ccio-dma.c

《嵌入式系统设计与实例开发实验教材二源码》Linux内核移植与编译实验

/*
** ccio-dma.c:
**	DMA management routines for first generation cache-coherent machines.
**	Program U2/Uturn in "Virtual Mode" and use the I/O MMU.
**
**	(c) Copyright 2000 Grant Grundler
**	(c) Copyright 2000 Ryan Bradetich
**	(c) Copyright 2000 Hewlett-Packard Company
**
** This program is free software; you can redistribute it and/or modify
** it under the terms of the GNU General Public License as published by
** the Free Software Foundation; either version 2 of the License, or
** (at your option) any later version.
**
**
**  "Real Mode" operation refers to U2/Uturn chip operation.
**  U2/Uturn were designed to perform coherency checks w/o using
**  the I/O MMU - basically what x86 does.
**
**  Philipp Rumpf has a "Real Mode" driver for PCX-W machines at:
**      CVSROOT=:pserver:anonymous@198.186.203.37:/cvsroot/linux-parisc
**      cvs -z3 co linux/arch/parisc/kernel/dma-rm.c
**
**  I've rewritten his code to work under TPG's tree. See ccio-rm-dma.c.
**
**  Drawbacks of using Real Mode are:
**	o outbound DMA is slower - U2 won't prefetch data (GSC+ XQL signal).
**      o Inbound DMA less efficient - U2 can't use DMA_FAST attribute.
**	o Ability to do scatter/gather in HW is lost.
**	o Doesn't work under PCX-U/U+ machines since they didn't follow
**        the coherency design originally worked out. Only PCX-W does.
*/

#include <linux/config.h>
#include <linux/types.h>
#include <linux/init.h>
#include <linux/mm.h>
#include <linux/spinlock.h>
#include <linux/slab.h>
#include <linux/string.h>
#include <linux/pci.h>

#include <asm/byteorder.h>
#include <asm/cache.h>		/* for L1_CACHE_BYTES */
#include <asm/uaccess.h>
#include <asm/pgalloc.h>
#include <asm/page.h>

#include <asm/io.h>
#include <asm/gsc.h>		/* for gsc_writeN()... */

/* 
** Choose "ccio" since that's what HP-UX calls it.
** Make it easier for folks to migrate from one to the other :^)
*/
#define MODULE_NAME "ccio"

/*
#define DEBUG_CCIO_RES
#define DEBUG_CCIO_RUN
#define DEBUG_CCIO_INIT
#define DUMP_RESMAP
*/

#include <linux/proc_fs.h>
#include <asm/runway.h>		/* for proc_runway_root */

#ifdef DEBUG_CCIO_INIT
#define DBG_INIT(x...)  printk(x)
#else
#define DBG_INIT(x...)
#endif

#ifdef DEBUG_CCIO_RUN
#define DBG_RUN(x...)   printk(x)
#else
#define DBG_RUN(x...)
#endif

#ifdef DEBUG_CCIO_RES
#define DBG_RES(x...)   printk(x)
#else
#define DBG_RES(x...)
#endif

#define CCIO_INLINE	/* inline */
#define WRITE_U32(value, addr) gsc_writel(value, (u32 *) (addr))

#define U2_IOA_RUNWAY 0x580
#define U2_BC_GSC     0x501
#define UTURN_IOA_RUNWAY 0x581
#define UTURN_BC_GSC     0x502
/* We *can't* support JAVA (T600). Venture there at your own risk. */

static void dump_resmap(void);

static int ccio_driver_callback(struct hp_device *, struct pa_iodc_driver *);

static struct pa_iodc_driver ccio_drivers_for[] = {

   {HPHW_IOA, U2_IOA_RUNWAY, 0x0, 0xb, 0, 0x10,
		DRIVER_CHECK_HVERSION +
		DRIVER_CHECK_SVERSION + DRIVER_CHECK_HWTYPE,
                MODULE_NAME, "U2 I/O MMU", (void *) ccio_driver_callback},

   {HPHW_IOA, UTURN_IOA_RUNWAY, 0x0, 0xb, 0, 0x10,
		DRIVER_CHECK_HVERSION +
		DRIVER_CHECK_SVERSION + DRIVER_CHECK_HWTYPE,
                MODULE_NAME, "Uturn I/O MMU", (void *) ccio_driver_callback},

/*
** FIXME: The following claims the GSC bus port, not the IOA.
**        And there are two busses below a single I/O TLB.
**
** These should go away once we have a real PA bus walk.
** Firmware wants to tell the PA bus walk code about the GSC ports
** since they are not "architected" PA I/O devices. Ie a PA bus walk
** wouldn't discover them. But the PA bus walk code could check
** the "fixed module table" to add such devices to an I/O Tree
** and proceed with the recursive, depth first bus walk.
*/
   {HPHW_BCPORT, U2_BC_GSC, 0x0, 0xc, 0, 0x10,
		DRIVER_CHECK_HVERSION +
		DRIVER_CHECK_SVERSION + DRIVER_CHECK_HWTYPE,
                MODULE_NAME, "U2 GSC+ BC", (void *) ccio_driver_callback},

   {HPHW_BCPORT, UTURN_BC_GSC, 0x0, 0xc, 0, 0x10,
		DRIVER_CHECK_HVERSION +
		DRIVER_CHECK_SVERSION + DRIVER_CHECK_HWTYPE,
                MODULE_NAME, "Uturn GSC+ BC", (void *) ccio_driver_callback},

   {0,0,0,0,0,0,
   0,
   (char *) NULL, (char *) NULL, (void *) NULL }
};


#define IS_U2(id) ( \
    (((id)->hw_type == HPHW_IOA) && ((id)->hversion == U2_IOA_RUNWAY)) || \
    (((id)->hw_type == HPHW_BCPORT) && ((id)->hversion == U2_BC_GSC))  \
)

#define IS_UTURN(id) ( \
    (((id)->hw_type == HPHW_IOA) && ((id)->hversion == UTURN_IOA_RUNWAY)) || \
    (((id)->hw_type == HPHW_BCPORT) && ((id)->hversion == UTURN_BC_GSC))  \
)


#define IOA_NORMAL_MODE      0x00020080 /* IO_CONTROL to turn on CCIO        */
#define CMD_TLB_DIRECT_WRITE 35         /* IO_COMMAND for I/O TLB Writes     */
#define CMD_TLB_PURGE        33         /* IO_COMMAND to Purge I/O TLB entry */

struct ioa_registers {
        /* Runway Supervisory Set */
        volatile int32_t    unused1[12];
        volatile uint32_t   io_command;             /* Offset 12 */
        volatile uint32_t   io_status;              /* Offset 13 */
        volatile uint32_t   io_control;             /* Offset 14 */
        volatile int32_t    unused2[1];

        /* Runway Auxiliary Register Set */
        volatile uint32_t   io_err_resp;            /* Offset  0 */
        volatile uint32_t   io_err_info;            /* Offset  1 */
        volatile uint32_t   io_err_req;             /* Offset  2 */
        volatile uint32_t   io_err_resp_hi;         /* Offset  3 */
        volatile uint32_t   io_tlb_entry_m;         /* Offset  4 */
        volatile uint32_t   io_tlb_entry_l;         /* Offset  5 */
        volatile uint32_t   unused3[1];
        volatile uint32_t   io_pdir_base;           /* Offset  7 */
        volatile uint32_t   io_io_low_hv;           /* Offset  8 */
        volatile uint32_t   io_io_high_hv;          /* Offset  9 */
        volatile uint32_t   unused4[1];
        volatile uint32_t   io_chain_id_mask;       /* Offset 11 */
        volatile uint32_t   unused5[2];
        volatile uint32_t   io_io_low;              /* Offset 14 */
        volatile uint32_t   io_io_high;             /* Offset 15 */
};


struct ccio_device {
	struct ccio_device    *next;  /* list of LBA's in system */
	struct hp_device      *iodc;  /* data about dev from firmware */
	spinlock_t            ccio_lock;

	struct ioa_registers   *ccio_hpa; /* base address */
	u64   *pdir_base;   /* physical base address */
	char  *res_map;     /* resource map, bit == pdir entry */

	int   res_hint;	/* next available IOVP - circular search */
	int   res_size;	/* size of resource map in bytes */
	int   chainid_shift; /* specify bit location of chain_id */
	int   flags;         /* state/functionality enabled */
#ifdef DELAYED_RESOURCE_CNT
	dma_addr_t res_delay[DELAYED_RESOURCE_CNT];
#endif

	/* STUFF We don't need in performance path */
	int  pdir_size;     /* in bytes, determined by IOV Space size */
	int  hw_rev;        /* HW revision of chip */
};


/* Ratio of Host MEM to IOV Space size */
static unsigned long ccio_mem_ratio = 4;
static struct ccio_device *ccio_list = NULL;

static int ccio_proc_info(char *buffer, char **start, off_t offset, int length);
static unsigned long ccio_used_bytes = 0;
static unsigned long ccio_used_pages = 0;
static int ccio_cujo_bug = 0;

static unsigned long ccio_alloc_size = 0;
static unsigned long ccio_free_size = 0;

/**************************************************************
*
*   I/O Pdir Resource Management
*
*   Bits set in the resource map are in use.
*   Each bit can represent a number of pages.
*   LSbs represent lower addresses (IOVA's).
*
*   This was was copied from sba_iommu.c. Don't try to unify
*   the two resource managers unless a way to have different
*   allocation policies is also adjusted. We'd like to avoid
*   I/O TLB thrashing by having resource allocation policy
*   match the I/O TLB replacement policy.
*
***************************************************************/
#define PAGES_PER_RANGE 1	/* could increase this to 4 or 8 if needed */
#define IOVP_SIZE PAGE_SIZE
#define IOVP_SHIFT PAGE_SHIFT
#define IOVP_MASK PAGE_MASK

/* Convert from IOVP to IOVA and vice versa. */
#define CCIO_IOVA(iovp,offset) ((iovp) | (offset))
#define CCIO_IOVP(iova) ((iova) & ~(IOVP_SIZE-1) )

#define PDIR_INDEX(iovp)    ((iovp)>>IOVP_SHIFT)
#define MKIOVP(pdir_idx)    ((long)(pdir_idx) << IOVP_SHIFT)
#define MKIOVA(iovp,offset) (dma_addr_t)((long)iovp | (long)offset)

/* CUJO20 KLUDGE start */
#define CUJO_20_BITMASK    0x0ffff000	/* upper nibble is a don't care   */
#define CUJO_20_STEP       0x10000000	/* inc upper nibble */
#define CUJO_20_BADPAGE1   0x01003000	/* pages that hpmc on raven U+    */
#define CUJO_20_BADPAGE2   0x01607000	/* pages that hpmc on firehawk U+ */
#define CUJO_20_BADHVERS   0x6821	/* low nibble 1 is cujo rev 2.0 */
#define CUJO_RAVEN_LOC     0xf1000000UL	/* cujo location on raven U+ */
#define CUJO_FIREHAWK_LOC  0xf1604000UL	/* cujo location on firehawk U+ */
/* CUJO20 KLUDGE end */

/*
** Don't worry about the 150% average search length on a miss.
** If the search wraps around, and passes the res_hint, it will
** cause the kernel to panic anyhow.
*/

/* ioa->res_hint = idx + (size >> 3); \ */

#define CCIO_SEARCH_LOOP(ioa, idx, mask, size)  \
       for(; res_ptr < res_end; ++res_ptr) \
       { \
               if(0 == ((*res_ptr) & mask)) { \
                       *res_ptr |= mask; \
                       idx = (int)((unsigned long)res_ptr - (unsigned long)ioa->res_map); \
                       ioa->res_hint = 0;\
                       goto resource_found; \
               } \
       }

#define CCIO_FIND_FREE_MAPPING(ioa, idx, mask, size)  { \
       u##size *res_ptr = (u##size *)&((ioa)->res_map[ioa->res_hint & ~((size >> 3) - 1)]); \
       u##size *res_end = (u##size *)&(ioa)->res_map[ioa->res_size]; \
       CCIO_SEARCH_LOOP(ioa, idx, mask, size); \
       res_ptr = (u##size *)&(ioa)->res_map[0]; \
       CCIO_SEARCH_LOOP(ioa, idx, mask, size); \
}

/*
** Find available bit in this ioa's resource map.
** Use a "circular" search:
**   o Most IOVA's are "temporary" - avg search time should be small.
** o keep a history of what happened for debugging
** o KISS.
**
** Perf optimizations:
** o search for log2(size) bits at a time.
** o search for available resource bits using byte/word/whatever.
** o use different search for "large" (eg > 4 pages) or "very large"
**   (eg > 16 pages) mappings.
*/
static int
ccio_alloc_range(struct ccio_device *ioa, size_t size)
{
	int res_idx;
	unsigned long mask, flags;
	unsigned int pages_needed = size >> PAGE_SHIFT;

	ASSERT(pages_needed);
	ASSERT((pages_needed * IOVP_SIZE) < DMA_CHUNK_SIZE);
	ASSERT(pages_needed < (BITS_PER_LONG - IOVP_SHIFT));

	mask = (unsigned long) -1L;
 	mask >>= BITS_PER_LONG - pages_needed;

	DBG_RES(__FUNCTION__ " size: %d pages_needed %d pages_mask 0x%08lx\n", 
		size, pages_needed, mask);

	spin_lock_irqsave(&ioa->ccio_lock, flags);

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

	if(pages_needed <= 8) {
		CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, 8);
	} else if(pages_needed <= 16) {
		CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, 16);
	} else if(pages_needed <= 32) {
		CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, 32);
#ifdef __LP64__
	} else if(pages_needed <= 64) {
		CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, 64)
#endif
	} else {
		panic(__FILE__ ":" __FUNCTION__ "() Too many pages to map.\n");
	}

#ifdef DUMP_RESMAP
	dump_resmap();
#endif
	panic(__FILE__ ":" __FUNCTION__ "() I/O MMU is out of mapping resources\n");
	
resource_found:
	
	DBG_RES(__FUNCTION__ " res_idx %d mask 0x%08lx res_hint: %d\n",
		res_idx, mask, ioa->res_hint);

	ccio_used_pages += pages_needed;
	ccio_used_bytes += ((pages_needed >> 3) ? (pages_needed >> 3) : 1);

	spin_unlock_irqrestore(&ioa->ccio_lock, flags);

#ifdef DUMP_RESMAP
	dump_resmap();
#endif

	/* 
	** return the bit address (convert from byte to bit).
	*/
	return (res_idx << 3);
}


#define CCIO_FREE_MAPPINGS(ioa, idx, mask, size) \
        u##size *res_ptr = (u##size *)&((ioa)->res_map[idx + (((size >> 3) - 1) & ~((size >> 3) - 1))]); \
        ASSERT((*res_ptr & mask) == mask); \
        *res_ptr &= ~mask;

/*
** clear bits in the ioa's resource map
*/
static void
ccio_free_range(struct ccio_device *ioa, dma_addr_t iova, size_t size)
{
	unsigned long mask, flags;
	unsigned long iovp = CCIO_IOVP(iova);
	unsigned int res_idx = PDIR_INDEX(iovp)>>3;
	unsigned int pages_mapped = (size >> IOVP_SHIFT) + !!(size & ~IOVP_MASK);

	ASSERT(pages_needed);
	ASSERT((pages_needed * IOVP_SIZE) < DMA_CHUNK_SIZE);
	ASSERT(pages_needed < (BITS_PER_LONG - IOVP_SHIFT));

	mask = (unsigned long) -1L;
 	mask >>= BITS_PER_LONG - pages_mapped;

	DBG_RES(__FUNCTION__ " res_idx: %d size: %d pages_mapped %d mask 0x%08lx\n", 
		res_idx, size, pages_mapped, mask);

	spin_lock_irqsave(&ioa->ccio_lock, flags);

	if(pages_mapped <= 8) {
		CCIO_FREE_MAPPINGS(ioa, res_idx, mask, 8);
	} else if(pages_mapped <= 16) {
		CCIO_FREE_MAPPINGS(ioa, res_idx, mask, 16);
	} else if(pages_mapped <= 32) {
		CCIO_FREE_MAPPINGS(ioa, res_idx, mask, 32);
#ifdef __LP64__
	} else if(pages_mapped <= 64) {
		CCIO_FREE_MAPPINGS(ioa, res_idx, mask, 64);
#endif
	} else {
		panic(__FILE__ ":" __FUNCTION__ "() Too many pages to unmap.\n");
	}
	
	ccio_used_pages -= (pages_mapped ? pages_mapped : 1);
	ccio_used_bytes -= ((pages_mapped >> 3) ? (pages_mapped >> 3) : 1);

	spin_unlock_irqrestore(&ioa->ccio_lock, flags);

#ifdef DUMP_RESMAP