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