_START_WORK:
321 return afu_ioctl_start_work(ctx, (struct cxl_ioctl_start_work __user *)arg);
322 case CXL_IOCTL_GET_PROCESS_ELEMENT:
323 return afu_ioctl_process_element(ctx, (__u32 __user *)arg);
324 case CXL_IOCTL_GET_AFU_ID:
325 return afu_ioctl_get_afu_id(ctx, (struct cxl_afu_id __user *)
326 arg);
327 }
328 return -EINVAL;
329}
330
331static long afu_compat_ioctl(struct file *file, unsigned int cmd,
332 unsigned long arg)
333{
334 return afu_ioctl(file, cmd, arg);
335}
336
337int afu_mmap(struct file *file, struct vm_area_struct *vm)
338{
339 struct cxl_context *ctx = file->private_data;
340
341 /* AFU must be started before we can MMIO */
342 if (ctx->status != STARTED)
343 return -EIO;
344
345 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
346 return -EIO;
347
348 return cxl_context_iomap(ctx, vm);
349}
350
351static inline bool ctx_event_pending(struct cxl_context *ctx)
352{
353 if (ctx->pending_irq || ctx->pending_fault || ctx->pending_afu_err)
354 return true;
355
356 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events))
357 return true;
358
359 return false;
360}
361
362__poll_t afu_poll(struct file *file, struct poll_table_struct *poll)
363{
364 struct cxl_context *ctx = file->private_data;
365 __poll_t mask = 0;
366 unsigned long flags;
367
368
369 poll_wait(file, &ctx->wq, poll);
370
371 pr_devel("afu_poll wait done pe: %i\n", ctx->pe);
372
373 spin_lock_irqsave(&ctx->lock, flags);
374 if (ctx_event_pending(ctx))
375 mask |= EPOLLIN | EPOLLRDNORM;
376 else if (ctx->status == CLOSED)
377 /* Only error on closed when there are no futher events pending
378 */
379 mask |= EPOLLERR;
380 spin_unlock_irqrestore(&ctx->lock, flags);
381
382 pr_devel("afu_poll pe: %i returning %#x\n", ctx->pe, mask);
383
384 return mask;
385}
386
387static ssize_t afu_driver_event_copy(struct cxl_context *ctx,
388 char __user *buf,
389 struct cxl_event *event,
390 struct cxl_event_afu_driver_reserved *pl)
391{
392 /* Check event */
393 if (!pl) {
394 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
395 return -EFAULT;
396 }
397
398 /* Check event size */
399 event->header.size += pl->data_size;
400 if (event->header.size > CXL_READ_MIN_SIZE) {
401 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
402 return -EFAULT;
403 }
404
405 /* Copy event header */
406 if (copy_to_user(buf, event, sizeof(struct cxl_event_header))) {
407 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
408 return -EFAULT;
409 }
410
411 /* Copy event data */
412 buf += sizeof(struct cxl_event_header);
413 if (copy_to_user(buf, &pl->data, pl->data_size)) {
414 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
415 return -EFAULT;
416 }
417
418 ctx->afu_driver_ops->event_delivered(ctx, pl, 0); /* Success */
419 return event->header.size;
420}
421
422ssize_t afu_read(struct file *file, char __user *buf, size_t count,
423 loff_t *off)
424{
425 struct cxl_context *ctx = file->private_data;
426 struct cxl_event_afu_driver_reserved *pl = NULL;
427 struct cxl_event event;
428 unsigned long flags;
429 int rc;
430 DEFINE_WAIT(wait);
431
432 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
433 return -EIO;
434
435 if (count < CXL_READ_MIN_SIZE)
436 return -EINVAL;
437
438 spin_lock_irqsave(&ctx->lock, flags);
439
440 for (;;) {
441 prepare_to_wait(&ctx->wq, &wait, TASK_INTERRUPTIBLE);
442 if (ctx_event_pending(ctx) || (ctx->status == CLOSED))
443 break;
444
445 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
446 rc = -EIO;
447 goto out;
448 }
449
450 if (file->f_flags & O_NONBLOCK) {
451 rc = -EAGAIN;
452 goto out;
453 }
454
455 if (signal_pending(current)) {
456 rc = -ERESTARTSYS;
457 goto out;
458 }
459
460 spin_unlock_irqrestore(&ctx->lock, flags);
461 pr_devel("afu_read going to sleep...\n");
462 schedule();
463 pr_devel("afu_read woken up\n");
464 spin_lock_irqsave(&ctx->lock, flags);
465 }
466
467 finish_wait(&ctx->wq, &wait);
468
469 memset(&event, 0, sizeof(event));
470 event.header.process_element = ctx->pe;
471 event.header.size = sizeof(struct cxl_event_header);
472 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) {
473 pr_devel("afu_read delivering AFU driver specific event\n");
474 pl = ctx->afu_driver_ops->fetch_event(ctx);
475 atomic_dec(&ctx->afu_driver_events);
476 event.header.type = CXL_EVENT_AFU_DRIVER;
477 } else if (ctx->pending_irq) {
478 pr_devel("afu_read delivering AFU interrupt\n");
479 event.header.size += sizeof(struct cxl_event_afu_interrupt);
480 event.header.type = CXL_EVENT_AFU_INTERRUPT;
481 event.irq.irq = find_first_bit(ctx->irq_bitmap, ctx->irq_count) + 1;
482 clear_bit(event.irq.irq - 1, ctx->irq_bitmap);
483 if (bitmap_empty(ctx->irq_bitmap, ctx->irq_count))
484 ctx->pending_irq = false;
485 } else if (ctx->pending_fault) {
486 pr_devel("afu_read delivering data storage fault\n");
487 event.header.size += sizeof(struct cxl_event_data_storage);
488 event.header.type = CXL_EVENT_DATA_STORAGE;
489 event.fault.addr = ctx->fault_addr;
490 event.fault.dsisr = ctx->fault_dsisr;
491 ctx->pending_fault = false;
492 } else if (ctx->pending_afu_err) {
493 pr_devel("afu_read delivering afu error\n");
494 event.header.size += sizeof(struct cxl_event_afu_error);
495 event.header.type = CXL_EVENT_AFU_ERROR;
496 event.afu_error.error = ctx->afu_err;
497 ctx->pending_afu_err = false;
498 } else if (ctx->status == CLOSED) {
499 pr_devel("afu_read fatal error\n");
500 spin_unlock_irqrestore(&ctx->lock, flags);
501 return -EIO;
502 } else
503 WARN(1, "afu_read must be buggy\n");
504
505 spin_unlock_irqrestore(&ctx->lock, flags);
506
507 if (event.header.type == CXL_EVENT_AFU_DRIVER)
508 return afu_driver_event_copy(ctx, buf, &event, pl);
509
510 if (copy_to_user(buf, &event, event.header.size))
511 return -EFAULT;
512 return event.header.size;
513
514out:
515 finish_wait(&ctx->wq, &wait);
516 spin_unlock_irqrestore(&ctx->lock, flags);
517 return rc;
518}
519
520/*
521 * Note: if this is updated, we need to update api.c to patch the new ones in
522 * too
523 */
524const struct file_operations afu_fops = {
525 .owner = THIS_MODULE,
526 .open = afu_open,
527 .poll = afu_poll,
528 .read = afu_read,
529 .release = afu_release,
530 .unlocked_ioctl = afu_ioctl,
531 .compat_ioctl = afu_compat_ioctl,
532 .mmap = afu_mmap,
533};
534
535static const struct file_operations afu_master_fops = {
536 .owner = THIS_MODULE,
537 .open = afu_master_open,
538 .poll = afu_poll,
539 .read = afu_read,
540 .release = afu_release,
541 .unlocked_ioctl = afu_ioctl,
542 .compat_ioctl = afu_compat_ioctl,
543 .mmap = afu_mmap,
544};
545
546
547static char *cxl_devnode(const struct device *dev, umode_t *mode)
548{
549 if (cpu_has_feature(CPU_FTR_HVMODE) &&
550 CXL_DEVT_IS_CARD(dev->devt)) {
551 /*
552 * These minor numbers will eventually be used to program the
553 * PSL and AFUs once we have dynamic reprogramming support
554 */
555 return NULL;
556 }
557 return kasprintf(GFP_KERNEL, "cxl/%s", dev_name(dev));
558}
559
560static const struct class cxl_class = {
561 .name = "cxl",
562 .devnode = cxl_devnode,
563};
564
565static int cxl_add_chardev(struct cxl_afu *afu, dev_t devt, struct cdev *cdev,
566 struct device **chardev, char *postfix, char *desc,
567 const struct file_operations *fops)
568{
569 struct device *dev;
570 int rc;
571
572 cdev_init(cdev, fops);
573 rc = cdev_add(cdev, devt, 1);
574 if (rc) {
575 dev_err(&afu->dev, "Unable to add %s chardev: %i\n", desc, rc);
576 return rc;
577 }
578
579 dev = device_create(&cxl_class, &afu->dev, devt, afu,
580 "afu%i.%i%s", afu->adapter->adapter_num, afu->slice, postfix);
581 if (IS_ERR(dev)) {
582 rc = PTR_ERR(dev);
583 dev_err(&afu->dev, "Unable to create %s chardev in sysfs: %i\n", desc, rc);
584 goto err;
585 }
586
587 *chardev = dev;
588
589 return 0;
590err:
591 cdev_del(cdev);
592 return rc;
593}
594
595int cxl_chardev_d_afu_add(struct cxl_afu *afu)
596{
597 return cxl_add_chardev(afu, CXL_AFU_MKDEV_D(afu), &afu->afu_cdev_d,
598 &afu->chardev_d, "d", "dedicated",
599 &afu_master_fops); /* Uses master fops */
600}
601
602int cxl_chardev_m_afu_add(struct cxl_afu *afu)
603{
604 return cxl_add_chardev(afu, CXL_AFU_MKDEV_M(afu), &afu->afu_cdev_m,
605 &afu->chardev_m, "m", "master",
606 &afu_master_fops);
607}
608
609int cxl_chardev_s_afu_add(struct cxl_afu *afu)
610{
611 return cxl_add_chardev(afu, CXL_AFU_MKDEV_S(afu), &afu->afu_cdev_s,
612 &afu->chardev_s, "s", "shared",
613 &afu_fops);
614}
615
616void cxl_chardev_afu_remove(struct cxl_afu *afu)
617{
618 if (afu->chardev_d) {
619 cdev_del(&afu->afu_cdev_d);
620 device_unregister(afu->chardev_d);
621 afu->chardev_d = NULL;
622 }
623 if (afu->chardev_m) {
624 cdev_del(&afu->afu_cdev_m);
625 device_unregister(afu->chardev_m);
626 afu->chardev_m = NULL;
627 }
628 if (afu->chardev_s) {
629 cdev_del(&afu->afu_cdev_s);
630 device_unregister(afu->chardev_s);
631 afu->chardev_s = NULL;
632 }
633}
634
635int cxl_register_afu(struct cxl_afu *afu)
636{
637 afu->dev.class = &cxl_class;
638
639 return device_register(&afu->dev);
640}
641
642int cxl_register_adapter(struct cxl *adapter)
643{
644 adapter->dev.class = &cxl_class;
645
646 /*
647 * Future: When we support dynamically reprogramming the PSL & AFU we
648 * will expose the interface to do that via a chardev:
649 * adapter->dev.devt = CXL_CARD_MKDEV(adapter);
650 */
651
652 return device_register(&adapter->dev);
653}
654
655dev_t cxl_get_dev(void)
656{
657 return cxl_dev;
658}
659
660int __init cxl_file_init(void)
661{
662 int rc;
663
664 /*
665 * If these change we really need to update API. Either change some
666 * flags or update API version number CXL_API_VERSION.
667 */
668 BUILD_BUG_ON(CXL_API_VERSION != 3);
669 BUILD_BUG_ON(sizeof(struct cxl_ioctl_start_work) != 64);
670 BUILD_BUG_ON(sizeof(struct cxl_event_header) != 8);
671 BUILD_BUG_ON(sizeof(struct cxl_event_afu_interrupt) != 8);
672 BUILD_BUG_ON(sizeof(struct cxl_event_data_storage) != 32);
673 BUILD_BUG_ON(sizeof(struct cxl_event_afu_error) != 16);
674
675 if ((rc = alloc_chrdev_region(&cxl_dev, 0, CXL_NUM_MINORS, "cxl"))) {
676 pr_err("Unable to allocate CXL major number: %i\n", rc);
677 return rc;
678 }
679
680 pr_devel("CXL device allocated, MAJOR %i\n", MAJOR(cxl_dev));
681
682 rc = class_register(&cxl_class);
683 if (rc) {
684 pr_err("Unable to create CXL class\n");
685 goto err;
686 }
687
688 return 0;
689
690err:
691 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
692 return rc;
693}
694
695void cxl_file_exit(void)
696{
697 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
698 class_unregister(&cxl_class);
699}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright 2014 IBM Corp.
4 */
5
6#include <linux/spinlock.h>
7#include <linux/module.h>
8#include <linux/export.h>
9#include <linux/kernel.h>
10#include <linux/bitmap.h>
11#include <linux/sched/signal.h>
12#include <linux/poll.h>
13#include <linux/pid.h>
14#include <linux/fs.h>
15#include <linux/mm.h>
16#include <linux/slab.h>
17#include <linux/sched/mm.h>
18#include <linux/mmu_context.h>
19#include <asm/cputable.h>
20#include <asm/current.h>
21#include <asm/copro.h>
22
23#include "cxl.h"
24#include "trace.h"
25
26#define CXL_NUM_MINORS 256 /* Total to reserve */
27
28#define CXL_AFU_MINOR_D(afu) (CXL_CARD_MINOR(afu->adapter) + 1 + (3 * afu->slice))
29#define CXL_AFU_MINOR_M(afu) (CXL_AFU_MINOR_D(afu) + 1)
30#define CXL_AFU_MINOR_S(afu) (CXL_AFU_MINOR_D(afu) + 2)
31#define CXL_AFU_MKDEV_D(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_D(afu))
32#define CXL_AFU_MKDEV_M(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_M(afu))
33#define CXL_AFU_MKDEV_S(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_S(afu))
34
35#define CXL_DEVT_AFU(dev) ((MINOR(dev) % CXL_DEV_MINORS - 1) / 3)
36
37#define CXL_DEVT_IS_CARD(dev) (MINOR(dev) % CXL_DEV_MINORS == 0)
38
39static dev_t cxl_dev;
40
41static struct class *cxl_class;
42
43static int __afu_open(struct inode *inode, struct file *file, bool master)
44{
45 struct cxl *adapter;
46 struct cxl_afu *afu;
47 struct cxl_context *ctx;
48 int adapter_num = CXL_DEVT_ADAPTER(inode->i_rdev);
49 int slice = CXL_DEVT_AFU(inode->i_rdev);
50 int rc = -ENODEV;
51
52 pr_devel("afu_open afu%i.%i\n", slice, adapter_num);
53
54 if (!(adapter = get_cxl_adapter(adapter_num)))
55 return -ENODEV;
56
57 if (slice > adapter->slices)
58 goto err_put_adapter;
59
60 spin_lock(&adapter->afu_list_lock);
61 if (!(afu = adapter->afu[slice])) {
62 spin_unlock(&adapter->afu_list_lock);
63 goto err_put_adapter;
64 }
65
66 /*
67 * taking a ref to the afu so that it doesn't go away
68 * for rest of the function. This ref is released before
69 * we return.
70 */
71 cxl_afu_get(afu);
72 spin_unlock(&adapter->afu_list_lock);
73
74 if (!afu->current_mode)
75 goto err_put_afu;
76
77 if (!cxl_ops->link_ok(adapter, afu)) {
78 rc = -EIO;
79 goto err_put_afu;
80 }
81
82 if (!(ctx = cxl_context_alloc())) {
83 rc = -ENOMEM;
84 goto err_put_afu;
85 }
86
87 rc = cxl_context_init(ctx, afu, master);
88 if (rc)
89 goto err_put_afu;
90
91 cxl_context_set_mapping(ctx, inode->i_mapping);
92
93 pr_devel("afu_open pe: %i\n", ctx->pe);
94 file->private_data = ctx;
95
96 /* indicate success */
97 rc = 0;
98
99err_put_afu:
100 /* release the ref taken earlier */
101 cxl_afu_put(afu);
102err_put_adapter:
103 put_device(&adapter->dev);
104 return rc;
105}
106
107int afu_open(struct inode *inode, struct file *file)
108{
109 return __afu_open(inode, file, false);
110}
111
112static int afu_master_open(struct inode *inode, struct file *file)
113{
114 return __afu_open(inode, file, true);
115}
116
117int 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
147static 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, sizeof(work)))
159 return -EFAULT;
160
161 mutex_lock(&ctx->status_mutex);
162 if (ctx->status != OPENED) {
163 rc = -EIO;
164 goto out;
165 }
166
167 /*
168 * if any of the reserved fields are set or any of the unused
169 * flags are set it's invalid
170 */
171 if (work.reserved1 || work.reserved2 || work.reserved3 ||
172 work.reserved4 || work.reserved5 ||
173 (work.flags & ~CXL_START_WORK_ALL)) {
174 rc = -EINVAL;
175 goto out;
176 }
177
178 if (!(work.flags & CXL_START_WORK_NUM_IRQS))
179 work.num_interrupts = ctx->afu->pp_irqs;
180 else if ((work.num_interrupts < ctx->afu->pp_irqs) ||
181 (work.num_interrupts > ctx->afu->irqs_max)) {
182 rc = -EINVAL;
183 goto out;
184 }
185
186 if ((rc = afu_register_irqs(ctx, work.num_interrupts)))
187 goto out;
188
189 if (work.flags & CXL_START_WORK_AMR)
190 amr = work.amr & mfspr(SPRN_UAMOR);
191
192 if (work.flags & CXL_START_WORK_TID)
193 ctx->assign_tidr = true;
194
195 ctx->mmio_err_ff = !!(work.flags & CXL_START_WORK_ERR_FF);
196
197 /*
198 * Increment the mapped context count for adapter. This also checks
199 * if adapter_context_lock is taken.
200 */
201 rc = cxl_adapter_context_get(ctx->afu->adapter);
202 if (rc) {
203 afu_release_irqs(ctx, ctx);
204 goto out;
205 }
206
207 /*
208 * We grab the PID here and not in the file open to allow for the case
209 * where a process (master, some daemon, etc) has opened the chardev on
210 * behalf of another process, so the AFU's mm gets bound to the process
211 * that performs this ioctl and not the process that opened the file.
212 * Also we grab the PID of the group leader so that if the task that
213 * has performed the attach operation exits the mm context of the
214 * process is still accessible.
215 */
216 ctx->pid = get_task_pid(current, PIDTYPE_PID);
217
218 /* acquire a reference to the task's mm */
219 ctx->mm = get_task_mm(current);
220
221 /* ensure this mm_struct can't be freed */
222 cxl_context_mm_count_get(ctx);
223
224 if (ctx->mm) {
225 /* decrement the use count from above */
226 mmput(ctx->mm);
227 /* make TLBIs for this context global */
228 mm_context_add_copro(ctx->mm);
229 }
230
231 /*
232 * Increment driver use count. Enables global TLBIs for hash
233 * and callbacks to handle the segment table
234 */
235 cxl_ctx_get();
236
237 /*
238 * A barrier is needed to make sure all TLBIs are global
239 * before we attach and the context starts being used by the
240 * adapter.
241 *
242 * Needed after mm_context_add_copro() for radix and
243 * cxl_ctx_get() for hash/p8.
244 *
245 * The barrier should really be mb(), since it involves a
246 * device. However, it's only useful when we have local
247 * vs. global TLBIs, i.e SMP=y. So keep smp_mb().
248 */
249 smp_mb();
250
251 trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr);
252
253 if ((rc = cxl_ops->attach_process(ctx, false, work.work_element_descriptor,
254 amr))) {
255 afu_release_irqs(ctx, ctx);
256 cxl_adapter_context_put(ctx->afu->adapter);
257 put_pid(ctx->pid);
258 ctx->pid = NULL;
259 cxl_ctx_put();
260 cxl_context_mm_count_put(ctx);
261 if (ctx->mm)
262 mm_context_remove_copro(ctx->mm);
263 goto out;
264 }
265
266 rc = 0;
267 if (work.flags & CXL_START_WORK_TID) {
268 work.tid = ctx->tidr;
269 if (copy_to_user(uwork, &work, sizeof(work)))
270 rc = -EFAULT;
271 }
272
273 ctx->status = STARTED;
274
275out:
276 mutex_unlock(&ctx->status_mutex);
277 return rc;
278}
279
280static long afu_ioctl_process_element(struct cxl_context *ctx,
281 int __user *upe)
282{
283 pr_devel("%s: pe: %i\n", __func__, ctx->pe);
284
285 if (copy_to_user(upe, &ctx->external_pe, sizeof(__u32)))
286 return -EFAULT;
287
288 return 0;
289}
290
291static long afu_ioctl_get_afu_id(struct cxl_context *ctx,
292 struct cxl_afu_id __user *upafuid)
293{
294 struct cxl_afu_id afuid = { 0 };
295
296 afuid.card_id = ctx->afu->adapter->adapter_num;
297 afuid.afu_offset = ctx->afu->slice;
298 afuid.afu_mode = ctx->afu->current_mode;
299
300 /* set the flag bit in case the afu is a slave */
301 if (ctx->afu->current_mode == CXL_MODE_DIRECTED && !ctx->master)
302 afuid.flags |= CXL_AFUID_FLAG_SLAVE;
303
304 if (copy_to_user(upafuid, &afuid, sizeof(afuid)))
305 return -EFAULT;
306
307 return 0;
308}
309
310long afu_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
311{
312 struct cxl_context *ctx = file->private_data;
313
314 if (ctx->status == CLOSED)
315 return -EIO;
316
317 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
318 return -EIO;
319
320 pr_devel("afu_ioctl\n");
321 switch (cmd) {
322 case CXL_IOCTL_START_WORK:
323 return afu_ioctl_start_work(ctx, (struct cxl_ioctl_start_work __user *)arg);
324 case CXL_IOCTL_GET_PROCESS_ELEMENT:
325 return afu_ioctl_process_element(ctx, (__u32 __user *)arg);
326 case CXL_IOCTL_GET_AFU_ID:
327 return afu_ioctl_get_afu_id(ctx, (struct cxl_afu_id __user *)
328 arg);
329 }
330 return -EINVAL;
331}
332
333static long afu_compat_ioctl(struct file *file, unsigned int cmd,
334 unsigned long arg)
335{
336 return afu_ioctl(file, cmd, arg);
337}
338
339int afu_mmap(struct file *file, struct vm_area_struct *vm)
340{
341 struct cxl_context *ctx = file->private_data;
342
343 /* AFU must be started before we can MMIO */
344 if (ctx->status != STARTED)
345 return -EIO;
346
347 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
348 return -EIO;
349
350 return cxl_context_iomap(ctx, vm);
351}
352
353static inline bool ctx_event_pending(struct cxl_context *ctx)
354{
355 if (ctx->pending_irq || ctx->pending_fault || ctx->pending_afu_err)
356 return true;
357
358 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events))
359 return true;
360
361 return false;
362}
363
364__poll_t afu_poll(struct file *file, struct poll_table_struct *poll)
365{
366 struct cxl_context *ctx = file->private_data;
367 __poll_t mask = 0;
368 unsigned long flags;
369
370
371 poll_wait(file, &ctx->wq, poll);
372
373 pr_devel("afu_poll wait done pe: %i\n", ctx->pe);
374
375 spin_lock_irqsave(&ctx->lock, flags);
376 if (ctx_event_pending(ctx))
377 mask |= EPOLLIN | EPOLLRDNORM;
378 else if (ctx->status == CLOSED)
379 /* Only error on closed when there are no futher events pending
380 */
381 mask |= EPOLLERR;
382 spin_unlock_irqrestore(&ctx->lock, flags);
383
384 pr_devel("afu_poll pe: %i returning %#x\n", ctx->pe, mask);
385
386 return mask;
387}
388
389static ssize_t afu_driver_event_copy(struct cxl_context *ctx,
390 char __user *buf,
391 struct cxl_event *event,
392 struct cxl_event_afu_driver_reserved *pl)
393{
394 /* Check event */
395 if (!pl) {
396 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
397 return -EFAULT;
398 }
399
400 /* Check event size */
401 event->header.size += pl->data_size;
402 if (event->header.size > CXL_READ_MIN_SIZE) {
403 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
404 return -EFAULT;
405 }
406
407 /* Copy event header */
408 if (copy_to_user(buf, event, sizeof(struct cxl_event_header))) {
409 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
410 return -EFAULT;
411 }
412
413 /* Copy event data */
414 buf += sizeof(struct cxl_event_header);
415 if (copy_to_user(buf, &pl->data, pl->data_size)) {
416 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
417 return -EFAULT;
418 }
419
420 ctx->afu_driver_ops->event_delivered(ctx, pl, 0); /* Success */
421 return event->header.size;
422}
423
424ssize_t afu_read(struct file *file, char __user *buf, size_t count,
425 loff_t *off)
426{
427 struct cxl_context *ctx = file->private_data;
428 struct cxl_event_afu_driver_reserved *pl = NULL;
429 struct cxl_event event;
430 unsigned long flags;
431 int rc;
432 DEFINE_WAIT(wait);
433
434 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
435 return -EIO;
436
437 if (count < CXL_READ_MIN_SIZE)
438 return -EINVAL;
439
440 spin_lock_irqsave(&ctx->lock, flags);
441
442 for (;;) {
443 prepare_to_wait(&ctx->wq, &wait, TASK_INTERRUPTIBLE);
444 if (ctx_event_pending(ctx) || (ctx->status == CLOSED))
445 break;
446
447 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
448 rc = -EIO;
449 goto out;
450 }
451
452 if (file->f_flags & O_NONBLOCK) {
453 rc = -EAGAIN;
454 goto out;
455 }
456
457 if (signal_pending(current)) {
458 rc = -ERESTARTSYS;
459 goto out;
460 }
461
462 spin_unlock_irqrestore(&ctx->lock, flags);
463 pr_devel("afu_read going to sleep...\n");
464 schedule();
465 pr_devel("afu_read woken up\n");
466 spin_lock_irqsave(&ctx->lock, flags);
467 }
468
469 finish_wait(&ctx->wq, &wait);
470
471 memset(&event, 0, sizeof(event));
472 event.header.process_element = ctx->pe;
473 event.header.size = sizeof(struct cxl_event_header);
474 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) {
475 pr_devel("afu_read delivering AFU driver specific event\n");
476 pl = ctx->afu_driver_ops->fetch_event(ctx);
477 atomic_dec(&ctx->afu_driver_events);
478 event.header.type = CXL_EVENT_AFU_DRIVER;
479 } else if (ctx->pending_irq) {
480 pr_devel("afu_read delivering AFU interrupt\n");
481 event.header.size += sizeof(struct cxl_event_afu_interrupt);
482 event.header.type = CXL_EVENT_AFU_INTERRUPT;
483 event.irq.irq = find_first_bit(ctx->irq_bitmap, ctx->irq_count) + 1;
484 clear_bit(event.irq.irq - 1, ctx->irq_bitmap);
485 if (bitmap_empty(ctx->irq_bitmap, ctx->irq_count))
486 ctx->pending_irq = false;
487 } else if (ctx->pending_fault) {
488 pr_devel("afu_read delivering data storage fault\n");
489 event.header.size += sizeof(struct cxl_event_data_storage);
490 event.header.type = CXL_EVENT_DATA_STORAGE;
491 event.fault.addr = ctx->fault_addr;
492 event.fault.dsisr = ctx->fault_dsisr;
493 ctx->pending_fault = false;
494 } else if (ctx->pending_afu_err) {
495 pr_devel("afu_read delivering afu error\n");
496 event.header.size += sizeof(struct cxl_event_afu_error);
497 event.header.type = CXL_EVENT_AFU_ERROR;
498 event.afu_error.error = ctx->afu_err;
499 ctx->pending_afu_err = false;
500 } else if (ctx->status == CLOSED) {
501 pr_devel("afu_read fatal error\n");
502 spin_unlock_irqrestore(&ctx->lock, flags);
503 return -EIO;
504 } else
505 WARN(1, "afu_read must be buggy\n");
506
507 spin_unlock_irqrestore(&ctx->lock, flags);
508
509 if (event.header.type == CXL_EVENT_AFU_DRIVER)
510 return afu_driver_event_copy(ctx, buf, &event, pl);
511
512 if (copy_to_user(buf, &event, event.header.size))
513 return -EFAULT;
514 return event.header.size;
515
516out:
517 finish_wait(&ctx->wq, &wait);
518 spin_unlock_irqrestore(&ctx->lock, flags);
519 return rc;
520}
521
522/*
523 * Note: if this is updated, we need to update api.c to patch the new ones in
524 * too
525 */
526const struct file_operations afu_fops = {
527 .owner = THIS_MODULE,
528 .open = afu_open,
529 .poll = afu_poll,
530 .read = afu_read,
531 .release = afu_release,
532 .unlocked_ioctl = afu_ioctl,
533 .compat_ioctl = afu_compat_ioctl,
534 .mmap = afu_mmap,
535};
536
537static const struct file_operations afu_master_fops = {
538 .owner = THIS_MODULE,
539 .open = afu_master_open,
540 .poll = afu_poll,
541 .read = afu_read,
542 .release = afu_release,
543 .unlocked_ioctl = afu_ioctl,
544 .compat_ioctl = afu_compat_ioctl,
545 .mmap = afu_mmap,
546};
547
548
549static char *cxl_devnode(struct device *dev, umode_t *mode)
550{
551 if (cpu_has_feature(CPU_FTR_HVMODE) &&
552 CXL_DEVT_IS_CARD(dev->devt)) {
553 /*
554 * These minor numbers will eventually be used to program the
555 * PSL and AFUs once we have dynamic reprogramming support
556 */
557 return NULL;
558 }
559 return kasprintf(GFP_KERNEL, "cxl/%s", dev_name(dev));
560}
561
562extern struct class *cxl_class;
563
564static int cxl_add_chardev(struct cxl_afu *afu, dev_t devt, struct cdev *cdev,
565 struct device **chardev, char *postfix, char *desc,
566 const struct file_operations *fops)
567{
568 struct device *dev;
569 int rc;
570
571 cdev_init(cdev, fops);
572 rc = cdev_add(cdev, devt, 1);
573 if (rc) {
574 dev_err(&afu->dev, "Unable to add %s chardev: %i\n", desc, rc);
575 return rc;
576 }
577
578 dev = device_create(cxl_class, &afu->dev, devt, afu,
579 "afu%i.%i%s", afu->adapter->adapter_num, afu->slice, postfix);
580 if (IS_ERR(dev)) {
581 rc = PTR_ERR(dev);
582 dev_err(&afu->dev, "Unable to create %s chardev in sysfs: %i\n", desc, rc);
583 goto err;
584 }
585
586 *chardev = dev;
587
588 return 0;
589err:
590 cdev_del(cdev);
591 return rc;
592}
593
594int cxl_chardev_d_afu_add(struct cxl_afu *afu)
595{
596 return cxl_add_chardev(afu, CXL_AFU_MKDEV_D(afu), &afu->afu_cdev_d,
597 &afu->chardev_d, "d", "dedicated",
598 &afu_master_fops); /* Uses master fops */
599}
600
601int cxl_chardev_m_afu_add(struct cxl_afu *afu)
602{
603 return cxl_add_chardev(afu, CXL_AFU_MKDEV_M(afu), &afu->afu_cdev_m,
604 &afu->chardev_m, "m", "master",
605 &afu_master_fops);
606}
607
608int cxl_chardev_s_afu_add(struct cxl_afu *afu)
609{
610 return cxl_add_chardev(afu, CXL_AFU_MKDEV_S(afu), &afu->afu_cdev_s,
611 &afu->chardev_s, "s", "shared",
612 &afu_fops);
613}
614
615void cxl_chardev_afu_remove(struct cxl_afu *afu)
616{
617 if (afu->chardev_d) {
618 cdev_del(&afu->afu_cdev_d);
619 device_unregister(afu->chardev_d);
620 afu->chardev_d = NULL;
621 }
622 if (afu->chardev_m) {
623 cdev_del(&afu->afu_cdev_m);
624 device_unregister(afu->chardev_m);
625 afu->chardev_m = NULL;
626 }
627 if (afu->chardev_s) {
628 cdev_del(&afu->afu_cdev_s);
629 device_unregister(afu->chardev_s);
630 afu->chardev_s = NULL;
631 }
632}
633
634int cxl_register_afu(struct cxl_afu *afu)
635{
636 afu->dev.class = cxl_class;
637
638 return device_register(&afu->dev);
639}
640
641int cxl_register_adapter(struct cxl *adapter)
642{
643 adapter->dev.class = cxl_class;
644
645 /*
646 * Future: When we support dynamically reprogramming the PSL & AFU we
647 * will expose the interface to do that via a chardev:
648 * adapter->dev.devt = CXL_CARD_MKDEV(adapter);
649 */
650
651 return device_register(&adapter->dev);
652}
653
654dev_t cxl_get_dev(void)
655{
656 return cxl_dev;
657}
658
659int __init cxl_file_init(void)
660{
661 int rc;
662
663 /*
664 * If these change we really need to update API. Either change some
665 * flags or update API version number CXL_API_VERSION.
666 */
667 BUILD_BUG_ON(CXL_API_VERSION != 3);
668 BUILD_BUG_ON(sizeof(struct cxl_ioctl_start_work) != 64);
669 BUILD_BUG_ON(sizeof(struct cxl_event_header) != 8);
670 BUILD_BUG_ON(sizeof(struct cxl_event_afu_interrupt) != 8);
671 BUILD_BUG_ON(sizeof(struct cxl_event_data_storage) != 32);
672 BUILD_BUG_ON(sizeof(struct cxl_event_afu_error) != 16);
673
674 if ((rc = alloc_chrdev_region(&cxl_dev, 0, CXL_NUM_MINORS, "cxl"))) {
675 pr_err("Unable to allocate CXL major number: %i\n", rc);
676 return rc;
677 }
678
679 pr_devel("CXL device allocated, MAJOR %i\n", MAJOR(cxl_dev));
680
681 cxl_class = class_create(THIS_MODULE, "cxl");
682 if (IS_ERR(cxl_class)) {
683 pr_err("Unable to create CXL class\n");
684 rc = PTR_ERR(cxl_class);
685 goto err;
686 }
687 cxl_class->devnode = cxl_devnode;
688
689 return 0;
690
691err:
692 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
693 return rc;
694}
695
696void cxl_file_exit(void)
697{
698 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
699 class_destroy(cxl_class);
700}