1 /* via_dmablit.c -- PCI DMA BitBlt support for the VIA Unichrome/Pro
3 * Copyright (C) 2005 Thomas Hellstrom, All Rights Reserved.
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:
12 * The above copyright notice and this permission notice (including the
13 * next paragraph) shall be included in all copies or substantial portions
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.
26 * Partially based on code obtained from Digeo Inc.
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
33 * Unmaps the DMA mappings.
34 * FIXME: Is this a NoOp on x86? Also
35 * FIXME: What happens if this one is called and a pending blit has previously done
36 * the same DMA mappings?
39 #include "dev/drm/drmP.h"
40 #include "dev/drm/via_drm.h"
41 #include "dev/drm/via_drv.h"
42 #include "dev/drm/via_dmablit.h"
44 #define VIA_PGDN(x) (((unsigned long)(x)) & ~PAGE_MASK)
45 #define VIA_PGOFF(x) (((unsigned long)(x)) & PAGE_MASK)
46 #define VIA_PFN(x) ((unsigned long)(x) >> PAGE_SHIFT)
48 typedef struct _drm_via_descriptor {
53 } drm_via_descriptor_t;
55 static void via_dmablit_timer(void *arg);
58 * Unmap a DMA mapping.
61 via_unmap_blit_from_device(drm_via_sg_info_t *vsg)
63 int num_desc = vsg->num_desc;
64 unsigned cur_descriptor_page = num_desc / vsg->descriptors_per_page;
65 unsigned descriptor_this_page = num_desc % vsg->descriptors_per_page;
66 drm_via_descriptor_t *desc_ptr = vsg->desc_pages[cur_descriptor_page] +
68 dma_addr_t next = vsg->chain_start;
71 if (descriptor_this_page-- == 0) {
72 cur_descriptor_page--;
73 descriptor_this_page = vsg->descriptors_per_page - 1;
74 desc_ptr = vsg->desc_pages[cur_descriptor_page] +
77 next = (dma_addr_t) desc_ptr->next;
84 * If mode = 0, count how many descriptors are needed.
85 * If mode = 1, Map the DMA pages for the device, put together and map also the descriptors.
86 * Descriptors are run in reverse order by the hardware because we are not allowed to update the
87 * 'next' field without syncing calls when the descriptor is already mapped.
90 via_map_blit_for_device(const drm_via_dmablit_t *xfer,
91 drm_via_sg_info_t *vsg, int mode)
93 unsigned cur_descriptor_page = 0;
94 unsigned num_descriptors_this_page = 0;
95 unsigned char *mem_addr = xfer->mem_addr;
96 unsigned char *cur_mem;
97 unsigned char *first_addr = (unsigned char *)VIA_PGDN(mem_addr);
98 uint32_t fb_addr = xfer->fb_addr;
100 unsigned long line_len;
101 unsigned remaining_len;
104 dma_addr_t next = 0 | VIA_DMA_DPR_EC;
105 drm_via_descriptor_t *desc_ptr = NULL;
108 desc_ptr = vsg->desc_pages[cur_descriptor_page];
110 for (cur_line = 0; cur_line < xfer->num_lines; ++cur_line) {
112 line_len = xfer->line_length;
116 while (line_len > 0) {
118 remaining_len = min(PAGE_SIZE - VIA_PGOFF(cur_mem),
120 line_len -= remaining_len;
125 vsg->pages[VIA_PFN(cur_mem) -
126 VIA_PFN(first_addr)]) + VIA_PGOFF(cur_mem);
127 desc_ptr->dev_addr = cur_fb;
129 desc_ptr->size = remaining_len;
130 desc_ptr->next = (uint32_t) next;
132 next = vtophys(desc_ptr);
135 if (++num_descriptors_this_page >= vsg->descriptors_per_page) {
136 num_descriptors_this_page = 0;
137 desc_ptr = vsg->desc_pages[++cur_descriptor_page];
142 cur_mem += remaining_len;
143 cur_fb += remaining_len;
146 mem_addr += xfer->mem_stride;
147 fb_addr += xfer->fb_stride;
151 vsg->chain_start = next;
152 vsg->state = dr_via_device_mapped;
154 vsg->num_desc = num_desc;
159 * Function that frees up all resources for a blit. It is usable even if the
160 * blit info has only been partially built as long as the status enum is consistent
161 * with the actual status of the used resources.
164 via_free_sg_info(drm_via_sg_info_t *vsg)
170 case dr_via_device_mapped:
171 via_unmap_blit_from_device(vsg);
172 case dr_via_desc_pages_alloc:
173 for (i=0; i<vsg->num_desc_pages; ++i) {
174 if (vsg->desc_pages[i] != NULL)
175 free(vsg->desc_pages[i], DRM_MEM_PAGES);
177 free(vsg->desc_pages, DRM_MEM_DRIVER);
178 case dr_via_pages_locked:
179 for (i=0; i < vsg->num_pages; ++i) {
180 page = vsg->pages[i];
182 vm_page_unwire(page, 0);
183 vm_page_unlock(page);
185 case dr_via_pages_alloc:
186 free(vsg->pages, DRM_MEM_DRIVER);
188 vsg->state = dr_via_sg_init;
190 free(vsg->bounce_buffer, DRM_MEM_DRIVER);
191 vsg->bounce_buffer = NULL;
192 vsg->free_on_sequence = 0;
197 * Fire a blit engine.
200 via_fire_dmablit(struct drm_device *dev, drm_via_sg_info_t *vsg, int engine)
202 drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
204 VIA_WRITE(VIA_PCI_DMA_MAR0 + engine*0x10, 0);
205 VIA_WRITE(VIA_PCI_DMA_DAR0 + engine*0x10, 0);
206 VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DD | VIA_DMA_CSR_TD |
208 VIA_WRITE(VIA_PCI_DMA_MR0 + engine*0x04, VIA_DMA_MR_CM | VIA_DMA_MR_TDIE);
209 VIA_WRITE(VIA_PCI_DMA_BCR0 + engine*0x10, 0);
210 VIA_WRITE(VIA_PCI_DMA_DPR0 + engine*0x10, vsg->chain_start);
211 DRM_WRITEMEMORYBARRIER();
212 VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DE | VIA_DMA_CSR_TS);
213 (void)VIA_READ(VIA_PCI_DMA_CSR0 + engine*0x04);
218 * Obtain a page pointer array and lock all pages into system memory. A segmentation violation will
219 * occur here if the calling user does not have access to the submitted address.
222 via_lock_all_dma_pages(drm_via_sg_info_t *vsg, drm_via_dmablit_t *xfer)
224 unsigned long first_pfn = VIA_PFN(xfer->mem_addr);
228 vsg->num_pages = VIA_PFN(xfer->mem_addr +
229 (xfer->num_lines * xfer->mem_stride -1)) - first_pfn + 1;
231 if (NULL == (vsg->pages = malloc(sizeof(vm_page_t) * vsg->num_pages,
232 DRM_MEM_DRIVER, M_NOWAIT)))
235 vsg->state = dr_via_pages_alloc;
237 if (vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
238 (vm_offset_t)xfer->mem_addr, vsg->num_pages * PAGE_SIZE,
239 VM_PROT_READ | VM_PROT_WRITE, vsg->pages, vsg->num_pages) < 0)
242 for (i = 0; i < vsg->num_pages; i++) {
249 vsg->state = dr_via_pages_locked;
251 DRM_DEBUG("DMA pages locked\n");
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 contingous.
263 via_alloc_desc_pages(drm_via_sg_info_t *vsg)
267 vsg->descriptors_per_page = PAGE_SIZE / sizeof(drm_via_descriptor_t);
268 vsg->num_desc_pages = (vsg->num_desc + vsg->descriptors_per_page - 1) /
269 vsg->descriptors_per_page;
271 if (NULL == (vsg->desc_pages = malloc(vsg->num_desc_pages *
272 sizeof(void *), DRM_MEM_DRIVER, M_NOWAIT | M_ZERO)))
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 *)malloc(PAGE_SIZE, DRM_MEM_PAGES,
282 DRM_DEBUG("Allocated %d pages for %d descriptors.\n",
283 vsg->num_desc_pages, vsg->num_desc);
290 via_abort_dmablit(struct drm_device *dev, int engine)
292 drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
294 VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TA);
299 via_dmablit_engine_off(struct drm_device *dev, int engine)
301 drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
303 VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TD | VIA_DMA_CSR_DD);
308 * The dmablit part of the IRQ handler. Trying to do only reasonably fast things here.
309 * The rest, like unmapping and freeing memory for done blits is done in a separate workqueue
310 * task. Basically the task of the interrupt handler is to submit a new blit to the engine, while
311 * the workqueue task takes care of processing associated with the old blit.
314 via_dmablit_handler(struct drm_device *dev, int engine, int from_irq)
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;
322 DRM_DEBUG("DMA blit handler called. engine = %d, from_irq = %d, blitq = 0x%lx\n",
323 engine, from_irq, (unsigned long) blitq);
325 mtx_lock(&blitq->blit_lock);
327 done_transfer = blitq->is_active &&
328 (( status = VIA_READ(VIA_PCI_DMA_CSR0 + engine*0x04)) & VIA_DMA_CSR_TD);
329 done_transfer = done_transfer || ( blitq->aborting && !(status & VIA_DMA_CSR_DE));
334 blitq->blits[cur]->aborted = blitq->aborting;
335 blitq->done_blit_handle++;
336 DRM_WAKEUP(&blitq->blit_queue[cur]);
339 if (cur >= VIA_NUM_BLIT_SLOTS)
344 * Clear transfer done flag.
347 VIA_WRITE(VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TD);
349 blitq->is_active = 0;
352 taskqueue_enqueue(taskqueue_swi, &blitq->wq);
354 } else if (blitq->is_active && (ticks >= blitq->end)) {
357 * Abort transfer after one second.
360 via_abort_dmablit(dev, engine);
362 blitq->end = ticks + DRM_HZ;
365 if (!blitq->is_active) {
366 if (blitq->num_outstanding) {
367 via_fire_dmablit(dev, blitq->blits[cur], engine);
368 blitq->is_active = 1;
370 blitq->num_outstanding--;
371 blitq->end = ticks + DRM_HZ;
373 if (!callout_pending(&blitq->poll_timer))
374 callout_reset(&blitq->poll_timer,
375 1, (timeout_t *)via_dmablit_timer,
378 if (callout_pending(&blitq->poll_timer)) {
379 callout_stop(&blitq->poll_timer);
381 via_dmablit_engine_off(dev, engine);
385 mtx_unlock(&blitq->blit_lock);
390 * Check whether this blit is still active, performing necessary locking.
393 via_dmablit_active(drm_via_blitq_t *blitq, int engine, uint32_t handle, wait_queue_head_t **queue)
398 mtx_lock(&blitq->blit_lock);
401 * Allow for handle wraparounds.
403 active = ((blitq->done_blit_handle - handle) > (1 << 23)) &&
404 ((blitq->cur_blit_handle - handle) <= (1 << 23));
406 if (queue && active) {
407 slot = handle - blitq->done_blit_handle + blitq->cur -1;
408 if (slot >= VIA_NUM_BLIT_SLOTS) {
409 slot -= VIA_NUM_BLIT_SLOTS;
411 *queue = blitq->blit_queue + slot;
414 mtx_unlock(&blitq->blit_lock);
421 * Sync. Wait for at least three seconds for the blit to be performed.
424 via_dmablit_sync(struct drm_device *dev, uint32_t handle, int engine)
427 drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
428 drm_via_blitq_t *blitq = dev_priv->blit_queues + engine;
429 wait_queue_head_t *queue;
432 if (via_dmablit_active(blitq, engine, handle, &queue)) {
433 DRM_WAIT_ON(ret, *queue, 3 * DRM_HZ,
434 !via_dmablit_active(blitq, engine, handle, NULL));
436 DRM_DEBUG("DMA blit sync handle 0x%x engine %d returned %d\n",
437 handle, engine, ret);
444 * A timer that regularly polls the blit engine in cases where we don't have interrupts:
445 * a) Broken hardware (typically those that don't have any video capture facility).
446 * b) Blit abort. The hardware doesn't send an interrupt when a blit is aborted.
447 * The timer and hardware IRQ's can and do work in parallel. If the hardware has
448 * irqs, it will shorten the latency somewhat.
451 via_dmablit_timer(void *arg)
453 drm_via_blitq_t *blitq = (drm_via_blitq_t *)arg;
454 struct drm_device *dev = blitq->dev;
456 (blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues);
458 DRM_DEBUG("Polling timer called for engine %d, jiffies %lu\n", engine,
459 (unsigned long) jiffies);
461 via_dmablit_handler(dev, engine, 0);
463 if (!callout_pending(&blitq->poll_timer)) {
464 callout_schedule(&blitq->poll_timer, 1);
467 * Rerun handler to delete timer if engines are off, and
468 * to shorten abort latency. This is a little nasty.
471 via_dmablit_handler(dev, engine, 0);
478 * Workqueue task that frees data and mappings associated with a blit.
479 * Also wakes up waiting processes. Each of these tasks handles one
480 * blit engine only and may not be called on each interrupt.
483 via_dmablit_workqueue(void *arg, int pending)
485 drm_via_blitq_t *blitq = (drm_via_blitq_t *)arg;
486 struct drm_device *dev = blitq->dev;
487 drm_via_sg_info_t *cur_sg;
491 DRM_DEBUG("task called for blit engine %ld\n",(unsigned long)
492 (blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues));
494 mtx_lock(&blitq->blit_lock);
496 while(blitq->serviced != blitq->cur) {
498 cur_released = blitq->serviced++;
500 DRM_DEBUG("Releasing blit slot %d\n", cur_released);
502 if (blitq->serviced >= VIA_NUM_BLIT_SLOTS)
505 cur_sg = blitq->blits[cur_released];
508 mtx_unlock(&blitq->blit_lock);
510 DRM_WAKEUP(&blitq->busy_queue);
512 via_free_sg_info(cur_sg);
513 free(cur_sg, DRM_MEM_DRIVER);
515 mtx_lock(&blitq->blit_lock);
518 mtx_unlock(&blitq->blit_lock);
523 * Init all blit engines. Currently we use two, but some hardware have 4.
526 via_init_dmablit(struct drm_device *dev)
529 drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
530 drm_via_blitq_t *blitq;
532 for (i=0; i< VIA_NUM_BLIT_ENGINES; ++i) {
533 blitq = dev_priv->blit_queues + i;
535 blitq->cur_blit_handle = 0;
536 blitq->done_blit_handle = 0;
540 blitq->num_free = VIA_NUM_BLIT_SLOTS - 1;
541 blitq->num_outstanding = 0;
542 blitq->is_active = 0;
544 mtx_init(&blitq->blit_lock, "via_blit_lk", NULL, MTX_DEF);
545 for (j=0; j<VIA_NUM_BLIT_SLOTS; ++j) {
546 DRM_INIT_WAITQUEUE(blitq->blit_queue + j);
548 DRM_INIT_WAITQUEUE(&blitq->busy_queue);
549 TASK_INIT(&blitq->wq, 0, via_dmablit_workqueue, blitq);
550 callout_init(&blitq->poll_timer, 0);
556 * Build all info and do all mappings required for a blit.
559 via_build_sg_info(struct drm_device *dev, drm_via_sg_info_t *vsg,
560 drm_via_dmablit_t *xfer)
564 vsg->bounce_buffer = NULL;
566 vsg->state = dr_via_sg_init;
568 if (xfer->num_lines <= 0 || xfer->line_length <= 0) {
569 DRM_ERROR("Zero size bitblt.\n");
574 * Below check is a driver limitation, not a hardware one. We
575 * don't want to lock unused pages, and don't want to incoporate the
576 * extra logic of avoiding them. Make sure there are no.
577 * (Not a big limitation anyway.)
579 if ((xfer->mem_stride - xfer->line_length) > 2 * PAGE_SIZE) {
580 DRM_ERROR("Too large system memory stride. Stride: %d, "
581 "Length: %d\n", xfer->mem_stride, xfer->line_length);
585 if ((xfer->mem_stride == xfer->line_length) &&
586 (xfer->fb_stride == xfer->line_length)) {
587 xfer->mem_stride *= xfer->num_lines;
588 xfer->line_length = xfer->mem_stride;
589 xfer->fb_stride = xfer->mem_stride;
594 * Don't lock an arbitrary large number of pages, since that causes a
597 if (xfer->num_lines > 2048 ||
598 (xfer->num_lines*xfer->mem_stride > (2048*2048*4))) {
599 DRM_ERROR("Too large PCI DMA bitblt.\n");
604 * we allow a negative fb stride to allow flipping of images in
607 if (xfer->mem_stride < xfer->line_length ||
608 abs(xfer->fb_stride) < xfer->line_length) {
609 DRM_ERROR("Invalid frame-buffer / memory stride.\n");
614 * A hardware bug seems to be worked around if system memory addresses
615 * start on 16 byte boundaries. This seems a bit restrictive however.
616 * VIA is contacted about this. Meanwhile, impose the following
620 if ((((unsigned long)xfer->mem_addr & 3) !=
621 ((unsigned long)xfer->fb_addr & 3)) ||
622 ((xfer->num_lines > 1) && ((xfer->mem_stride & 3) !=
623 (xfer->fb_stride & 3)))) {
624 DRM_ERROR("Invalid DRM bitblt alignment.\n");
628 if ((((unsigned long)xfer->mem_addr & 15) ||
629 ((unsigned long)xfer->fb_addr & 3)) ||
630 ((xfer->num_lines > 1) &&
631 ((xfer->mem_stride & 15) || (xfer->fb_stride & 3)))) {
632 DRM_ERROR("Invalid DRM bitblt alignment.\n");
637 if (0 != (ret = via_lock_all_dma_pages(vsg, xfer))) {
638 DRM_ERROR("Could not lock DMA pages.\n");
639 via_free_sg_info(vsg);
643 via_map_blit_for_device(xfer, vsg, 0);
644 if (0 != (ret = via_alloc_desc_pages(vsg))) {
645 DRM_ERROR("Could not allocate DMA descriptor pages.\n");
646 via_free_sg_info(vsg);
649 via_map_blit_for_device(xfer, vsg, 1);
656 * Reserve one free slot in the blit queue. Will wait for one second for one
657 * to become available. Otherwise -EBUSY is returned.
660 via_dmablit_grab_slot(drm_via_blitq_t *blitq, int engine)
662 struct drm_device *dev = blitq->dev;
665 DRM_DEBUG("Num free is %d\n", blitq->num_free);
666 mtx_lock(&blitq->blit_lock);
667 while(blitq->num_free == 0) {
668 mtx_unlock(&blitq->blit_lock);
670 DRM_WAIT_ON(ret, blitq->busy_queue, DRM_HZ,
671 blitq->num_free > 0);
673 return (-EINTR == ret) ? -EAGAIN : ret;
676 mtx_lock(&blitq->blit_lock);
680 mtx_unlock(&blitq->blit_lock);
687 * Hand back a free slot if we changed our mind.
690 via_dmablit_release_slot(drm_via_blitq_t *blitq)
693 mtx_lock(&blitq->blit_lock);
695 mtx_unlock(&blitq->blit_lock);
696 DRM_WAKEUP( &blitq->busy_queue );
701 * Grab a free slot. Build blit info and queue a blit.
704 via_dmablit(struct drm_device *dev, drm_via_dmablit_t *xfer)
706 drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
707 drm_via_sg_info_t *vsg;
708 drm_via_blitq_t *blitq;
712 if (dev_priv == NULL) {
713 DRM_ERROR("Called without initialization.\n");
717 engine = (xfer->to_fb) ? 0 : 1;
718 blitq = dev_priv->blit_queues + engine;
719 if (0 != (ret = via_dmablit_grab_slot(blitq, engine))) {
722 if (NULL == (vsg = malloc(sizeof(*vsg), DRM_MEM_DRIVER,
723 M_NOWAIT | M_ZERO))) {
724 via_dmablit_release_slot(blitq);
727 if (0 != (ret = via_build_sg_info(dev, vsg, xfer))) {
728 via_dmablit_release_slot(blitq);
729 free(vsg, DRM_MEM_DRIVER);
732 mtx_lock(&blitq->blit_lock);
734 blitq->blits[blitq->head++] = vsg;
735 if (blitq->head >= VIA_NUM_BLIT_SLOTS)
737 blitq->num_outstanding++;
738 xfer->sync.sync_handle = ++blitq->cur_blit_handle;
740 mtx_unlock(&blitq->blit_lock);
741 xfer->sync.engine = engine;
743 via_dmablit_handler(dev, engine, 0);
750 * Sync on a previously submitted blit. Note that the X server use signals
751 * extensively, and that there is a very big probability that this IOCTL will
752 * be interrupted by a signal. In that case it returns with -EAGAIN for the
753 * signal to be delivered. The caller should then reissue the IOCTL. This is
754 * similar to what is being done for drmGetLock().
757 via_dma_blit_sync( struct drm_device *dev, void *data,
758 struct drm_file *file_priv )
760 drm_via_blitsync_t *sync = data;
763 if (sync->engine >= VIA_NUM_BLIT_ENGINES)
766 err = via_dmablit_sync(dev, sync->sync_handle, sync->engine);
776 * Queue a blit and hand back a handle to be used for sync. This IOCTL may be
777 * interrupted by a signal while waiting for a free slot in the blit queue.
778 * In that case it returns with -EAGAIN and should be reissued. See the above
782 via_dma_blit( struct drm_device *dev, void *data, struct drm_file *file_priv )
784 drm_via_dmablit_t *xfer = data;
787 err = via_dmablit(dev, xfer);