1 /**************************************************************************
2  *
3  * Copyright © 2009 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 "drmP.h"
29 #include "vmwgfx_drv.h"
30 
31 #define VMW_FENCE_WRAP (1 << 24)
32 
33 irqreturn_t vmw_irq_handler(DRM_IRQ_ARGS)
34 {
35 	struct drm_device *dev = (struct drm_device *)arg;
36 	struct vmw_private *dev_priv = vmw_priv(dev);
37 	uint32_t status;
38 
39 	spin_lock(&dev_priv->irq_lock);
40 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
41 	spin_unlock(&dev_priv->irq_lock);
42 
43 	if (status & SVGA_IRQFLAG_ANY_FENCE)
44 		wake_up_all(&dev_priv->fence_queue);
45 	if (status & SVGA_IRQFLAG_FIFO_PROGRESS)
46 		wake_up_all(&dev_priv->fifo_queue);
47 
48 	if (likely(status)) {
49 		outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
50 		return IRQ_HANDLED;
51 	}
52 
53 	return IRQ_NONE;
54 }
55 
56 static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t sequence)
57 {
58 	uint32_t busy;
59 
60 	mutex_lock(&dev_priv->hw_mutex);
61 	busy = vmw_read(dev_priv, SVGA_REG_BUSY);
62 	mutex_unlock(&dev_priv->hw_mutex);
63 
64 	return (busy == 0);
65 }
66 
67 
68 bool vmw_fence_signaled(struct vmw_private *dev_priv,
69 			uint32_t sequence)
70 {
71 	__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
72 	struct vmw_fifo_state *fifo_state;
73 	bool ret;
74 
75 	if (likely(dev_priv->last_read_sequence - sequence < VMW_FENCE_WRAP))
76 		return true;
77 
78 	dev_priv->last_read_sequence = ioread32(fifo_mem + SVGA_FIFO_FENCE);
79 	if (likely(dev_priv->last_read_sequence - sequence < VMW_FENCE_WRAP))
80 		return true;
81 
82 	fifo_state = &dev_priv->fifo;
83 	if (!(fifo_state->capabilities & SVGA_FIFO_CAP_FENCE) &&
84 	    vmw_fifo_idle(dev_priv, sequence))
85 		return true;
86 
87 	/**
88 	 * Below is to signal stale fences that have wrapped.
89 	 * First, block fence submission.
90 	 */
91 
92 	down_read(&fifo_state->rwsem);
93 
94 	/**
95 	 * Then check if the sequence is higher than what we've actually
96 	 * emitted. Then the fence is stale and signaled.
97 	 */
98 
99 	ret = ((dev_priv->fence_seq - sequence) > VMW_FENCE_WRAP);
100 	up_read(&fifo_state->rwsem);
101 
102 	return ret;
103 }
104 
105 int vmw_fallback_wait(struct vmw_private *dev_priv,
106 		      bool lazy,
107 		      bool fifo_idle,
108 		      uint32_t sequence,
109 		      bool interruptible,
110 		      unsigned long timeout)
111 {
112 	struct vmw_fifo_state *fifo_state = &dev_priv->fifo;
113 
114 	uint32_t count = 0;
115 	uint32_t signal_seq;
116 	int ret;
117 	unsigned long end_jiffies = jiffies + timeout;
118 	bool (*wait_condition)(struct vmw_private *, uint32_t);
119 	DEFINE_WAIT(__wait);
120 
121 	wait_condition = (fifo_idle) ? &vmw_fifo_idle :
122 		&vmw_fence_signaled;
123 
124 	/**
125 	 * Block command submission while waiting for idle.
126 	 */
127 
128 	if (fifo_idle)
129 		down_read(&fifo_state->rwsem);
130 	signal_seq = dev_priv->fence_seq;
131 	ret = 0;
132 
133 	for (;;) {
134 		prepare_to_wait(&dev_priv->fence_queue, &__wait,
135 				(interruptible) ?
136 				TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
137 		if (wait_condition(dev_priv, sequence))
138 			break;
139 		if (time_after_eq(jiffies, end_jiffies)) {
140 			DRM_ERROR("SVGA device lockup.\n");
141 			break;
142 		}
143 		if (lazy)
144 			schedule_timeout(1);
145 		else if ((++count & 0x0F) == 0) {
146 			/**
147 			 * FIXME: Use schedule_hr_timeout here for
148 			 * newer kernels and lower CPU utilization.
149 			 */
150 
151 			__set_current_state(TASK_RUNNING);
152 			schedule();
153 			__set_current_state((interruptible) ?
154 					    TASK_INTERRUPTIBLE :
155 					    TASK_UNINTERRUPTIBLE);
156 		}
157 		if (interruptible && signal_pending(current)) {
158 			ret = -ERESTARTSYS;
159 			break;
160 		}
161 	}
162 	finish_wait(&dev_priv->fence_queue, &__wait);
163 	if (ret == 0 && fifo_idle) {
164 		__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
165 		iowrite32(signal_seq, fifo_mem + SVGA_FIFO_FENCE);
166 	}
167 	wake_up_all(&dev_priv->fence_queue);
168 	if (fifo_idle)
169 		up_read(&fifo_state->rwsem);
170 
171 	return ret;
172 }
173 
174 int vmw_wait_fence(struct vmw_private *dev_priv,
175 		   bool lazy, uint32_t sequence,
176 		   bool interruptible, unsigned long timeout)
177 {
178 	long ret;
179 	unsigned long irq_flags;
180 	struct vmw_fifo_state *fifo = &dev_priv->fifo;
181 
182 	if (likely(dev_priv->last_read_sequence - sequence < VMW_FENCE_WRAP))
183 		return 0;
184 
185 	if (likely(vmw_fence_signaled(dev_priv, sequence)))
186 		return 0;
187 
188 	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
189 
190 	if (!(fifo->capabilities & SVGA_FIFO_CAP_FENCE))
191 		return vmw_fallback_wait(dev_priv, lazy, true, sequence,
192 					 interruptible, timeout);
193 
194 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
195 		return vmw_fallback_wait(dev_priv, lazy, false, sequence,
196 					 interruptible, timeout);
197 
198 	mutex_lock(&dev_priv->hw_mutex);
199 	if (atomic_add_return(1, &dev_priv->fence_queue_waiters) > 0) {
200 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
201 		outl(SVGA_IRQFLAG_ANY_FENCE,
202 		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
203 		vmw_write(dev_priv, SVGA_REG_IRQMASK,
204 			  vmw_read(dev_priv, SVGA_REG_IRQMASK) |
205 			  SVGA_IRQFLAG_ANY_FENCE);
206 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
207 	}
208 	mutex_unlock(&dev_priv->hw_mutex);
209 
210 	if (interruptible)
211 		ret = wait_event_interruptible_timeout
212 		    (dev_priv->fence_queue,
213 		     vmw_fence_signaled(dev_priv, sequence),
214 		     timeout);
215 	else
216 		ret = wait_event_timeout
217 		    (dev_priv->fence_queue,
218 		     vmw_fence_signaled(dev_priv, sequence),
219 		     timeout);
220 
221 	if (unlikely(ret == 0))
222 		ret = -EBUSY;
223 	else if (likely(ret > 0))
224 		ret = 0;
225 
226 	mutex_lock(&dev_priv->hw_mutex);
227 	if (atomic_dec_and_test(&dev_priv->fence_queue_waiters)) {
228 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
229 		vmw_write(dev_priv, SVGA_REG_IRQMASK,
230 			  vmw_read(dev_priv, SVGA_REG_IRQMASK) &
231 			  ~SVGA_IRQFLAG_ANY_FENCE);
232 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
233 	}
234 	mutex_unlock(&dev_priv->hw_mutex);
235 
236 	return ret;
237 }
238 
239 void vmw_irq_preinstall(struct drm_device *dev)
240 {
241 	struct vmw_private *dev_priv = vmw_priv(dev);
242 	uint32_t status;
243 
244 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
245 		return;
246 
247 	spin_lock_init(&dev_priv->irq_lock);
248 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
249 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
250 }
251 
252 int vmw_irq_postinstall(struct drm_device *dev)
253 {
254 	return 0;
255 }
256 
257 void vmw_irq_uninstall(struct drm_device *dev)
258 {
259 	struct vmw_private *dev_priv = vmw_priv(dev);
260 	uint32_t status;
261 
262 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
263 		return;
264 
265 	mutex_lock(&dev_priv->hw_mutex);
266 	vmw_write(dev_priv, SVGA_REG_IRQMASK, 0);
267 	mutex_unlock(&dev_priv->hw_mutex);
268 
269 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
270 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
271 }
272 
273 #define VMW_FENCE_WAIT_TIMEOUT 3*HZ;
274 
275 int vmw_fence_wait_ioctl(struct drm_device *dev, void *data,
276 			 struct drm_file *file_priv)
277 {
278 	struct drm_vmw_fence_wait_arg *arg =
279 	    (struct drm_vmw_fence_wait_arg *)data;
280 	unsigned long timeout;
281 
282 	if (!arg->cookie_valid) {
283 		arg->cookie_valid = 1;
284 		arg->kernel_cookie = jiffies + VMW_FENCE_WAIT_TIMEOUT;
285 	}
286 
287 	timeout = jiffies;
288 	if (time_after_eq(timeout, (unsigned long)arg->kernel_cookie))
289 		return -EBUSY;
290 
291 	timeout = (unsigned long)arg->kernel_cookie - timeout;
292 	return vmw_wait_fence(vmw_priv(dev), true, arg->sequence, true, timeout);
293 }
294