1 /* via_dmablit.c -- PCI DMA BitBlt support for the VIA Unichrome/Pro
2  *
3  * Copyright (C) 2005 Thomas Hellstrom, All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sub license,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the
13  * next paragraph) shall be included in all copies or substantial portions
14  * of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors:
25  *    Thomas Hellstrom.
26  *    Partially based on code obtained from Digeo Inc.
27  */
28 
29 
30 /*
31  * Unmaps the DMA mappings.
32  * FIXME: Is this a NoOp on x86? Also
33  * FIXME: What happens if this one is called and a pending blit has previously done
34  * the same DMA mappings?
35  */
36 
37 #include <linux/pagemap.h>
38 #include <linux/pci.h>
39 #include <linux/slab.h>
40 #include <linux/vmalloc.h>
41 
42 #include <drm/drm_device.h>
43 #include <drm/via_drm.h>
44 
45 #include "via_dmablit.h"
46 #include "via_drv.h"
47 
48 #define VIA_PGDN(x)	     (((unsigned long)(x)) & PAGE_MASK)
49 #define VIA_PGOFF(x)	    (((unsigned long)(x)) & ~PAGE_MASK)
50 #define VIA_PFN(x)	      ((unsigned long)(x) >> PAGE_SHIFT)
51 
52 typedef struct _drm_via_descriptor {
53 	uint32_t mem_addr;
54 	uint32_t dev_addr;
55 	uint32_t size;
56 	uint32_t next;
57 } drm_via_descriptor_t;
58 
59 
60 /*
61  * Unmap a DMA mapping.
62  */
63 
64 
65 
66 static void
via_unmap_blit_from_device(struct pci_dev * pdev,drm_via_sg_info_t * vsg)67 via_unmap_blit_from_device(struct pci_dev *pdev, drm_via_sg_info_t *vsg)
68 {
69 	int num_desc = vsg->num_desc;
70 	unsigned cur_descriptor_page = num_desc / vsg->descriptors_per_page;
71 	unsigned descriptor_this_page = num_desc % vsg->descriptors_per_page;
72 	drm_via_descriptor_t *desc_ptr = vsg->desc_pages[cur_descriptor_page] +
73 		descriptor_this_page;
74 	dma_addr_t next = vsg->chain_start;
75 
76 	while (num_desc--) {
77 		if (descriptor_this_page-- == 0) {
78 			cur_descriptor_page--;
79 			descriptor_this_page = vsg->descriptors_per_page - 1;
80 			desc_ptr = vsg->desc_pages[cur_descriptor_page] +
81 				descriptor_this_page;
82 		}
83 		dma_unmap_single(&pdev->dev, next, sizeof(*desc_ptr), DMA_TO_DEVICE);
84 		dma_unmap_page(&pdev->dev, desc_ptr->mem_addr, desc_ptr->size, vsg->direction);
85 		next = (dma_addr_t) desc_ptr->next;
86 		desc_ptr--;
87 	}
88 }
89 
90 /*
91  * If mode = 0, count how many descriptors are needed.
92  * If mode = 1, Map the DMA pages for the device, put together and map also the descriptors.
93  * Descriptors are run in reverse order by the hardware because we are not allowed to update the
94  * 'next' field without syncing calls when the descriptor is already mapped.
95  */
96 
97 static void
via_map_blit_for_device(struct pci_dev * pdev,const drm_via_dmablit_t * xfer,drm_via_sg_info_t * vsg,int mode)98 via_map_blit_for_device(struct pci_dev *pdev,
99 		   const drm_via_dmablit_t *xfer,
100 		   drm_via_sg_info_t *vsg,
101 		   int mode)
102 {
103 	unsigned cur_descriptor_page = 0;
104 	unsigned num_descriptors_this_page = 0;
105 	unsigned char *mem_addr = xfer->mem_addr;
106 	unsigned char *cur_mem;
107 	unsigned char *first_addr = (unsigned char *)VIA_PGDN(mem_addr);
108 	uint32_t fb_addr = xfer->fb_addr;
109 	uint32_t cur_fb;
110 	unsigned long line_len;
111 	unsigned remaining_len;
112 	int num_desc = 0;
113 	int cur_line;
114 	dma_addr_t next = 0 | VIA_DMA_DPR_EC;
115 	drm_via_descriptor_t *desc_ptr = NULL;
116 
117 	if (mode == 1)
118 		desc_ptr = vsg->desc_pages[cur_descriptor_page];
119 
120 	for (cur_line = 0; cur_line < xfer->num_lines; ++cur_line) {
121 
122 		line_len = xfer->line_length;
123 		cur_fb = fb_addr;
124 		cur_mem = mem_addr;
125 
126 		while (line_len > 0) {
127 
128 			remaining_len = min(PAGE_SIZE-VIA_PGOFF(cur_mem), line_len);
129 			line_len -= remaining_len;
130 
131 			if (mode == 1) {
132 				desc_ptr->mem_addr =
133 					dma_map_page(&pdev->dev,
134 						     vsg->pages[VIA_PFN(cur_mem) -
135 								VIA_PFN(first_addr)],
136 						     VIA_PGOFF(cur_mem), remaining_len,
137 						     vsg->direction);
138 				desc_ptr->dev_addr = cur_fb;
139 
140 				desc_ptr->size = remaining_len;
141 				desc_ptr->next = (uint32_t) next;
142 				next = dma_map_single(&pdev->dev, desc_ptr, sizeof(*desc_ptr),
143 						      DMA_TO_DEVICE);
144 				desc_ptr++;
145 				if (++num_descriptors_this_page >= vsg->descriptors_per_page) {
146 					num_descriptors_this_page = 0;
147 					desc_ptr = vsg->desc_pages[++cur_descriptor_page];
148 				}
149 			}
150 
151 			num_desc++;
152 			cur_mem += remaining_len;
153 			cur_fb += remaining_len;
154 		}
155 
156 		mem_addr += xfer->mem_stride;
157 		fb_addr += xfer->fb_stride;
158 	}
159 
160 	if (mode == 1) {
161 		vsg->chain_start = next;
162 		vsg->state = dr_via_device_mapped;
163 	}
164 	vsg->num_desc = num_desc;
165 }
166 
167 /*
168  * Function that frees up all resources for a blit. It is usable even if the
169  * blit info has only been partially built as long as the status enum is consistent
170  * with the actual status of the used resources.
171  */
172 
173 
174 static void
via_free_sg_info(struct pci_dev * pdev,drm_via_sg_info_t * vsg)175 via_free_sg_info(struct pci_dev *pdev, drm_via_sg_info_t *vsg)
176 {
177 	int i;
178 
179 	switch (vsg->state) {
180 	case dr_via_device_mapped:
181 		via_unmap_blit_from_device(pdev, vsg);
182 		fallthrough;
183 	case dr_via_desc_pages_alloc:
184 		for (i = 0; i < vsg->num_desc_pages; ++i) {
185 			if (vsg->desc_pages[i] != NULL)
186 				free_page((unsigned long)vsg->desc_pages[i]);
187 		}
188 		kfree(vsg->desc_pages);
189 		fallthrough;
190 	case dr_via_pages_locked:
191 		unpin_user_pages_dirty_lock(vsg->pages, vsg->num_pages,
192 					   (vsg->direction == DMA_FROM_DEVICE));
193 		fallthrough;
194 	case dr_via_pages_alloc:
195 		vfree(vsg->pages);
196 		fallthrough;
197 	default:
198 		vsg->state = dr_via_sg_init;
199 	}
200 	vfree(vsg->bounce_buffer);
201 	vsg->bounce_buffer = NULL;
202 	vsg->free_on_sequence = 0;
203 }
204 
205 /*
206  * Fire a blit engine.
207  */
208 
209 static void
via_fire_dmablit(struct drm_device * dev,drm_via_sg_info_t * vsg,int engine)210 via_fire_dmablit(struct drm_device *dev, drm_via_sg_info_t *vsg, int engine)
211 {
212 	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
213 
214 	via_write(dev_priv, VIA_PCI_DMA_MAR0 + engine*0x10, 0);
215 	via_write(dev_priv, VIA_PCI_DMA_DAR0 + engine*0x10, 0);
216 	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DD | VIA_DMA_CSR_TD |
217 		  VIA_DMA_CSR_DE);
218 	via_write(dev_priv, VIA_PCI_DMA_MR0  + engine*0x04, VIA_DMA_MR_CM | VIA_DMA_MR_TDIE);
219 	via_write(dev_priv, VIA_PCI_DMA_BCR0 + engine*0x10, 0);
220 	via_write(dev_priv, VIA_PCI_DMA_DPR0 + engine*0x10, vsg->chain_start);
221 	wmb();
222 	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DE | VIA_DMA_CSR_TS);
223 	via_read(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04);
224 }
225 
226 /*
227  * Obtain a page pointer array and lock all pages into system memory. A segmentation violation will
228  * occur here if the calling user does not have access to the submitted address.
229  */
230 
231 static int
via_lock_all_dma_pages(drm_via_sg_info_t * vsg,drm_via_dmablit_t * xfer)232 via_lock_all_dma_pages(drm_via_sg_info_t *vsg,  drm_via_dmablit_t *xfer)
233 {
234 	int ret;
235 	unsigned long first_pfn = VIA_PFN(xfer->mem_addr);
236 	vsg->num_pages = VIA_PFN(xfer->mem_addr + (xfer->num_lines * xfer->mem_stride - 1)) -
237 		first_pfn + 1;
238 
239 	vsg->pages = vzalloc(array_size(sizeof(struct page *), vsg->num_pages));
240 	if (NULL == vsg->pages)
241 		return -ENOMEM;
242 	ret = pin_user_pages_fast((unsigned long)xfer->mem_addr,
243 			vsg->num_pages,
244 			vsg->direction == DMA_FROM_DEVICE ? FOLL_WRITE : 0,
245 			vsg->pages);
246 	if (ret != vsg->num_pages) {
247 		if (ret < 0)
248 			return ret;
249 		vsg->state = dr_via_pages_locked;
250 		return -EINVAL;
251 	}
252 	vsg->state = dr_via_pages_locked;
253 	DRM_DEBUG("DMA pages locked\n");
254 	return 0;
255 }
256 
257 /*
258  * Allocate DMA capable memory for the blit descriptor chain, and an array that keeps track of the
259  * pages we allocate. We don't want to use kmalloc for the descriptor chain because it may be
260  * quite large for some blits, and pages don't need to be contiguous.
261  */
262 
263 static int
via_alloc_desc_pages(drm_via_sg_info_t * vsg)264 via_alloc_desc_pages(drm_via_sg_info_t *vsg)
265 {
266 	int i;
267 
268 	vsg->descriptors_per_page = PAGE_SIZE / sizeof(drm_via_descriptor_t);
269 	vsg->num_desc_pages = (vsg->num_desc + vsg->descriptors_per_page - 1) /
270 		vsg->descriptors_per_page;
271 
272 	if (NULL ==  (vsg->desc_pages = kcalloc(vsg->num_desc_pages, sizeof(void *), GFP_KERNEL)))
273 		return -ENOMEM;
274 
275 	vsg->state = dr_via_desc_pages_alloc;
276 	for (i = 0; i < vsg->num_desc_pages; ++i) {
277 		if (NULL == (vsg->desc_pages[i] =
278 			     (drm_via_descriptor_t *) __get_free_page(GFP_KERNEL)))
279 			return -ENOMEM;
280 	}
281 	DRM_DEBUG("Allocated %d pages for %d descriptors.\n", vsg->num_desc_pages,
282 		  vsg->num_desc);
283 	return 0;
284 }
285 
286 static void
via_abort_dmablit(struct drm_device * dev,int engine)287 via_abort_dmablit(struct drm_device *dev, int engine)
288 {
289 	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
290 
291 	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TA);
292 }
293 
294 static void
via_dmablit_engine_off(struct drm_device * dev,int engine)295 via_dmablit_engine_off(struct drm_device *dev, int engine)
296 {
297 	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
298 
299 	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TD | VIA_DMA_CSR_DD);
300 }
301 
302 
303 
304 /*
305  * The dmablit part of the IRQ handler. Trying to do only reasonably fast things here.
306  * The rest, like unmapping and freeing memory for done blits is done in a separate workqueue
307  * task. Basically the task of the interrupt handler is to submit a new blit to the engine, while
308  * the workqueue task takes care of processing associated with the old blit.
309  */
310 
311 void
via_dmablit_handler(struct drm_device * dev,int engine,int from_irq)312 via_dmablit_handler(struct drm_device *dev, int engine, int from_irq)
313 {
314 	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
315 	drm_via_blitq_t *blitq = dev_priv->blit_queues + engine;
316 	int cur;
317 	int done_transfer;
318 	unsigned long irqsave = 0;
319 	uint32_t status = 0;
320 
321 	DRM_DEBUG("DMA blit handler called. engine = %d, from_irq = %d, blitq = 0x%lx\n",
322 		  engine, from_irq, (unsigned long) blitq);
323 
324 	if (from_irq)
325 		spin_lock(&blitq->blit_lock);
326 	else
327 		spin_lock_irqsave(&blitq->blit_lock, irqsave);
328 
329 	done_transfer = blitq->is_active &&
330 	  ((status = via_read(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04)) & VIA_DMA_CSR_TD);
331 	done_transfer = done_transfer || (blitq->aborting && !(status & VIA_DMA_CSR_DE));
332 
333 	cur = blitq->cur;
334 	if (done_transfer) {
335 
336 		blitq->blits[cur]->aborted = blitq->aborting;
337 		blitq->done_blit_handle++;
338 		wake_up(blitq->blit_queue + cur);
339 
340 		cur++;
341 		if (cur >= VIA_NUM_BLIT_SLOTS)
342 			cur = 0;
343 		blitq->cur = cur;
344 
345 		/*
346 		 * Clear transfer done flag.
347 		 */
348 
349 		via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04,  VIA_DMA_CSR_TD);
350 
351 		blitq->is_active = 0;
352 		blitq->aborting = 0;
353 		schedule_work(&blitq->wq);
354 
355 	} else if (blitq->is_active && time_after_eq(jiffies, blitq->end)) {
356 
357 		/*
358 		 * Abort transfer after one second.
359 		 */
360 
361 		via_abort_dmablit(dev, engine);
362 		blitq->aborting = 1;
363 		blitq->end = jiffies + HZ;
364 	}
365 
366 	if (!blitq->is_active) {
367 		if (blitq->num_outstanding) {
368 			via_fire_dmablit(dev, blitq->blits[cur], engine);
369 			blitq->is_active = 1;
370 			blitq->cur = cur;
371 			blitq->num_outstanding--;
372 			blitq->end = jiffies + HZ;
373 			if (!timer_pending(&blitq->poll_timer))
374 				mod_timer(&blitq->poll_timer, jiffies + 1);
375 		} else {
376 			if (timer_pending(&blitq->poll_timer))
377 				del_timer(&blitq->poll_timer);
378 			via_dmablit_engine_off(dev, engine);
379 		}
380 	}
381 
382 	if (from_irq)
383 		spin_unlock(&blitq->blit_lock);
384 	else
385 		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
386 }
387 
388 
389 
390 /*
391  * Check whether this blit is still active, performing necessary locking.
392  */
393 
394 static int
via_dmablit_active(drm_via_blitq_t * blitq,int engine,uint32_t handle,wait_queue_head_t ** queue)395 via_dmablit_active(drm_via_blitq_t *blitq, int engine, uint32_t handle, wait_queue_head_t **queue)
396 {
397 	unsigned long irqsave;
398 	uint32_t slot;
399 	int active;
400 
401 	spin_lock_irqsave(&blitq->blit_lock, irqsave);
402 
403 	/*
404 	 * Allow for handle wraparounds.
405 	 */
406 
407 	active = ((blitq->done_blit_handle - handle) > (1 << 23)) &&
408 		((blitq->cur_blit_handle - handle) <= (1 << 23));
409 
410 	if (queue && active) {
411 		slot = handle - blitq->done_blit_handle + blitq->cur - 1;
412 		if (slot >= VIA_NUM_BLIT_SLOTS)
413 			slot -= VIA_NUM_BLIT_SLOTS;
414 		*queue = blitq->blit_queue + slot;
415 	}
416 
417 	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
418 
419 	return active;
420 }
421 
422 /*
423  * Sync. Wait for at least three seconds for the blit to be performed.
424  */
425 
426 static int
via_dmablit_sync(struct drm_device * dev,uint32_t handle,int engine)427 via_dmablit_sync(struct drm_device *dev, uint32_t handle, int engine)
428 {
429 
430 	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
431 	drm_via_blitq_t *blitq = dev_priv->blit_queues + engine;
432 	wait_queue_head_t *queue;
433 	int ret = 0;
434 
435 	if (via_dmablit_active(blitq, engine, handle, &queue)) {
436 		VIA_WAIT_ON(ret, *queue, 3 * HZ,
437 			    !via_dmablit_active(blitq, engine, handle, NULL));
438 	}
439 	DRM_DEBUG("DMA blit sync handle 0x%x engine %d returned %d\n",
440 		  handle, engine, ret);
441 
442 	return ret;
443 }
444 
445 
446 /*
447  * A timer that regularly polls the blit engine in cases where we don't have interrupts:
448  * a) Broken hardware (typically those that don't have any video capture facility).
449  * b) Blit abort. The hardware doesn't send an interrupt when a blit is aborted.
450  * The timer and hardware IRQ's can and do work in parallel. If the hardware has
451  * irqs, it will shorten the latency somewhat.
452  */
453 
454 
455 
456 static void
via_dmablit_timer(struct timer_list * t)457 via_dmablit_timer(struct timer_list *t)
458 {
459 	drm_via_blitq_t *blitq = from_timer(blitq, t, poll_timer);
460 	struct drm_device *dev = blitq->dev;
461 	int engine = (int)
462 		(blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues);
463 
464 	DRM_DEBUG("Polling timer called for engine %d, jiffies %lu\n", engine,
465 		  (unsigned long) jiffies);
466 
467 	via_dmablit_handler(dev, engine, 0);
468 
469 	if (!timer_pending(&blitq->poll_timer)) {
470 		mod_timer(&blitq->poll_timer, jiffies + 1);
471 
472 	       /*
473 		* Rerun handler to delete timer if engines are off, and
474 		* to shorten abort latency. This is a little nasty.
475 		*/
476 
477 	       via_dmablit_handler(dev, engine, 0);
478 
479 	}
480 }
481 
482 
483 
484 
485 /*
486  * Workqueue task that frees data and mappings associated with a blit.
487  * Also wakes up waiting processes. Each of these tasks handles one
488  * blit engine only and may not be called on each interrupt.
489  */
490 
491 
492 static void
via_dmablit_workqueue(struct work_struct * work)493 via_dmablit_workqueue(struct work_struct *work)
494 {
495 	drm_via_blitq_t *blitq = container_of(work, drm_via_blitq_t, wq);
496 	struct drm_device *dev = blitq->dev;
497 	struct pci_dev *pdev = to_pci_dev(dev->dev);
498 	unsigned long irqsave;
499 	drm_via_sg_info_t *cur_sg;
500 	int cur_released;
501 
502 
503 	DRM_DEBUG("Workqueue task called for blit engine %ld\n", (unsigned long)
504 		  (blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues));
505 
506 	spin_lock_irqsave(&blitq->blit_lock, irqsave);
507 
508 	while (blitq->serviced != blitq->cur) {
509 
510 		cur_released = blitq->serviced++;
511 
512 		DRM_DEBUG("Releasing blit slot %d\n", cur_released);
513 
514 		if (blitq->serviced >= VIA_NUM_BLIT_SLOTS)
515 			blitq->serviced = 0;
516 
517 		cur_sg = blitq->blits[cur_released];
518 		blitq->num_free++;
519 
520 		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
521 
522 		wake_up(&blitq->busy_queue);
523 
524 		via_free_sg_info(pdev, cur_sg);
525 		kfree(cur_sg);
526 
527 		spin_lock_irqsave(&blitq->blit_lock, irqsave);
528 	}
529 
530 	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
531 }
532 
533 
534 /*
535  * Init all blit engines. Currently we use two, but some hardware have 4.
536  */
537 
538 
539 void
via_init_dmablit(struct drm_device * dev)540 via_init_dmablit(struct drm_device *dev)
541 {
542 	int i, j;
543 	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
544 	struct pci_dev *pdev = to_pci_dev(dev->dev);
545 	drm_via_blitq_t *blitq;
546 
547 	pci_set_master(pdev);
548 
549 	for (i = 0; i < VIA_NUM_BLIT_ENGINES; ++i) {
550 		blitq = dev_priv->blit_queues + i;
551 		blitq->dev = dev;
552 		blitq->cur_blit_handle = 0;
553 		blitq->done_blit_handle = 0;
554 		blitq->head = 0;
555 		blitq->cur = 0;
556 		blitq->serviced = 0;
557 		blitq->num_free = VIA_NUM_BLIT_SLOTS - 1;
558 		blitq->num_outstanding = 0;
559 		blitq->is_active = 0;
560 		blitq->aborting = 0;
561 		spin_lock_init(&blitq->blit_lock);
562 		for (j = 0; j < VIA_NUM_BLIT_SLOTS; ++j)
563 			init_waitqueue_head(blitq->blit_queue + j);
564 		init_waitqueue_head(&blitq->busy_queue);
565 		INIT_WORK(&blitq->wq, via_dmablit_workqueue);
566 		timer_setup(&blitq->poll_timer, via_dmablit_timer, 0);
567 	}
568 }
569 
570 /*
571  * Build all info and do all mappings required for a blit.
572  */
573 
574 
575 static int
via_build_sg_info(struct drm_device * dev,drm_via_sg_info_t * vsg,drm_via_dmablit_t * xfer)576 via_build_sg_info(struct drm_device *dev, drm_via_sg_info_t *vsg, drm_via_dmablit_t *xfer)
577 {
578 	struct pci_dev *pdev = to_pci_dev(dev->dev);
579 	int draw = xfer->to_fb;
580 	int ret = 0;
581 
582 	vsg->direction = (draw) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
583 	vsg->bounce_buffer = NULL;
584 
585 	vsg->state = dr_via_sg_init;
586 
587 	if (xfer->num_lines <= 0 || xfer->line_length <= 0) {
588 		DRM_ERROR("Zero size bitblt.\n");
589 		return -EINVAL;
590 	}
591 
592 	/*
593 	 * Below check is a driver limitation, not a hardware one. We
594 	 * don't want to lock unused pages, and don't want to incoporate the
595 	 * extra logic of avoiding them. Make sure there are no.
596 	 * (Not a big limitation anyway.)
597 	 */
598 
599 	if ((xfer->mem_stride - xfer->line_length) > 2*PAGE_SIZE) {
600 		DRM_ERROR("Too large system memory stride. Stride: %d, "
601 			  "Length: %d\n", xfer->mem_stride, xfer->line_length);
602 		return -EINVAL;
603 	}
604 
605 	if ((xfer->mem_stride == xfer->line_length) &&
606 	   (xfer->fb_stride == xfer->line_length)) {
607 		xfer->mem_stride *= xfer->num_lines;
608 		xfer->line_length = xfer->mem_stride;
609 		xfer->fb_stride = xfer->mem_stride;
610 		xfer->num_lines = 1;
611 	}
612 
613 	/*
614 	 * Don't lock an arbitrary large number of pages, since that causes a
615 	 * DOS security hole.
616 	 */
617 
618 	if (xfer->num_lines > 2048 || (xfer->num_lines*xfer->mem_stride > (2048*2048*4))) {
619 		DRM_ERROR("Too large PCI DMA bitblt.\n");
620 		return -EINVAL;
621 	}
622 
623 	/*
624 	 * we allow a negative fb stride to allow flipping of images in
625 	 * transfer.
626 	 */
627 
628 	if (xfer->mem_stride < xfer->line_length ||
629 		abs(xfer->fb_stride) < xfer->line_length) {
630 		DRM_ERROR("Invalid frame-buffer / memory stride.\n");
631 		return -EINVAL;
632 	}
633 
634 	/*
635 	 * A hardware bug seems to be worked around if system memory addresses start on
636 	 * 16 byte boundaries. This seems a bit restrictive however. VIA is contacted
637 	 * about this. Meanwhile, impose the following restrictions:
638 	 */
639 
640 #ifdef VIA_BUGFREE
641 	if ((((unsigned long)xfer->mem_addr & 3) != ((unsigned long)xfer->fb_addr & 3)) ||
642 	    ((xfer->num_lines > 1) && ((xfer->mem_stride & 3) != (xfer->fb_stride & 3)))) {
643 		DRM_ERROR("Invalid DRM bitblt alignment.\n");
644 		return -EINVAL;
645 	}
646 #else
647 	if ((((unsigned long)xfer->mem_addr & 15) ||
648 	      ((unsigned long)xfer->fb_addr & 3)) ||
649 	   ((xfer->num_lines > 1) &&
650 	   ((xfer->mem_stride & 15) || (xfer->fb_stride & 3)))) {
651 		DRM_ERROR("Invalid DRM bitblt alignment.\n");
652 		return -EINVAL;
653 	}
654 #endif
655 
656 	if (0 != (ret = via_lock_all_dma_pages(vsg, xfer))) {
657 		DRM_ERROR("Could not lock DMA pages.\n");
658 		via_free_sg_info(pdev, vsg);
659 		return ret;
660 	}
661 
662 	via_map_blit_for_device(pdev, xfer, vsg, 0);
663 	if (0 != (ret = via_alloc_desc_pages(vsg))) {
664 		DRM_ERROR("Could not allocate DMA descriptor pages.\n");
665 		via_free_sg_info(pdev, vsg);
666 		return ret;
667 	}
668 	via_map_blit_for_device(pdev, xfer, vsg, 1);
669 
670 	return 0;
671 }
672 
673 
674 /*
675  * Reserve one free slot in the blit queue. Will wait for one second for one
676  * to become available. Otherwise -EBUSY is returned.
677  */
678 
679 static int
via_dmablit_grab_slot(drm_via_blitq_t * blitq,int engine)680 via_dmablit_grab_slot(drm_via_blitq_t *blitq, int engine)
681 {
682 	int ret = 0;
683 	unsigned long irqsave;
684 
685 	DRM_DEBUG("Num free is %d\n", blitq->num_free);
686 	spin_lock_irqsave(&blitq->blit_lock, irqsave);
687 	while (blitq->num_free == 0) {
688 		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
689 
690 		VIA_WAIT_ON(ret, blitq->busy_queue, HZ, blitq->num_free > 0);
691 		if (ret)
692 			return (-EINTR == ret) ? -EAGAIN : ret;
693 
694 		spin_lock_irqsave(&blitq->blit_lock, irqsave);
695 	}
696 
697 	blitq->num_free--;
698 	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
699 
700 	return 0;
701 }
702 
703 /*
704  * Hand back a free slot if we changed our mind.
705  */
706 
707 static void
via_dmablit_release_slot(drm_via_blitq_t * blitq)708 via_dmablit_release_slot(drm_via_blitq_t *blitq)
709 {
710 	unsigned long irqsave;
711 
712 	spin_lock_irqsave(&blitq->blit_lock, irqsave);
713 	blitq->num_free++;
714 	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
715 	wake_up(&blitq->busy_queue);
716 }
717 
718 /*
719  * Grab a free slot. Build blit info and queue a blit.
720  */
721 
722 
723 static int
via_dmablit(struct drm_device * dev,drm_via_dmablit_t * xfer)724 via_dmablit(struct drm_device *dev, drm_via_dmablit_t *xfer)
725 {
726 	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
727 	drm_via_sg_info_t *vsg;
728 	drm_via_blitq_t *blitq;
729 	int ret;
730 	int engine;
731 	unsigned long irqsave;
732 
733 	if (dev_priv == NULL) {
734 		DRM_ERROR("Called without initialization.\n");
735 		return -EINVAL;
736 	}
737 
738 	engine = (xfer->to_fb) ? 0 : 1;
739 	blitq = dev_priv->blit_queues + engine;
740 	if (0 != (ret = via_dmablit_grab_slot(blitq, engine)))
741 		return ret;
742 	if (NULL == (vsg = kmalloc(sizeof(*vsg), GFP_KERNEL))) {
743 		via_dmablit_release_slot(blitq);
744 		return -ENOMEM;
745 	}
746 	if (0 != (ret = via_build_sg_info(dev, vsg, xfer))) {
747 		via_dmablit_release_slot(blitq);
748 		kfree(vsg);
749 		return ret;
750 	}
751 	spin_lock_irqsave(&blitq->blit_lock, irqsave);
752 
753 	blitq->blits[blitq->head++] = vsg;
754 	if (blitq->head >= VIA_NUM_BLIT_SLOTS)
755 		blitq->head = 0;
756 	blitq->num_outstanding++;
757 	xfer->sync.sync_handle = ++blitq->cur_blit_handle;
758 
759 	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
760 	xfer->sync.engine = engine;
761 
762 	via_dmablit_handler(dev, engine, 0);
763 
764 	return 0;
765 }
766 
767 /*
768  * Sync on a previously submitted blit. Note that the X server use signals extensively, and
769  * that there is a very big probability that this IOCTL will be interrupted by a signal. In that
770  * case it returns with -EAGAIN for the signal to be delivered.
771  * The caller should then reissue the IOCTL. This is similar to what is being done for drmGetLock().
772  */
773 
774 int
via_dma_blit_sync(struct drm_device * dev,void * data,struct drm_file * file_priv)775 via_dma_blit_sync(struct drm_device *dev, void *data, struct drm_file *file_priv)
776 {
777 	drm_via_blitsync_t *sync = data;
778 	int err;
779 
780 	if (sync->engine >= VIA_NUM_BLIT_ENGINES)
781 		return -EINVAL;
782 
783 	err = via_dmablit_sync(dev, sync->sync_handle, sync->engine);
784 
785 	if (-EINTR == err)
786 		err = -EAGAIN;
787 
788 	return err;
789 }
790 
791 
792 /*
793  * Queue a blit and hand back a handle to be used for sync. This IOCTL may be interrupted by a signal
794  * while waiting for a free slot in the blit queue. In that case it returns with -EAGAIN and should
795  * be reissued. See the above IOCTL code.
796  */
797 
798 int
via_dma_blit(struct drm_device * dev,void * data,struct drm_file * file_priv)799 via_dma_blit(struct drm_device *dev, void *data, struct drm_file *file_priv)
800 {
801 	drm_via_dmablit_t *xfer = data;
802 	int err;
803 
804 	err = via_dmablit(dev, xfer);
805 
806 	return err;
807 }
808