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 
vmw_irq_handler(DRM_IRQ_ARGS)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 
vmw_fifo_idle(struct vmw_private * dev_priv,uint32_t sequence)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 
vmw_update_sequence(struct vmw_private * dev_priv,struct vmw_fifo_state * fifo_state)67 void vmw_update_sequence(struct vmw_private *dev_priv,
68 			 struct vmw_fifo_state *fifo_state)
69 {
70 	__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
71 
72 	uint32_t sequence = ioread32(fifo_mem + SVGA_FIFO_FENCE);
73 
74 	if (dev_priv->last_read_sequence != sequence) {
75 		dev_priv->last_read_sequence = sequence;
76 		vmw_fence_pull(&fifo_state->fence_queue, sequence);
77 	}
78 }
79 
vmw_fence_signaled(struct vmw_private * dev_priv,uint32_t sequence)80 bool vmw_fence_signaled(struct vmw_private *dev_priv,
81 			uint32_t sequence)
82 {
83 	struct vmw_fifo_state *fifo_state;
84 	bool ret;
85 
86 	if (likely(dev_priv->last_read_sequence - sequence < VMW_FENCE_WRAP))
87 		return true;
88 
89 	fifo_state = &dev_priv->fifo;
90 	vmw_update_sequence(dev_priv, fifo_state);
91 	if (likely(dev_priv->last_read_sequence - sequence < VMW_FENCE_WRAP))
92 		return true;
93 
94 	if (!(fifo_state->capabilities & SVGA_FIFO_CAP_FENCE) &&
95 	    vmw_fifo_idle(dev_priv, sequence))
96 		return true;
97 
98 	/**
99 	 * Then check if the sequence is higher than what we've actually
100 	 * emitted. Then the fence is stale and signaled.
101 	 */
102 
103 	ret = ((atomic_read(&dev_priv->fence_seq) - sequence)
104 	       > VMW_FENCE_WRAP);
105 
106 	return ret;
107 }
108 
vmw_fallback_wait(struct vmw_private * dev_priv,bool lazy,bool fifo_idle,uint32_t sequence,bool interruptible,unsigned long timeout)109 int vmw_fallback_wait(struct vmw_private *dev_priv,
110 		      bool lazy,
111 		      bool fifo_idle,
112 		      uint32_t sequence,
113 		      bool interruptible,
114 		      unsigned long timeout)
115 {
116 	struct vmw_fifo_state *fifo_state = &dev_priv->fifo;
117 
118 	uint32_t count = 0;
119 	uint32_t signal_seq;
120 	int ret;
121 	unsigned long end_jiffies = jiffies + timeout;
122 	bool (*wait_condition)(struct vmw_private *, uint32_t);
123 	DEFINE_WAIT(__wait);
124 
125 	wait_condition = (fifo_idle) ? &vmw_fifo_idle :
126 		&vmw_fence_signaled;
127 
128 	/**
129 	 * Block command submission while waiting for idle.
130 	 */
131 
132 	if (fifo_idle)
133 		down_read(&fifo_state->rwsem);
134 	signal_seq = atomic_read(&dev_priv->fence_seq);
135 	ret = 0;
136 
137 	for (;;) {
138 		prepare_to_wait(&dev_priv->fence_queue, &__wait,
139 				(interruptible) ?
140 				TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE);
141 		if (wait_condition(dev_priv, sequence))
142 			break;
143 		if (time_after_eq(jiffies, end_jiffies)) {
144 			DRM_ERROR("SVGA device lockup.\n");
145 			break;
146 		}
147 		if (lazy)
148 			schedule_timeout(1);
149 		else if ((++count & 0x0F) == 0) {
150 			/**
151 			 * FIXME: Use schedule_hr_timeout here for
152 			 * newer kernels and lower CPU utilization.
153 			 */
154 
155 			__set_current_state(TASK_RUNNING);
156 			schedule();
157 			__set_current_state((interruptible) ?
158 					    TASK_INTERRUPTIBLE :
159 					    TASK_UNINTERRUPTIBLE);
160 		}
161 		if (interruptible && signal_pending(current)) {
162 			ret = -ERESTARTSYS;
163 			break;
164 		}
165 	}
166 	finish_wait(&dev_priv->fence_queue, &__wait);
167 	if (ret == 0 && fifo_idle) {
168 		__le32 __iomem *fifo_mem = dev_priv->mmio_virt;
169 		iowrite32(signal_seq, fifo_mem + SVGA_FIFO_FENCE);
170 	}
171 	wake_up_all(&dev_priv->fence_queue);
172 	if (fifo_idle)
173 		up_read(&fifo_state->rwsem);
174 
175 	return ret;
176 }
177 
vmw_wait_fence(struct vmw_private * dev_priv,bool lazy,uint32_t sequence,bool interruptible,unsigned long timeout)178 int vmw_wait_fence(struct vmw_private *dev_priv,
179 		   bool lazy, uint32_t sequence,
180 		   bool interruptible, unsigned long timeout)
181 {
182 	long ret;
183 	unsigned long irq_flags;
184 	struct vmw_fifo_state *fifo = &dev_priv->fifo;
185 
186 	if (likely(dev_priv->last_read_sequence - sequence < VMW_FENCE_WRAP))
187 		return 0;
188 
189 	if (likely(vmw_fence_signaled(dev_priv, sequence)))
190 		return 0;
191 
192 	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
193 
194 	if (!(fifo->capabilities & SVGA_FIFO_CAP_FENCE))
195 		return vmw_fallback_wait(dev_priv, lazy, true, sequence,
196 					 interruptible, timeout);
197 
198 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
199 		return vmw_fallback_wait(dev_priv, lazy, false, sequence,
200 					 interruptible, timeout);
201 
202 	mutex_lock(&dev_priv->hw_mutex);
203 	if (atomic_add_return(1, &dev_priv->fence_queue_waiters) > 0) {
204 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
205 		outl(SVGA_IRQFLAG_ANY_FENCE,
206 		     dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
207 		vmw_write(dev_priv, SVGA_REG_IRQMASK,
208 			  vmw_read(dev_priv, SVGA_REG_IRQMASK) |
209 			  SVGA_IRQFLAG_ANY_FENCE);
210 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
211 	}
212 	mutex_unlock(&dev_priv->hw_mutex);
213 
214 	if (interruptible)
215 		ret = wait_event_interruptible_timeout
216 		    (dev_priv->fence_queue,
217 		     vmw_fence_signaled(dev_priv, sequence),
218 		     timeout);
219 	else
220 		ret = wait_event_timeout
221 		    (dev_priv->fence_queue,
222 		     vmw_fence_signaled(dev_priv, sequence),
223 		     timeout);
224 
225 	if (unlikely(ret == 0))
226 		ret = -EBUSY;
227 	else if (likely(ret > 0))
228 		ret = 0;
229 
230 	mutex_lock(&dev_priv->hw_mutex);
231 	if (atomic_dec_and_test(&dev_priv->fence_queue_waiters)) {
232 		spin_lock_irqsave(&dev_priv->irq_lock, irq_flags);
233 		vmw_write(dev_priv, SVGA_REG_IRQMASK,
234 			  vmw_read(dev_priv, SVGA_REG_IRQMASK) &
235 			  ~SVGA_IRQFLAG_ANY_FENCE);
236 		spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags);
237 	}
238 	mutex_unlock(&dev_priv->hw_mutex);
239 
240 	return ret;
241 }
242 
vmw_irq_preinstall(struct drm_device * dev)243 void vmw_irq_preinstall(struct drm_device *dev)
244 {
245 	struct vmw_private *dev_priv = vmw_priv(dev);
246 	uint32_t status;
247 
248 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
249 		return;
250 
251 	spin_lock_init(&dev_priv->irq_lock);
252 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
253 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
254 }
255 
vmw_irq_postinstall(struct drm_device * dev)256 int vmw_irq_postinstall(struct drm_device *dev)
257 {
258 	return 0;
259 }
260 
vmw_irq_uninstall(struct drm_device * dev)261 void vmw_irq_uninstall(struct drm_device *dev)
262 {
263 	struct vmw_private *dev_priv = vmw_priv(dev);
264 	uint32_t status;
265 
266 	if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK))
267 		return;
268 
269 	mutex_lock(&dev_priv->hw_mutex);
270 	vmw_write(dev_priv, SVGA_REG_IRQMASK, 0);
271 	mutex_unlock(&dev_priv->hw_mutex);
272 
273 	status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
274 	outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT);
275 }
276 
277 #define VMW_FENCE_WAIT_TIMEOUT 3*HZ;
278 
vmw_fence_wait_ioctl(struct drm_device * dev,void * data,struct drm_file * file_priv)279 int vmw_fence_wait_ioctl(struct drm_device *dev, void *data,
280 			 struct drm_file *file_priv)
281 {
282 	struct drm_vmw_fence_wait_arg *arg =
283 	    (struct drm_vmw_fence_wait_arg *)data;
284 	unsigned long timeout;
285 
286 	if (!arg->cookie_valid) {
287 		arg->cookie_valid = 1;
288 		arg->kernel_cookie = jiffies + VMW_FENCE_WAIT_TIMEOUT;
289 	}
290 
291 	timeout = jiffies;
292 	if (time_after_eq(timeout, (unsigned long)arg->kernel_cookie))
293 		return -EBUSY;
294 
295 	timeout = (unsigned long)arg->kernel_cookie - timeout;
296 	return vmw_wait_fence(vmw_priv(dev), true, arg->sequence, true, timeout);
297 }
298