pci_iommu.c

讲述linux的初始化过程

	pci_unmap_single(pdev, dma_addr, size, PCI_DMA_BIDIRECTIONAL);
	free_pages((unsigned long)cpu_addr, get_order(size));

	DBGA2("pci_free_consistent: [%x,%lx] from %p\n",
	      dma_addr, size, __builtin_return_address(0));
}


/* Classify the elements of the scatterlist.  Write dma_address
   of each element with:
	0   : Followers all physically adjacent.
	1   : Followers all virtually adjacent.
	-1  : Not leader, physically adjacent to previous.
	-2  : Not leader, virtually adjacent to previous.
   Write dma_length of each leader with the combined lengths of
   the mergable followers.  */

static inline void
sg_classify(struct scatterlist *sg, struct scatterlist *end, int virt_ok)
{
	unsigned long next_vaddr;
	struct scatterlist *leader;
	long leader_flag, leader_length;

	leader = sg;
	leader_flag = 0;
	leader_length = leader->length;
	next_vaddr = (unsigned long)leader->address + leader_length;

	for (++sg; sg < end; ++sg) {
		unsigned long addr, len;
		addr = (unsigned long) sg->address;
		len = sg->length;

		if (next_vaddr == addr) {
			sg->dma_address = -1;
			leader_length += len;
		} else if (((next_vaddr | addr) & ~PAGE_MASK) == 0 && virt_ok) {
			sg->dma_address = -2;
			leader_flag = 1;
			leader_length += len;
		} else {
			leader->dma_address = leader_flag;
			leader->dma_length = leader_length;
			leader = sg;
			leader_flag = 0;
			leader_length = len;
		}

		next_vaddr = addr + len;
	}

	leader->dma_address = leader_flag;
	leader->dma_length = leader_length;
}

/* Given a scatterlist leader, choose an allocation method and fill
   in the blanks.  */

static inline int
sg_fill(struct scatterlist *leader, struct scatterlist *end,
	struct scatterlist *out, struct pci_iommu_arena *arena,
	dma_addr_t max_dma)
{
	unsigned long paddr = virt_to_phys(leader->address);
	long size = leader->dma_length;
	struct scatterlist *sg;
	unsigned long *ptes;
	long npages, dma_ofs, i;

#if !DEBUG_NODIRECT
	/* If everything is physically contiguous, and the addresses
	   fall into the direct-map window, use it.  */
	if (leader->dma_address == 0
	    && paddr + size + __direct_map_base - 1 <= max_dma
	    && paddr + size <= __direct_map_size) {
		out->dma_address = paddr + __direct_map_base;
		out->dma_length = size;

		DBGA("    sg_fill: [%p,%lx] -> direct %x\n",
		     leader->address, size, out->dma_address);

		return 0;
	}
#endif

	/* Otherwise, we'll use the iommu to make the pages virtually
	   contiguous.  */

	paddr &= ~PAGE_MASK;
	npages = calc_npages(paddr + size);
	dma_ofs = iommu_arena_alloc(arena, npages);
	if (dma_ofs < 0)
		return -1;

	out->dma_address = arena->dma_base + dma_ofs*PAGE_SIZE + paddr;
	out->dma_length = size;

	DBGA("    sg_fill: [%p,%lx] -> sg %x np %ld\n",
	     leader->address, size, out->dma_address, npages);

	/* All virtually contiguous.  We need to find the length of each
	   physically contiguous subsegment to fill in the ptes.  */
	ptes = &arena->ptes[dma_ofs];
	sg = leader;
	do {
#if DEBUG_ALLOC > 0
		struct scatterlist *last_sg = sg;
#endif

		size = sg->length;
		paddr = virt_to_phys(sg->address);

		while (sg+1 < end && (int) sg[1].dma_address == -1) {
			size += sg[1].length;
			sg++;
		}

		npages = calc_npages((paddr & ~PAGE_MASK) + size);

		paddr &= PAGE_MASK;
		for (i = 0; i < npages; ++i, paddr += PAGE_SIZE)
			*ptes++ = mk_iommu_pte(paddr);

#if DEBUG_ALLOC > 0
		DBGA("    (%ld) [%p,%x] np %ld\n",
		     last_sg - leader, last_sg->address,
		     last_sg->length, npages);
		while (++last_sg <= sg) {
			DBGA("        (%ld) [%p,%x] cont\n",
			     last_sg - leader, last_sg->address,
			     last_sg->length);
		}
#endif
	} while (++sg < end && (int) sg->dma_address < 0);

	return 1;
}

int
pci_map_sg(struct pci_dev *pdev, struct scatterlist *sg, int nents,
	   int direction)
{
	struct scatterlist *start, *end, *out;
	struct pci_controler *hose;
	struct pci_iommu_arena *arena;
	dma_addr_t max_dma;

	if (direction == PCI_DMA_NONE)
		BUG();

	/* Fast path single entry scatterlists.  */
	if (nents == 1) {
		sg->dma_length = sg->length;
		sg->dma_address
		  = pci_map_single(pdev, sg->address, sg->length, direction);
		return sg->dma_address != 0;
	}

	start = sg;
	end = sg + nents;

	/* First, prepare information about the entries.  */
	sg_classify(sg, end, alpha_mv.mv_pci_tbi != 0);

	/* Second, figure out where we're going to map things.  */
	if (alpha_mv.mv_pci_tbi) {
		hose = pdev ? pdev->sysdata : pci_isa_hose;
		max_dma = pdev ? pdev->dma_mask : 0x00ffffff;
		arena = hose->sg_pci;
		if (!arena || arena->dma_base + arena->size > max_dma)
			arena = hose->sg_isa;
	} else {
		max_dma = -1;
		arena = NULL;
		hose = NULL;
	}

	/* Third, iterate over the scatterlist leaders and allocate
	   dma space as needed.  */
	for (out = sg; sg < end; ++sg) {
		if ((int) sg->dma_address < 0)
			continue;
		if (sg_fill(sg, end, out, arena, max_dma) < 0)
			goto error;
		out++;
	}

	/* Mark the end of the list for pci_unmap_sg.  */
	if (out < end)
		out->dma_length = 0;

	if (out - start == 0)
		printk(KERN_INFO "pci_map_sg failed: no entries?\n");
	DBGA("pci_map_sg: %ld entries\n", out - start);

	return out - start;

error:
	printk(KERN_INFO "pci_map_sg failed: "
	       "could not allocate dma page tables\n");

	/* Some allocation failed while mapping the scatterlist
	   entries.  Unmap them now.  */
	if (out > start)
		pci_unmap_sg(pdev, start, out - start, direction);
	return 0;
}


/* Unmap a set of streaming mode DMA translations.  Again, cpu read
   rules concerning calls here are the same as for pci_unmap_single()
   above.  */

void
pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sg, int nents,
	     int direction)
{
	struct pci_controler *hose;
	struct pci_iommu_arena *arena;
	struct scatterlist *end;
	dma_addr_t max_dma;
	dma_addr_t fbeg, fend;

	if (direction == PCI_DMA_NONE)
		BUG();

	if (! alpha_mv.mv_pci_tbi)
		return;

	hose = pdev ? pdev->sysdata : pci_isa_hose;
	max_dma = pdev ? pdev->dma_mask : 0x00ffffff;
	arena = hose->sg_pci;
	if (!arena || arena->dma_base + arena->size > max_dma)
		arena = hose->sg_isa;

	fbeg = -1, fend = 0;
	for (end = sg + nents; sg < end; ++sg) {
		unsigned long addr, size;
		long npages, ofs;
		dma_addr_t tend;

		addr = sg->dma_address;
		size = sg->dma_length;
		if (!size)
			break;

#if !DEBUG_NODIRECT
		if (addr >= __direct_map_base
		    && addr < __direct_map_base + __direct_map_size) {
			/* Nothing to do.  */
			DBGA("    (%ld) direct [%lx,%lx]\n",
			      sg - end + nents, addr, size);
			continue;
		}
#endif

		DBGA("    (%ld) sg [%lx,%lx]\n",
		     sg - end + nents, addr, size);

		npages = calc_npages((addr & ~PAGE_MASK) + size);
		ofs = (addr - arena->dma_base) >> PAGE_SHIFT;
		iommu_arena_free(arena, ofs, npages);

		tend = addr + size - 1;
		if (fbeg > addr) fbeg = addr;
		if (fend < tend) fend = tend;
	}

        /*
	   If we're freeing ptes above the `next_entry' pointer (they
           may have snuck back into the TLB since the last wrap flush),
           we need to flush the TLB before reallocating the latter.
	*/
	if ((fend - arena->dma_base) >> PAGE_SHIFT >= arena->next_entry)
		alpha_mv.mv_pci_tbi(hose, fbeg, fend);

	DBGA("pci_unmap_sg: %d entries\n", nents - (end - sg));
}

/* Return whether the given PCI device DMA address mask can be
   supported properly.  */

int
pci_dma_supported(struct pci_dev *pdev, dma_addr_t mask)
{
	struct pci_controler *hose;
	struct pci_iommu_arena *arena;

#if !DEBUG_NODIRECT
	/* If there exists a direct map, and the mask fits either
	   MAX_DMA_ADDRESS defined such that GFP_DMA does something
	   useful, or the total system memory as shifted by the
	   map base.  */
	if (__direct_map_size != 0
	    && (__direct_map_base + MAX_DMA_ADDRESS-IDENT_ADDR-1 <= mask
		|| __direct_map_base + (max_low_pfn<<PAGE_SHIFT)-1 <= mask))
		return 1;
#endif

	/* Check that we have a scatter-gather arena that fits.  */
	hose = pdev ? pdev->sysdata : pci_isa_hose;
	arena = hose->sg_isa;
	if (arena && arena->dma_base + arena->size - 1 <= mask)
		return 1;
	arena = hose->sg_pci;
	if (arena && arena->dma_base + arena->size - 1 <= mask)
		return 1;

	return 0;
}