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/*
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 <linux/sched/mm.h>
22#include <linux/mmu_context.h>
23#include <asm/cputable.h>
24#include <asm/current.h>
25#include <asm/copro.h>
26
27#include "cxl.h"
28#include "trace.h"
29
30#define CXL_NUM_MINORS 256 /* Total to reserve */
31
32#define CXL_AFU_MINOR_D(afu) (CXL_CARD_MINOR(afu->adapter) + 1 + (3 * afu->slice))
33#define CXL_AFU_MINOR_M(afu) (CXL_AFU_MINOR_D(afu) + 1)
34#define CXL_AFU_MINOR_S(afu) (CXL_AFU_MINOR_D(afu) + 2)
35#define CXL_AFU_MKDEV_D(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_D(afu))
36#define CXL_AFU_MKDEV_M(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_M(afu))
37#define CXL_AFU_MKDEV_S(afu) MKDEV(MAJOR(cxl_dev), CXL_AFU_MINOR_S(afu))
38
39#define CXL_DEVT_AFU(dev) ((MINOR(dev) % CXL_DEV_MINORS - 1) / 3)
40
41#define CXL_DEVT_IS_CARD(dev) (MINOR(dev) % CXL_DEV_MINORS == 0)
42
43static dev_t cxl_dev;
44
45static struct class *cxl_class;
46
47static int __afu_open(struct inode *inode, struct file *file, bool master)
48{
49 struct cxl *adapter;
50 struct cxl_afu *afu;
51 struct cxl_context *ctx;
52 int adapter_num = CXL_DEVT_ADAPTER(inode->i_rdev);
53 int slice = CXL_DEVT_AFU(inode->i_rdev);
54 int rc = -ENODEV;
55
56 pr_devel("afu_open afu%i.%i\n", slice, adapter_num);
57
58 if (!(adapter = get_cxl_adapter(adapter_num)))
59 return -ENODEV;
60
61 if (slice > adapter->slices)
62 goto err_put_adapter;
63
64 spin_lock(&adapter->afu_list_lock);
65 if (!(afu = adapter->afu[slice])) {
66 spin_unlock(&adapter->afu_list_lock);
67 goto err_put_adapter;
68 }
69
70 /*
71 * taking a ref to the afu so that it doesn't go away
72 * for rest of the function. This ref is released before
73 * we return.
74 */
75 cxl_afu_get(afu);
76 spin_unlock(&adapter->afu_list_lock);
77
78 if (!afu->current_mode)
79 goto err_put_afu;
80
81 if (!cxl_ops->link_ok(adapter, afu)) {
82 rc = -EIO;
83 goto err_put_afu;
84 }
85
86 if (!(ctx = cxl_context_alloc())) {
87 rc = -ENOMEM;
88 goto err_put_afu;
89 }
90
91 rc = cxl_context_init(ctx, afu, master);
92 if (rc)
93 goto err_put_afu;
94
95 cxl_context_set_mapping(ctx, inode->i_mapping);
96
97 pr_devel("afu_open pe: %i\n", ctx->pe);
98 file->private_data = ctx;
99
100 /* indicate success */
101 rc = 0;
102
103err_put_afu:
104 /* release the ref taken earlier */
105 cxl_afu_put(afu);
106err_put_adapter:
107 put_device(&adapter->dev);
108 return rc;
109}
110
111int afu_open(struct inode *inode, struct file *file)
112{
113 return __afu_open(inode, file, false);
114}
115
116static int afu_master_open(struct inode *inode, struct file *file)
117{
118 return __afu_open(inode, file, true);
119}
120
121int afu_release(struct inode *inode, struct file *file)
122{
123 struct cxl_context *ctx = file->private_data;
124
125 pr_devel("%s: closing cxl file descriptor. pe: %i\n",
126 __func__, ctx->pe);
127 cxl_context_detach(ctx);
128
129
130 /*
131 * Delete the context's mapping pointer, unless it's created by the
132 * kernel API, in which case leave it so it can be freed by reclaim_ctx()
133 */
134 if (!ctx->kernelapi) {
135 mutex_lock(&ctx->mapping_lock);
136 ctx->mapping = NULL;
137 mutex_unlock(&ctx->mapping_lock);
138 }
139
140 /*
141 * At this this point all bottom halfs have finished and we should be
142 * getting no more IRQs from the hardware for this context. Once it's
143 * removed from the IDR (and RCU synchronised) it's safe to free the
144 * sstp and context.
145 */
146 cxl_context_free(ctx);
147
148 return 0;
149}
150
151static long afu_ioctl_start_work(struct cxl_context *ctx,
152 struct cxl_ioctl_start_work __user *uwork)
153{
154 struct cxl_ioctl_start_work work;
155 u64 amr = 0;
156 int rc;
157
158 pr_devel("%s: pe: %i\n", __func__, ctx->pe);
159
160 /* Do this outside the status_mutex to avoid a circular dependency with
161 * the locking in cxl_mmap_fault() */
162 if (copy_from_user(&work, uwork, sizeof(work)))
163 return -EFAULT;
164
165 mutex_lock(&ctx->status_mutex);
166 if (ctx->status != OPENED) {
167 rc = -EIO;
168 goto out;
169 }
170
171 /*
172 * if any of the reserved fields are set or any of the unused
173 * flags are set it's invalid
174 */
175 if (work.reserved1 || work.reserved2 || work.reserved3 ||
176 work.reserved4 || work.reserved5 ||
177 (work.flags & ~CXL_START_WORK_ALL)) {
178 rc = -EINVAL;
179 goto out;
180 }
181
182 if (!(work.flags & CXL_START_WORK_NUM_IRQS))
183 work.num_interrupts = ctx->afu->pp_irqs;
184 else if ((work.num_interrupts < ctx->afu->pp_irqs) ||
185 (work.num_interrupts > ctx->afu->irqs_max)) {
186 rc = -EINVAL;
187 goto out;
188 }
189
190 if ((rc = afu_register_irqs(ctx, work.num_interrupts)))
191 goto out;
192
193 if (work.flags & CXL_START_WORK_AMR)
194 amr = work.amr & mfspr(SPRN_UAMOR);
195
196 if (work.flags & CXL_START_WORK_TID)
197 ctx->assign_tidr = true;
198
199 ctx->mmio_err_ff = !!(work.flags & CXL_START_WORK_ERR_FF);
200
201 /*
202 * Increment the mapped context count for adapter. This also checks
203 * if adapter_context_lock is taken.
204 */
205 rc = cxl_adapter_context_get(ctx->afu->adapter);
206 if (rc) {
207 afu_release_irqs(ctx, ctx);
208 goto out;
209 }
210
211 /*
212 * We grab the PID here and not in the file open to allow for the case
213 * where a process (master, some daemon, etc) has opened the chardev on
214 * behalf of another process, so the AFU's mm gets bound to the process
215 * that performs this ioctl and not the process that opened the file.
216 * Also we grab the PID of the group leader so that if the task that
217 * has performed the attach operation exits the mm context of the
218 * process is still accessible.
219 */
220 ctx->pid = get_task_pid(current, PIDTYPE_PID);
221
222 /* acquire a reference to the task's mm */
223 ctx->mm = get_task_mm(current);
224
225 /* ensure this mm_struct can't be freed */
226 cxl_context_mm_count_get(ctx);
227
228 if (ctx->mm) {
229 /* decrement the use count from above */
230 mmput(ctx->mm);
231 /* make TLBIs for this context global */
232 mm_context_add_copro(ctx->mm);
233 }
234
235 /*
236 * Increment driver use count. Enables global TLBIs for hash
237 * and callbacks to handle the segment table
238 */
239 cxl_ctx_get();
240
241 /*
242 * A barrier is needed to make sure all TLBIs are global
243 * before we attach and the context starts being used by the
244 * adapter.
245 *
246 * Needed after mm_context_add_copro() for radix and
247 * cxl_ctx_get() for hash/p8.
248 *
249 * The barrier should really be mb(), since it involves a
250 * device. However, it's only useful when we have local
251 * vs. global TLBIs, i.e SMP=y. So keep smp_mb().
252 */
253 smp_mb();
254
255 trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr);
256
257 if ((rc = cxl_ops->attach_process(ctx, false, work.work_element_descriptor,
258 amr))) {
259 afu_release_irqs(ctx, ctx);
260 cxl_adapter_context_put(ctx->afu->adapter);
261 put_pid(ctx->pid);
262 ctx->pid = NULL;
263 cxl_ctx_put();
264 cxl_context_mm_count_put(ctx);
265 if (ctx->mm)
266 mm_context_remove_copro(ctx->mm);
267 goto out;
268 }
269
270 rc = 0;
271 if (work.flags & CXL_START_WORK_TID) {
272 work.tid = ctx->tidr;
273 if (copy_to_user(uwork, &work, sizeof(work)))
274 rc = -EFAULT;
275 }
276
277 ctx->status = STARTED;
278
279out:
280 mutex_unlock(&ctx->status_mutex);
281 return rc;
282}
283
284static long afu_ioctl_process_element(struct cxl_context *ctx,
285 int __user *upe)
286{
287 pr_devel("%s: pe: %i\n", __func__, ctx->pe);
288
289 if (copy_to_user(upe, &ctx->external_pe, sizeof(__u32)))
290 return -EFAULT;
291
292 return 0;
293}
294
295static long afu_ioctl_get_afu_id(struct cxl_context *ctx,
296 struct cxl_afu_id __user *upafuid)
297{
298 struct cxl_afu_id afuid = { 0 };
299
300 afuid.card_id = ctx->afu->adapter->adapter_num;
301 afuid.afu_offset = ctx->afu->slice;
302 afuid.afu_mode = ctx->afu->current_mode;
303
304 /* set the flag bit in case the afu is a slave */
305 if (ctx->afu->current_mode == CXL_MODE_DIRECTED && !ctx->master)
306 afuid.flags |= CXL_AFUID_FLAG_SLAVE;
307
308 if (copy_to_user(upafuid, &afuid, sizeof(afuid)))
309 return -EFAULT;
310
311 return 0;
312}
313
314long afu_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
315{
316 struct cxl_context *ctx = file->private_data;
317
318 if (ctx->status == CLOSED)
319 return -EIO;
320
321 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
322 return -EIO;
323
324 pr_devel("afu_ioctl\n");
325 switch (cmd) {
326 case CXL_IOCTL_START_WORK:
327 return afu_ioctl_start_work(ctx, (struct cxl_ioctl_start_work __user *)arg);
328 case CXL_IOCTL_GET_PROCESS_ELEMENT:
329 return afu_ioctl_process_element(ctx, (__u32 __user *)arg);
330 case CXL_IOCTL_GET_AFU_ID:
331 return afu_ioctl_get_afu_id(ctx, (struct cxl_afu_id __user *)
332 arg);
333 }
334 return -EINVAL;
335}
336
337static long afu_compat_ioctl(struct file *file, unsigned int cmd,
338 unsigned long arg)
339{
340 return afu_ioctl(file, cmd, arg);
341}
342
343int afu_mmap(struct file *file, struct vm_area_struct *vm)
344{
345 struct cxl_context *ctx = file->private_data;
346
347 /* AFU must be started before we can MMIO */
348 if (ctx->status != STARTED)
349 return -EIO;
350
351 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
352 return -EIO;
353
354 return cxl_context_iomap(ctx, vm);
355}
356
357static inline bool ctx_event_pending(struct cxl_context *ctx)
358{
359 if (ctx->pending_irq || ctx->pending_fault || ctx->pending_afu_err)
360 return true;
361
362 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events))
363 return true;
364
365 return false;
366}
367
368__poll_t afu_poll(struct file *file, struct poll_table_struct *poll)
369{
370 struct cxl_context *ctx = file->private_data;
371 __poll_t mask = 0;
372 unsigned long flags;
373
374
375 poll_wait(file, &ctx->wq, poll);
376
377 pr_devel("afu_poll wait done pe: %i\n", ctx->pe);
378
379 spin_lock_irqsave(&ctx->lock, flags);
380 if (ctx_event_pending(ctx))
381 mask |= EPOLLIN | EPOLLRDNORM;
382 else if (ctx->status == CLOSED)
383 /* Only error on closed when there are no futher events pending
384 */
385 mask |= EPOLLERR;
386 spin_unlock_irqrestore(&ctx->lock, flags);
387
388 pr_devel("afu_poll pe: %i returning %#x\n", ctx->pe, mask);
389
390 return mask;
391}
392
393static ssize_t afu_driver_event_copy(struct cxl_context *ctx,
394 char __user *buf,
395 struct cxl_event *event,
396 struct cxl_event_afu_driver_reserved *pl)
397{
398 /* Check event */
399 if (!pl) {
400 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
401 return -EFAULT;
402 }
403
404 /* Check event size */
405 event->header.size += pl->data_size;
406 if (event->header.size > CXL_READ_MIN_SIZE) {
407 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
408 return -EFAULT;
409 }
410
411 /* Copy event header */
412 if (copy_to_user(buf, event, sizeof(struct cxl_event_header))) {
413 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
414 return -EFAULT;
415 }
416
417 /* Copy event data */
418 buf += sizeof(struct cxl_event_header);
419 if (copy_to_user(buf, &pl->data, pl->data_size)) {
420 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
421 return -EFAULT;
422 }
423
424 ctx->afu_driver_ops->event_delivered(ctx, pl, 0); /* Success */
425 return event->header.size;
426}
427
428ssize_t afu_read(struct file *file, char __user *buf, size_t count,
429 loff_t *off)
430{
431 struct cxl_context *ctx = file->private_data;
432 struct cxl_event_afu_driver_reserved *pl = NULL;
433 struct cxl_event event;
434 unsigned long flags;
435 int rc;
436 DEFINE_WAIT(wait);
437
438 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
439 return -EIO;
440
441 if (count < CXL_READ_MIN_SIZE)
442 return -EINVAL;
443
444 spin_lock_irqsave(&ctx->lock, flags);
445
446 for (;;) {
447 prepare_to_wait(&ctx->wq, &wait, TASK_INTERRUPTIBLE);
448 if (ctx_event_pending(ctx) || (ctx->status == CLOSED))
449 break;
450
451 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
452 rc = -EIO;
453 goto out;
454 }
455
456 if (file->f_flags & O_NONBLOCK) {
457 rc = -EAGAIN;
458 goto out;
459 }
460
461 if (signal_pending(current)) {
462 rc = -ERESTARTSYS;
463 goto out;
464 }
465
466 spin_unlock_irqrestore(&ctx->lock, flags);
467 pr_devel("afu_read going to sleep...\n");
468 schedule();
469 pr_devel("afu_read woken up\n");
470 spin_lock_irqsave(&ctx->lock, flags);
471 }
472
473 finish_wait(&ctx->wq, &wait);
474
475 memset(&event, 0, sizeof(event));
476 event.header.process_element = ctx->pe;
477 event.header.size = sizeof(struct cxl_event_header);
478 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) {
479 pr_devel("afu_read delivering AFU driver specific event\n");
480 pl = ctx->afu_driver_ops->fetch_event(ctx);
481 atomic_dec(&ctx->afu_driver_events);
482 event.header.type = CXL_EVENT_AFU_DRIVER;
483 } else if (ctx->pending_irq) {
484 pr_devel("afu_read delivering AFU interrupt\n");
485 event.header.size += sizeof(struct cxl_event_afu_interrupt);
486 event.header.type = CXL_EVENT_AFU_INTERRUPT;
487 event.irq.irq = find_first_bit(ctx->irq_bitmap, ctx->irq_count) + 1;
488 clear_bit(event.irq.irq - 1, ctx->irq_bitmap);
489 if (bitmap_empty(ctx->irq_bitmap, ctx->irq_count))
490 ctx->pending_irq = false;
491 } else if (ctx->pending_fault) {
492 pr_devel("afu_read delivering data storage fault\n");
493 event.header.size += sizeof(struct cxl_event_data_storage);
494 event.header.type = CXL_EVENT_DATA_STORAGE;
495 event.fault.addr = ctx->fault_addr;
496 event.fault.dsisr = ctx->fault_dsisr;
497 ctx->pending_fault = false;
498 } else if (ctx->pending_afu_err) {
499 pr_devel("afu_read delivering afu error\n");
500 event.header.size += sizeof(struct cxl_event_afu_error);
501 event.header.type = CXL_EVENT_AFU_ERROR;
502 event.afu_error.error = ctx->afu_err;
503 ctx->pending_afu_err = false;
504 } else if (ctx->status == CLOSED) {
505 pr_devel("afu_read fatal error\n");
506 spin_unlock_irqrestore(&ctx->lock, flags);
507 return -EIO;
508 } else
509 WARN(1, "afu_read must be buggy\n");
510
511 spin_unlock_irqrestore(&ctx->lock, flags);
512
513 if (event.header.type == CXL_EVENT_AFU_DRIVER)
514 return afu_driver_event_copy(ctx, buf, &event, pl);
515
516 if (copy_to_user(buf, &event, event.header.size))
517 return -EFAULT;
518 return event.header.size;
519
520out:
521 finish_wait(&ctx->wq, &wait);
522 spin_unlock_irqrestore(&ctx->lock, flags);
523 return rc;
524}
525
526/*
527 * Note: if this is updated, we need to update api.c to patch the new ones in
528 * too
529 */
530const struct file_operations afu_fops = {
531 .owner = THIS_MODULE,
532 .open = afu_open,
533 .poll = afu_poll,
534 .read = afu_read,
535 .release = afu_release,
536 .unlocked_ioctl = afu_ioctl,
537 .compat_ioctl = afu_compat_ioctl,
538 .mmap = afu_mmap,
539};
540
541static const struct file_operations afu_master_fops = {
542 .owner = THIS_MODULE,
543 .open = afu_master_open,
544 .poll = afu_poll,
545 .read = afu_read,
546 .release = afu_release,
547 .unlocked_ioctl = afu_ioctl,
548 .compat_ioctl = afu_compat_ioctl,
549 .mmap = afu_mmap,
550};
551
552
553static char *cxl_devnode(struct device *dev, umode_t *mode)
554{
555 if (cpu_has_feature(CPU_FTR_HVMODE) &&
556 CXL_DEVT_IS_CARD(dev->devt)) {
557 /*
558 * These minor numbers will eventually be used to program the
559 * PSL and AFUs once we have dynamic reprogramming support
560 */
561 return NULL;
562 }
563 return kasprintf(GFP_KERNEL, "cxl/%s", dev_name(dev));
564}
565
566extern struct class *cxl_class;
567
568static int cxl_add_chardev(struct cxl_afu *afu, dev_t devt, struct cdev *cdev,
569 struct device **chardev, char *postfix, char *desc,
570 const struct file_operations *fops)
571{
572 struct device *dev;
573 int rc;
574
575 cdev_init(cdev, fops);
576 if ((rc = cdev_add(cdev, devt, 1))) {
577 dev_err(&afu->dev, "Unable to add %s chardev: %i\n", desc, rc);
578 return rc;
579 }
580
581 dev = device_create(cxl_class, &afu->dev, devt, afu,
582 "afu%i.%i%s", afu->adapter->adapter_num, afu->slice, postfix);
583 if (IS_ERR(dev)) {
584 dev_err(&afu->dev, "Unable to create %s chardev in sysfs: %i\n", desc, rc);
585 rc = PTR_ERR(dev);
586 goto err;
587 }
588
589 *chardev = dev;
590
591 return 0;
592err:
593 cdev_del(cdev);
594 return rc;
595}
596
597int cxl_chardev_d_afu_add(struct cxl_afu *afu)
598{
599 return cxl_add_chardev(afu, CXL_AFU_MKDEV_D(afu), &afu->afu_cdev_d,
600 &afu->chardev_d, "d", "dedicated",
601 &afu_master_fops); /* Uses master fops */
602}
603
604int cxl_chardev_m_afu_add(struct cxl_afu *afu)
605{
606 return cxl_add_chardev(afu, CXL_AFU_MKDEV_M(afu), &afu->afu_cdev_m,
607 &afu->chardev_m, "m", "master",
608 &afu_master_fops);
609}
610
611int cxl_chardev_s_afu_add(struct cxl_afu *afu)
612{
613 return cxl_add_chardev(afu, CXL_AFU_MKDEV_S(afu), &afu->afu_cdev_s,
614 &afu->chardev_s, "s", "shared",
615 &afu_fops);
616}
617
618void cxl_chardev_afu_remove(struct cxl_afu *afu)
619{
620 if (afu->chardev_d) {
621 cdev_del(&afu->afu_cdev_d);
622 device_unregister(afu->chardev_d);
623 afu->chardev_d = NULL;
624 }
625 if (afu->chardev_m) {
626 cdev_del(&afu->afu_cdev_m);
627 device_unregister(afu->chardev_m);
628 afu->chardev_m = NULL;
629 }
630 if (afu->chardev_s) {
631 cdev_del(&afu->afu_cdev_s);
632 device_unregister(afu->chardev_s);
633 afu->chardev_s = NULL;
634 }
635}
636
637int cxl_register_afu(struct cxl_afu *afu)
638{
639 afu->dev.class = cxl_class;
640
641 return device_register(&afu->dev);
642}
643
644int cxl_register_adapter(struct cxl *adapter)
645{
646 adapter->dev.class = cxl_class;
647
648 /*
649 * Future: When we support dynamically reprogramming the PSL & AFU we
650 * will expose the interface to do that via a chardev:
651 * adapter->dev.devt = CXL_CARD_MKDEV(adapter);
652 */
653
654 return device_register(&adapter->dev);
655}
656
657dev_t cxl_get_dev(void)
658{
659 return cxl_dev;
660}
661
662int __init cxl_file_init(void)
663{
664 int rc;
665
666 /*
667 * If these change we really need to update API. Either change some
668 * flags or update API version number CXL_API_VERSION.
669 */
670 BUILD_BUG_ON(CXL_API_VERSION != 3);
671 BUILD_BUG_ON(sizeof(struct cxl_ioctl_start_work) != 64);
672 BUILD_BUG_ON(sizeof(struct cxl_event_header) != 8);
673 BUILD_BUG_ON(sizeof(struct cxl_event_afu_interrupt) != 8);
674 BUILD_BUG_ON(sizeof(struct cxl_event_data_storage) != 32);
675 BUILD_BUG_ON(sizeof(struct cxl_event_afu_error) != 16);
676
677 if ((rc = alloc_chrdev_region(&cxl_dev, 0, CXL_NUM_MINORS, "cxl"))) {
678 pr_err("Unable to allocate CXL major number: %i\n", rc);
679 return rc;
680 }
681
682 pr_devel("CXL device allocated, MAJOR %i\n", MAJOR(cxl_dev));
683
684 cxl_class = class_create(THIS_MODULE, "cxl");
685 if (IS_ERR(cxl_class)) {
686 pr_err("Unable to create CXL class\n");
687 rc = PTR_ERR(cxl_class);
688 goto err;
689 }
690 cxl_class->devnode = cxl_devnode;
691
692 return 0;
693
694err:
695 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
696 return rc;
697}
698
699void cxl_file_exit(void)
700{
701 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
702 class_destroy(cxl_class);
703}