1 /*	$NetBSD: vmwgfx_irq.c,v 1.6 2022/10/25 23:36:21 riastradh Exp $	*/
2 
3 // SPDX-License-Identifier: GPL-2.0 OR MIT
4 /**************************************************************************
5  *
6  * Copyright 2009-2015 VMware, Inc., Palo Alto, CA., USA
7  *
8  * Permission is hereby granted, free of charge, to any person obtaining a
9  * copy of this software and associated documentation files (the
10  * "Software"), to deal in the Software without restriction, including
11  * without limitation the rights to use, copy, modify, merge, publish,
12  * distribute, sub license, and/or sell copies of the Software, and to
13  * permit persons to whom the Software is furnished to do so, subject to
14  * the following conditions:
15  *
16  * The above copyright notice and this permission notice (including the
17  * next paragraph) shall be included in all copies or substantial portions
18  * of the Software.
19  *
20  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
21  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
23  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
24  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
25  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
26  * USE OR OTHER DEALINGS IN THE SOFTWARE.
27  *
28  **************************************************************************/
29 
30 #include <sys/cdefs.h>
31 __KERNEL_RCSID(0, "$NetBSD: vmwgfx_irq.c,v 1.6 2022/10/25 23:36:21 riastradh Exp $");
32 
33 #include <linux/sched/signal.h>
34 
35 #include <drm/drm_irq.h>
36 
37 #include "vmwgfx_drv.h"
38 
39 #define VMW_FENCE_WRAP (1 << 24)
40 
41 /**
42  * vmw_thread_fn - Deferred (process context) irq handler
43  *
44  * @irq: irq number
45  * @arg: Closure argument. Pointer to a struct drm_device cast to void *
46  *
47  * This function implements the deferred part of irq processing.
48  * The function is guaranteed to run at least once after the
49  * vmw_irq_handler has returned with IRQ_WAKE_THREAD.
50  *
51  */
52 #ifdef __NetBSD__
53 static void
vmw_thread_fn(struct work * work,void * arg)54 vmw_thread_fn(struct work *work, void *arg)
55 #else
56 static irqreturn_t vmw_thread_fn(int irq, void *arg)
57 #endif
58 {
59 	struct drm_device *dev = (struct drm_device *)arg;
60 	struct vmw_private *dev_priv = vmw_priv(dev);
61 	irqreturn_t ret = IRQ_NONE;
62 
63 #ifdef __NetBSD__
64 	atomic_store_relaxed(&dev_priv->irqthread_scheduled, false);
65 #endif
66 
67 	if (test_and_clear_bit(VMW_IRQTHREAD_FENCE,
68 			       dev_priv->irqthread_pending)) {
69 		spin_lock(&dev_priv->fence_lock);
70 		vmw_fences_update(dev_priv->fman);
71 		DRM_SPIN_WAKEUP_ALL(&dev_priv->fence_queue,
72 		    &dev_priv->fence_lock);
73 		spin_unlock(&dev_priv->fence_lock);
74 		ret = IRQ_HANDLED;
75 	}
76 
77 	if (test_and_clear_bit(VMW_IRQTHREAD_CMDBUF,
78 			       dev_priv->irqthread_pending)) {
79 		vmw_cmdbuf_irqthread(dev_priv->cman);
80 		ret = IRQ_HANDLED;
81 	}
82 
83 #ifndef __NetBSD__
84 	return ret;
85 #endif
86 }
87 
88 /**
89  * vmw_irq_handler irq handler
90  *
91  * @irq: irq number
92  * @arg: Closure argument. Pointer to a struct drm_device cast to void *
93  *
94  * This function implements the quick part of irq processing.
95  * The function performs fast actions like clearing the device interrupt
96  * flags and also reasonably quick actions like waking processes waiting for
97  * FIFO space. Other IRQ actions are deferred to the IRQ thread.
98  */
vmw_irq_handler(int irq,void * arg)99 static irqreturn_t vmw_irq_handler(int irq, void *arg)
100 {
101 	struct drm_device *dev = (struct drm_device *)arg;
102 	struct vmw_private *dev_priv = vmw_priv(dev);
103 	uint32_t status, masked_status;
104 	irqreturn_t ret = IRQ_HANDLED;
105 
106 #ifdef __NetBSD__
107 	status = bus_space_read_4(dev_priv->iot, dev_priv->ioh,
108 	    VMWGFX_IRQSTATUS_PORT);
109 #else
110 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
111 #endif
112 	masked_status = status & READ_ONCE(dev_priv->irq_mask);
113 
114 	if (likely(status))
115 #ifdef __NetBSD__
116 		bus_space_write_4(dev_priv->iot, dev_priv->ioh,
117 		    VMWGFX_IRQSTATUS_PORT, status);
118 #else
119 		outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
120 #endif
121 
122 	if (!status)
123 		return IRQ_NONE;
124 
125 	if (masked_status & SVGA_IRQFLAG_FIFO_PROGRESS) {
126 		spin_lock(&dev_priv->fifo_lock);
127 		DRM_SPIN_WAKEUP_ALL(&dev_priv->fifo_queue,
128 		    &dev_priv->fifo_lock);
129 		spin_unlock(&dev_priv->fifo_lock);
130 	}
131 
132 	if ((masked_status & (SVGA_IRQFLAG_ANY_FENCE |
133 			      SVGA_IRQFLAG_FENCE_GOAL)) &&
134 	    !test_and_set_bit(VMW_IRQTHREAD_FENCE, dev_priv->irqthread_pending))
135 		ret = IRQ_WAKE_THREAD;
136 
137 	if ((masked_status & (SVGA_IRQFLAG_COMMAND_BUFFER |
138 			      SVGA_IRQFLAG_ERROR)) &&
139 	    !test_and_set_bit(VMW_IRQTHREAD_CMDBUF,
140 			      dev_priv->irqthread_pending))
141 		ret = IRQ_WAKE_THREAD;
142 
143 #ifdef __NetBSD__
144 	if (ret == IRQ_WAKE_THREAD) {
145 		if (atomic_swap_uint(&dev_priv->irqthread_scheduled, 1) == 0) {
146 			workqueue_enqueue(dev_priv->irqthread_wq,
147 			    &dev_priv->irqthread_work, NULL);
148 		}
149 		ret = IRQ_HANDLED;
150 	}
151 #endif
152 
153 	return ret;
154 }
155 
vmw_fifo_idle(struct vmw_private * dev_priv,uint32_t seqno)156 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno)
157 {
158 
159 	return (vmw_read(dev_priv, SVGA_REG_BUSY) == 0);
160 }
161 
vmw_update_seqno(struct vmw_private * dev_priv,struct vmw_fifo_state * fifo_state)162 void vmw_update_seqno(struct vmw_private *dev_priv,
163 			 struct vmw_fifo_state *fifo_state)
164 {
165 	u32 *fifo_mem = dev_priv->mmio_virt;
166 	uint32_t seqno = vmw_mmio_read(fifo_mem + SVGA_FIFO_FENCE);
167 
168 	assert_spin_locked(&dev_priv->fence_lock);
169 
170 	if (dev_priv->last_read_seqno != seqno) {
171 		dev_priv->last_read_seqno = seqno;
172 		vmw_marker_pull(&fifo_state->marker_queue, seqno);
173 		vmw_fences_update(dev_priv->fman);
174 	}
175 }
176 
vmw_seqno_passed(struct vmw_private * dev_priv,uint32_t seqno)177 bool vmw_seqno_passed(struct vmw_private *dev_priv,
178 			 uint32_t seqno)
179 {
180 	struct vmw_fifo_state *fifo_state;
181 	bool ret;
182 
183 	assert_spin_locked(&dev_priv->fence_lock);
184 
185 	if (likely(dev_priv->last_read_seqno - seqno < VMW_FENCE_WRAP))
186 		return true;
187 
188 	fifo_state = &dev_priv->fifo;
189 	vmw_update_seqno(dev_priv, fifo_state);
190 	if (likely(dev_priv->last_read_seqno - seqno < VMW_FENCE_WRAP))
191 		return true;
192 
193 	if (!(fifo_state->capabilities & SVGA_FIFO_CAP_FENCE) &&
194 	    vmw_fifo_idle(dev_priv, seqno))
195 		return true;
196 
197 	/**
198 	 * Then check if the seqno is higher than what we've actually
199 	 * emitted. Then the fence is stale and signaled.
200 	 */
201 
202 	ret = ((atomic_read(&dev_priv->marker_seq) - seqno)
203 	       > VMW_FENCE_WRAP);
204 
205 	return ret;
206 }
207 
vmw_fallback_wait(struct vmw_private * dev_priv,bool lazy,bool fifo_idle,uint32_t seqno,bool interruptible,unsigned long timeout)208 int vmw_fallback_wait(struct vmw_private *dev_priv,
209 		      bool lazy,
210 		      bool fifo_idle,
211 		      uint32_t seqno,
212 		      bool interruptible,
213 		      unsigned long timeout)
214 {
215 	struct vmw_fifo_state *fifo_state = &dev_priv->fifo;
216 
217 	uint32_t count = 0;
218 	uint32_t signal_seq;
219 	int ret;
220 	unsigned long end_jiffies = jiffies + timeout;
221 	bool (*wait_condition)(struct vmw_private *, uint32_t);
222 #ifndef __NetBSD__
223 	DEFINE_WAIT(__wait);
224 #endif
225 
226 	wait_condition = (fifo_idle) ? &vmw_fifo_idle :
227 		&vmw_seqno_passed;
228 
229 	/**
230 	 * Block command submission while waiting for idle.
231 	 */
232 
233 	if (fifo_idle) {
234 		down_read(&fifo_state->rwsem);
235 		if (dev_priv->cman) {
236 			ret = vmw_cmdbuf_idle(dev_priv->cman, interruptible,
237 					      10*HZ);
238 			if (ret)
239 				goto out_err;
240 		}
241 	}
242 
243 	spin_lock(&dev_priv->fence_lock);
244 
245 	signal_seq = atomic_read(&dev_priv->marker_seq);
246 	ret = 0;
247 
248 	for (;;) {
249 #ifdef __NetBSD__
250 		if (!lazy) {
251 			if (wait_condition(dev_priv, seqno))
252 				break;
253 			spin_unlock(&dev_priv->fence_lock);
254 			if ((++count & 0xf) == 0)
255 				yield();
256 			spin_lock(&dev_priv->fence_lock);
257 		} else if (interruptible) {
258 			DRM_SPIN_TIMED_WAIT_UNTIL(ret, &dev_priv->fence_queue,
259 			    &dev_priv->fence_lock, /*timeout*/1,
260 			    wait_condition(dev_priv, seqno));
261 		} else {
262 			DRM_SPIN_TIMED_WAIT_NOINTR_UNTIL(ret,
263 			    &dev_priv->fence_queue,
264 			    &dev_priv->fence_lock, /*timeout*/1,
265 			    wait_condition(dev_priv, seqno));
266 		}
267 		if (ret) {	/* success or error but not timeout */
268 			if (ret > 0) /* success */
269 				ret = 0;
270 			break;
271 		}
272 		if (time_after_eq(jiffies, end_jiffies)) {
273 			DRM_ERROR("SVGA device lockup.\n");
274 			break;
275 		}
276 #else
277 		prepare_to_wait(&dev_priv->fence_queue, &__wait,
278 				(interruptible) ?
279 				TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
280 		if (wait_condition(dev_priv, seqno))
281 			break;
282 		if (time_after_eq(jiffies, end_jiffies)) {
283 			DRM_ERROR("SVGA device lockup.\n");
284 			break;
285 		}
286 		if (lazy)
287 			schedule_timeout(1);
288 		else if ((++count & 0x0F) == 0) {
289 			/**
290 			 * FIXME: Use schedule_hr_timeout here for
291 			 * newer kernels and lower CPU utilization.
292 			 */
293 
294 			__set_current_state(TASK_RUNNING);
295 			schedule();
296 			__set_current_state((interruptible) ?
297 					    TASK_INTERRUPTIBLE :
298 					    TASK_UNINTERRUPTIBLE);
299 		}
300 		if (interruptible && signal_pending(current)) {
301 			ret = -ERESTARTSYS;
302 			break;
303 		}
304 #endif
305 	}
306 #ifndef __NetBSD__
307 	finish_wait(&dev_priv->fence_queue, &__wait);
308 #endif
309 	if (ret == 0 && fifo_idle) {
310 		u32 *fifo_mem = dev_priv->mmio_virt;
311 
312 		vmw_mmio_write(signal_seq, fifo_mem + SVGA_FIFO_FENCE);
313 	}
314 #ifdef __NetBSD__
315 	DRM_SPIN_WAKEUP_ALL(&dev_priv->fence_queue, &dev_priv->fence_lock);
316 	spin_unlock(&dev_priv->fence_lock);
317 #else
318 	wake_up_all(&dev_priv->fence_queue);
319 #endif
320 out_err:
321 	if (fifo_idle)
322 		up_read(&fifo_state->rwsem);
323 
324 	return ret;
325 }
326 
vmw_generic_waiter_add(struct vmw_private * dev_priv,u32 flag,int * waiter_count)327 void vmw_generic_waiter_add(struct vmw_private *dev_priv,
328 			    u32 flag, int *waiter_count)
329 {
330 	spin_lock_bh(&dev_priv->waiter_lock);
331 	if ((*waiter_count)++ == 0) {
332 #ifdef __NetBSD__
333 		bus_space_write_4(dev_priv->iot, dev_priv->ioh,
334 		    VMWGFX_IRQSTATUS_PORT, flag);
335 #else
336 		outl(flag, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
337 #endif
338 		dev_priv->irq_mask |= flag;
339 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
340 	}
341 	spin_unlock_bh(&dev_priv->waiter_lock);
342 }
343 
vmw_generic_waiter_remove(struct vmw_private * dev_priv,u32 flag,int * waiter_count)344 void vmw_generic_waiter_remove(struct vmw_private *dev_priv,
345 			       u32 flag, int *waiter_count)
346 {
347 	spin_lock_bh(&dev_priv->waiter_lock);
348 	if (--(*waiter_count) == 0) {
349 		dev_priv->irq_mask &= ~flag;
350 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
351 	}
352 	spin_unlock_bh(&dev_priv->waiter_lock);
353 }
354 
vmw_seqno_waiter_add(struct vmw_private * dev_priv)355 void vmw_seqno_waiter_add(struct vmw_private *dev_priv)
356 {
357 	vmw_generic_waiter_add(dev_priv, SVGA_IRQFLAG_ANY_FENCE,
358 			       &dev_priv->fence_queue_waiters);
359 }
360 
vmw_seqno_waiter_remove(struct vmw_private * dev_priv)361 void vmw_seqno_waiter_remove(struct vmw_private *dev_priv)
362 {
363 	vmw_generic_waiter_remove(dev_priv, SVGA_IRQFLAG_ANY_FENCE,
364 				  &dev_priv->fence_queue_waiters);
365 }
366 
vmw_goal_waiter_add(struct vmw_private * dev_priv)367 void vmw_goal_waiter_add(struct vmw_private *dev_priv)
368 {
369 	vmw_generic_waiter_add(dev_priv, SVGA_IRQFLAG_FENCE_GOAL,
370 			       &dev_priv->goal_queue_waiters);
371 }
372 
vmw_goal_waiter_remove(struct vmw_private * dev_priv)373 void vmw_goal_waiter_remove(struct vmw_private *dev_priv)
374 {
375 	vmw_generic_waiter_remove(dev_priv, SVGA_IRQFLAG_FENCE_GOAL,
376 				  &dev_priv->goal_queue_waiters);
377 }
378 
vmw_wait_seqno(struct vmw_private * dev_priv,bool lazy,uint32_t seqno,bool interruptible,unsigned long timeout)379 int vmw_wait_seqno(struct vmw_private *dev_priv,
380 		      bool lazy, uint32_t seqno,
381 		      bool interruptible, unsigned long timeout)
382 {
383 	long ret;
384 	struct vmw_fifo_state *fifo = &dev_priv->fifo;
385 
386 	spin_lock(&dev_priv->fence_lock);
387 	if (likely(dev_priv->last_read_seqno - seqno < VMW_FENCE_WRAP)) {
388 		spin_unlock(&dev_priv->fence_lock);
389 		return 0;
390 	}
391 
392 	if (likely(vmw_seqno_passed(dev_priv, seqno))) {
393 		spin_unlock(&dev_priv->fence_lock);
394 		return 0;
395 	}
396 
397 	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
398 
399 	if (!(fifo->capabilities & SVGA_FIFO_CAP_FENCE)) {
400 		spin_unlock(&dev_priv->fence_lock);
401 		return vmw_fallback_wait(dev_priv, lazy, true, seqno,
402 					 interruptible, timeout);
403 	}
404 
405 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK)) {
406 		spin_unlock(&dev_priv->fence_lock);
407 		return vmw_fallback_wait(dev_priv, lazy, false, seqno,
408 					 interruptible, timeout);
409 	}
410 
411 	vmw_seqno_waiter_add(dev_priv);
412 
413 	if (interruptible)
414 		DRM_SPIN_TIMED_WAIT_UNTIL(ret, &dev_priv->fence_queue,
415 		    &dev_priv->fence_lock, timeout,
416 		    vmw_seqno_passed(dev_priv, seqno));
417 	else
418 		DRM_SPIN_TIMED_WAIT_NOINTR_UNTIL(ret, &dev_priv->fence_queue,
419 		    &dev_priv->fence_lock, timeout,
420 		    vmw_seqno_passed(dev_priv, seqno));
421 
422 	vmw_seqno_waiter_remove(dev_priv);
423 
424 	spin_unlock(&dev_priv->fence_lock);
425 
426 	if (unlikely(ret == 0))
427 		ret = -EBUSY;
428 	else if (likely(ret > 0))
429 		ret = 0;
430 
431 	return ret;
432 }
433 
vmw_irq_preinstall(struct drm_device * dev)434 static void vmw_irq_preinstall(struct drm_device *dev)
435 {
436 	struct vmw_private *dev_priv = vmw_priv(dev);
437 	uint32_t status;
438 
439 #ifdef __NetBSD__
440 	status = bus_space_read_4(dev_priv->iot, dev_priv->ioh,
441 	    VMWGFX_IRQSTATUS_PORT);
442 	bus_space_write_4(dev_priv->iot, dev_priv->ioh, VMWGFX_IRQSTATUS_PORT,
443 	    status);
444 #else
445 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
446 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
447 #endif
448 }
449 
vmw_irq_uninstall(struct drm_device * dev)450 void vmw_irq_uninstall(struct drm_device *dev)
451 {
452 	struct vmw_private *dev_priv = vmw_priv(dev);
453 	uint32_t status;
454 
455 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
456 		return;
457 
458 	if (!dev->irq_enabled)
459 		return;
460 
461 	vmw_write(dev_priv, SVGA_REG_IRQMASK, 0);
462 
463 #ifdef __NetBSD__
464 	status = bus_space_read_4(dev_priv->iot, dev_priv->ioh,
465 	    VMWGFX_IRQSTATUS_PORT);
466 	bus_space_write_4(dev_priv->iot, dev_priv->ioh, VMWGFX_IRQSTATUS_PORT,
467 	    status);
468 #else
469 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
470 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
471 #endif
472 
473 	dev->irq_enabled = false;
474 #ifdef __NetBSD__
475 	int ret = drm_irq_uninstall(dev);
476 	KASSERT(ret == 0);
477 	workqueue_destroy(dev_priv->irqthread_wq);
478 #else
479 	free_irq(dev->irq, dev);
480 #endif
481 }
482 
483 /**
484  * vmw_irq_install - Install the irq handlers
485  *
486  * @dev:  Pointer to the drm device.
487  * @irq:  The irq number.
488  * Return:  Zero if successful. Negative number otherwise.
489  */
vmw_irq_install(struct drm_device * dev,int irq)490 int vmw_irq_install(struct drm_device *dev, int irq)
491 {
492 	int ret;
493 
494 	if (dev->irq_enabled)
495 		return -EBUSY;
496 
497 	vmw_irq_preinstall(dev);
498 
499 #ifdef __NetBSD__
500 	/* XXX errno NetBSD->Linux */
501 	ret = -workqueue_create(&vmw_priv(dev)->irqthread_wq, "vmwgfirq",
502 	    vmw_thread_fn, dev, PRI_NONE, IPL_DRM, WQ_MPSAFE);
503 	if (ret < 0)
504 		return ret;
505 	ret = drm_irq_install(dev);
506 	if (ret < 0) {
507 		workqueue_destroy(vmw_priv(dev)->irqthread_wq);
508 		vmw_priv(dev)->irqthread_wq = NULL;
509 	}
510 #else
511 	ret = request_threaded_irq(irq, vmw_irq_handler, vmw_thread_fn,
512 				   IRQF_SHARED, VMWGFX_DRIVER_NAME, dev);
513 #endif
514 	if (ret < 0)
515 		return ret;
516 
517 	dev->irq_enabled = true;
518 	dev->irq = irq;
519 
520 	return ret;
521 }
522