sba_iommu.c

linux-2.6.15.6

			for (; mask ; mask <<= o, bitshiftcnt += o) {
				if(0 == ((*res_ptr) & mask)) {
					*res_ptr |= mask;     /* mark resources busy! */
					pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
					pide <<= 3;	/* convert to bit address */
					pide += bitshiftcnt;
					ioc->res_bitshift = bitshiftcnt + bits_wanted;
					goto found_it;
				}
			}

			bitshiftcnt = 0;
			mask = base_mask;

		}

	} else {
		int qwords, bits, i;
		unsigned long *end;

		qwords = bits_wanted >> 6; /* /64 */
		bits = bits_wanted - (qwords * BITS_PER_LONG);

		end = res_end - qwords;

		for (; res_ptr < end; res_ptr++) {
			for (i = 0 ; i < qwords ; i++) {
				if (res_ptr[i] != 0)
					goto next_ptr;
			}
			if (bits && res_ptr[i] && (__ffs(res_ptr[i]) < bits))
				continue;

			/* Found it, mark it */
			for (i = 0 ; i < qwords ; i++)
				res_ptr[i] = ~0UL;
			res_ptr[i] |= RESMAP_MASK(bits);

			pide = ((unsigned long)res_ptr - (unsigned long)ioc->res_map);
			pide <<= 3;	/* convert to bit address */
			res_ptr += qwords;
			ioc->res_bitshift = bits;
			goto found_it;
next_ptr:
			;
		}
	}

not_found:
	prefetch(ioc->res_map);
	ioc->res_hint = (unsigned long *) ioc->res_map;
	ioc->res_bitshift = 0;
	spin_unlock_irqrestore(&ioc->res_lock, flags);
	return (pide);

found_it:
	ioc->res_hint = res_ptr;
	spin_unlock_irqrestore(&ioc->res_lock, flags);
	return (pide);
}


/**
 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @size: number of bytes to create a mapping for
 *
 * Given a size, find consecutive unmarked and then mark those bits in the
 * resource bit map.
 */
static int
sba_alloc_range(struct ioc *ioc, size_t size)
{
	unsigned int pages_needed = size >> iovp_shift;
#ifdef PDIR_SEARCH_TIMING
	unsigned long itc_start;
#endif
	unsigned long pide;

	ASSERT(pages_needed);
	ASSERT(0 == (size & ~iovp_mask));

#ifdef PDIR_SEARCH_TIMING
	itc_start = ia64_get_itc();
#endif
	/*
	** "seek and ye shall find"...praying never hurts either...
	*/
	pide = sba_search_bitmap(ioc, pages_needed, 1);
	if (unlikely(pide >= (ioc->res_size << 3))) {
		pide = sba_search_bitmap(ioc, pages_needed, 0);
		if (unlikely(pide >= (ioc->res_size << 3))) {
#if DELAYED_RESOURCE_CNT > 0
			unsigned long flags;

			/*
			** With delayed resource freeing, we can give this one more shot.  We're
			** getting close to being in trouble here, so do what we can to make this
			** one count.
			*/
			spin_lock_irqsave(&ioc->saved_lock, flags);
			if (ioc->saved_cnt > 0) {
				struct sba_dma_pair *d;
				int cnt = ioc->saved_cnt;

				d = &(ioc->saved[ioc->saved_cnt - 1]);

				spin_lock(&ioc->res_lock);
				while (cnt--) {
					sba_mark_invalid(ioc, d->iova, d->size);
					sba_free_range(ioc, d->iova, d->size);
					d--;
				}
				ioc->saved_cnt = 0;
				READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
				spin_unlock(&ioc->res_lock);
			}
			spin_unlock_irqrestore(&ioc->saved_lock, flags);

			pide = sba_search_bitmap(ioc, pages_needed, 0);
			if (unlikely(pide >= (ioc->res_size << 3)))
				panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n",
				      ioc->ioc_hpa);
#else
			panic(__FILE__ ": I/O MMU @ %p is out of mapping resources\n",
			      ioc->ioc_hpa);
#endif
		}
	}

#ifdef PDIR_SEARCH_TIMING
	ioc->avg_search[ioc->avg_idx++] = (ia64_get_itc() - itc_start) / pages_needed;
	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
#endif

	prefetchw(&(ioc->pdir_base[pide]));

#ifdef ASSERT_PDIR_SANITY
	/* verify the first enable bit is clear */
	if(0x00 != ((u8 *) ioc->pdir_base)[pide*PDIR_ENTRY_SIZE + 7]) {
		sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
	}
#endif

	DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
		__FUNCTION__, size, pages_needed, pide,
		(uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
		ioc->res_bitshift );

	return (pide);
}


/**
 * sba_free_range - unmark bits in IO PDIR resource bitmap
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @iova: IO virtual address which was previously allocated.
 * @size: number of bytes to create a mapping for
 *
 * clear bits in the ioc's resource map
 */
static SBA_INLINE void
sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
{
	unsigned long iovp = SBA_IOVP(ioc, iova);
	unsigned int pide = PDIR_INDEX(iovp);
	unsigned int ridx = pide >> 3;	/* convert bit to byte address */
	unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
	int bits_not_wanted = size >> iovp_shift;
	unsigned long m;

	/* Round up to power-of-two size: see AR2305 note above */
	bits_not_wanted = 1UL << get_iovp_order(bits_not_wanted << iovp_shift);
	for (; bits_not_wanted > 0 ; res_ptr++) {
		
		if (unlikely(bits_not_wanted > BITS_PER_LONG)) {

			/* these mappings start 64bit aligned */
			*res_ptr = 0UL;
			bits_not_wanted -= BITS_PER_LONG;
			pide += BITS_PER_LONG;

		} else {

			/* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
			m = RESMAP_MASK(bits_not_wanted) << (pide & (BITS_PER_LONG - 1));
			bits_not_wanted = 0;

			DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n", __FUNCTION__, (uint) iova, size,
		        	bits_not_wanted, m, pide, res_ptr, *res_ptr);

			ASSERT(m != 0);
			ASSERT(bits_not_wanted);
			ASSERT((*res_ptr & m) == m); /* verify same bits are set */
			*res_ptr &= ~m;
		}
	}
}


/**************************************************************
*
*   "Dynamic DMA Mapping" support (aka "Coherent I/O")
*
***************************************************************/

/**
 * sba_io_pdir_entry - fill in one IO PDIR entry
 * @pdir_ptr:  pointer to IO PDIR entry
 * @vba: Virtual CPU address of buffer to map
 *
 * SBA Mapping Routine
 *
 * Given a virtual address (vba, arg1) sba_io_pdir_entry()
 * loads the I/O PDIR entry pointed to by pdir_ptr (arg0).
 * Each IO Pdir entry consists of 8 bytes as shown below
 * (LSB == bit 0):
 *
 *  63                    40                                 11    7        0
 * +-+---------------------+----------------------------------+----+--------+
 * |V|        U            |            PPN[39:12]            | U  |   FF   |
 * +-+---------------------+----------------------------------+----+--------+
 *
 *  V  == Valid Bit
 *  U  == Unused
 * PPN == Physical Page Number
 *
 * The physical address fields are filled with the results of virt_to_phys()
 * on the vba.
 */

#if 1
#define sba_io_pdir_entry(pdir_ptr, vba) *pdir_ptr = ((vba & ~0xE000000000000FFFULL)	\
						      | 0x8000000000000000ULL)
#else
void SBA_INLINE
sba_io_pdir_entry(u64 *pdir_ptr, unsigned long vba)
{
	*pdir_ptr = ((vba & ~0xE000000000000FFFULL) | 0x80000000000000FFULL);
}
#endif

#ifdef ENABLE_MARK_CLEAN
/**
 * Since DMA is i-cache coherent, any (complete) pages that were written via
 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
 * flush them when they get mapped into an executable vm-area.
 */
static void
mark_clean (void *addr, size_t size)
{
	unsigned long pg_addr, end;

	pg_addr = PAGE_ALIGN((unsigned long) addr);
	end = (unsigned long) addr + size;
	while (pg_addr + PAGE_SIZE <= end) {
		struct page *page = virt_to_page((void *)pg_addr);
		set_bit(PG_arch_1, &page->flags);
		pg_addr += PAGE_SIZE;
	}
}
#endif

/**
 * sba_mark_invalid - invalidate one or more IO PDIR entries
 * @ioc: IO MMU structure which owns the pdir we are interested in.
 * @iova:  IO Virtual Address mapped earlier
 * @byte_cnt:  number of bytes this mapping covers.
 *
 * Marking the IO PDIR entry(ies) as Invalid and invalidate
 * corresponding IO TLB entry. The PCOM (Purge Command Register)
 * is to purge stale entries in the IO TLB when unmapping entries.
 *
 * The PCOM register supports purging of multiple pages, with a minium
 * of 1 page and a maximum of 2GB. Hardware requires the address be
 * aligned to the size of the range being purged. The size of the range
 * must be a power of 2. The "Cool perf optimization" in the
 * allocation routine helps keep that true.
 */
static SBA_INLINE void
sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
{
	u32 iovp = (u32) SBA_IOVP(ioc,iova);

	int off = PDIR_INDEX(iovp);

	/* Must be non-zero and rounded up */
	ASSERT(byte_cnt > 0);
	ASSERT(0 == (byte_cnt & ~iovp_mask));

#ifdef ASSERT_PDIR_SANITY
	/* Assert first pdir entry is set */
	if (!(ioc->pdir_base[off] >> 60)) {
		sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
	}
#endif

	if (byte_cnt <= iovp_size)
	{
		ASSERT(off < ioc->pdir_size);

		iovp |= iovp_shift;     /* set "size" field for PCOM */

#ifndef FULL_VALID_PDIR
		/*
		** clear I/O PDIR entry "valid" bit
		** Do NOT clear the rest - save it for debugging.
		** We should only clear bits that have previously
		** been enabled.
		*/
		ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
#else
		/*
  		** If we want to maintain the PDIR as valid, put in
		** the spill page so devices prefetching won't
		** cause a hard fail.
		*/
		ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
#endif
	} else {
		u32 t = get_iovp_order(byte_cnt) + iovp_shift;

		iovp |= t;
		ASSERT(t <= 31);   /* 2GB! Max value of "size" field */

		do {
			/* verify this pdir entry is enabled */
			ASSERT(ioc->pdir_base[off]  >> 63);
#ifndef FULL_VALID_PDIR
			/* clear I/O Pdir entry "valid" bit first */
			ioc->pdir_base[off] &= ~(0x80000000000000FFULL);
#else
			ioc->pdir_base[off] = (0x80000000000000FFULL | prefetch_spill_page);
#endif
			off++;
			byte_cnt -= iovp_size;
		} while (byte_cnt > 0);
	}

	WRITE_REG(iovp | ioc->ibase, ioc->ioc_hpa+IOC_PCOM);
}

/**
 * sba_map_single - map one buffer and return IOVA for DMA
 * @dev: instance of PCI owned by the driver that's asking.
 * @addr:  driver buffer to map.
 * @size:  number of bytes to map in driver buffer.
 * @dir:  R/W or both.
 *
 * See Documentation/DMA-mapping.txt
 */
dma_addr_t
sba_map_single(struct device *dev, void *addr, size_t size, int dir)
{
	struct ioc *ioc;
	dma_addr_t iovp;
	dma_addr_t offset;
	u64 *pdir_start;
	int pide;
#ifdef ASSERT_PDIR_SANITY
	unsigned long flags;
#endif
#ifdef ALLOW_IOV_BYPASS
	unsigned long pci_addr = virt_to_phys(addr);
#endif

#ifdef ALLOW_IOV_BYPASS
	ASSERT(to_pci_dev(dev)->dma_mask);
	/*
 	** Check if the PCI device can DMA to ptr... if so, just return ptr
 	*/
	if (likely((pci_addr & ~to_pci_dev(dev)->dma_mask) == 0)) {
		/*
 		** Device is bit capable of DMA'ing to the buffer...
		** just return the PCI address of ptr
 		*/
		DBG_BYPASS("sba_map_single() bypass mask/addr: 0x%lx/0x%lx\n",
		           to_pci_dev(dev)->dma_mask, pci_addr);
		return pci_addr;
	}
#endif
	ioc = GET_IOC(dev);
	ASSERT(ioc);

	prefetch(ioc->res_hint);

	ASSERT(size > 0);
	ASSERT(size <= DMA_CHUNK_SIZE);

	/* save offset bits */
	offset = ((dma_addr_t) (long) addr) & ~iovp_mask;

	/* round up to nearest iovp_size */
	size = (size + offset + ~iovp_mask) & iovp_mask;

#ifdef ASSERT_PDIR_SANITY
	spin_lock_irqsave(&ioc->res_lock, flags);
	if (sba_check_pdir(ioc,"Check before sba_map_single()"))
		panic("Sanity check failed");
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif

	pide = sba_alloc_range(ioc, size);

	iovp = (dma_addr_t) pide << iovp_shift;

	DBG_RUN("%s() 0x%p -> 0x%lx\n",
		__FUNCTION__, addr, (long) iovp | offset);

	pdir_start = &(ioc->pdir_base[pide]);

	while (size > 0) {
		ASSERT(((u8 *)pdir_start)[7] == 0); /* verify availability */
		sba_io_pdir_entry(pdir_start, (unsigned long) addr);

		DBG_RUN("     pdir 0x%p %lx\n", pdir_start, *pdir_start);

		addr += iovp_size;
		size -= iovp_size;
		pdir_start++;
	}
	/* force pdir update */
	wmb();

	/* form complete address */
#ifdef ASSERT_PDIR_SANITY
	spin_lock_irqsave(&ioc->res_lock, flags);
	sba_check_pdir(ioc,"Check after sba_map_single()");
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif
	return SBA_IOVA(ioc, iovp, offset);
}

#ifdef ENABLE_MARK_CLEAN
static SBA_INLINE void
sba_mark_clean(struct ioc *ioc, dma_addr_t iova, size_t size)
{
	u32	iovp = (u32) SBA_IOVP(ioc,iova);
	int	off = PDIR_INDEX(iovp);
	void	*addr;

	if (size <= iovp_size) {
		addr = phys_to_virt(ioc->pdir_base[off] &
		                    ~0xE000000000000FFFULL);
		mark_clean(addr, size);
	} else {
		do {
			addr = phys_to_virt(ioc->pdir_base[off] &
			                    ~0xE000000000000FFFULL);
			mark_clean(addr, min(size, iovp_size));
			off++;
			size -= iovp_size;
		} while (size > 0);
	}
}
#endif

/**
 * sba_unmap_single - unmap one IOVA and free resources
 * @dev: instance of PCI owned by the driver that's asking.
 * @iova:  IOVA of driver buffer previously mapped.
 * @size:  number of bytes mapped in driver buffer.
 * @dir:  R/W or both.
 *
 * See Documentation/DMA-mapping.txt
 */
void sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, int dir)
{
	struct ioc *ioc;
#if DELAYED_RESOURCE_CNT > 0
	struct sba_dma_pair *d;
#endif
	unsigned long flags;
	dma_addr_t offset;

	ioc = GET_IOC(dev);
	ASSERT(ioc);

#ifdef ALLOW_IOV_BYPASS
	if (likely((iova & ioc->imask) != ioc->ibase)) {
		/*
		** Address does not fall w/in IOVA, must be bypassing
		*/
		DBG_BYPASS("sba_unmap_single() bypass addr: 0x%lx\n", iova);

#ifdef ENABLE_MARK_CLEAN
		if (dir == DMA_FROM_DEVICE) {
			mark_clean(phys_to_virt(iova), size);
		}
#endif
		return;
	}
#endif
	offset = iova & ~iovp_mask;

	DBG_RUN("%s() iovp 0x%lx/%x\n",
		__FUNCTION__, (long) iova, size);

	iova ^= offset;        /* clear offset bits */
	size += offset;
	size = ROUNDUP(size, iovp_size);

#ifdef ENABLE_MARK_CLEAN
	if (dir == DMA_FROM_DEVICE)
		sba_mark_clean(ioc, iova, size);
#endif

#if DELAYED_RESOURCE_CNT > 0
	spin_lock_irqsave(&ioc->saved_lock, flags);
	d = &(ioc->saved[ioc->saved_cnt]);
	d->iova = iova;
	d->size = size;
	if (unlikely(++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT)) {
		int cnt = ioc->saved_cnt;
		spin_lock(&ioc->res_lock);
		while (cnt--) {
			sba_mark_invalid(ioc, d->iova, d->size);
			sba_free_range(ioc, d->iova, d->size);
			d--;
		}
		ioc->saved_cnt = 0;
		READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
		spin_unlock(&ioc->res_lock);
	}
	spin_unlock_irqrestore(&ioc->saved_lock, flags);
#else /* DELAYED_RESOURCE_CNT == 0 */
	spin_lock_irqsave(&ioc->res_lock, flags);
	sba_mark_invalid(ioc, iova, size);
	sba_free_range(ioc, iova, size);
	READ_REG(ioc->ioc_hpa+IOC_PCOM);	/* flush purges */
	spin_unlock_irqrestore(&ioc->res_lock, flags);
#endif /* DELAYED_RESOURCE_CNT == 0 */
}