sba_iommu.c

linux-2.4.29操作系统的源码

		mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;

		DBG_RES("%s() o %ld %p", __FUNCTION__, o, res_ptr);
		while(res_ptr < res_end)
		{ 
			DBG_RES("    %p %lx %lx\n", res_ptr, mask, *res_ptr);
			ASSERT(0 != mask);
			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;
				break;
			}
			mask >>= o;
			bitshiftcnt += o;
			if (0 == mask) {
				mask = RESMAP_MASK(bits_wanted);
				bitshiftcnt=0;
				res_ptr++;
			}
		}
		/* look in the same word on the next pass */
		ioc->res_bitshift = bitshiftcnt + bits_wanted;
	}

	/* wrapped ? */
	if (res_end <= res_ptr) {
		ioc->res_hint = (unsigned long *) ioc->res_map;
		ioc->res_bitshift = 0;
	} else {
		ioc->res_hint = res_ptr;
	}
	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 CONFIG_PROC_FS
	unsigned long cr_start = mfctl(16);
#endif
	unsigned long pide;

	ASSERT(pages_needed);
	ASSERT((pages_needed * IOVP_SIZE) <= DMA_CHUNK_SIZE);
	ASSERT(pages_needed <= BITS_PER_LONG);
	ASSERT(0 == (size & ~IOVP_MASK));

	/*
	** "seek and ye shall find"...praying never hurts either...
	** ggg sacrifices another 710 to the computer gods.
	*/

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

#ifdef ASSERT_PDIR_SANITY
	/* verify the first enable bit is clear */
	if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 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 );

#ifdef CONFIG_PROC_FS
	{
		unsigned long cr_end = mfctl(16);
		unsigned long tmp = cr_end - cr_start;
		/* check for roll over */
		cr_start = (cr_end < cr_start) ?  -(tmp) : (tmp);
	}
	ioc->avg_search[ioc->avg_idx++] = cr_start;
	ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;

	ioc->used_pages += pages_needed;
#endif

	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;

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

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

#ifdef CONFIG_PROC_FS
	ioc->used_pages -= bits_not_wanted;
#endif

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


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

#define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)


typedef unsigned long space_t;
#define KERNEL_SPACE 0

/**
 * sba_io_pdir_entry - fill in one IO PDIR entry
 * @pdir_ptr:  pointer to IO PDIR entry
 * @sid: process Space ID
 * @vba: Virtual CPU address of buffer to map
 *
 * SBA Mapping Routine
 *
 * Given a virtual address (vba, arg2) and space id, (sid, 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 (MSB == bit 0):
 *
 *  0                    19                                 51   55       63
 * +-+---------------------+----------------------------------+----+--------+
 * |V|        U            |            PPN[43:12]            | U  |   VI   |
 * +-+---------------------+----------------------------------+----+--------+
 *
 *  V  == Valid Bit
 *  U  == Unused
 * PPN == Physical Page Number
 * VI  == Virtual Index (aka Coherent Index)
 *
 * The physical address fields are filled with the results of the LPA
 * instruction.  The virtual index field is filled with the results of
 * of the LCI (Load Coherence Index) instruction.  The 8 bits used for
 * the virtual index are bits 12:19 of the value returned by LCI.
 *
 * We need to pre-swap the bytes since PCX-W is Big Endian.
 */


void SBA_INLINE
sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba)
{
	u64 pa; /* physical address */
	register unsigned ci; /* coherent index */

	/* We currently only support kernel addresses.
	 * fdc instr below will need to reload sr1 with KERNEL_SPACE
	 * once we try to support direct DMA to user space.
	 */
	ASSERT(sid == KERNEL_SPACE);

	pa = virt_to_phys(vba);
	pa &= ~4095ULL;			/* clear out offset bits */

	mtsp(sid,1);
	asm("lci 0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
	pa |= (ci >> 12) & 0xff;  /* move CI (8 bits) into lowest byte */

	pa |= 0x8000000000000000ULL;	/* set "valid" bit */
	*pdir_ptr = cpu_to_le64(pa);	/* swap and store into I/O Pdir */

	/*
	 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
	 * (bit #61, big endian), we have to flush and sync every time
	 * IO-PDIR is changed in Ike/Astro.
	 */
	if (ioc_needs_fdc) {
		asm volatile("fdc 0(%%sr1,%0)\n\tsync" : : "r" (pdir_ptr));
	}
}


/**
 * 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 Ike 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);

	/* Even though this is a big-endian machine, the entries
	** in the iopdir are little endian. That's why we clear the byte
	** at +7 instead of at +0.
	*/
	int off = PDIR_INDEX(iovp)*sizeof(u64)+7;

	/* 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 (0x80 != (((u8 *) ioc->pdir_base)[off])) {
		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 */

		/*
		** 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.
		*/
		((u8 *)(ioc->pdir_base))[off] = 0;
	} else {
		u32 t = get_order(byte_cnt) + PAGE_SHIFT;

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

		do {
			/* verify this pdir entry is enabled */
			ASSERT(0x80 == (((u8 *) ioc->pdir_base)[off] & 0x80));
			/* clear I/O Pdir entry "valid" bit first */
			((u8 *)(ioc->pdir_base))[off] = 0;
			off += sizeof(u64);
			byte_cnt -= IOVP_SIZE;
		} while (byte_cnt > 0);
	}

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

/**
 * sba_dma_supported - PCI driver can query DMA support
 * @dev: instance of PCI owned by the driver that's asking
 * @mask:  number of address bits this PCI device can handle
 *
 * See Documentation/DMA-mapping.txt
 */
static int
sba_dma_supported( struct pci_dev *dev, u64 mask)
{
	if (dev == NULL) {
		printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
		BUG();
		return(0);
	}

	/* only support 32-bit PCI devices - no DAC support (yet) */
	return((int) (mask == 0xffffffffUL));
}


/**
 * 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.
 * @direction:  R/W or both.
 *
 * See Documentation/DMA-mapping.txt
 */
static dma_addr_t
sba_map_single(struct pci_dev *dev, void *addr, size_t size, int direction)
{
	struct ioc *ioc;
	unsigned long flags; 
	dma_addr_t iovp;
	dma_addr_t offset;
	u64 *pdir_start;
	int pide;

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

	ASSERT(dev->sysdata);
	ioc = GET_IOC(dev);
	ASSERT(ioc);

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

	/* round up to nearest IOVP_SIZE */
	size = (size + offset + ~IOVP_MASK) & IOVP_MASK;

	spin_lock_irqsave(&ioc->res_lock, flags);
#ifdef ASSERT_PDIR_SANITY
	sba_check_pdir(ioc,"Check before sba_map_single()");
#endif

#ifdef CONFIG_PROC_FS
	ioc->msingle_calls++;
	ioc->msingle_pages += size >> IOVP_SHIFT;
#endif
	pide = sba_alloc_range(ioc, size);
	iovp = (dma_addr_t) pide << IOVP_SHIFT;

	DBG_RUN("%s() 0x%p -> 0x%lx",
		__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, KERNEL_SPACE, (unsigned long) addr);

		DBG_RUN(" pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
			pdir_start,
			(u8) (((u8 *) pdir_start)[7]),
			(u8) (((u8 *) pdir_start)[6]),
			(u8) (((u8 *) pdir_start)[5]),
			(u8) (((u8 *) pdir_start)[4]),
			(u8) (((u8 *) pdir_start)[3]),
			(u8) (((u8 *) pdir_start)[2]),
			(u8) (((u8 *) pdir_start)[1]),
			(u8) (((u8 *) pdir_start)[0])
			);

		addr += IOVP_SIZE;
		size -= IOVP_SIZE;
		pdir_start++;
	}
	/* form complete address */
#ifdef ASSERT_PDIR_SANITY
	sba_check_pdir(ioc,"Check after sba_map_single()");
#endif
	spin_unlock_irqrestore(&ioc->res_lock, flags);
	return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
}


/**
 * 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.
 * @direction:  R/W or both.
 *
 * See Documentation/DMA-mapping.txt
 */
static void
sba_unmap_single(struct pci_dev *dev, dma_addr_t iova, size_t size, int direction)
{
	struct ioc *ioc;
#if DELAYED_RESOURCE_CNT > 0
	struct sba_dma_pair *d;
#endif
	unsigned long flags; 
	dma_addr_t offset;

	ASSERT(dev->sysdata);
	ioc = GET_IOC(dev);
	ASSERT(ioc);

	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);

	spin_lock_irqsave(&ioc->res_lock, flags);

#ifdef CONFIG_PROC_FS
	ioc->usingle_calls++;
	ioc->usingle_pages += size >> IOVP_SHIFT;
#endif

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

	/* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
	** For Astro based systems this isn't a big deal WRT performance.
	** As long as 2.4 kernels copyin/copyout data from/to userspace,
	** we don't need the syncdma. The issue here is I/O MMU cachelines
	** are *not* coherent in all cases.  May be hwrev dependent.
	** Need to investigate more.
	asm volatile("syncdma");	
	*/
}


/**
 * sba_alloc_consistent - allocate/map shared mem for DMA
 * @hwdev: instance of PCI owned by the driver that's asking.
 * @size:  number of bytes mapped in driver buffer.
 * @dma_handle:  IOVA of new buffer.
 *
 * See Documentation/DMA-mapping.txt
 */
static void *
sba_alloc_consistent(struct pci_dev *hwdev, size_t size, dma_addr_t *dma_handle)
{
	void *ret;

	if (!hwdev) {
		/* only support PCI */
		*dma_handle = 0;
		return 0;
	}

        ret = (void *) __get_free_pages(GFP_ATOMIC, get_order(size));

	if (ret) {
		memset(ret, 0, size);
		*dma_handle = sba_map_single(hwdev, ret, size, 0);
	}

	return ret;
}


/**
 * sba_free_consistent - free/unmap shared mem for DMA
 * @hwdev: instance of PCI owned by the driver that's asking.
 * @size:  number of bytes mapped in driver buffer.
 * @vaddr:  virtual address IOVA of "consistent" buffer.
 * @dma_handler:  IO virtual address of "consistent" buffer.
 *
 * See Documentation/DMA-mapping.txt
 */
static void
sba_free_consistent(struct pci_dev *hwdev, size_t size, void *vaddr, dma_addr_t dma_handle)
{
	sba_unmap_single(hwdev, dma_handle, size, 0);
	free_pages((unsigned long) vaddr, get_order(size));
}


/*
** Since 0 is a valid pdir_base index value, can't use that
** to determine if a value is valid or not. Use a flag to indicate
** the SG list entry contains a valid pdir index.
*/
#define PIDE_FLAG 0x80000000UL

#ifdef DEBUG_LARGE_SG_ENTRIES
int dump_run_sg = 0;
#endif