xref: /src/sys/external/bsd/drm2/dist/drm/via/via_dmablit.c

/*	$NetBSD: via_dmablit.c,v 1.9 2021/12/18 23:45:44 riastradh Exp $	*/

/* via_dmablit.c -- PCI DMA BitBlt support for the VIA Unichrome/Pro
 *
 * Copyright (C) 2005 Thomas Hellstrom, All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sub license,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 * Authors:
 *    Thomas Hellstrom.
 *    Partially based on code obtained from Digeo Inc.
 */


/*
 * Unmaps the DMA mappings.
 * FIXME: Is this a NoOp on x86? Also
 * FIXME: What happens if this one is called and a pending blit has previously done
 * the same DMA mappings?
 */

#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: via_dmablit.c,v 1.9 2021/12/18 23:45:44 riastradh Exp $");

#include <linux/pagemap.h>
#include <linux/pci.h>
#include <linux/slab.h>
#include <linux/vmalloc.h>

#include <drm/drm_device.h>
#include <drm/via_drm.h>

#include "via_dmablit.h"
#include "via_drv.h"

#define VIA_PGDN(x)	     (((unsigned long)(x)) & PAGE_MASK)
#define VIA_PGOFF(x)	    (((unsigned long)(x)) & ~PAGE_MASK)
#define VIA_PFN(x)	      ((unsigned long)(x) >> PAGE_SHIFT)

typedef struct _drm_via_descriptor {
	uint32_t mem_addr;
	uint32_t dev_addr;
	uint32_t size;
	uint32_t next;
} drm_via_descriptor_t;


/*
 * Unmap a DMA mapping.
 */



static void
via_unmap_blit_from_device(struct drm_device *dev, struct pci_dev *pdev,
    drm_via_sg_info_t *vsg)
{
#ifdef __NetBSD__
	bus_dmamap_unload(dev->dmat, vsg->dmamap);
#else
	int num_desc = vsg->num_desc;
	unsigned cur_descriptor_page = num_desc / vsg->descriptors_per_page;
	unsigned descriptor_this_page = num_desc % vsg->descriptors_per_page;
	drm_via_descriptor_t *desc_ptr = vsg->desc_pages[cur_descriptor_page] +
		descriptor_this_page;
	dma_addr_t next = vsg->chain_start;

	while (num_desc--) {
		if (descriptor_this_page-- == 0) {
			cur_descriptor_page--;
			descriptor_this_page = vsg->descriptors_per_page - 1;
			desc_ptr = vsg->desc_pages[cur_descriptor_page] +
				descriptor_this_page;
		}
		dma_unmap_single(&pdev->dev, next, sizeof(*desc_ptr), DMA_TO_DEVICE);
		dma_unmap_page(&pdev->dev, desc_ptr->mem_addr, desc_ptr->size, vsg->direction);
		next = (dma_addr_t) desc_ptr->next;
		desc_ptr--;
	}
#endif
}

/*
 * If mode = 0, count how many descriptors are needed.
 * If mode = 1, Map the DMA pages for the device, put together and map also the descriptors.
 * Descriptors are run in reverse order by the hardware because we are not allowed to update the
 * 'next' field without syncing calls when the descriptor is already mapped.
 */

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)
{
	unsigned cur_descriptor_page = 0;
	unsigned num_descriptors_this_page = 0;
	unsigned char *mem_addr = xfer->mem_addr;
	unsigned char *cur_mem;
#ifndef __NetBSD__
	unsigned char *first_addr = (unsigned char *)VIA_PGDN(mem_addr);
#endif
	uint32_t fb_addr = xfer->fb_addr;
	uint32_t cur_fb;
	unsigned long line_len;
	unsigned remaining_len;
	int num_desc = 0;
	int cur_line;
	dma_addr_t next = 0 | VIA_DMA_DPR_EC;
	drm_via_descriptor_t *desc_ptr = NULL;

	if (mode == 1)
		desc_ptr = vsg->desc_pages[cur_descriptor_page];

	for (cur_line = 0; cur_line < xfer->num_lines; ++cur_line) {

		line_len = xfer->line_length;
		cur_fb = fb_addr;
		cur_mem = mem_addr;

		while (line_len > 0) {

			remaining_len = min(PAGE_SIZE-VIA_PGOFF(cur_mem), line_len);
			line_len -= remaining_len;

			if (mode == 1) {
#ifdef __NetBSD__
				const vaddr_t cur_va = (vaddr_t)cur_mem;
				const bus_dma_segment_t *const seg =
				    &vsg->dmamap->dm_segs[atop(cur_va)];
				desc_ptr->mem_addr =
				    seg->ds_addr + trunc_page(cur_va);
#else
				desc_ptr->mem_addr =
					dma_map_page(&pdev->dev,
						     vsg->pages[VIA_PFN(cur_mem) -
								VIA_PFN(first_addr)],
						     VIA_PGOFF(cur_mem), remaining_len,
						     vsg->direction);
#endif
				desc_ptr->dev_addr = cur_fb;

				desc_ptr->size = remaining_len;
				desc_ptr->next = (uint32_t) next;
#ifdef __NetBSD__
				next = vsg->desc_dmamap
				    ->dm_segs[cur_descriptor_page].ds_addr
				    + num_descriptors_this_page;
#else
				next = dma_map_single(&pdev->dev, desc_ptr, sizeof(*desc_ptr),
						      DMA_TO_DEVICE);
#endif
				desc_ptr++;
				if (++num_descriptors_this_page >= vsg->descriptors_per_page) {
					num_descriptors_this_page = 0;
					desc_ptr = vsg->desc_pages[++cur_descriptor_page];
				}
			}

			num_desc++;
			cur_mem += remaining_len;
			cur_fb += remaining_len;
		}

		mem_addr += xfer->mem_stride;
		fb_addr += xfer->fb_stride;
	}

	if (mode == 1) {
		vsg->chain_start = next;
		vsg->state = dr_via_device_mapped;
	}
	vsg->num_desc = num_desc;
}

/*
 * Function that frees up all resources for a blit. It is usable even if the
 * blit info has only been partially built as long as the status enum is consistent
 * with the actual status of the used resources.
 */


static void
via_free_sg_info(struct drm_device *dev, struct pci_dev *pdev,
    drm_via_sg_info_t *vsg)
{
	int i;

	switch (vsg->state) {
	case dr_via_device_mapped:
		via_unmap_blit_from_device(pdev, vsg);
		/* fall through */
	case dr_via_desc_pages_alloc:
#ifdef __NetBSD__
		bus_dmamap_unload(dev->dmat, vsg->desc_dmamap);
		bus_dmamap_destroy(dev->dmat, vsg->desc_dmamap);
		bus_dmamem_unmap(dev->dmat, vsg->desc_kva,
		    vsg->num_desc_pages << PAGE_SHIFT);
		bus_dmamem_free(dev->dmat, vsg->desc_segs, vsg->num_desc_segs);
		kfree(vsg->desc_segs);
#else
		for (i = 0; i < vsg->num_desc_pages; ++i) {
			if (vsg->desc_pages[i] != NULL)
				free_page((unsigned long)vsg->desc_pages[i]);
		}
#endif
		kfree(vsg->desc_pages);
		/* fall through */
	case dr_via_pages_locked:
		unpin_user_pages_dirty_lock(vsg->pages, vsg->num_pages,
					   (vsg->direction == DMA_FROM_DEVICE));
		/* fall through */
	case dr_via_pages_alloc:
#ifdef __NetBSD__
		bus_dmamap_destroy(dev->dmat, vsg->dmamap);
#else
		vfree(vsg->pages);
#endif
		/* fall through */
	default:
		vsg->state = dr_via_sg_init;
	}
	vsg->free_on_sequence = 0;
}

/*
 * Fire a blit engine.
 */

static void
via_fire_dmablit(struct drm_device *dev, drm_via_sg_info_t *vsg, int engine)
{
	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;

	via_write(dev_priv, VIA_PCI_DMA_MAR0 + engine*0x10, 0);
	via_write(dev_priv, VIA_PCI_DMA_DAR0 + engine*0x10, 0);
	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DD | VIA_DMA_CSR_TD |
		  VIA_DMA_CSR_DE);
	via_write(dev_priv, VIA_PCI_DMA_MR0  + engine*0x04, VIA_DMA_MR_CM | VIA_DMA_MR_TDIE);
	via_write(dev_priv, VIA_PCI_DMA_BCR0 + engine*0x10, 0);
	via_write(dev_priv, VIA_PCI_DMA_DPR0 + engine*0x10, vsg->chain_start);
	wmb();
	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_DE | VIA_DMA_CSR_TS);
	via_read(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04);
}

/*
 * Obtain a page pointer array and lock all pages into system memory. A segmentation violation will
 * occur here if the calling user does not have access to the submitted address.
 */

static int
via_lock_all_dma_pages(struct drm_device *dev, drm_via_sg_info_t *vsg,
    drm_via_dmablit_t *xfer)
{
	int ret;
#ifdef __NetBSD__
	const bus_size_t nbytes = roundup2(xfer->num_lines * xfer->mem_stride,
	    PAGE_SIZE);
	const bus_size_t npages = nbytes >> PAGE_SHIFT;
	struct iovec iov = {
		.iov_base	= xfer->mem_addr,
		.iov_len	= nbytes,
	};
	struct uio uio = {
		.uio_iov	= &iov,
		.uio_iovcnt	= 1,
		.uio_offset	= 0,
		.uio_resid	= nbytes,
		.uio_rw		= xfer->to_fb ? UIO_WRITE : UIO_READ,
		.uio_vmspace	= curproc->p_vmspace,
	};

	/*
	 * XXX Lock out anyone else from doing this?  Add a
	 * dr_via_pages_loading state?  Just rely on the giant lock?
	 */
	/* XXX errno NetBSD->Linux */
	ret = -bus_dmamap_create(dev->dmat, nbytes, npages, nbytes, PAGE_SIZE,
	    BUS_DMA_WAITOK, &vsg->dmamap);
	if (ret) {
		DRM_ERROR("bus_dmamap_create failed: %d\n", ret);
		return ret;
	}
	ret = -bus_dmamap_load_uio(dev->dmat, vsg->dmamap, &uio,
	    BUS_DMA_WAITOK | (xfer->to_fb? BUS_DMA_WRITE : BUS_DMA_READ));
	if (ret) {
		DRM_ERROR("bus_dmamap_load failed: %d\n", ret);
		bus_dmamap_destroy(dev->dmat, vsg->dmamap);
		return ret;
	}
	vsg->num_pages = npages;
#else
	unsigned long first_pfn = VIA_PFN(xfer->mem_addr);
	vsg->num_pages = VIA_PFN(xfer->mem_addr + (xfer->num_lines * xfer->mem_stride - 1)) -
		first_pfn + 1;

	vsg->pages = vzalloc(array_size(sizeof(struct page *), vsg->num_pages));
	if (NULL == vsg->pages)
		return -ENOMEM;
	ret = pin_user_pages_fast((unsigned long)xfer->mem_addr,
			vsg->num_pages,
			vsg->direction == DMA_FROM_DEVICE ? FOLL_WRITE : 0,
			vsg->pages);
	if (ret != vsg->num_pages) {
		if (ret < 0)
			return ret;
		vsg->state = dr_via_pages_locked;
		return -EINVAL;
	}
#endif
	vsg->state = dr_via_pages_locked;
	DRM_DEBUG("DMA pages locked\n");
	return 0;
}

/*
 * Allocate DMA capable memory for the blit descriptor chain, and an array that keeps track of the
 * pages we allocate. We don't want to use kmalloc for the descriptor chain because it may be
 * quite large for some blits, and pages don't need to be contiguous.
 */

static int
via_alloc_desc_pages(struct drm_device *dev, drm_via_sg_info_t *vsg)
{
	int i;
#ifdef __NetBSD__
	int ret;
#endif

	vsg->descriptors_per_page = PAGE_SIZE / sizeof(drm_via_descriptor_t);
	vsg->num_desc_pages = (vsg->num_desc + vsg->descriptors_per_page - 1) /
		vsg->descriptors_per_page;

	if (NULL ==  (vsg->desc_pages = kcalloc(vsg->num_desc_pages, sizeof(void *), GFP_KERNEL)))
		return -ENOMEM;

#ifdef __NetBSD__
	vsg->desc_segs = kcalloc(vsg->num_desc_pages, sizeof(*vsg->desc_segs),
	    GFP_KERNEL);
	if (vsg->desc_segs == NULL) {
		kfree(vsg->desc_pages);
		return -ENOMEM;
	}
	/* XXX errno NetBSD->Linux */
	ret = -bus_dmamem_alloc(dev->dmat, vsg->num_desc_pages << PAGE_SHIFT,
	    PAGE_SIZE, 0, vsg->desc_segs, vsg->num_pages, &vsg->num_desc_segs,
	    BUS_DMA_WAITOK);
	if (ret) {
		kfree(vsg->desc_segs);
		kfree(vsg->desc_pages);
		return -ENOMEM;
	}
	/* XXX No nice way to scatter/gather map bus_dmamem.  */
	/* XXX errno NetBSD->Linux */
	ret = -bus_dmamem_map(dev->dmat, vsg->desc_segs, vsg->num_desc_segs,
	    vsg->num_desc_pages << PAGE_SHIFT, &vsg->desc_kva, BUS_DMA_WAITOK);
	if (ret) {
		bus_dmamem_free(dev->dmat, vsg->desc_segs, vsg->num_desc_segs);
		kfree(vsg->desc_segs);
		kfree(vsg->desc_pages);
		return -ENOMEM;
	}
	/* XXX errno NetBSD->Linux */
	ret = -bus_dmamap_create(dev->dmat, vsg->num_desc_pages << PAGE_SHIFT,
	    vsg->num_desc_pages, PAGE_SIZE, 0, BUS_DMA_WAITOK,
	    &vsg->desc_dmamap);
	if (ret) {
		bus_dmamem_unmap(dev->dmat, vsg->desc_kva,
		    vsg->num_desc_pages << PAGE_SHIFT);
		bus_dmamem_free(dev->dmat, vsg->desc_segs, vsg->num_desc_segs);
		kfree(vsg->desc_segs);
		kfree(vsg->desc_pages);
		return -ENOMEM;
	}
	ret = -bus_dmamap_load(dev->dmat, vsg->desc_dmamap, vsg->desc_kva,
	    vsg->num_desc_pages << PAGE_SHIFT, NULL, BUS_DMA_WAITOK);
	if (ret) {
		bus_dmamap_destroy(dev->dmat, vsg->desc_dmamap);
		bus_dmamem_unmap(dev->dmat, vsg->desc_kva,
		    vsg->num_desc_pages << PAGE_SHIFT);
		bus_dmamem_free(dev->dmat, vsg->desc_segs, vsg->num_desc_segs);
		kfree(vsg->desc_segs);
		kfree(vsg->desc_pages);
		return -ENOMEM;
	}
	for (i = 0; i < vsg->num_desc_pages; i++)
		vsg->desc_pages[i] = (void *)
		    ((char *)vsg->desc_kva + (i * PAGE_SIZE));
	vsg->state = dr_via_desc_pages_alloc;
#else
	vsg->state = dr_via_desc_pages_alloc;
	for (i = 0; i < vsg->num_desc_pages; ++i) {
		if (NULL == (vsg->desc_pages[i] =
			     (drm_via_descriptor_t *) __get_free_page(GFP_KERNEL)))
			return -ENOMEM;
	}
#endif
	DRM_DEBUG("Allocated %d pages for %d descriptors.\n", vsg->num_desc_pages,
		  vsg->num_desc);
	return 0;
}

static void
via_abort_dmablit(struct drm_device *dev, int engine)
{
	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;

	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TA);
}

static void
via_dmablit_engine_off(struct drm_device *dev, int engine)
{
	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;

	via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04, VIA_DMA_CSR_TD | VIA_DMA_CSR_DD);
}



/*
 * The dmablit part of the IRQ handler. Trying to do only reasonably fast things here.
 * The rest, like unmapping and freeing memory for done blits is done in a separate workqueue
 * task. Basically the task of the interrupt handler is to submit a new blit to the engine, while
 * the workqueue task takes care of processing associated with the old blit.
 */

void
via_dmablit_handler(struct drm_device *dev, int engine, int from_irq)
{
	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
	drm_via_blitq_t *blitq = dev_priv->blit_queues + engine;
	int cur;
	int done_transfer;
	unsigned long irqsave = 0;
	uint32_t status = 0;

	DRM_DEBUG("DMA blit handler called. engine = %d, from_irq = %d, blitq = 0x%lx\n",
		  engine, from_irq, (unsigned long) blitq);

	if (from_irq)
		spin_lock(&blitq->blit_lock);
	else
		spin_lock_irqsave(&blitq->blit_lock, irqsave);

	done_transfer = blitq->is_active &&
	  ((status = via_read(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04)) & VIA_DMA_CSR_TD);
	done_transfer = done_transfer || (blitq->aborting && !(status & VIA_DMA_CSR_DE));

	cur = blitq->cur;
	if (done_transfer) {

		blitq->blits[cur]->aborted = blitq->aborting;
		blitq->done_blit_handle++;
#ifdef __NetBSD__
		DRM_SPIN_WAKEUP_ALL(&blitq->blit_queue[cur],
		    &blitq->blit_lock);
#else
		wake_up(blitq->blit_queue + cur);
#endif

		cur++;
		if (cur >= VIA_NUM_BLIT_SLOTS)
			cur = 0;
		blitq->cur = cur;

		/*
		 * Clear transfer done flag.
		 */

		via_write(dev_priv, VIA_PCI_DMA_CSR0 + engine*0x04,  VIA_DMA_CSR_TD);

		blitq->is_active = 0;
		blitq->aborting = 0;
		schedule_work(&blitq->wq);

	} else if (blitq->is_active && time_after_eq(jiffies, blitq->end)) {

		/*
		 * Abort transfer after one second.
		 */

		via_abort_dmablit(dev, engine);
		blitq->aborting = 1;
		blitq->end = jiffies + HZ;
	}

	if (!blitq->is_active) {
		if (blitq->num_outstanding) {
			via_fire_dmablit(dev, blitq->blits[cur], engine);
			blitq->is_active = 1;
			blitq->cur = cur;
			blitq->num_outstanding--;
			blitq->end = jiffies + HZ;
			if (!timer_pending(&blitq->poll_timer))
				mod_timer(&blitq->poll_timer, jiffies + 1);
		} else {
			if (timer_pending(&blitq->poll_timer))
				del_timer(&blitq->poll_timer);
			via_dmablit_engine_off(dev, engine);
		}
	}

	if (from_irq)
		spin_unlock(&blitq->blit_lock);
	else
		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
}



/*
 * Check whether this blit is still active, performing necessary locking.
 */

static int
#ifdef __NetBSD__
via_dmablit_active(drm_via_blitq_t *blitq, int engine, uint32_t handle, drm_waitqueue_t **queue)
#else
via_dmablit_active(drm_via_blitq_t *blitq, int engine, uint32_t handle, wait_queue_head_t **queue)
#endif
{
#ifndef __NetBSD__
	unsigned long irqsave;
#endif
	uint32_t slot;
	int active;

#ifndef __NetBSD__
	spin_lock_irqsave(&blitq->blit_lock, irqsave);
#endif

	/*
	 * Allow for handle wraparounds.
	 */

	active = ((blitq->done_blit_handle - handle) > (1 << 23)) &&
		((blitq->cur_blit_handle - handle) <= (1 << 23));

	if (queue && active) {
		slot = handle - blitq->done_blit_handle + blitq->cur - 1;
		if (slot >= VIA_NUM_BLIT_SLOTS)
			slot -= VIA_NUM_BLIT_SLOTS;
		*queue = blitq->blit_queue + slot;
	}

#ifndef __NetBSD__
	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
#endif

	return active;
}

/*
 * Sync. Wait for at least three seconds for the blit to be performed.
 */

static int
via_dmablit_sync(struct drm_device *dev, uint32_t handle, int engine)
{

	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
	drm_via_blitq_t *blitq = dev_priv->blit_queues + engine;
#ifdef __NetBSD__
	drm_waitqueue_t *queue;
#else
	wait_queue_head_t *queue;
#endif
	int ret = 0;

#ifdef __NetBSD__
	spin_lock(&blitq->blit_lock);
	if (via_dmablit_active(blitq, engine, handle, &queue)) {
		DRM_SPIN_WAIT_ON(ret, queue, &blitq->blit_lock, 3*HZ,
		    !via_dmablit_active(blitq, engine, handle, NULL));
	}
	spin_unlock(&blitq->blit_lock);
#else
	if (via_dmablit_active(blitq, engine, handle, &queue)) {
		VIA_WAIT_ON(ret, *queue, 3 * HZ,
			    !via_dmablit_active(blitq, engine, handle, NULL));
	}
#endif
	DRM_DEBUG("DMA blit sync handle 0x%x engine %d returned %d\n",
		  handle, engine, ret);

	return ret;
}


/*
 * A timer that regularly polls the blit engine in cases where we don't have interrupts:
 * a) Broken hardware (typically those that don't have any video capture facility).
 * b) Blit abort. The hardware doesn't send an interrupt when a blit is aborted.
 * The timer and hardware IRQ's can and do work in parallel. If the hardware has
 * irqs, it will shorten the latency somewhat.
 */



static void
via_dmablit_timer(struct timer_list *t)
{
	drm_via_blitq_t *blitq = from_timer(blitq, t, poll_timer);
	struct drm_device *dev = blitq->dev;
	int engine = (int)
		(blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues);

	DRM_DEBUG("Polling timer called for engine %d, jiffies %lu\n", engine,
		  (unsigned long) jiffies);

	via_dmablit_handler(dev, engine, 0);

	if (!timer_pending(&blitq->poll_timer)) {
		mod_timer(&blitq->poll_timer, jiffies + 1);

	       /*
		* Rerun handler to delete timer if engines are off, and
		* to shorten abort latency. This is a little nasty.
		*/

	       via_dmablit_handler(dev, engine, 0);

	}
}




/*
 * Workqueue task that frees data and mappings associated with a blit.
 * Also wakes up waiting processes. Each of these tasks handles one
 * blit engine only and may not be called on each interrupt.
 */


static void
via_dmablit_workqueue(struct work_struct *work)
{
	drm_via_blitq_t *blitq = container_of(work, drm_via_blitq_t, wq);
	struct drm_device *dev = blitq->dev;
	unsigned long irqsave;
	drm_via_sg_info_t *cur_sg;
	int cur_released;


	DRM_DEBUG("Workqueue task called for blit engine %ld\n", (unsigned long)
		  (blitq - ((drm_via_private_t *)dev->dev_private)->blit_queues));

	spin_lock_irqsave(&blitq->blit_lock, irqsave);

	while (blitq->serviced != blitq->cur) {

		cur_released = blitq->serviced++;

		DRM_DEBUG("Releasing blit slot %d\n", cur_released);

		if (blitq->serviced >= VIA_NUM_BLIT_SLOTS)
			blitq->serviced = 0;

		cur_sg = blitq->blits[cur_released];
		blitq->num_free++;

#ifdef __NetBSD__
		DRM_SPIN_WAKEUP_ONE(&blitq->busy_queue, &blitq->blit_lock);
#endif

		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);

#ifndef __NetBSD__
		wake_up(&blitq->busy_queue);
#endif

#ifdef __NetBSD__
		/* Transfer completed.  Sync it.  */
		bus_dmamap_sync(dev->dmat, cur_sg->dmamap, 0,
		    cur_sg->num_pages << PAGE_SHIFT,
		    (cur_sg->direction == DMA_FROM_DEVICE
			? BUS_DMASYNC_POSTREAD
			: BUS_DMASYNC_POSTWRITE));
#endif
		via_free_sg_info(dev, dev->pdev, cur_sg);
		kfree(cur_sg);

		spin_lock_irqsave(&blitq->blit_lock, irqsave);
	}

	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
}


/*
 * Init all blit engines. Currently we use two, but some hardware have 4.
 */


void
via_init_dmablit(struct drm_device *dev)
{
	int i, j;
	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
	drm_via_blitq_t *blitq;

	pci_set_master(dev->pdev);

	for (i = 0; i < VIA_NUM_BLIT_ENGINES; ++i) {
		blitq = dev_priv->blit_queues + i;
		blitq->dev = dev;
		blitq->cur_blit_handle = 0;
		blitq->done_blit_handle = 0;
		blitq->head = 0;
		blitq->cur = 0;
		blitq->serviced = 0;
		blitq->num_free = VIA_NUM_BLIT_SLOTS - 1;
		blitq->num_outstanding = 0;
		blitq->is_active = 0;
		blitq->aborting = 0;
		spin_lock_init(&blitq->blit_lock);
#ifdef __NetBSD__
		for (j = 0; j < VIA_NUM_BLIT_SLOTS; ++j)
			DRM_INIT_WAITQUEUE(blitq->blit_queue + j, "viablt");
		DRM_INIT_WAITQUEUE(&blitq->busy_queue, "viabusy");
#else
		for (j = 0; j < VIA_NUM_BLIT_SLOTS; ++j)
			init_waitqueue_head(blitq->blit_queue + j);
		init_waitqueue_head(&blitq->busy_queue);
#endif
		INIT_WORK(&blitq->wq, via_dmablit_workqueue);
		timer_setup(&blitq->poll_timer, via_dmablit_timer, 0);
	}
}

/*
 * Build all info and do all mappings required for a blit.
 */


static int
via_build_sg_info(struct drm_device *dev, drm_via_sg_info_t *vsg, drm_via_dmablit_t *xfer)
{
	int draw = xfer->to_fb;
	int ret = 0;

	vsg->direction = (draw) ? DMA_TO_DEVICE : DMA_FROM_DEVICE;

	vsg->state = dr_via_sg_init;

	if (xfer->num_lines <= 0 || xfer->line_length <= 0) {
		DRM_ERROR("Zero size bitblt.\n");
		return -EINVAL;
	}

	/*
	 * Below check is a driver limitation, not a hardware one. We
	 * don't want to lock unused pages, and don't want to incoporate the
	 * extra logic of avoiding them. Make sure there are no.
	 * (Not a big limitation anyway.)
	 */

	if ((xfer->mem_stride - xfer->line_length) > 2*PAGE_SIZE) {
		DRM_ERROR("Too large system memory stride. Stride: %d, "
			  "Length: %d\n", xfer->mem_stride, xfer->line_length);
		return -EINVAL;
	}

	if ((xfer->mem_stride == xfer->line_length) &&
	   (xfer->fb_stride == xfer->line_length)) {
		xfer->mem_stride *= xfer->num_lines;
		xfer->line_length = xfer->mem_stride;
		xfer->fb_stride = xfer->mem_stride;
		xfer->num_lines = 1;
	}

	/*
	 * Don't lock an arbitrary large number of pages, since that causes a
	 * DOS security hole.
	 */

	if (xfer->num_lines > 2048 || (xfer->num_lines*xfer->mem_stride > (2048*2048*4))) {
		DRM_ERROR("Too large PCI DMA bitblt.\n");
		return -EINVAL;
	}

	/*
	 * we allow a negative fb stride to allow flipping of images in
	 * transfer.
	 */

	if (xfer->mem_stride < xfer->line_length ||
		abs(xfer->fb_stride) < xfer->line_length) {
		DRM_ERROR("Invalid frame-buffer / memory stride.\n");
		return -EINVAL;
	}

	/*
	 * A hardware bug seems to be worked around if system memory addresses start on
	 * 16 byte boundaries. This seems a bit restrictive however. VIA is contacted
	 * about this. Meanwhile, impose the following restrictions:
	 */

#ifdef VIA_BUGFREE
	if ((((unsigned long)xfer->mem_addr & 3) != ((unsigned long)xfer->fb_addr & 3)) ||
	    ((xfer->num_lines > 1) && ((xfer->mem_stride & 3) != (xfer->fb_stride & 3)))) {
		DRM_ERROR("Invalid DRM bitblt alignment.\n");
		return -EINVAL;
	}
#else
	if ((((unsigned long)xfer->mem_addr & 15) ||
	      ((unsigned long)xfer->fb_addr & 3)) ||
	   ((xfer->num_lines > 1) &&
	   ((xfer->mem_stride & 15) || (xfer->fb_stride & 3)))) {
		DRM_ERROR("Invalid DRM bitblt alignment.\n");
		return -EINVAL;
	}
#endif

	if (0 != (ret = via_lock_all_dma_pages(dev, vsg, xfer))) {
		DRM_ERROR("Could not lock DMA pages.\n");
		via_free_sg_info(dev, dev->pdev, vsg);
		return ret;
	}

	via_map_blit_for_device(dev->pdev, xfer, vsg, 0);
	if (0 != (ret = via_alloc_desc_pages(dev, vsg))) {
		DRM_ERROR("Could not allocate DMA descriptor pages.\n");
		via_free_sg_info(dev, dev->pdev, vsg);
		return ret;
	}
	via_map_blit_for_device(dev->pdev, xfer, vsg, 1);

	return 0;
}


/*
 * Reserve one free slot in the blit queue. Will wait for one second for one
 * to become available. Otherwise -EBUSY is returned.
 */

static int
via_dmablit_grab_slot(drm_via_blitq_t *blitq, int engine)
{
	int ret = 0;
	unsigned long irqsave;

	DRM_DEBUG("Num free is %d\n", blitq->num_free);
	spin_lock_irqsave(&blitq->blit_lock, irqsave);
	while (blitq->num_free == 0) {
#ifdef __NetBSD__
		DRM_SPIN_WAIT_ON(ret, &blitq->busy_queue, &blitq->blit_lock,
		    HZ,
		    blitq->num_free > 0);
		/* Map -EINTR to -EAGAIN.  */
		if (ret == -EINTR)
			ret = -EAGAIN;
		/* Bail on failure.  */
		if (ret) {
			spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
			return ret;
		}
#else
		spin_unlock_irqrestore(&blitq->blit_lock, irqsave);

		VIA_WAIT_ON(ret, blitq->busy_queue, HZ, blitq->num_free > 0);
		if (ret)
			return (-EINTR == ret) ? -EAGAIN : ret;

		spin_lock_irqsave(&blitq->blit_lock, irqsave);
#endif
	}

	blitq->num_free--;
	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);

	return 0;
}

/*
 * Hand back a free slot if we changed our mind.
 */

static void
via_dmablit_release_slot(drm_via_blitq_t *blitq)
{
	unsigned long irqsave;

	spin_lock_irqsave(&blitq->blit_lock, irqsave);
	blitq->num_free++;
#ifdef __NetBSD__
	DRM_SPIN_WAKEUP_ONE(&blitq->busy_queue, &blitq->blit_lock);
#endif
	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
#ifndef __NetBSD__
	wake_up(&blitq->busy_queue);
#endif
}

/*
 * Grab a free slot. Build blit info and queue a blit.
 */


static int
via_dmablit(struct drm_device *dev, drm_via_dmablit_t *xfer)
{
	drm_via_private_t *dev_priv = (drm_via_private_t *)dev->dev_private;
	drm_via_sg_info_t *vsg;
	drm_via_blitq_t *blitq;
	int ret;
	int engine;
	unsigned long irqsave;

	if (dev_priv == NULL) {
		DRM_ERROR("Called without initialization.\n");
		return -EINVAL;
	}

	engine = (xfer->to_fb) ? 0 : 1;
	blitq = dev_priv->blit_queues + engine;
	if (0 != (ret = via_dmablit_grab_slot(blitq, engine)))
		return ret;
	if (NULL == (vsg = kmalloc(sizeof(*vsg), GFP_KERNEL))) {
		via_dmablit_release_slot(blitq);
		return -ENOMEM;
	}
	if (0 != (ret = via_build_sg_info(dev, vsg, xfer))) {
		via_dmablit_release_slot(blitq);
		kfree(vsg);
		return ret;
	}
#ifdef __NetBSD__
	/* Prepare to begin a DMA transfer.  */
	bus_dmamap_sync(dev->dmat, vsg->dmamap, 0,
	    vsg->num_pages << PAGE_SHIFT,
	    (vsg->direction == DMA_FROM_DEVICE
		? BUS_DMASYNC_PREREAD
		: BUS_DMASYNC_PREWRITE));
#endif
	spin_lock_irqsave(&blitq->blit_lock, irqsave);

	blitq->blits[blitq->head++] = vsg;
	if (blitq->head >= VIA_NUM_BLIT_SLOTS)
		blitq->head = 0;
	blitq->num_outstanding++;
	xfer->sync.sync_handle = ++blitq->cur_blit_handle;

	spin_unlock_irqrestore(&blitq->blit_lock, irqsave);
	xfer->sync.engine = engine;

	via_dmablit_handler(dev, engine, 0);

	return 0;
}

/*
 * Sync on a previously submitted blit. Note that the X server use signals extensively, and
 * that there is a very big probability that this IOCTL will be interrupted by a signal. In that
 * case it returns with -EAGAIN for the signal to be delivered.
 * The caller should then reissue the IOCTL. This is similar to what is being done for drmGetLock().
 */

int
via_dma_blit_sync(struct drm_device *dev, void *data, struct drm_file *file_priv)
{
	drm_via_blitsync_t *sync = data;
	int err;

	if (sync->engine >= VIA_NUM_BLIT_ENGINES)
		return -EINVAL;

	err = via_dmablit_sync(dev, sync->sync_handle, sync->engine);

	if (-EINTR == err)
		err = -EAGAIN;

	return err;
}


/*
 * Queue a blit and hand back a handle to be used for sync. This IOCTL may be interrupted by a signal
 * while waiting for a free slot in the blit queue. In that case it returns with -EAGAIN and should
 * be reissued. See the above IOCTL code.
 */

int
via_dma_blit(struct drm_device *dev, void *data, struct drm_file *file_priv)
{
	drm_via_dmablit_t *xfer = data;
	int err;

	err = via_dmablit(dev, xfer);

	return err;
}