mshv_eventfd.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * eventfd support for mshv
  4 *
  5 * Heavily inspired from KVM implementation of irqfd/ioeventfd. The basic
  6 * framework code is taken from the kvm implementation.
  7 *
  8 * All credits to kvm developers.
  9 */
 10
 11#include <linux/syscalls.h>
 12#include <linux/wait.h>
 13#include <linux/poll.h>
 14#include <linux/file.h>
 15#include <linux/list.h>
 16#include <linux/workqueue.h>
 17#include <linux/eventfd.h>
 18
 19#if IS_ENABLED(CONFIG_X86_64)
 20#include <asm/apic.h>
 21#endif
 22#include <asm/mshyperv.h>
 23
 24#include "mshv_eventfd.h"
 25#include "mshv.h"
 26#include "mshv_root.h"
 27
 28static struct workqueue_struct *irqfd_cleanup_wq;
 29
 30void mshv_register_irq_ack_notifier(struct mshv_partition *partition,
 31				    struct mshv_irq_ack_notifier *mian)
 32{
 33	mutex_lock(&partition->pt_irq_lock);
 34	hlist_add_head_rcu(&mian->link, &partition->irq_ack_notifier_list);
 35	mutex_unlock(&partition->pt_irq_lock);
 36}
 37
 38void mshv_unregister_irq_ack_notifier(struct mshv_partition *partition,
 39				      struct mshv_irq_ack_notifier *mian)
 40{
 41	mutex_lock(&partition->pt_irq_lock);
 42	hlist_del_init_rcu(&mian->link);
 43	mutex_unlock(&partition->pt_irq_lock);
 44	synchronize_rcu();
 45}
 46
 47bool mshv_notify_acked_gsi(struct mshv_partition *partition, int gsi)
 48{
 49	struct mshv_irq_ack_notifier *mian;
 50	bool acked = false;
 51
 52	rcu_read_lock();
 53	hlist_for_each_entry_rcu(mian, &partition->irq_ack_notifier_list,
 54				 link) {
 55		if (mian->irq_ack_gsi == gsi) {
 56			mian->irq_acked(mian);
 57			acked = true;
 58		}
 59	}
 60	rcu_read_unlock();
 61
 62	return acked;
 63}
 64
 65#if IS_ENABLED(CONFIG_ARM64)
 66static inline bool hv_should_clear_interrupt(enum hv_interrupt_type type)
 67{
 68	return false;
 69}
 70#elif IS_ENABLED(CONFIG_X86_64)
 71static inline bool hv_should_clear_interrupt(enum hv_interrupt_type type)
 72{
 73	return type == HV_X64_INTERRUPT_TYPE_EXTINT;
 74}
 75#endif
 76
 77static void mshv_irqfd_resampler_ack(struct mshv_irq_ack_notifier *mian)
 78{
 79	struct mshv_irqfd_resampler *resampler;
 80	struct mshv_partition *partition;
 81	struct mshv_irqfd *irqfd;
 82	int idx;
 83
 84	resampler = container_of(mian, struct mshv_irqfd_resampler,
 85				 rsmplr_notifier);
 86	partition = resampler->rsmplr_partn;
 87
 88	idx = srcu_read_lock(&partition->pt_irq_srcu);
 89
 90	hlist_for_each_entry_srcu(irqfd, &resampler->rsmplr_irqfd_list,
 91				 irqfd_resampler_hnode,
 92				 srcu_read_lock_held(&partition->pt_irq_srcu)) {
 93		if (hv_should_clear_interrupt(irqfd->irqfd_lapic_irq.lapic_control.interrupt_type))
 94			hv_call_clear_virtual_interrupt(partition->pt_id);
 95
 96		eventfd_signal(irqfd->irqfd_resamplefd);
 97	}
 98
 99	srcu_read_unlock(&partition->pt_irq_srcu, idx);
100}
101
102#if IS_ENABLED(CONFIG_X86_64)
103static bool
104mshv_vp_irq_vector_injected(union hv_vp_register_page_interrupt_vectors iv,
105			    u32 vector)
106{
107	int i;
108
109	for (i = 0; i < iv.vector_count; i++) {
110		if (iv.vector[i] == vector)
111			return true;
112	}
113
114	return false;
115}
116
117static int mshv_vp_irq_try_set_vector(struct mshv_vp *vp, u32 vector)
118{
119	union hv_vp_register_page_interrupt_vectors iv, new_iv;
120
121	iv = vp->vp_register_page->interrupt_vectors;
122	new_iv = iv;
123
124	if (mshv_vp_irq_vector_injected(iv, vector))
125		return 0;
126
127	if (iv.vector_count >= HV_VP_REGISTER_PAGE_MAX_VECTOR_COUNT)
128		return -ENOSPC;
129
130	new_iv.vector[new_iv.vector_count++] = vector;
131
132	if (cmpxchg(&vp->vp_register_page->interrupt_vectors.as_uint64,
133		    iv.as_uint64, new_iv.as_uint64) != iv.as_uint64)
134		return -EAGAIN;
135
136	return 0;
137}
138
139static int mshv_vp_irq_set_vector(struct mshv_vp *vp, u32 vector)
140{
141	int ret;
142
143	do {
144		ret = mshv_vp_irq_try_set_vector(vp, vector);
145	} while (ret == -EAGAIN && !need_resched());
146
147	return ret;
148}
149
150/*
151 * Try to raise irq for guest via shared vector array. hyp does the actual
152 * inject of the interrupt.
153 */
154static int mshv_try_assert_irq_fast(struct mshv_irqfd *irqfd)
155{
156	struct mshv_partition *partition = irqfd->irqfd_partn;
157	struct mshv_lapic_irq *irq = &irqfd->irqfd_lapic_irq;
158	struct mshv_vp *vp;
159
160	if (!(ms_hyperv.ext_features &
161	      HV_VP_DISPATCH_INTERRUPT_INJECTION_AVAILABLE))
162		return -EOPNOTSUPP;
163
164	if (hv_scheduler_type != HV_SCHEDULER_TYPE_ROOT)
165		return -EOPNOTSUPP;
166
167#if IS_ENABLED(CONFIG_X86)
168	if (irq->lapic_control.logical_dest_mode)
169		return -EOPNOTSUPP;
170#endif
171
172	vp = partition->pt_vp_array[irq->lapic_apic_id];
173
174	if (!vp->vp_register_page)
175		return -EOPNOTSUPP;
176
177	if (mshv_vp_irq_set_vector(vp, irq->lapic_vector))
178		return -EINVAL;
179
180	if (vp->run.flags.root_sched_dispatched &&
181	    vp->vp_register_page->interrupt_vectors.as_uint64)
182		return -EBUSY;
183
184	wake_up(&vp->run.vp_suspend_queue);
185
186	return 0;
187}
188#else /* CONFIG_X86_64 */
189static int mshv_try_assert_irq_fast(struct mshv_irqfd *irqfd)
190{
191	return -EOPNOTSUPP;
192}
193#endif
194
195static void mshv_assert_irq_slow(struct mshv_irqfd *irqfd)
196{
197	struct mshv_partition *partition = irqfd->irqfd_partn;
198	struct mshv_lapic_irq *irq = &irqfd->irqfd_lapic_irq;
199	unsigned int seq;
200	int idx;
201
202#if IS_ENABLED(CONFIG_X86)
203	WARN_ON(irqfd->irqfd_resampler &&
204		!irq->lapic_control.level_triggered);
205#endif
206
207	idx = srcu_read_lock(&partition->pt_irq_srcu);
208	if (irqfd->irqfd_girq_ent.guest_irq_num) {
209		if (!irqfd->irqfd_girq_ent.girq_entry_valid) {
210			srcu_read_unlock(&partition->pt_irq_srcu, idx);
211			return;
212		}
213
214		do {
215			seq = read_seqcount_begin(&irqfd->irqfd_irqe_sc);
216		} while (read_seqcount_retry(&irqfd->irqfd_irqe_sc, seq));
217	}
218
219	hv_call_assert_virtual_interrupt(irqfd->irqfd_partn->pt_id,
220					 irq->lapic_vector, irq->lapic_apic_id,
221					 irq->lapic_control);
222	srcu_read_unlock(&partition->pt_irq_srcu, idx);
223}
224
225static void mshv_irqfd_resampler_shutdown(struct mshv_irqfd *irqfd)
226{
227	struct mshv_irqfd_resampler *rp = irqfd->irqfd_resampler;
228	struct mshv_partition *pt = rp->rsmplr_partn;
229
230	mutex_lock(&pt->irqfds_resampler_lock);
231
232	hlist_del_rcu(&irqfd->irqfd_resampler_hnode);
233	synchronize_srcu(&pt->pt_irq_srcu);
234
235	if (hlist_empty(&rp->rsmplr_irqfd_list)) {
236		hlist_del(&rp->rsmplr_hnode);
237		mshv_unregister_irq_ack_notifier(pt, &rp->rsmplr_notifier);
238		kfree(rp);
239	}
240
241	mutex_unlock(&pt->irqfds_resampler_lock);
242}
243
244/*
245 * Race-free decouple logic (ordering is critical)
246 */
247static void mshv_irqfd_shutdown(struct work_struct *work)
248{
249	struct mshv_irqfd *irqfd =
250			container_of(work, struct mshv_irqfd, irqfd_shutdown);
251	u64 cnt;
252
253	/*
254	 * Synchronize with the wait-queue and unhook ourselves to prevent
255	 * further events.
256	 */
257	eventfd_ctx_remove_wait_queue(irqfd->irqfd_eventfd_ctx, &irqfd->irqfd_wait, &cnt);
258
259	if (irqfd->irqfd_resampler) {
260		mshv_irqfd_resampler_shutdown(irqfd);
261		eventfd_ctx_put(irqfd->irqfd_resamplefd);
262	}
263
264	/*
265	 * It is now safe to release the object's resources
266	 */
267	eventfd_ctx_put(irqfd->irqfd_eventfd_ctx);
268	kfree(irqfd);
269}
270
271/* assumes partition->pt_irqfds_lock is held */
272static bool mshv_irqfd_is_active(struct mshv_irqfd *irqfd)
273{
274	return !hlist_unhashed(&irqfd->irqfd_hnode);
275}
276
277/*
278 * Mark the irqfd as inactive and schedule it for removal
279 *
280 * assumes partition->pt_irqfds_lock is held
281 */
282static void mshv_irqfd_deactivate(struct mshv_irqfd *irqfd)
283{
284	if (!mshv_irqfd_is_active(irqfd))
285		return;
286
287	hlist_del(&irqfd->irqfd_hnode);
288
289	queue_work(irqfd_cleanup_wq, &irqfd->irqfd_shutdown);
290}
291
292/*
293 * Called with wqh->lock held and interrupts disabled
294 */
295static int mshv_irqfd_wakeup(wait_queue_entry_t *wait, unsigned int mode,
296			     int sync, void *key)
297{
298	struct mshv_irqfd *irqfd = container_of(wait, struct mshv_irqfd,
299						irqfd_wait);
300	unsigned long flags = (unsigned long)key;
301	int idx;
302	unsigned int seq;
303	struct mshv_partition *pt = irqfd->irqfd_partn;
304	int ret = 0;
305
306	if (flags & POLLIN) {
307		u64 cnt;
308
309		eventfd_ctx_do_read(irqfd->irqfd_eventfd_ctx, &cnt);
310		idx = srcu_read_lock(&pt->pt_irq_srcu);
311		do {
312			seq = read_seqcount_begin(&irqfd->irqfd_irqe_sc);
313		} while (read_seqcount_retry(&irqfd->irqfd_irqe_sc, seq));
314
315		/* An event has been signaled, raise an interrupt */
316		ret = mshv_try_assert_irq_fast(irqfd);
317		if (ret)
318			mshv_assert_irq_slow(irqfd);
319
320		srcu_read_unlock(&pt->pt_irq_srcu, idx);
321
322		ret = 1;
323	}
324
325	if (flags & POLLHUP) {
326		/* The eventfd is closing, detach from the partition */
327		unsigned long flags;
328
329		spin_lock_irqsave(&pt->pt_irqfds_lock, flags);
330
331		/*
332		 * We must check if someone deactivated the irqfd before
333		 * we could acquire the pt_irqfds_lock since the item is
334		 * deactivated from the mshv side before it is unhooked from
335		 * the wait-queue.  If it is already deactivated, we can
336		 * simply return knowing the other side will cleanup for us.
337		 * We cannot race against the irqfd going away since the
338		 * other side is required to acquire wqh->lock, which we hold
339		 */
340		if (mshv_irqfd_is_active(irqfd))
341			mshv_irqfd_deactivate(irqfd);
342
343		spin_unlock_irqrestore(&pt->pt_irqfds_lock, flags);
344	}
345
346	return ret;
347}
348
349/* Must be called under pt_irqfds_lock */
350static void mshv_irqfd_update(struct mshv_partition *pt,
351			      struct mshv_irqfd *irqfd)
352{
353	write_seqcount_begin(&irqfd->irqfd_irqe_sc);
354	irqfd->irqfd_girq_ent = mshv_ret_girq_entry(pt,
355						    irqfd->irqfd_irqnum);
356	mshv_copy_girq_info(&irqfd->irqfd_girq_ent, &irqfd->irqfd_lapic_irq);
357	write_seqcount_end(&irqfd->irqfd_irqe_sc);
358}
359
360void mshv_irqfd_routing_update(struct mshv_partition *pt)
361{
362	struct mshv_irqfd *irqfd;
363
364	spin_lock_irq(&pt->pt_irqfds_lock);
365	hlist_for_each_entry(irqfd, &pt->pt_irqfds_list, irqfd_hnode)
366		mshv_irqfd_update(pt, irqfd);
367	spin_unlock_irq(&pt->pt_irqfds_lock);
368}
369
370static void mshv_irqfd_queue_proc(struct file *file, wait_queue_head_t *wqh,
371				  poll_table *polltbl)
372{
373	struct mshv_irqfd *irqfd =
374			container_of(polltbl, struct mshv_irqfd, irqfd_polltbl);
375
376	/*
377	 * TODO: Ensure there isn't already an exclusive, priority waiter, e.g.
378	 * that the irqfd isn't already bound to another partition.  Only the
379	 * first exclusive waiter encountered will be notified, and
380	 * add_wait_queue_priority() doesn't enforce exclusivity.
381	 */
382	irqfd->irqfd_wait.flags |= WQ_FLAG_EXCLUSIVE;
383	add_wait_queue_priority(wqh, &irqfd->irqfd_wait);
384}
385
386static int mshv_irqfd_assign(struct mshv_partition *pt,
387			     struct mshv_user_irqfd *args)
388{
389	struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
390	struct mshv_irqfd *irqfd, *tmp;
391	__poll_t events;
392	int ret;
393	int idx;
394
395	CLASS(fd, f)(args->fd);
396
397	irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
398	if (!irqfd)
399		return -ENOMEM;
400
401	irqfd->irqfd_partn = pt;
402	irqfd->irqfd_irqnum = args->gsi;
403	INIT_WORK(&irqfd->irqfd_shutdown, mshv_irqfd_shutdown);
404	seqcount_spinlock_init(&irqfd->irqfd_irqe_sc, &pt->pt_irqfds_lock);
405
406	if (fd_empty(f)) {
407		ret = -EBADF;
408		goto out;
409	}
410
411	eventfd = eventfd_ctx_fileget(fd_file(f));
412	if (IS_ERR(eventfd)) {
413		ret = PTR_ERR(eventfd);
414		goto fail;
415	}
416
417	irqfd->irqfd_eventfd_ctx = eventfd;
418
419	if (args->flags & BIT(MSHV_IRQFD_BIT_RESAMPLE)) {
420		struct mshv_irqfd_resampler *rp;
421
422		resamplefd = eventfd_ctx_fdget(args->resamplefd);
423		if (IS_ERR(resamplefd)) {
424			ret = PTR_ERR(resamplefd);
425			goto fail;
426		}
427
428		irqfd->irqfd_resamplefd = resamplefd;
429
430		mutex_lock(&pt->irqfds_resampler_lock);
431
432		hlist_for_each_entry(rp, &pt->irqfds_resampler_list,
433				     rsmplr_hnode) {
434			if (rp->rsmplr_notifier.irq_ack_gsi ==
435							 irqfd->irqfd_irqnum) {
436				irqfd->irqfd_resampler = rp;
437				break;
438			}
439		}
440
441		if (!irqfd->irqfd_resampler) {
442			rp = kzalloc(sizeof(*rp), GFP_KERNEL_ACCOUNT);
443			if (!rp) {
444				ret = -ENOMEM;
445				mutex_unlock(&pt->irqfds_resampler_lock);
446				goto fail;
447			}
448
449			rp->rsmplr_partn = pt;
450			INIT_HLIST_HEAD(&rp->rsmplr_irqfd_list);
451			rp->rsmplr_notifier.irq_ack_gsi = irqfd->irqfd_irqnum;
452			rp->rsmplr_notifier.irq_acked =
453						      mshv_irqfd_resampler_ack;
454
455			hlist_add_head(&rp->rsmplr_hnode,
456				       &pt->irqfds_resampler_list);
457			mshv_register_irq_ack_notifier(pt,
458						       &rp->rsmplr_notifier);
459			irqfd->irqfd_resampler = rp;
460		}
461
462		hlist_add_head_rcu(&irqfd->irqfd_resampler_hnode,
463				   &irqfd->irqfd_resampler->rsmplr_irqfd_list);
464
465		mutex_unlock(&pt->irqfds_resampler_lock);
466	}
467
468	/*
469	 * Install our own custom wake-up handling so we are notified via
470	 * a callback whenever someone signals the underlying eventfd
471	 */
472	init_waitqueue_func_entry(&irqfd->irqfd_wait, mshv_irqfd_wakeup);
473	init_poll_funcptr(&irqfd->irqfd_polltbl, mshv_irqfd_queue_proc);
474
475	spin_lock_irq(&pt->pt_irqfds_lock);
476#if IS_ENABLED(CONFIG_X86)
477	if (args->flags & BIT(MSHV_IRQFD_BIT_RESAMPLE) &&
478	    !irqfd->irqfd_lapic_irq.lapic_control.level_triggered) {
479		/*
480		 * Resample Fd must be for level triggered interrupt
481		 * Otherwise return with failure
482		 */
483		spin_unlock_irq(&pt->pt_irqfds_lock);
484		ret = -EINVAL;
485		goto fail;
486	}
487#endif
488	ret = 0;
489	hlist_for_each_entry(tmp, &pt->pt_irqfds_list, irqfd_hnode) {
490		if (irqfd->irqfd_eventfd_ctx != tmp->irqfd_eventfd_ctx)
491			continue;
492		/* This fd is used for another irq already. */
493		ret = -EBUSY;
494		spin_unlock_irq(&pt->pt_irqfds_lock);
495		goto fail;
496	}
497
498	idx = srcu_read_lock(&pt->pt_irq_srcu);
499	mshv_irqfd_update(pt, irqfd);
500	hlist_add_head(&irqfd->irqfd_hnode, &pt->pt_irqfds_list);
501	spin_unlock_irq(&pt->pt_irqfds_lock);
502
503	/*
504	 * Check if there was an event already pending on the eventfd
505	 * before we registered, and trigger it as if we didn't miss it.
506	 */
507	events = vfs_poll(fd_file(f), &irqfd->irqfd_polltbl);
508
509	if (events & POLLIN)
510		mshv_assert_irq_slow(irqfd);
511
512	srcu_read_unlock(&pt->pt_irq_srcu, idx);
513	return 0;
514
515fail:
516	if (irqfd->irqfd_resampler)
517		mshv_irqfd_resampler_shutdown(irqfd);
518
519	if (resamplefd && !IS_ERR(resamplefd))
520		eventfd_ctx_put(resamplefd);
521
522	if (eventfd && !IS_ERR(eventfd))
523		eventfd_ctx_put(eventfd);
524
525out:
526	kfree(irqfd);
527	return ret;
528}
529
530/*
531 * shutdown any irqfd's that match fd+gsi
532 */
533static int mshv_irqfd_deassign(struct mshv_partition *pt,
534			       struct mshv_user_irqfd *args)
535{
536	struct mshv_irqfd *irqfd;
537	struct hlist_node *n;
538	struct eventfd_ctx *eventfd;
539
540	eventfd = eventfd_ctx_fdget(args->fd);
541	if (IS_ERR(eventfd))
542		return PTR_ERR(eventfd);
543
544	hlist_for_each_entry_safe(irqfd, n, &pt->pt_irqfds_list,
545				  irqfd_hnode) {
546		if (irqfd->irqfd_eventfd_ctx == eventfd &&
547		    irqfd->irqfd_irqnum == args->gsi)
548
549			mshv_irqfd_deactivate(irqfd);
550	}
551
552	eventfd_ctx_put(eventfd);
553
554	/*
555	 * Block until we know all outstanding shutdown jobs have completed
556	 * so that we guarantee there will not be any more interrupts on this
557	 * gsi once this deassign function returns.
558	 */
559	flush_workqueue(irqfd_cleanup_wq);
560
561	return 0;
562}
563
564int mshv_set_unset_irqfd(struct mshv_partition *pt,
565			 struct mshv_user_irqfd *args)
566{
567	if (args->flags & ~MSHV_IRQFD_FLAGS_MASK)
568		return -EINVAL;
569
570	if (args->flags & BIT(MSHV_IRQFD_BIT_DEASSIGN))
571		return mshv_irqfd_deassign(pt, args);
572
573	return mshv_irqfd_assign(pt, args);
574}
575
576/*
577 * This function is called as the mshv VM fd is being released.
578 * Shutdown all irqfds that still remain open
579 */
580static void mshv_irqfd_release(struct mshv_partition *pt)
581{
582	struct mshv_irqfd *irqfd;
583	struct hlist_node *n;
584
585	spin_lock_irq(&pt->pt_irqfds_lock);
586
587	hlist_for_each_entry_safe(irqfd, n, &pt->pt_irqfds_list, irqfd_hnode)
588		mshv_irqfd_deactivate(irqfd);
589
590	spin_unlock_irq(&pt->pt_irqfds_lock);
591
592	/*
593	 * Block until we know all outstanding shutdown jobs have completed
594	 * since we do not take a mshv_partition* reference.
595	 */
596	flush_workqueue(irqfd_cleanup_wq);
597}
598
599int mshv_irqfd_wq_init(void)
600{
601	irqfd_cleanup_wq = alloc_workqueue("mshv-irqfd-cleanup", WQ_PERCPU, 0);
602	if (!irqfd_cleanup_wq)
603		return -ENOMEM;
604
605	return 0;
606}
607
608void mshv_irqfd_wq_cleanup(void)
609{
610	destroy_workqueue(irqfd_cleanup_wq);
611}
612
613/*
614 * --------------------------------------------------------------------
615 * ioeventfd: translate a MMIO memory write to an eventfd signal.
616 *
617 * userspace can register a MMIO address with an eventfd for receiving
618 * notification when the memory has been touched.
619 * --------------------------------------------------------------------
620 */
621
622static void ioeventfd_release(struct mshv_ioeventfd *p, u64 partition_id)
623{
624	if (p->iovntfd_doorbell_id > 0)
625		mshv_unregister_doorbell(partition_id, p->iovntfd_doorbell_id);
626	eventfd_ctx_put(p->iovntfd_eventfd);
627	kfree(p);
628}
629
630/* MMIO writes trigger an event if the addr/val match */
631static void ioeventfd_mmio_write(int doorbell_id, void *data)
632{
633	struct mshv_partition *partition = (struct mshv_partition *)data;
634	struct mshv_ioeventfd *p;
635
636	rcu_read_lock();
637	hlist_for_each_entry_rcu(p, &partition->ioeventfds_list, iovntfd_hnode)
638		if (p->iovntfd_doorbell_id == doorbell_id) {
639			eventfd_signal(p->iovntfd_eventfd);
640			break;
641		}
642
643	rcu_read_unlock();
644}
645
646static bool ioeventfd_check_collision(struct mshv_partition *pt,
647				      struct mshv_ioeventfd *p)
648	__must_hold(&pt->mutex)
649{
650	struct mshv_ioeventfd *_p;
651
652	hlist_for_each_entry(_p, &pt->ioeventfds_list, iovntfd_hnode)
653		if (_p->iovntfd_addr == p->iovntfd_addr &&
654		    _p->iovntfd_length == p->iovntfd_length &&
655		    (_p->iovntfd_wildcard || p->iovntfd_wildcard ||
656		     _p->iovntfd_datamatch == p->iovntfd_datamatch))
657			return true;
658
659	return false;
660}
661
662static int mshv_assign_ioeventfd(struct mshv_partition *pt,
663				 struct mshv_user_ioeventfd *args)
664	__must_hold(&pt->mutex)
665{
666	struct mshv_ioeventfd *p;
667	struct eventfd_ctx *eventfd;
668	u64 doorbell_flags = 0;
669	int ret;
670
671	/* This mutex is currently protecting ioeventfd.items list */
672	WARN_ON_ONCE(!mutex_is_locked(&pt->pt_mutex));
673
674	if (args->flags & BIT(MSHV_IOEVENTFD_BIT_PIO))
675		return -EOPNOTSUPP;
676
677	/* must be natural-word sized */
678	switch (args->len) {
679	case 0:
680		doorbell_flags = HV_DOORBELL_FLAG_TRIGGER_SIZE_ANY;
681		break;
682	case 1:
683		doorbell_flags = HV_DOORBELL_FLAG_TRIGGER_SIZE_BYTE;
684		break;
685	case 2:
686		doorbell_flags = HV_DOORBELL_FLAG_TRIGGER_SIZE_WORD;
687		break;
688	case 4:
689		doorbell_flags = HV_DOORBELL_FLAG_TRIGGER_SIZE_DWORD;
690		break;
691	case 8:
692		doorbell_flags = HV_DOORBELL_FLAG_TRIGGER_SIZE_QWORD;
693		break;
694	default:
695		return -EINVAL;
696	}
697
698	/* check for range overflow */
699	if (args->addr + args->len < args->addr)
700		return -EINVAL;
701
702	/* check for extra flags that we don't understand */
703	if (args->flags & ~MSHV_IOEVENTFD_FLAGS_MASK)
704		return -EINVAL;
705
706	eventfd = eventfd_ctx_fdget(args->fd);
707	if (IS_ERR(eventfd))
708		return PTR_ERR(eventfd);
709
710	p = kzalloc(sizeof(*p), GFP_KERNEL);
711	if (!p) {
712		ret = -ENOMEM;
713		goto fail;
714	}
715
716	p->iovntfd_addr = args->addr;
717	p->iovntfd_length  = args->len;
718	p->iovntfd_eventfd = eventfd;
719
720	/* The datamatch feature is optional, otherwise this is a wildcard */
721	if (args->flags & BIT(MSHV_IOEVENTFD_BIT_DATAMATCH)) {
722		p->iovntfd_datamatch = args->datamatch;
723	} else {
724		p->iovntfd_wildcard = true;
725		doorbell_flags |= HV_DOORBELL_FLAG_TRIGGER_ANY_VALUE;
726	}
727
728	if (ioeventfd_check_collision(pt, p)) {
729		ret = -EEXIST;
730		goto unlock_fail;
731	}
732
733	ret = mshv_register_doorbell(pt->pt_id, ioeventfd_mmio_write,
734				     (void *)pt, p->iovntfd_addr,
735				     p->iovntfd_datamatch, doorbell_flags);
736	if (ret < 0)
737		goto unlock_fail;
738
739	p->iovntfd_doorbell_id = ret;
740
741	hlist_add_head_rcu(&p->iovntfd_hnode, &pt->ioeventfds_list);
742
743	return 0;
744
745unlock_fail:
746	kfree(p);
747
748fail:
749	eventfd_ctx_put(eventfd);
750
751	return ret;
752}
753
754static int mshv_deassign_ioeventfd(struct mshv_partition *pt,
755				   struct mshv_user_ioeventfd *args)
756	__must_hold(&pt->mutex)
757{
758	struct mshv_ioeventfd *p;
759	struct eventfd_ctx *eventfd;
760	struct hlist_node *n;
761	int ret = -ENOENT;
762
763	/* This mutex is currently protecting ioeventfd.items list */
764	WARN_ON_ONCE(!mutex_is_locked(&pt->pt_mutex));
765
766	eventfd = eventfd_ctx_fdget(args->fd);
767	if (IS_ERR(eventfd))
768		return PTR_ERR(eventfd);
769
770	hlist_for_each_entry_safe(p, n, &pt->ioeventfds_list, iovntfd_hnode) {
771		bool wildcard = !(args->flags & BIT(MSHV_IOEVENTFD_BIT_DATAMATCH));
772
773		if (p->iovntfd_eventfd != eventfd  ||
774		    p->iovntfd_addr != args->addr  ||
775		    p->iovntfd_length != args->len ||
776		    p->iovntfd_wildcard != wildcard)
777			continue;
778
779		if (!p->iovntfd_wildcard &&
780		    p->iovntfd_datamatch != args->datamatch)
781			continue;
782
783		hlist_del_rcu(&p->iovntfd_hnode);
784		synchronize_rcu();
785		ioeventfd_release(p, pt->pt_id);
786		ret = 0;
787		break;
788	}
789
790	eventfd_ctx_put(eventfd);
791
792	return ret;
793}
794
795int mshv_set_unset_ioeventfd(struct mshv_partition *pt,
796			     struct mshv_user_ioeventfd *args)
797	__must_hold(&pt->mutex)
798{
799	if ((args->flags & ~MSHV_IOEVENTFD_FLAGS_MASK) ||
800	    mshv_field_nonzero(*args, rsvd))
801		return -EINVAL;
802
803	/* PIO not yet implemented */
804	if (args->flags & BIT(MSHV_IOEVENTFD_BIT_PIO))
805		return -EOPNOTSUPP;
806
807	if (args->flags & BIT(MSHV_IOEVENTFD_BIT_DEASSIGN))
808		return mshv_deassign_ioeventfd(pt, args);
809
810	return mshv_assign_ioeventfd(pt, args);
811}
812
813void mshv_eventfd_init(struct mshv_partition *pt)
814{
815	spin_lock_init(&pt->pt_irqfds_lock);
816	INIT_HLIST_HEAD(&pt->pt_irqfds_list);
817
818	INIT_HLIST_HEAD(&pt->irqfds_resampler_list);
819	mutex_init(&pt->irqfds_resampler_lock);
820
821	INIT_HLIST_HEAD(&pt->ioeventfds_list);
822}
823
824void mshv_eventfd_release(struct mshv_partition *pt)
825{
826	struct hlist_head items;
827	struct hlist_node *n;
828	struct mshv_ioeventfd *p;
829
830	hlist_move_list(&pt->ioeventfds_list, &items);
831	synchronize_rcu();
832
833	hlist_for_each_entry_safe(p, n, &items, iovntfd_hnode) {
834		hlist_del(&p->iovntfd_hnode);
835		ioeventfd_release(p, pt->pt_id);
836	}
837
838	mshv_irqfd_release(pt);
839}