Linux Audio

Check our new training course

Loading...
v4.6
  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 <drm/drmP.h>
 38#include <drm/via_drm.h>
 39#include "via_drv.h"
 40#include "via_dmablit.h"
 41
 42#include <linux/pagemap.h>
 
 43#include <linux/slab.h>
 
 
 
 
 
 
 
 44
 45#define VIA_PGDN(x)	     (((unsigned long)(x)) & PAGE_MASK)
 46#define VIA_PGOFF(x)	    (((unsigned long)(x)) & ~PAGE_MASK)
 47#define VIA_PFN(x)	      ((unsigned long)(x) >> PAGE_SHIFT)
 48
 49typedef struct _drm_via_descriptor {
 50	uint32_t mem_addr;
 51	uint32_t dev_addr;
 52	uint32_t size;
 53	uint32_t next;
 54} drm_via_descriptor_t;
 55
 56
 57/*
 58 * Unmap a DMA mapping.
 59 */
 60
 61
 62
 63static void
 64via_unmap_blit_from_device(struct pci_dev *pdev, drm_via_sg_info_t *vsg)
 65{
 66	int num_desc = vsg->num_desc;
 67	unsigned cur_descriptor_page = num_desc / vsg->descriptors_per_page;
 68	unsigned descriptor_this_page = num_desc % vsg->descriptors_per_page;
 69	drm_via_descriptor_t *desc_ptr = vsg->desc_pages[cur_descriptor_page] +
 70		descriptor_this_page;
 71	dma_addr_t next = vsg->chain_start;
 72
 73	while (num_desc--) {
 74		if (descriptor_this_page-- == 0) {
 75			cur_descriptor_page--;
 76			descriptor_this_page = vsg->descriptors_per_page - 1;
 77			desc_ptr = vsg->desc_pages[cur_descriptor_page] +
 78				descriptor_this_page;
 79		}
 80		dma_unmap_single(&pdev->dev, next, sizeof(*desc_ptr), DMA_TO_DEVICE);
 81		dma_unmap_page(&pdev->dev, desc_ptr->mem_addr, desc_ptr->size, vsg->direction);
 82		next = (dma_addr_t) desc_ptr->next;
 83		desc_ptr--;
 84	}
 85}
 86
 87/*
 88 * If mode = 0, count how many descriptors are needed.
 89 * If mode = 1, Map the DMA pages for the device, put together and map also the descriptors.
 90 * Descriptors are run in reverse order by the hardware because we are not allowed to update the
 91 * 'next' field without syncing calls when the descriptor is already mapped.
 92 */
 93
 94static void
 95via_map_blit_for_device(struct pci_dev *pdev,
 96		   const drm_via_dmablit_t *xfer,
 97		   drm_via_sg_info_t *vsg,
 98		   int mode)
 99{
100	unsigned cur_descriptor_page = 0;
101	unsigned num_descriptors_this_page = 0;
102	unsigned char *mem_addr = xfer->mem_addr;
103	unsigned char *cur_mem;
104	unsigned char *first_addr = (unsigned char *)VIA_PGDN(mem_addr);
105	uint32_t fb_addr = xfer->fb_addr;
106	uint32_t cur_fb;
107	unsigned long line_len;
108	unsigned remaining_len;
109	int num_desc = 0;
110	int cur_line;
111	dma_addr_t next = 0 | VIA_DMA_DPR_EC;
112	drm_via_descriptor_t *desc_ptr = NULL;
113
114	if (mode == 1)
115		desc_ptr = vsg->desc_pages[cur_descriptor_page];
116
117	for (cur_line = 0; cur_line < xfer->num_lines; ++cur_line) {
118
119		line_len = xfer->line_length;
120		cur_fb = fb_addr;
121		cur_mem = mem_addr;
122
123		while (line_len > 0) {
124
125			remaining_len = min(PAGE_SIZE-VIA_PGOFF(cur_mem), line_len);
126			line_len -= remaining_len;
127
128			if (mode == 1) {
129				desc_ptr->mem_addr =
130					dma_map_page(&pdev->dev,
131						     vsg->pages[VIA_PFN(cur_mem) -
132								VIA_PFN(first_addr)],
133						     VIA_PGOFF(cur_mem), remaining_len,
134						     vsg->direction);
135				desc_ptr->dev_addr = cur_fb;
136
137				desc_ptr->size = remaining_len;
138				desc_ptr->next = (uint32_t) next;
139				next = dma_map_single(&pdev->dev, desc_ptr, sizeof(*desc_ptr),
140						      DMA_TO_DEVICE);
141				desc_ptr++;
142				if (++num_descriptors_this_page >= vsg->descriptors_per_page) {
143					num_descriptors_this_page = 0;
144					desc_ptr = vsg->desc_pages[++cur_descriptor_page];
145				}
146			}
147
148			num_desc++;
149			cur_mem += remaining_len;
150			cur_fb += remaining_len;
151		}
152
153		mem_addr += xfer->mem_stride;
154		fb_addr += xfer->fb_stride;
155	}
156
157	if (mode == 1) {
158		vsg->chain_start = next;
159		vsg->state = dr_via_device_mapped;
160	}
161	vsg->num_desc = num_desc;
162}
163
164/*
165 * Function that frees up all resources for a blit. It is usable even if the
166 * blit info has only been partially built as long as the status enum is consistent
167 * with the actual status of the used resources.
168 */
169
170
171static void
172via_free_sg_info(struct pci_dev *pdev, drm_via_sg_info_t *vsg)
173{
174	struct page *page;
175	int i;
176
177	switch (vsg->state) {
178	case dr_via_device_mapped:
179		via_unmap_blit_from_device(pdev, vsg);
 
180	case dr_via_desc_pages_alloc:
181		for (i = 0; i < vsg->num_desc_pages; ++i) {
182			if (vsg->desc_pages[i] != NULL)
183				free_page((unsigned long)vsg->desc_pages[i]);
184		}
185		kfree(vsg->desc_pages);
 
186	case dr_via_pages_locked:
187		for (i = 0; i < vsg->num_pages; ++i) {
188			if (NULL != (page = vsg->pages[i])) {
189				if (!PageReserved(page) && (DMA_FROM_DEVICE == vsg->direction))
190					SetPageDirty(page);
191				put_page(page);
192			}
193		}
194	case dr_via_pages_alloc:
195		vfree(vsg->pages);
 
196	default:
197		vsg->state = dr_via_sg_init;
198	}
199	vfree(vsg->bounce_buffer);
200	vsg->bounce_buffer = NULL;
201	vsg->free_on_sequence = 0;
202}
203
204/*
205 * Fire a blit engine.
206 */
207
208static void
209via_fire_dmablit(struct drm_device *dev, drm_via_sg_info_t *vsg, int engine)
210{
211	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
212
213	VIA_WRITE(VIA_PCI_DMA_MAR0 + engine*0x10, 0);
214	VIA_WRITE(VIA_PCI_DMA_DAR0 + engine*0x10, 0);
215	VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DD | VIA_DMA_CSR_TD |
216		  VIA_DMA_CSR_DE);
217	VIA_WRITE(VIA_PCI_DMA_MR0  + engine*0x04, VIA_DMA_MR_CM | VIA_DMA_MR_TDIE);
218	VIA_WRITE(VIA_PCI_DMA_BCR0 + engine*0x10, 0);
219	VIA_WRITE(VIA_PCI_DMA_DPR0 + engine*0x10, vsg->chain_start);
220	wmb();
221	VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DE | VIA_DMA_CSR_TS);
222	VIA_READ(VIA_PCI_DMA_CSR0 + engine*0x04);
223}
224
225/*
226 * Obtain a page pointer array and lock all pages into system memory. A segmentation violation will
227 * occur here if the calling user does not have access to the submitted address.
228 */
229
230static int
231via_lock_all_dma_pages(drm_via_sg_info_t *vsg,  drm_via_dmablit_t *xfer)
232{
233	int ret;
234	unsigned long first_pfn = VIA_PFN(xfer->mem_addr);
235	vsg->num_pages = VIA_PFN(xfer->mem_addr + (xfer->num_lines * xfer->mem_stride - 1)) -
236		first_pfn + 1;
237
238	vsg->pages = vzalloc(sizeof(struct page *) * vsg->num_pages);
239	if (NULL == vsg->pages)
240		return -ENOMEM;
241	down_read(&current->mm->mmap_sem);
242	ret = get_user_pages((unsigned long)xfer->mem_addr,
243			     vsg->num_pages,
244			     (vsg->direction == DMA_FROM_DEVICE),
245			     0, vsg->pages, NULL);
246
247	up_read(&current->mm->mmap_sem);
248	if (ret != vsg->num_pages) {
249		if (ret < 0)
250			return ret;
251		vsg->state = dr_via_pages_locked;
252		return -EINVAL;
253	}
254	vsg->state = dr_via_pages_locked;
255	DRM_DEBUG("DMA pages locked\n");
256	return 0;
257}
258
259/*
260 * Allocate DMA capable memory for the blit descriptor chain, and an array that keeps track of the
261 * pages we allocate. We don't want to use kmalloc for the descriptor chain because it may be
262 * quite large for some blits, and pages don't need to be contiguous.
263 */
264
265static int
266via_alloc_desc_pages(drm_via_sg_info_t *vsg)
267{
268	int i;
269
270	vsg->descriptors_per_page = PAGE_SIZE / sizeof(drm_via_descriptor_t);
271	vsg->num_desc_pages = (vsg->num_desc + vsg->descriptors_per_page - 1) /
272		vsg->descriptors_per_page;
273
274	if (NULL ==  (vsg->desc_pages = kcalloc(vsg->num_desc_pages, sizeof(void *), GFP_KERNEL)))
275		return -ENOMEM;
276
277	vsg->state = dr_via_desc_pages_alloc;
278	for (i = 0; i < vsg->num_desc_pages; ++i) {
279		if (NULL == (vsg->desc_pages[i] =
280			     (drm_via_descriptor_t *) __get_free_page(GFP_KERNEL)))
281			return -ENOMEM;
282	}
283	DRM_DEBUG("Allocated %d pages for %d descriptors.\n", vsg->num_desc_pages,
284		  vsg->num_desc);
285	return 0;
286}
287
288static void
289via_abort_dmablit(struct drm_device *dev, int engine)
290{
291	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
292
293	VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TA);
294}
295
296static void
297via_dmablit_engine_off(struct drm_device *dev, int engine)
298{
299	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
300
301	VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TD | VIA_DMA_CSR_DD);
302}
303
304
305
306/*
307 * The dmablit part of the IRQ handler. Trying to do only reasonably fast things here.
308 * The rest, like unmapping and freeing memory for done blits is done in a separate workqueue
309 * task. Basically the task of the interrupt handler is to submit a new blit to the engine, while
310 * the workqueue task takes care of processing associated with the old blit.
311 */
312
313void
314via_dmablit_handler(struct drm_device *dev, int engine, int from_irq)
315{
316	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
317	drm_via_blitq_t *blitq = dev_priv->blit_queues + engine;
318	int cur;
319	int done_transfer;
320	unsigned long irqsave = 0;
321	uint32_t status = 0;
322
323	DRM_DEBUG("DMA blit handler called. engine = %d, from_irq = %d, blitq = 0x%lx\n",
324		  engine, from_irq, (unsigned long) blitq);
325
326	if (from_irq)
327		spin_lock(&blitq->blit_lock);
328	else
329		spin_lock_irqsave(&blitq->blit_lock, irqsave);
330
331	done_transfer = blitq->is_active &&
332	  ((status = VIA_READ(VIA_PCI_DMA_CSR0 + engine*0x04)) & VIA_DMA_CSR_TD);
333	done_transfer = done_transfer || (blitq->aborting && !(status & VIA_DMA_CSR_DE));
334
335	cur = blitq->cur;
336	if (done_transfer) {
337
338		blitq->blits[cur]->aborted = blitq->aborting;
339		blitq->done_blit_handle++;
340		wake_up(blitq->blit_queue + cur);
341
342		cur++;
343		if (cur >= VIA_NUM_BLIT_SLOTS)
344			cur = 0;
345		blitq->cur = cur;
346
347		/*
348		 * Clear transfer done flag.
349		 */
350
351		VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04,  VIA_DMA_CSR_TD);
352
353		blitq->is_active = 0;
354		blitq->aborting = 0;
355		schedule_work(&blitq->wq);
356
357	} else if (blitq->is_active && time_after_eq(jiffies, blitq->end)) {
358
359		/*
360		 * Abort transfer after one second.
361		 */
362
363		via_abort_dmablit(dev, engine);
364		blitq->aborting = 1;
365		blitq->end = jiffies + HZ;
366	}
367
368	if (!blitq->is_active) {
369		if (blitq->num_outstanding) {
370			via_fire_dmablit(dev, blitq->blits[cur], engine);
371			blitq->is_active = 1;
372			blitq->cur = cur;
373			blitq->num_outstanding--;
374			blitq->end = jiffies + HZ;
375			if (!timer_pending(&blitq->poll_timer))
376				mod_timer(&blitq->poll_timer, jiffies + 1);
377		} else {
378			if (timer_pending(&blitq->poll_timer))
379				del_timer(&blitq->poll_timer);
380			via_dmablit_engine_off(dev, engine);
381		}
382	}
383
384	if (from_irq)
385		spin_unlock(&blitq->blit_lock);
386	else
387		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
388}
389
390
391
392/*
393 * Check whether this blit is still active, performing necessary locking.
394 */
395
396static int
397via_dmablit_active(drm_via_blitq_t *blitq, int engine, uint32_t handle, wait_queue_head_t **queue)
398{
399	unsigned long irqsave;
400	uint32_t slot;
401	int active;
402
403	spin_lock_irqsave(&blitq->blit_lock, irqsave);
404
405	/*
406	 * Allow for handle wraparounds.
407	 */
408
409	active = ((blitq->done_blit_handle - handle) > (1 << 23)) &&
410		((blitq->cur_blit_handle - handle) <= (1 << 23));
411
412	if (queue && active) {
413		slot = handle - blitq->done_blit_handle + blitq->cur - 1;
414		if (slot >= VIA_NUM_BLIT_SLOTS)
415			slot -= VIA_NUM_BLIT_SLOTS;
416		*queue = blitq->blit_queue + slot;
417	}
418
419	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
420
421	return active;
422}
423
424/*
425 * Sync. Wait for at least three seconds for the blit to be performed.
426 */
427
428static int
429via_dmablit_sync(struct drm_device *dev, uint32_t handle, int engine)
430{
431
432	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
433	drm_via_blitq_t *blitq = dev_priv->blit_queues + engine;
434	wait_queue_head_t *queue;
435	int ret = 0;
436
437	if (via_dmablit_active(blitq, engine, handle, &queue)) {
438		DRM_WAIT_ON(ret, *queue, 3 * HZ,
439			    !via_dmablit_active(blitq, engine, handle, NULL));
440	}
441	DRM_DEBUG("DMA blit sync handle 0x%x engine %d returned %d\n",
442		  handle, engine, ret);
443
444	return ret;
445}
446
447
448/*
449 * A timer that regularly polls the blit engine in cases where we don't have interrupts:
450 * a) Broken hardware (typically those that don't have any video capture facility).
451 * b) Blit abort. The hardware doesn't send an interrupt when a blit is aborted.
452 * The timer and hardware IRQ's can and do work in parallel. If the hardware has
453 * irqs, it will shorten the latency somewhat.
454 */
455
456
457
458static void
459via_dmablit_timer(unsigned long data)
460{
461	drm_via_blitq_t *blitq = (drm_via_blitq_t *) data;
462	struct drm_device *dev = blitq->dev;
463	int engine = (int)
464		(blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues);
465
466	DRM_DEBUG("Polling timer called for engine %d, jiffies %lu\n", engine,
467		  (unsigned long) jiffies);
468
469	via_dmablit_handler(dev, engine, 0);
470
471	if (!timer_pending(&blitq->poll_timer)) {
472		mod_timer(&blitq->poll_timer, jiffies + 1);
473
474	       /*
475		* Rerun handler to delete timer if engines are off, and
476		* to shorten abort latency. This is a little nasty.
477		*/
478
479	       via_dmablit_handler(dev, engine, 0);
480
481	}
482}
483
484
485
486
487/*
488 * Workqueue task that frees data and mappings associated with a blit.
489 * Also wakes up waiting processes. Each of these tasks handles one
490 * blit engine only and may not be called on each interrupt.
491 */
492
493
494static void
495via_dmablit_workqueue(struct work_struct *work)
496{
497	drm_via_blitq_t *blitq = container_of(work, drm_via_blitq_t, wq);
498	struct drm_device *dev = blitq->dev;
499	unsigned long irqsave;
500	drm_via_sg_info_t *cur_sg;
501	int cur_released;
502
503
504	DRM_DEBUG("Workqueue task called for blit engine %ld\n", (unsigned long)
505		  (blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues));
506
507	spin_lock_irqsave(&blitq->blit_lock, irqsave);
508
509	while (blitq->serviced != blitq->cur) {
510
511		cur_released = blitq->serviced++;
512
513		DRM_DEBUG("Releasing blit slot %d\n", cur_released);
514
515		if (blitq->serviced >= VIA_NUM_BLIT_SLOTS)
516			blitq->serviced = 0;
517
518		cur_sg = blitq->blits[cur_released];
519		blitq->num_free++;
520
521		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
522
523		wake_up(&blitq->busy_queue);
524
525		via_free_sg_info(dev->pdev, cur_sg);
526		kfree(cur_sg);
527
528		spin_lock_irqsave(&blitq->blit_lock, irqsave);
529	}
530
531	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
532}
533
534
535/*
536 * Init all blit engines. Currently we use two, but some hardware have 4.
537 */
538
539
540void
541via_init_dmablit(struct drm_device *dev)
542{
543	int i, j;
544	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
545	drm_via_blitq_t *blitq;
546
547	pci_set_master(dev->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		setup_timer(&blitq->poll_timer, via_dmablit_timer,
567				(unsigned long)blitq);
568	}
569}
570
571/*
572 * Build all info and do all mappings required for a blit.
573 */
574
575
576static int
577via_build_sg_info(struct drm_device *dev, drm_via_sg_info_t *vsg, drm_via_dmablit_t *xfer)
578{
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(dev->pdev, vsg);
659		return ret;
660	}
661
662	via_map_blit_for_device(dev->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(dev->pdev, vsg);
666		return ret;
667	}
668	via_map_blit_for_device(dev->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
679static int
680via_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		DRM_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
707static void
708via_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
723static int
724via_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
774int
775via_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
798int
799via_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}
v5.9
  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
 52typedef 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
 66static void
 67via_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
 97static void
 98via_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
174static void
175via_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
209static void
210via_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
231static int
232via_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
263static int
264via_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
286static void
287via_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
294static void
295via_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
311void
312via_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
394static int
395via_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
426static int
427via_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
456static void
457via_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
492static void
493via_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	unsigned long irqsave;
498	drm_via_sg_info_t *cur_sg;
499	int cur_released;
500
501
502	DRM_DEBUG("Workqueue task called for blit engine %ld\n", (unsigned long)
503		  (blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues));
504
505	spin_lock_irqsave(&blitq->blit_lock, irqsave);
506
507	while (blitq->serviced != blitq->cur) {
508
509		cur_released = blitq->serviced++;
510
511		DRM_DEBUG("Releasing blit slot %d\n", cur_released);
512
513		if (blitq->serviced >= VIA_NUM_BLIT_SLOTS)
514			blitq->serviced = 0;
515
516		cur_sg = blitq->blits[cur_released];
517		blitq->num_free++;
518
519		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
520
521		wake_up(&blitq->busy_queue);
522
523		via_free_sg_info(dev->pdev, cur_sg);
524		kfree(cur_sg);
525
526		spin_lock_irqsave(&blitq->blit_lock, irqsave);
527	}
528
529	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
530}
531
532
533/*
534 * Init all blit engines. Currently we use two, but some hardware have 4.
535 */
536
537
538void
539via_init_dmablit(struct drm_device *dev)
540{
541	int i, j;
542	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
543	drm_via_blitq_t *blitq;
544
545	pci_set_master(dev->pdev);
546
547	for (i = 0; i < VIA_NUM_BLIT_ENGINES; ++i) {
548		blitq = dev_priv->blit_queues + i;
549		blitq->dev = dev;
550		blitq->cur_blit_handle = 0;
551		blitq->done_blit_handle = 0;
552		blitq->head = 0;
553		blitq->cur = 0;
554		blitq->serviced = 0;
555		blitq->num_free = VIA_NUM_BLIT_SLOTS - 1;
556		blitq->num_outstanding = 0;
557		blitq->is_active = 0;
558		blitq->aborting = 0;
559		spin_lock_init(&blitq->blit_lock);
560		for (j = 0; j < VIA_NUM_BLIT_SLOTS; ++j)
561			init_waitqueue_head(blitq->blit_queue + j);
562		init_waitqueue_head(&blitq->busy_queue);
563		INIT_WORK(&blitq->wq, via_dmablit_workqueue);
564		timer_setup(&blitq->poll_timer, via_dmablit_timer, 0);
 
565	}
566}
567
568/*
569 * Build all info and do all mappings required for a blit.
570 */
571
572
573static int
574via_build_sg_info(struct drm_device *dev, drm_via_sg_info_t *vsg, drm_via_dmablit_t *xfer)
575{
576	int draw = xfer->to_fb;
577	int ret = 0;
578
579	vsg->direction = (draw) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;
580	vsg->bounce_buffer = NULL;
581
582	vsg->state = dr_via_sg_init;
583
584	if (xfer->num_lines <= 0 || xfer->line_length <= 0) {
585		DRM_ERROR("Zero size bitblt.\n");
586		return -EINVAL;
587	}
588
589	/*
590	 * Below check is a driver limitation, not a hardware one. We
591	 * don't want to lock unused pages, and don't want to incoporate the
592	 * extra logic of avoiding them. Make sure there are no.
593	 * (Not a big limitation anyway.)
594	 */
595
596	if ((xfer->mem_stride - xfer->line_length) > 2*PAGE_SIZE) {
597		DRM_ERROR("Too large system memory stride. Stride: %d, "
598			  "Length: %d\n", xfer->mem_stride, xfer->line_length);
599		return -EINVAL;
600	}
601
602	if ((xfer->mem_stride == xfer->line_length) &&
603	   (xfer->fb_stride == xfer->line_length)) {
604		xfer->mem_stride *= xfer->num_lines;
605		xfer->line_length = xfer->mem_stride;
606		xfer->fb_stride = xfer->mem_stride;
607		xfer->num_lines = 1;
608	}
609
610	/*
611	 * Don't lock an arbitrary large number of pages, since that causes a
612	 * DOS security hole.
613	 */
614
615	if (xfer->num_lines > 2048 || (xfer->num_lines*xfer->mem_stride > (2048*2048*4))) {
616		DRM_ERROR("Too large PCI DMA bitblt.\n");
617		return -EINVAL;
618	}
619
620	/*
621	 * we allow a negative fb stride to allow flipping of images in
622	 * transfer.
623	 */
624
625	if (xfer->mem_stride < xfer->line_length ||
626		abs(xfer->fb_stride) < xfer->line_length) {
627		DRM_ERROR("Invalid frame-buffer / memory stride.\n");
628		return -EINVAL;
629	}
630
631	/*
632	 * A hardware bug seems to be worked around if system memory addresses start on
633	 * 16 byte boundaries. This seems a bit restrictive however. VIA is contacted
634	 * about this. Meanwhile, impose the following restrictions:
635	 */
636
637#ifdef VIA_BUGFREE
638	if ((((unsigned long)xfer->mem_addr & 3) != ((unsigned long)xfer->fb_addr & 3)) ||
639	    ((xfer->num_lines > 1) && ((xfer->mem_stride & 3) != (xfer->fb_stride & 3)))) {
640		DRM_ERROR("Invalid DRM bitblt alignment.\n");
641		return -EINVAL;
642	}
643#else
644	if ((((unsigned long)xfer->mem_addr & 15) ||
645	      ((unsigned long)xfer->fb_addr & 3)) ||
646	   ((xfer->num_lines > 1) &&
647	   ((xfer->mem_stride & 15) || (xfer->fb_stride & 3)))) {
648		DRM_ERROR("Invalid DRM bitblt alignment.\n");
649		return -EINVAL;
650	}
651#endif
652
653	if (0 != (ret = via_lock_all_dma_pages(vsg, xfer))) {
654		DRM_ERROR("Could not lock DMA pages.\n");
655		via_free_sg_info(dev->pdev, vsg);
656		return ret;
657	}
658
659	via_map_blit_for_device(dev->pdev, xfer, vsg, 0);
660	if (0 != (ret = via_alloc_desc_pages(vsg))) {
661		DRM_ERROR("Could not allocate DMA descriptor pages.\n");
662		via_free_sg_info(dev->pdev, vsg);
663		return ret;
664	}
665	via_map_blit_for_device(dev->pdev, xfer, vsg, 1);
666
667	return 0;
668}
669
670
671/*
672 * Reserve one free slot in the blit queue. Will wait for one second for one
673 * to become available. Otherwise -EBUSY is returned.
674 */
675
676static int
677via_dmablit_grab_slot(drm_via_blitq_t *blitq, int engine)
678{
679	int ret = 0;
680	unsigned long irqsave;
681
682	DRM_DEBUG("Num free is %d\n", blitq->num_free);
683	spin_lock_irqsave(&blitq->blit_lock, irqsave);
684	while (blitq->num_free == 0) {
685		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
686
687		VIA_WAIT_ON(ret, blitq->busy_queue, HZ, blitq->num_free > 0);
688		if (ret)
689			return (-EINTR == ret) ? -EAGAIN : ret;
690
691		spin_lock_irqsave(&blitq->blit_lock, irqsave);
692	}
693
694	blitq->num_free--;
695	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
696
697	return 0;
698}
699
700/*
701 * Hand back a free slot if we changed our mind.
702 */
703
704static void
705via_dmablit_release_slot(drm_via_blitq_t *blitq)
706{
707	unsigned long irqsave;
708
709	spin_lock_irqsave(&blitq->blit_lock, irqsave);
710	blitq->num_free++;
711	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
712	wake_up(&blitq->busy_queue);
713}
714
715/*
716 * Grab a free slot. Build blit info and queue a blit.
717 */
718
719
720static int
721via_dmablit(struct drm_device *dev, drm_via_dmablit_t *xfer)
722{
723	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
724	drm_via_sg_info_t *vsg;
725	drm_via_blitq_t *blitq;
726	int ret;
727	int engine;
728	unsigned long irqsave;
729
730	if (dev_priv == NULL) {
731		DRM_ERROR("Called without initialization.\n");
732		return -EINVAL;
733	}
734
735	engine = (xfer->to_fb) ? 0 : 1;
736	blitq = dev_priv->blit_queues + engine;
737	if (0 != (ret = via_dmablit_grab_slot(blitq, engine)))
738		return ret;
739	if (NULL == (vsg = kmalloc(sizeof(*vsg), GFP_KERNEL))) {
740		via_dmablit_release_slot(blitq);
741		return -ENOMEM;
742	}
743	if (0 != (ret = via_build_sg_info(dev, vsg, xfer))) {
744		via_dmablit_release_slot(blitq);
745		kfree(vsg);
746		return ret;
747	}
748	spin_lock_irqsave(&blitq->blit_lock, irqsave);
749
750	blitq->blits[blitq->head++] = vsg;
751	if (blitq->head >= VIA_NUM_BLIT_SLOTS)
752		blitq->head = 0;
753	blitq->num_outstanding++;
754	xfer->sync.sync_handle = ++blitq->cur_blit_handle;
755
756	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
757	xfer->sync.engine = engine;
758
759	via_dmablit_handler(dev, engine, 0);
760
761	return 0;
762}
763
764/*
765 * Sync on a previously submitted blit. Note that the X server use signals extensively, and
766 * that there is a very big probability that this IOCTL will be interrupted by a signal. In that
767 * case it returns with -EAGAIN for the signal to be delivered.
768 * The caller should then reissue the IOCTL. This is similar to what is being done for drmGetLock().
769 */
770
771int
772via_dma_blit_sync(struct drm_device *dev, void *data, struct drm_file *file_priv)
773{
774	drm_via_blitsync_t *sync = data;
775	int err;
776
777	if (sync->engine >= VIA_NUM_BLIT_ENGINES)
778		return -EINVAL;
779
780	err = via_dmablit_sync(dev, sync->sync_handle, sync->engine);
781
782	if (-EINTR == err)
783		err = -EAGAIN;
784
785	return err;
786}
787
788
789/*
790 * Queue a blit and hand back a handle to be used for sync. This IOCTL may be interrupted by a signal
791 * while waiting for a free slot in the blit queue. In that case it returns with -EAGAIN and should
792 * be reissued. See the above IOCTL code.
793 */
794
795int
796via_dma_blit(struct drm_device *dev, void *data, struct drm_file *file_priv)
797{
798	drm_via_dmablit_t *xfer = data;
799	int err;
800
801	err = via_dmablit(dev, xfer);
802
803	return err;
804}