1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /**************************************************************************
3  *
4  * Copyright 2009-2015 VMware, Inc., Palo Alto, CA., USA
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 
28 #include <linux/sched/signal.h>
29 
30 #include "vmwgfx_drv.h"
31 
32 #define VMW_FENCE_WRAP (1 << 24)
33 
34 /**
35  * vmw_thread_fn - Deferred (process context) irq handler
36  *
37  * @irq: irq number
38  * @arg: Closure argument. Pointer to a struct drm_device cast to void *
39  *
40  * This function implements the deferred part of irq processing.
41  * The function is guaranteed to run at least once after the
42  * vmw_irq_handler has returned with IRQ_WAKE_THREAD.
43  *
44  */
45 static irqreturn_t vmw_thread_fn(int irq, void *arg)
46 {
47 	struct drm_device *dev = (struct drm_device *)arg;
48 	struct vmw_private *dev_priv = vmw_priv(dev);
49 	irqreturn_t ret = IRQ_NONE;
50 
51 	if (test_and_clear_bit(VMW_IRQTHREAD_FENCE,
52 			       dev_priv->irqthread_pending)) {
53 		vmw_fences_update(dev_priv->fman);
54 		wake_up_all(&dev_priv->fence_queue);
55 		ret = IRQ_HANDLED;
56 	}
57 
58 	if (test_and_clear_bit(VMW_IRQTHREAD_CMDBUF,
59 			       dev_priv->irqthread_pending)) {
60 		vmw_cmdbuf_irqthread(dev_priv->cman);
61 		ret = IRQ_HANDLED;
62 	}
63 
64 	return ret;
65 }
66 
67 /**
68  * vmw_irq_handler irq handler
69  *
70  * @irq: irq number
71  * @arg: Closure argument. Pointer to a struct drm_device cast to void *
72  *
73  * This function implements the quick part of irq processing.
74  * The function performs fast actions like clearing the device interrupt
75  * flags and also reasonably quick actions like waking processes waiting for
76  * FIFO space. Other IRQ actions are deferred to the IRQ thread.
77  */
78 static irqreturn_t vmw_irq_handler(int irq, void *arg)
79 {
80 	struct drm_device *dev = (struct drm_device *)arg;
81 	struct vmw_private *dev_priv = vmw_priv(dev);
82 	uint32_t status, masked_status;
83 	irqreturn_t ret = IRQ_HANDLED;
84 
85 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
86 	masked_status = status & READ_ONCE(dev_priv->irq_mask);
87 
88 	if (likely(status))
89 		outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
90 
91 	if (!status)
92 		return IRQ_NONE;
93 
94 	if (masked_status & SVGA_IRQFLAG_FIFO_PROGRESS)
95 		wake_up_all(&dev_priv->fifo_queue);
96 
97 	if ((masked_status & (SVGA_IRQFLAG_ANY_FENCE |
98 			      SVGA_IRQFLAG_FENCE_GOAL)) &&
99 	    !test_and_set_bit(VMW_IRQTHREAD_FENCE, dev_priv->irqthread_pending))
100 		ret = IRQ_WAKE_THREAD;
101 
102 	if ((masked_status & (SVGA_IRQFLAG_COMMAND_BUFFER |
103 			      SVGA_IRQFLAG_ERROR)) &&
104 	    !test_and_set_bit(VMW_IRQTHREAD_CMDBUF,
105 			      dev_priv->irqthread_pending))
106 		ret = IRQ_WAKE_THREAD;
107 
108 	return ret;
109 }
110 
111 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno)
112 {
113 
114 	return (vmw_read(dev_priv, SVGA_REG_BUSY) == 0);
115 }
116 
117 void vmw_update_seqno(struct vmw_private *dev_priv,
118 			 struct vmw_fifo_state *fifo_state)
119 {
120 	u32 *fifo_mem = dev_priv->mmio_virt;
121 	uint32_t seqno = vmw_mmio_read(fifo_mem + SVGA_FIFO_FENCE);
122 
123 	if (dev_priv->last_read_seqno != seqno) {
124 		dev_priv->last_read_seqno = seqno;
125 		vmw_marker_pull(&fifo_state->marker_queue, seqno);
126 		vmw_fences_update(dev_priv->fman);
127 	}
128 }
129 
130 bool vmw_seqno_passed(struct vmw_private *dev_priv,
131 			 uint32_t seqno)
132 {
133 	struct vmw_fifo_state *fifo_state;
134 	bool ret;
135 
136 	if (likely(dev_priv->last_read_seqno - seqno < VMW_FENCE_WRAP))
137 		return true;
138 
139 	fifo_state = &dev_priv->fifo;
140 	vmw_update_seqno(dev_priv, fifo_state);
141 	if (likely(dev_priv->last_read_seqno - seqno < VMW_FENCE_WRAP))
142 		return true;
143 
144 	if (!(fifo_state->capabilities & SVGA_FIFO_CAP_FENCE) &&
145 	    vmw_fifo_idle(dev_priv, seqno))
146 		return true;
147 
148 	/**
149 	 * Then check if the seqno is higher than what we've actually
150 	 * emitted. Then the fence is stale and signaled.
151 	 */
152 
153 	ret = ((atomic_read(&dev_priv->marker_seq) - seqno)
154 	       > VMW_FENCE_WRAP);
155 
156 	return ret;
157 }
158 
159 int vmw_fallback_wait(struct vmw_private *dev_priv,
160 		      bool lazy,
161 		      bool fifo_idle,
162 		      uint32_t seqno,
163 		      bool interruptible,
164 		      unsigned long timeout)
165 {
166 	struct vmw_fifo_state *fifo_state = &dev_priv->fifo;
167 
168 	uint32_t count = 0;
169 	uint32_t signal_seq;
170 	int ret;
171 	unsigned long end_jiffies = jiffies + timeout;
172 	bool (*wait_condition)(struct vmw_private *, uint32_t);
173 	DEFINE_WAIT(__wait);
174 
175 	wait_condition = (fifo_idle) ? &vmw_fifo_idle :
176 		&vmw_seqno_passed;
177 
178 	/**
179 	 * Block command submission while waiting for idle.
180 	 */
181 
182 	if (fifo_idle) {
183 		down_read(&fifo_state->rwsem);
184 		if (dev_priv->cman) {
185 			ret = vmw_cmdbuf_idle(dev_priv->cman, interruptible,
186 					      10*HZ);
187 			if (ret)
188 				goto out_err;
189 		}
190 	}
191 
192 	signal_seq = atomic_read(&dev_priv->marker_seq);
193 	ret = 0;
194 
195 	for (;;) {
196 		prepare_to_wait(&dev_priv->fence_queue, &__wait,
197 				(interruptible) ?
198 				TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
199 		if (wait_condition(dev_priv, seqno))
200 			break;
201 		if (time_after_eq(jiffies, end_jiffies)) {
202 			DRM_ERROR("SVGA device lockup.\n");
203 			break;
204 		}
205 		if (lazy)
206 			schedule_timeout(1);
207 		else if ((++count & 0x0F) == 0) {
208 			/**
209 			 * FIXME: Use schedule_hr_timeout here for
210 			 * newer kernels and lower CPU utilization.
211 			 */
212 
213 			__set_current_state(TASK_RUNNING);
214 			schedule();
215 			__set_current_state((interruptible) ?
216 					    TASK_INTERRUPTIBLE :
217 					    TASK_UNINTERRUPTIBLE);
218 		}
219 		if (interruptible && signal_pending(current)) {
220 			ret = -ERESTARTSYS;
221 			break;
222 		}
223 	}
224 	finish_wait(&dev_priv->fence_queue, &__wait);
225 	if (ret == 0 && fifo_idle) {
226 		u32 *fifo_mem = dev_priv->mmio_virt;
227 
228 		vmw_mmio_write(signal_seq, fifo_mem + SVGA_FIFO_FENCE);
229 	}
230 	wake_up_all(&dev_priv->fence_queue);
231 out_err:
232 	if (fifo_idle)
233 		up_read(&fifo_state->rwsem);
234 
235 	return ret;
236 }
237 
238 void vmw_generic_waiter_add(struct vmw_private *dev_priv,
239 			    u32 flag, int *waiter_count)
240 {
241 	spin_lock_bh(&dev_priv->waiter_lock);
242 	if ((*waiter_count)++ == 0) {
243 		outl(flag, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
244 		dev_priv->irq_mask |= flag;
245 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
246 	}
247 	spin_unlock_bh(&dev_priv->waiter_lock);
248 }
249 
250 void vmw_generic_waiter_remove(struct vmw_private *dev_priv,
251 			       u32 flag, int *waiter_count)
252 {
253 	spin_lock_bh(&dev_priv->waiter_lock);
254 	if (--(*waiter_count) == 0) {
255 		dev_priv->irq_mask &= ~flag;
256 		vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask);
257 	}
258 	spin_unlock_bh(&dev_priv->waiter_lock);
259 }
260 
261 void vmw_seqno_waiter_add(struct vmw_private *dev_priv)
262 {
263 	vmw_generic_waiter_add(dev_priv, SVGA_IRQFLAG_ANY_FENCE,
264 			       &dev_priv->fence_queue_waiters);
265 }
266 
267 void vmw_seqno_waiter_remove(struct vmw_private *dev_priv)
268 {
269 	vmw_generic_waiter_remove(dev_priv, SVGA_IRQFLAG_ANY_FENCE,
270 				  &dev_priv->fence_queue_waiters);
271 }
272 
273 void vmw_goal_waiter_add(struct vmw_private *dev_priv)
274 {
275 	vmw_generic_waiter_add(dev_priv, SVGA_IRQFLAG_FENCE_GOAL,
276 			       &dev_priv->goal_queue_waiters);
277 }
278 
279 void vmw_goal_waiter_remove(struct vmw_private *dev_priv)
280 {
281 	vmw_generic_waiter_remove(dev_priv, SVGA_IRQFLAG_FENCE_GOAL,
282 				  &dev_priv->goal_queue_waiters);
283 }
284 
285 int vmw_wait_seqno(struct vmw_private *dev_priv,
286 		      bool lazy, uint32_t seqno,
287 		      bool interruptible, unsigned long timeout)
288 {
289 	long ret;
290 	struct vmw_fifo_state *fifo = &dev_priv->fifo;
291 
292 	if (likely(dev_priv->last_read_seqno - seqno < VMW_FENCE_WRAP))
293 		return 0;
294 
295 	if (likely(vmw_seqno_passed(dev_priv, seqno)))
296 		return 0;
297 
298 	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
299 
300 	if (!(fifo->capabilities & SVGA_FIFO_CAP_FENCE))
301 		return vmw_fallback_wait(dev_priv, lazy, true, seqno,
302 					 interruptible, timeout);
303 
304 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
305 		return vmw_fallback_wait(dev_priv, lazy, false, seqno,
306 					 interruptible, timeout);
307 
308 	vmw_seqno_waiter_add(dev_priv);
309 
310 	if (interruptible)
311 		ret = wait_event_interruptible_timeout
312 		    (dev_priv->fence_queue,
313 		     vmw_seqno_passed(dev_priv, seqno),
314 		     timeout);
315 	else
316 		ret = wait_event_timeout
317 		    (dev_priv->fence_queue,
318 		     vmw_seqno_passed(dev_priv, seqno),
319 		     timeout);
320 
321 	vmw_seqno_waiter_remove(dev_priv);
322 
323 	if (unlikely(ret == 0))
324 		ret = -EBUSY;
325 	else if (likely(ret > 0))
326 		ret = 0;
327 
328 	return ret;
329 }
330 
331 static void vmw_irq_preinstall(struct drm_device *dev)
332 {
333 	struct vmw_private *dev_priv = vmw_priv(dev);
334 	uint32_t status;
335 
336 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
337 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
338 }
339 
340 void vmw_irq_uninstall(struct drm_device *dev)
341 {
342 	struct vmw_private *dev_priv = vmw_priv(dev);
343 	uint32_t status;
344 
345 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
346 		return;
347 
348 	if (!dev->irq_enabled)
349 		return;
350 
351 	vmw_write(dev_priv, SVGA_REG_IRQMASK, 0);
352 
353 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
354 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
355 
356 	dev->irq_enabled = false;
357 	free_irq(dev->irq, dev);
358 }
359 
360 /**
361  * vmw_irq_install - Install the irq handlers
362  *
363  * @dev:  Pointer to the drm device.
364  * @irq:  The irq number.
365  * Return:  Zero if successful. Negative number otherwise.
366  */
367 int vmw_irq_install(struct drm_device *dev, int irq)
368 {
369 	int ret;
370 
371 	if (dev->irq_enabled)
372 		return -EBUSY;
373 
374 	vmw_irq_preinstall(dev);
375 
376 	ret = request_threaded_irq(irq, vmw_irq_handler, vmw_thread_fn,
377 				   IRQF_SHARED, VMWGFX_DRIVER_NAME, dev);
378 	if (ret < 0)
379 		return ret;
380 
381 	dev->irq_enabled = true;
382 	dev->irq = irq;
383 
384 	return ret;
385 }
386