2 * Copyright 2014 IBM Corp.
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.
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>
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>
30 #define CXL_NUM_MINORS 256 /* Total to reserve */
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))
39 #define CXL_DEVT_AFU(dev) ((MINOR(dev) % CXL_DEV_MINORS - 1) / 3)
41 #define CXL_DEVT_IS_CARD(dev) (MINOR(dev) % CXL_DEV_MINORS == 0)
45 static struct class *cxl_class;
47 static int __afu_open(struct inode *inode, struct file *file, bool master)
51 struct cxl_context *ctx;
52 int adapter_num = CXL_DEVT_ADAPTER(inode->i_rdev);
53 int slice = CXL_DEVT_AFU(inode->i_rdev);
56 pr_devel("afu_open afu%i.%i\n", slice, adapter_num);
58 if (!(adapter = get_cxl_adapter(adapter_num)))
61 if (slice > adapter->slices)
64 spin_lock(&adapter->afu_list_lock);
65 if (!(afu = adapter->afu[slice])) {
66 spin_unlock(&adapter->afu_list_lock);
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
76 spin_unlock(&adapter->afu_list_lock);
78 if (!afu->current_mode)
81 if (!cxl_ops->link_ok(adapter, afu)) {
86 if (!(ctx = cxl_context_alloc())) {
91 rc = cxl_context_init(ctx, afu, master);
95 cxl_context_set_mapping(ctx, inode->i_mapping);
97 pr_devel("afu_open pe: %i\n", ctx->pe);
98 file->private_data = ctx;
100 /* indicate success */
104 /* release the ref taken earlier */
107 put_device(&adapter->dev);
111 int afu_open(struct inode *inode, struct file *file)
113 return __afu_open(inode, file, false);
116 static int afu_master_open(struct inode *inode, struct file *file)
118 return __afu_open(inode, file, true);
121 int afu_release(struct inode *inode, struct file *file)
123 struct cxl_context *ctx = file->private_data;
125 pr_devel("%s: closing cxl file descriptor. pe: %i\n",
127 cxl_context_detach(ctx);
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()
134 if (!ctx->kernelapi) {
135 mutex_lock(&ctx->mapping_lock);
137 mutex_unlock(&ctx->mapping_lock);
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
146 cxl_context_free(ctx);
151 static long afu_ioctl_start_work(struct cxl_context *ctx,
152 struct cxl_ioctl_start_work __user *uwork)
154 struct cxl_ioctl_start_work work;
158 pr_devel("%s: pe: %i\n", __func__, ctx->pe);
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)))
165 mutex_lock(&ctx->status_mutex);
166 if (ctx->status != OPENED) {
172 * if any of the reserved fields are set or any of the unused
173 * flags are set it's invalid
175 if (work.reserved1 || work.reserved2 || work.reserved3 ||
176 work.reserved4 || work.reserved5 || work.reserved6 ||
177 (work.flags & ~CXL_START_WORK_ALL)) {
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)) {
189 if ((rc = afu_register_irqs(ctx, work.num_interrupts)))
192 if (work.flags & CXL_START_WORK_AMR)
193 amr = work.amr & mfspr(SPRN_UAMOR);
195 ctx->mmio_err_ff = !!(work.flags & CXL_START_WORK_ERR_FF);
198 * Increment the mapped context count for adapter. This also checks
199 * if adapter_context_lock is taken.
201 rc = cxl_adapter_context_get(ctx->afu->adapter);
203 afu_release_irqs(ctx, ctx);
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.
216 ctx->pid = get_task_pid(current, PIDTYPE_PID);
218 /* acquire a reference to the task's mm */
219 ctx->mm = get_task_mm(current);
221 /* ensure this mm_struct can't be freed */
222 cxl_context_mm_count_get(ctx);
225 /* decrement the use count from above */
227 /* make TLBIs for this context global */
228 mm_context_add_copro(ctx->mm);
232 * Increment driver use count. Enables global TLBIs for hash
233 * and callbacks to handle the segment table
238 * A barrier is needed to make sure all TLBIs are global
239 * before we attach and the context starts being used by the
242 * Needed after mm_context_add_copro() for radix and
243 * cxl_ctx_get() for hash/p8.
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().
251 trace_cxl_attach(ctx, work.work_element_descriptor, work.num_interrupts, amr);
253 if ((rc = cxl_ops->attach_process(ctx, false, work.work_element_descriptor,
255 afu_release_irqs(ctx, ctx);
256 cxl_adapter_context_put(ctx->afu->adapter);
260 cxl_context_mm_count_put(ctx);
262 mm_context_remove_copro(ctx->mm);
266 ctx->status = STARTED;
269 mutex_unlock(&ctx->status_mutex);
273 static long afu_ioctl_process_element(struct cxl_context *ctx,
276 pr_devel("%s: pe: %i\n", __func__, ctx->pe);
278 if (copy_to_user(upe, &ctx->external_pe, sizeof(__u32)))
284 static long afu_ioctl_get_afu_id(struct cxl_context *ctx,
285 struct cxl_afu_id __user *upafuid)
287 struct cxl_afu_id afuid = { 0 };
289 afuid.card_id = ctx->afu->adapter->adapter_num;
290 afuid.afu_offset = ctx->afu->slice;
291 afuid.afu_mode = ctx->afu->current_mode;
293 /* set the flag bit in case the afu is a slave */
294 if (ctx->afu->current_mode == CXL_MODE_DIRECTED && !ctx->master)
295 afuid.flags |= CXL_AFUID_FLAG_SLAVE;
297 if (copy_to_user(upafuid, &afuid, sizeof(afuid)))
303 long afu_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
305 struct cxl_context *ctx = file->private_data;
307 if (ctx->status == CLOSED)
310 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
313 pr_devel("afu_ioctl\n");
315 case CXL_IOCTL_START_WORK:
316 return afu_ioctl_start_work(ctx, (struct cxl_ioctl_start_work __user *)arg);
317 case CXL_IOCTL_GET_PROCESS_ELEMENT:
318 return afu_ioctl_process_element(ctx, (__u32 __user *)arg);
319 case CXL_IOCTL_GET_AFU_ID:
320 return afu_ioctl_get_afu_id(ctx, (struct cxl_afu_id __user *)
326 static long afu_compat_ioctl(struct file *file, unsigned int cmd,
329 return afu_ioctl(file, cmd, arg);
332 int afu_mmap(struct file *file, struct vm_area_struct *vm)
334 struct cxl_context *ctx = file->private_data;
336 /* AFU must be started before we can MMIO */
337 if (ctx->status != STARTED)
340 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
343 return cxl_context_iomap(ctx, vm);
346 static inline bool ctx_event_pending(struct cxl_context *ctx)
348 if (ctx->pending_irq || ctx->pending_fault || ctx->pending_afu_err)
351 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events))
357 __poll_t afu_poll(struct file *file, struct poll_table_struct *poll)
359 struct cxl_context *ctx = file->private_data;
364 poll_wait(file, &ctx->wq, poll);
366 pr_devel("afu_poll wait done pe: %i\n", ctx->pe);
368 spin_lock_irqsave(&ctx->lock, flags);
369 if (ctx_event_pending(ctx))
370 mask |= POLLIN | POLLRDNORM;
371 else if (ctx->status == CLOSED)
372 /* Only error on closed when there are no futher events pending
375 spin_unlock_irqrestore(&ctx->lock, flags);
377 pr_devel("afu_poll pe: %i returning %#x\n", ctx->pe, mask);
382 static ssize_t afu_driver_event_copy(struct cxl_context *ctx,
384 struct cxl_event *event,
385 struct cxl_event_afu_driver_reserved *pl)
389 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
393 /* Check event size */
394 event->header.size += pl->data_size;
395 if (event->header.size > CXL_READ_MIN_SIZE) {
396 ctx->afu_driver_ops->event_delivered(ctx, pl, -EINVAL);
400 /* Copy event header */
401 if (copy_to_user(buf, event, sizeof(struct cxl_event_header))) {
402 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
406 /* Copy event data */
407 buf += sizeof(struct cxl_event_header);
408 if (copy_to_user(buf, &pl->data, pl->data_size)) {
409 ctx->afu_driver_ops->event_delivered(ctx, pl, -EFAULT);
413 ctx->afu_driver_ops->event_delivered(ctx, pl, 0); /* Success */
414 return event->header.size;
417 ssize_t afu_read(struct file *file, char __user *buf, size_t count,
420 struct cxl_context *ctx = file->private_data;
421 struct cxl_event_afu_driver_reserved *pl = NULL;
422 struct cxl_event event;
427 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu))
430 if (count < CXL_READ_MIN_SIZE)
433 spin_lock_irqsave(&ctx->lock, flags);
436 prepare_to_wait(&ctx->wq, &wait, TASK_INTERRUPTIBLE);
437 if (ctx_event_pending(ctx) || (ctx->status == CLOSED))
440 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) {
445 if (file->f_flags & O_NONBLOCK) {
450 if (signal_pending(current)) {
455 spin_unlock_irqrestore(&ctx->lock, flags);
456 pr_devel("afu_read going to sleep...\n");
458 pr_devel("afu_read woken up\n");
459 spin_lock_irqsave(&ctx->lock, flags);
462 finish_wait(&ctx->wq, &wait);
464 memset(&event, 0, sizeof(event));
465 event.header.process_element = ctx->pe;
466 event.header.size = sizeof(struct cxl_event_header);
467 if (ctx->afu_driver_ops && atomic_read(&ctx->afu_driver_events)) {
468 pr_devel("afu_read delivering AFU driver specific event\n");
469 pl = ctx->afu_driver_ops->fetch_event(ctx);
470 atomic_dec(&ctx->afu_driver_events);
471 event.header.type = CXL_EVENT_AFU_DRIVER;
472 } else if (ctx->pending_irq) {
473 pr_devel("afu_read delivering AFU interrupt\n");
474 event.header.size += sizeof(struct cxl_event_afu_interrupt);
475 event.header.type = CXL_EVENT_AFU_INTERRUPT;
476 event.irq.irq = find_first_bit(ctx->irq_bitmap, ctx->irq_count) + 1;
477 clear_bit(event.irq.irq - 1, ctx->irq_bitmap);
478 if (bitmap_empty(ctx->irq_bitmap, ctx->irq_count))
479 ctx->pending_irq = false;
480 } else if (ctx->pending_fault) {
481 pr_devel("afu_read delivering data storage fault\n");
482 event.header.size += sizeof(struct cxl_event_data_storage);
483 event.header.type = CXL_EVENT_DATA_STORAGE;
484 event.fault.addr = ctx->fault_addr;
485 event.fault.dsisr = ctx->fault_dsisr;
486 ctx->pending_fault = false;
487 } else if (ctx->pending_afu_err) {
488 pr_devel("afu_read delivering afu error\n");
489 event.header.size += sizeof(struct cxl_event_afu_error);
490 event.header.type = CXL_EVENT_AFU_ERROR;
491 event.afu_error.error = ctx->afu_err;
492 ctx->pending_afu_err = false;
493 } else if (ctx->status == CLOSED) {
494 pr_devel("afu_read fatal error\n");
495 spin_unlock_irqrestore(&ctx->lock, flags);
498 WARN(1, "afu_read must be buggy\n");
500 spin_unlock_irqrestore(&ctx->lock, flags);
502 if (event.header.type == CXL_EVENT_AFU_DRIVER)
503 return afu_driver_event_copy(ctx, buf, &event, pl);
505 if (copy_to_user(buf, &event, event.header.size))
507 return event.header.size;
510 finish_wait(&ctx->wq, &wait);
511 spin_unlock_irqrestore(&ctx->lock, flags);
516 * Note: if this is updated, we need to update api.c to patch the new ones in
519 const struct file_operations afu_fops = {
520 .owner = THIS_MODULE,
524 .release = afu_release,
525 .unlocked_ioctl = afu_ioctl,
526 .compat_ioctl = afu_compat_ioctl,
530 static const struct file_operations afu_master_fops = {
531 .owner = THIS_MODULE,
532 .open = afu_master_open,
535 .release = afu_release,
536 .unlocked_ioctl = afu_ioctl,
537 .compat_ioctl = afu_compat_ioctl,
542 static char *cxl_devnode(struct device *dev, umode_t *mode)
544 if (cpu_has_feature(CPU_FTR_HVMODE) &&
545 CXL_DEVT_IS_CARD(dev->devt)) {
547 * These minor numbers will eventually be used to program the
548 * PSL and AFUs once we have dynamic reprogramming support
552 return kasprintf(GFP_KERNEL, "cxl/%s", dev_name(dev));
555 extern struct class *cxl_class;
557 static int cxl_add_chardev(struct cxl_afu *afu, dev_t devt, struct cdev *cdev,
558 struct device **chardev, char *postfix, char *desc,
559 const struct file_operations *fops)
564 cdev_init(cdev, fops);
565 if ((rc = cdev_add(cdev, devt, 1))) {
566 dev_err(&afu->dev, "Unable to add %s chardev: %i\n", desc, rc);
570 dev = device_create(cxl_class, &afu->dev, devt, afu,
571 "afu%i.%i%s", afu->adapter->adapter_num, afu->slice, postfix);
573 dev_err(&afu->dev, "Unable to create %s chardev in sysfs: %i\n", desc, rc);
586 int cxl_chardev_d_afu_add(struct cxl_afu *afu)
588 return cxl_add_chardev(afu, CXL_AFU_MKDEV_D(afu), &afu->afu_cdev_d,
589 &afu->chardev_d, "d", "dedicated",
590 &afu_master_fops); /* Uses master fops */
593 int cxl_chardev_m_afu_add(struct cxl_afu *afu)
595 return cxl_add_chardev(afu, CXL_AFU_MKDEV_M(afu), &afu->afu_cdev_m,
596 &afu->chardev_m, "m", "master",
600 int cxl_chardev_s_afu_add(struct cxl_afu *afu)
602 return cxl_add_chardev(afu, CXL_AFU_MKDEV_S(afu), &afu->afu_cdev_s,
603 &afu->chardev_s, "s", "shared",
607 void cxl_chardev_afu_remove(struct cxl_afu *afu)
609 if (afu->chardev_d) {
610 cdev_del(&afu->afu_cdev_d);
611 device_unregister(afu->chardev_d);
612 afu->chardev_d = NULL;
614 if (afu->chardev_m) {
615 cdev_del(&afu->afu_cdev_m);
616 device_unregister(afu->chardev_m);
617 afu->chardev_m = NULL;
619 if (afu->chardev_s) {
620 cdev_del(&afu->afu_cdev_s);
621 device_unregister(afu->chardev_s);
622 afu->chardev_s = NULL;
626 int cxl_register_afu(struct cxl_afu *afu)
628 afu->dev.class = cxl_class;
630 return device_register(&afu->dev);
633 int cxl_register_adapter(struct cxl *adapter)
635 adapter->dev.class = cxl_class;
638 * Future: When we support dynamically reprogramming the PSL & AFU we
639 * will expose the interface to do that via a chardev:
640 * adapter->dev.devt = CXL_CARD_MKDEV(adapter);
643 return device_register(&adapter->dev);
646 dev_t cxl_get_dev(void)
651 int __init cxl_file_init(void)
656 * If these change we really need to update API. Either change some
657 * flags or update API version number CXL_API_VERSION.
659 BUILD_BUG_ON(CXL_API_VERSION != 3);
660 BUILD_BUG_ON(sizeof(struct cxl_ioctl_start_work) != 64);
661 BUILD_BUG_ON(sizeof(struct cxl_event_header) != 8);
662 BUILD_BUG_ON(sizeof(struct cxl_event_afu_interrupt) != 8);
663 BUILD_BUG_ON(sizeof(struct cxl_event_data_storage) != 32);
664 BUILD_BUG_ON(sizeof(struct cxl_event_afu_error) != 16);
666 if ((rc = alloc_chrdev_region(&cxl_dev, 0, CXL_NUM_MINORS, "cxl"))) {
667 pr_err("Unable to allocate CXL major number: %i\n", rc);
671 pr_devel("CXL device allocated, MAJOR %i\n", MAJOR(cxl_dev));
673 cxl_class = class_create(THIS_MODULE, "cxl");
674 if (IS_ERR(cxl_class)) {
675 pr_err("Unable to create CXL class\n");
676 rc = PTR_ERR(cxl_class);
679 cxl_class->devnode = cxl_devnode;
684 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
688 void cxl_file_exit(void)
690 unregister_chrdev_region(cxl_dev, CXL_NUM_MINORS);
691 class_destroy(cxl_class);