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