1/*
   2** ccio-dma.c:
   3**	DMA management routines for first generation cache-coherent machines.
   4**	Program U2/Uturn in "Virtual Mode" and use the I/O MMU.
   5**
   6**	(c) Copyright 2000 Grant Grundler
   7**	(c) Copyright 2000 Ryan Bradetich
   8**	(c) Copyright 2000 Hewlett-Packard Company
   9**
  10** This program is free software; you can redistribute it and/or modify
  11** it under the terms of the GNU General Public License as published by
  12** the Free Software Foundation; either version 2 of the License, or
  13** (at your option) any later version.
  14**
  15**
  16**  "Real Mode" operation refers to U2/Uturn chip operation.
  17**  U2/Uturn were designed to perform coherency checks w/o using
  18**  the I/O MMU - basically what x86 does.
  19**
  20**  Philipp Rumpf has a "Real Mode" driver for PCX-W machines at:
  21**      CVSROOT=:pserver:anonymous@198.186.203.37:/cvsroot/linux-parisc
  22**      cvs -z3 co linux/arch/parisc/kernel/dma-rm.c
  23**
  24**  I've rewritten his code to work under TPG's tree. See ccio-rm-dma.c.
  25**
  26**  Drawbacks of using Real Mode are:
  27**	o outbound DMA is slower - U2 won't prefetch data (GSC+ XQL signal).
  28**      o Inbound DMA less efficient - U2 can't use DMA_FAST attribute.
  29**	o Ability to do scatter/gather in HW is lost.
  30**	o Doesn't work under PCX-U/U+ machines since they didn't follow
  31**        the coherency design originally worked out. Only PCX-W does.
  32*/
  33
  34#include <linux/types.h>
  35#include <linux/kernel.h>
  36#include <linux/init.h>
  37#include <linux/mm.h>
  38#include <linux/spinlock.h>
  39#include <linux/slab.h>
  40#include <linux/string.h>
  41#include <linux/pci.h>
  42#include <linux/reboot.h>
  43#include <linux/proc_fs.h>
  44#include <linux/seq_file.h>
  45#include <linux/scatterlist.h>
  46#include <linux/iommu-helper.h>
 
  47
  48#include <asm/byteorder.h>
  49#include <asm/cache.h>		/* for L1_CACHE_BYTES */
  50#include <asm/uaccess.h>
  51#include <asm/page.h>
  52#include <asm/dma.h>
  53#include <asm/io.h>
  54#include <asm/hardware.h>       /* for register_module() */
  55#include <asm/parisc-device.h>
  56
  57/* 
  58** Choose "ccio" since that's what HP-UX calls it.
  59** Make it easier for folks to migrate from one to the other :^)
  60*/
  61#define MODULE_NAME "ccio"
  62
  63#undef DEBUG_CCIO_RES
  64#undef DEBUG_CCIO_RUN
  65#undef DEBUG_CCIO_INIT
  66#undef DEBUG_CCIO_RUN_SG
  67
  68#ifdef CONFIG_PROC_FS
  69/* depends on proc fs support. But costs CPU performance. */
  70#undef CCIO_COLLECT_STATS
  71#endif
  72
  73#include <asm/runway.h>		/* for proc_runway_root */
  74
  75#ifdef DEBUG_CCIO_INIT
  76#define DBG_INIT(x...)  printk(x)
  77#else
  78#define DBG_INIT(x...)
  79#endif
  80
  81#ifdef DEBUG_CCIO_RUN
  82#define DBG_RUN(x...)   printk(x)
  83#else
  84#define DBG_RUN(x...)
  85#endif
  86
  87#ifdef DEBUG_CCIO_RES
  88#define DBG_RES(x...)   printk(x)
  89#else
  90#define DBG_RES(x...)
  91#endif
  92
  93#ifdef DEBUG_CCIO_RUN_SG
  94#define DBG_RUN_SG(x...) printk(x)
  95#else
  96#define DBG_RUN_SG(x...)
  97#endif
  98
  99#define CCIO_INLINE	inline
 100#define WRITE_U32(value, addr) __raw_writel(value, addr)
 101#define READ_U32(addr) __raw_readl(addr)
 102
 103#define U2_IOA_RUNWAY 0x580
 104#define U2_BC_GSC     0x501
 105#define UTURN_IOA_RUNWAY 0x581
 106#define UTURN_BC_GSC     0x502
 107
 108#define IOA_NORMAL_MODE      0x00020080 /* IO_CONTROL to turn on CCIO        */
 109#define CMD_TLB_DIRECT_WRITE 35         /* IO_COMMAND for I/O TLB Writes     */
 110#define CMD_TLB_PURGE        33         /* IO_COMMAND to Purge I/O TLB entry */
 111
 
 
 112struct ioa_registers {
 113        /* Runway Supervisory Set */
 114        int32_t    unused1[12];
 115        uint32_t   io_command;             /* Offset 12 */
 116        uint32_t   io_status;              /* Offset 13 */
 117        uint32_t   io_control;             /* Offset 14 */
 118        int32_t    unused2[1];
 119
 120        /* Runway Auxiliary Register Set */
 121        uint32_t   io_err_resp;            /* Offset  0 */
 122        uint32_t   io_err_info;            /* Offset  1 */
 123        uint32_t   io_err_req;             /* Offset  2 */
 124        uint32_t   io_err_resp_hi;         /* Offset  3 */
 125        uint32_t   io_tlb_entry_m;         /* Offset  4 */
 126        uint32_t   io_tlb_entry_l;         /* Offset  5 */
 127        uint32_t   unused3[1];
 128        uint32_t   io_pdir_base;           /* Offset  7 */
 129        uint32_t   io_io_low_hv;           /* Offset  8 */
 130        uint32_t   io_io_high_hv;          /* Offset  9 */
 131        uint32_t   unused4[1];
 132        uint32_t   io_chain_id_mask;       /* Offset 11 */
 133        uint32_t   unused5[2];
 134        uint32_t   io_io_low;              /* Offset 14 */
 135        uint32_t   io_io_high;             /* Offset 15 */
 136};
 137
 138/*
 139** IOA Registers
 140** -------------
 141**
 142** Runway IO_CONTROL Register (+0x38)
 143** 
 144** The Runway IO_CONTROL register controls the forwarding of transactions.
 145**
 146** | 0  ...  13  |  14 15 | 16 ... 21 | 22 | 23 24 |  25 ... 31 |
 147** |    HV       |   TLB  |  reserved | HV | mode  |  reserved  |
 148**
 149** o mode field indicates the address translation of transactions
 150**   forwarded from Runway to GSC+:
 151**       Mode Name     Value        Definition
 152**       Off (default)   0          Opaque to matching addresses.
 153**       Include         1          Transparent for matching addresses.
 154**       Peek            3          Map matching addresses.
 155**
 156**       + "Off" mode: Runway transactions which match the I/O range
 157**         specified by the IO_IO_LOW/IO_IO_HIGH registers will be ignored.
 158**       + "Include" mode: all addresses within the I/O range specified
 159**         by the IO_IO_LOW and IO_IO_HIGH registers are transparently
 160**         forwarded. This is the I/O Adapter's normal operating mode.
 161**       + "Peek" mode: used during system configuration to initialize the
 162**         GSC+ bus. Runway Write_Shorts in the address range specified by
 163**         IO_IO_LOW and IO_IO_HIGH are forwarded through the I/O Adapter
 164**         *AND* the GSC+ address is remapped to the Broadcast Physical
 165**         Address space by setting the 14 high order address bits of the
 166**         32 bit GSC+ address to ones.
 167**
 168** o TLB field affects transactions which are forwarded from GSC+ to Runway.
 169**   "Real" mode is the poweron default.
 170** 
 171**   TLB Mode  Value  Description
 172**   Real        0    No TLB translation. Address is directly mapped and the
 173**                    virtual address is composed of selected physical bits.
 174**   Error       1    Software fills the TLB manually.
 175**   Normal      2    IOA fetches IO TLB misses from IO PDIR (in host memory).
 176**
 177**
 178** IO_IO_LOW_HV	  +0x60 (HV dependent)
 179** IO_IO_HIGH_HV  +0x64 (HV dependent)
 180** IO_IO_LOW      +0x78	(Architected register)
 181** IO_IO_HIGH     +0x7c	(Architected register)
 182**
 183** IO_IO_LOW and IO_IO_HIGH set the lower and upper bounds of the
 184** I/O Adapter address space, respectively.
 185**
 186** 0  ... 7 | 8 ... 15 |  16   ...   31 |
 187** 11111111 | 11111111 |      address   |
 188**
 189** Each LOW/HIGH pair describes a disjoint address space region.
 190** (2 per GSC+ port). Each incoming Runway transaction address is compared
 191** with both sets of LOW/HIGH registers. If the address is in the range
 192** greater than or equal to IO_IO_LOW and less than IO_IO_HIGH the transaction
 193** for forwarded to the respective GSC+ bus.
 194** Specify IO_IO_LOW equal to or greater than IO_IO_HIGH to avoid specifying
 195** an address space region.
 196**
 197** In order for a Runway address to reside within GSC+ extended address space:
 198**	Runway Address [0:7]    must identically compare to 8'b11111111
 199**	Runway Address [8:11]   must be equal to IO_IO_LOW(_HV)[16:19]
 200** 	Runway Address [12:23]  must be greater than or equal to
 201**	           IO_IO_LOW(_HV)[20:31] and less than IO_IO_HIGH(_HV)[20:31].
 202**	Runway Address [24:39]  is not used in the comparison.
 203**
 204** When the Runway transaction is forwarded to GSC+, the GSC+ address is
 205** as follows:
 206**	GSC+ Address[0:3]	4'b1111
 207**	GSC+ Address[4:29]	Runway Address[12:37]
 208**	GSC+ Address[30:31]	2'b00
 209**
 210** All 4 Low/High registers must be initialized (by PDC) once the lower bus
 211** is interrogated and address space is defined. The operating system will
 212** modify the architectural IO_IO_LOW and IO_IO_HIGH registers following
 213** the PDC initialization.  However, the hardware version dependent IO_IO_LOW
 214** and IO_IO_HIGH registers should not be subsequently altered by the OS.
 215** 
 216** Writes to both sets of registers will take effect immediately, bypassing
 217** the queues, which ensures that subsequent Runway transactions are checked
 218** against the updated bounds values. However reads are queued, introducing
 219** the possibility of a read being bypassed by a subsequent write to the same
 220** register. This sequence can be avoided by having software wait for read
 221** returns before issuing subsequent writes.
 222*/
 223
 224struct ioc {
 225	struct ioa_registers __iomem *ioc_regs;  /* I/O MMU base address */
 226	u8  *res_map;	                /* resource map, bit == pdir entry */
 227	u64 *pdir_base;	                /* physical base address */
 228	u32 pdir_size; 			/* bytes, function of IOV Space size */
 229	u32 res_hint;	                /* next available IOVP - 
 230					   circular search */
 231	u32 res_size;		    	/* size of resource map in bytes */
 232	spinlock_t res_lock;
 233
 234#ifdef CCIO_COLLECT_STATS
 235#define CCIO_SEARCH_SAMPLE 0x100
 236	unsigned long avg_search[CCIO_SEARCH_SAMPLE];
 237	unsigned long avg_idx;		  /* current index into avg_search */
 238	unsigned long used_pages;
 239	unsigned long msingle_calls;
 240	unsigned long msingle_pages;
 241	unsigned long msg_calls;
 242	unsigned long msg_pages;
 243	unsigned long usingle_calls;
 244	unsigned long usingle_pages;
 245	unsigned long usg_calls;
 246	unsigned long usg_pages;
 247#endif
 248	unsigned short cujo20_bug;
 249
 250	/* STUFF We don't need in performance path */
 251	u32 chainid_shift; 		/* specify bit location of chain_id */
 252	struct ioc *next;		/* Linked list of discovered iocs */
 253	const char *name;		/* device name from firmware */
 254	unsigned int hw_path;           /* the hardware path this ioc is associatd with */
 255	struct pci_dev *fake_pci_dev;   /* the fake pci_dev for non-pci devs */
 256	struct resource mmio_region[2]; /* The "routed" MMIO regions */
 257};
 258
 259static struct ioc *ioc_list;
 260static int ioc_count;
 261
 262/**************************************************************
 263*
 264*   I/O Pdir Resource Management
 265*
 266*   Bits set in the resource map are in use.
 267*   Each bit can represent a number of pages.
 268*   LSbs represent lower addresses (IOVA's).
 269*
 270*   This was was copied from sba_iommu.c. Don't try to unify
 271*   the two resource managers unless a way to have different
 272*   allocation policies is also adjusted. We'd like to avoid
 273*   I/O TLB thrashing by having resource allocation policy
 274*   match the I/O TLB replacement policy.
 275*
 276***************************************************************/
 277#define IOVP_SIZE PAGE_SIZE
 278#define IOVP_SHIFT PAGE_SHIFT
 279#define IOVP_MASK PAGE_MASK
 280
 281/* Convert from IOVP to IOVA and vice versa. */
 282#define CCIO_IOVA(iovp,offset) ((iovp) | (offset))
 283#define CCIO_IOVP(iova) ((iova) & IOVP_MASK)
 284
 285#define PDIR_INDEX(iovp)    ((iovp)>>IOVP_SHIFT)
 286#define MKIOVP(pdir_idx)    ((long)(pdir_idx) << IOVP_SHIFT)
 287#define MKIOVA(iovp,offset) (dma_addr_t)((long)iovp | (long)offset)
 288
 289/*
 290** Don't worry about the 150% average search length on a miss.
 291** If the search wraps around, and passes the res_hint, it will
 292** cause the kernel to panic anyhow.
 293*/
 294#define CCIO_SEARCH_LOOP(ioc, res_idx, mask, size)  \
 295       for(; res_ptr < res_end; ++res_ptr) { \
 296		int ret;\
 297		unsigned int idx;\
 298		idx = (unsigned int)((unsigned long)res_ptr - (unsigned long)ioc->res_map); \
 299		ret = iommu_is_span_boundary(idx << 3, pages_needed, 0, boundary_size);\
 300		if ((0 == (*res_ptr & mask)) && !ret) { \
 301			*res_ptr |= mask; \
 302			res_idx = idx;\
 303			ioc->res_hint = res_idx + (size >> 3); \
 304			goto resource_found; \
 305		} \
 306	}
 307
 308#define CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, size) \
 309       u##size *res_ptr = (u##size *)&((ioc)->res_map[ioa->res_hint & ~((size >> 3) - 1)]); \
 310       u##size *res_end = (u##size *)&(ioc)->res_map[ioa->res_size]; \
 311       CCIO_SEARCH_LOOP(ioc, res_idx, mask, size); \
 312       res_ptr = (u##size *)&(ioc)->res_map[0]; \
 313       CCIO_SEARCH_LOOP(ioa, res_idx, mask, size);
 314
 315/*
 316** Find available bit in this ioa's resource map.
 317** Use a "circular" search:
 318**   o Most IOVA's are "temporary" - avg search time should be small.
 319** o keep a history of what happened for debugging
 320** o KISS.
 321**
 322** Perf optimizations:
 323** o search for log2(size) bits at a time.
 324** o search for available resource bits using byte/word/whatever.
 325** o use different search for "large" (eg > 4 pages) or "very large"
 326**   (eg > 16 pages) mappings.
 327*/
 328
 329/**
 330 * ccio_alloc_range - Allocate pages in the ioc's resource map.
 331 * @ioc: The I/O Controller.
 332 * @pages_needed: The requested number of pages to be mapped into the
 333 * I/O Pdir...
 334 *
 335 * This function searches the resource map of the ioc to locate a range
 336 * of available pages for the requested size.
 337 */
 338static int
 339ccio_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
 340{
 341	unsigned int pages_needed = size >> IOVP_SHIFT;
 342	unsigned int res_idx;
 343	unsigned long boundary_size;
 344#ifdef CCIO_COLLECT_STATS
 345	unsigned long cr_start = mfctl(16);
 346#endif
 347	
 348	BUG_ON(pages_needed == 0);
 349	BUG_ON((pages_needed * IOVP_SIZE) > DMA_CHUNK_SIZE);
 350     
 351	DBG_RES("%s() size: %d pages_needed %d\n", 
 352		__func__, size, pages_needed);
 353
 354	/*
 355	** "seek and ye shall find"...praying never hurts either...
 356	** ggg sacrifices another 710 to the computer gods.
 357	*/
 358
 359	boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1,
 360			      1ULL << IOVP_SHIFT) >> IOVP_SHIFT;
 361
 362	if (pages_needed <= 8) {
 363		/*
 364		 * LAN traffic will not thrash the TLB IFF the same NIC
 365		 * uses 8 adjacent pages to map separate payload data.
 366		 * ie the same byte in the resource bit map.
 367		 */
 368#if 0
 369		/* FIXME: bit search should shift it's way through
 370		 * an unsigned long - not byte at a time. As it is now,
 371		 * we effectively allocate this byte to this mapping.
 372		 */
 373		unsigned long mask = ~(~0UL >> pages_needed);
 374		CCIO_FIND_FREE_MAPPING(ioc, res_idx, mask, 8);
 375#else
 376		CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xff, 8);
 377#endif
 378	} else if (pages_needed <= 16) {
 379		CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xffff, 16);
 380	} else if (pages_needed <= 32) {
 381		CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~(unsigned int)0, 32);
 382#ifdef __LP64__
 383	} else if (pages_needed <= 64) {
 384		CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~0UL, 64);
 385#endif
 386	} else {
 387		panic("%s: %s() Too many pages to map. pages_needed: %u\n",
 388		       __FILE__,  __func__, pages_needed);
 389	}
 390
 391	panic("%s: %s() I/O MMU is out of mapping resources.\n", __FILE__,
 392	      __func__);
 393	
 394resource_found:
 395	
 396	DBG_RES("%s() res_idx %d res_hint: %d\n",
 397		__func__, res_idx, ioc->res_hint);
 398
 399#ifdef CCIO_COLLECT_STATS
 400	{
 401		unsigned long cr_end = mfctl(16);
 402		unsigned long tmp = cr_end - cr_start;
 403		/* check for roll over */
 404		cr_start = (cr_end < cr_start) ?  -(tmp) : (tmp);
 405	}
 406	ioc->avg_search[ioc->avg_idx++] = cr_start;
 407	ioc->avg_idx &= CCIO_SEARCH_SAMPLE - 1;
 408	ioc->used_pages += pages_needed;
 409#endif
 410	/* 
 411	** return the bit address.
 412	*/
 413	return res_idx << 3;
 414}
 415
 416#define CCIO_FREE_MAPPINGS(ioc, res_idx, mask, size) \
 417        u##size *res_ptr = (u##size *)&((ioc)->res_map[res_idx]); \
 418        BUG_ON((*res_ptr & mask) != mask); \
 419        *res_ptr &= ~(mask);
 420
 421/**
 422 * ccio_free_range - Free pages from the ioc's resource map.
 423 * @ioc: The I/O Controller.
 424 * @iova: The I/O Virtual Address.
 425 * @pages_mapped: The requested number of pages to be freed from the
 426 * I/O Pdir.
 427 *
 428 * This function frees the resouces allocated for the iova.
 429 */
 430static void
 431ccio_free_range(struct ioc *ioc, dma_addr_t iova, unsigned long pages_mapped)
 432{
 433	unsigned long iovp = CCIO_IOVP(iova);
 434	unsigned int res_idx = PDIR_INDEX(iovp) >> 3;
 435
 436	BUG_ON(pages_mapped == 0);
 437	BUG_ON((pages_mapped * IOVP_SIZE) > DMA_CHUNK_SIZE);
 438	BUG_ON(pages_mapped > BITS_PER_LONG);
 439
 440	DBG_RES("%s():  res_idx: %d pages_mapped %d\n", 
 441		__func__, res_idx, pages_mapped);
 442
 443#ifdef CCIO_COLLECT_STATS
 444	ioc->used_pages -= pages_mapped;
 445#endif
 446
 447	if(pages_mapped <= 8) {
 448#if 0
 449		/* see matching comments in alloc_range */
 450		unsigned long mask = ~(~0UL >> pages_mapped);
 451		CCIO_FREE_MAPPINGS(ioc, res_idx, mask, 8);
 452#else
 453		CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffUL, 8);
 454#endif
 455	} else if(pages_mapped <= 16) {
 456		CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffffUL, 16);
 457	} else if(pages_mapped <= 32) {
 458		CCIO_FREE_MAPPINGS(ioc, res_idx, ~(unsigned int)0, 32);
 459#ifdef __LP64__
 460	} else if(pages_mapped <= 64) {
 461		CCIO_FREE_MAPPINGS(ioc, res_idx, ~0UL, 64);
 462#endif
 463	} else {
 464		panic("%s:%s() Too many pages to unmap.\n", __FILE__,
 465		      __func__);
 466	}
 467}
 468
 469/****************************************************************
 470**
 471**          CCIO dma_ops support routines
 472**
 473*****************************************************************/
 474
 475typedef unsigned long space_t;
 476#define KERNEL_SPACE 0
 477
 478/*
 479** DMA "Page Type" and Hints 
 480** o if SAFE_DMA isn't set, mapping is for FAST_DMA. SAFE_DMA should be
 481**   set for subcacheline DMA transfers since we don't want to damage the
 482**   other part of a cacheline.
 483** o SAFE_DMA must be set for "memory" allocated via pci_alloc_consistent().
 484**   This bit tells U2 to do R/M/W for partial cachelines. "Streaming"
 485**   data can avoid this if the mapping covers full cache lines.
 486** o STOP_MOST is needed for atomicity across cachelines.
 487**   Apparently only "some EISA devices" need this.
 488**   Using CONFIG_ISA is hack. Only the IOA with EISA under it needs
 489**   to use this hint iff the EISA devices needs this feature.
 490**   According to the U2 ERS, STOP_MOST enabled pages hurt performance.
 491** o PREFETCH should *not* be set for cases like Multiple PCI devices
 492**   behind GSCtoPCI (dino) bus converter. Only one cacheline per GSC
 493**   device can be fetched and multiply DMA streams will thrash the
 494**   prefetch buffer and burn memory bandwidth. See 6.7.3 "Prefetch Rules
 495**   and Invalidation of Prefetch Entries".
 496**
 497** FIXME: the default hints need to be per GSC device - not global.
 498** 
 499** HP-UX dorks: linux device driver programming model is totally different
 500**    than HP-UX's. HP-UX always sets HINT_PREFETCH since it's drivers
 501**    do special things to work on non-coherent platforms...linux has to
 502**    be much more careful with this.
 503*/
 504#define IOPDIR_VALID    0x01UL
 505#define HINT_SAFE_DMA   0x02UL	/* used for pci_alloc_consistent() pages */
 506#ifdef CONFIG_EISA
 507#define HINT_STOP_MOST  0x04UL	/* LSL support */
 508#else
 509#define HINT_STOP_MOST  0x00UL	/* only needed for "some EISA devices" */
 510#endif
 511#define HINT_UDPATE_ENB 0x08UL  /* not used/supported by U2 */
 512#define HINT_PREFETCH   0x10UL	/* for outbound pages which are not SAFE */
 513
 514
 515/*
 516** Use direction (ie PCI_DMA_TODEVICE) to pick hint.
 517** ccio_alloc_consistent() depends on this to get SAFE_DMA
 518** when it passes in BIDIRECTIONAL flag.
 519*/
 520static u32 hint_lookup[] = {
 521	[PCI_DMA_BIDIRECTIONAL]	= HINT_STOP_MOST | HINT_SAFE_DMA | IOPDIR_VALID,
 522	[PCI_DMA_TODEVICE]	= HINT_STOP_MOST | HINT_PREFETCH | IOPDIR_VALID,
 523	[PCI_DMA_FROMDEVICE]	= HINT_STOP_MOST | IOPDIR_VALID,
 524};
 525
 526/**
 527 * ccio_io_pdir_entry - Initialize an I/O Pdir.
 528 * @pdir_ptr: A pointer into I/O Pdir.
 529 * @sid: The Space Identifier.
 530 * @vba: The virtual address.
 531 * @hints: The DMA Hint.
 532 *
 533 * Given a virtual address (vba, arg2) and space id, (sid, arg1),
 534 * load the I/O PDIR entry pointed to by pdir_ptr (arg0). Each IO Pdir
 535 * entry consists of 8 bytes as shown below (MSB == bit 0):
 536 *
 537 *
 538 * WORD 0:
 539 * +------+----------------+-----------------------------------------------+
 540 * | Phys | Virtual Index  |               Phys                            |
 541 * | 0:3  |     0:11       |               4:19                            |
 542 * |4 bits|   12 bits      |              16 bits                          |
 543 * +------+----------------+-----------------------------------------------+
 544 * WORD 1:
 545 * +-----------------------+-----------------------------------------------+
 546 * |      Phys    |  Rsvd  | Prefetch |Update |Rsvd  |Lock  |Safe  |Valid  |
 547 * |     20:39    |        | Enable   |Enable |      |Enable|DMA   |       |
 548 * |    20 bits   | 5 bits | 1 bit    |1 bit  |2 bits|1 bit |1 bit |1 bit  |
 549 * +-----------------------+-----------------------------------------------+
 550 *
 551 * The virtual index field is filled with the results of the LCI
 552 * (Load Coherence Index) instruction.  The 8 bits used for the virtual
 553 * index are bits 12:19 of the value returned by LCI.
 554 */ 
 555static void CCIO_INLINE
 556ccio_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
 557		   unsigned long hints)
 558{
 559	register unsigned long pa;
 560	register unsigned long ci; /* coherent index */
 561
 562	/* We currently only support kernel addresses */
 563	BUG_ON(sid != KERNEL_SPACE);
 564
 565	mtsp(sid,1);
 566
 567	/*
 568	** WORD 1 - low order word
 569	** "hints" parm includes the VALID bit!
 570	** "dep" clobbers the physical address offset bits as well.
 571	*/
 572	pa = virt_to_phys(vba);
 573	asm volatile("depw  %1,31,12,%0" : "+r" (pa) : "r" (hints));
 574	((u32 *)pdir_ptr)[1] = (u32) pa;
 575
 576	/*
 577	** WORD 0 - high order word
 578	*/
 579
 580#ifdef __LP64__
 581	/*
 582	** get bits 12:15 of physical address
 583	** shift bits 16:31 of physical address
 584	** and deposit them
 585	*/
 586	asm volatile ("extrd,u %1,15,4,%0" : "=r" (ci) : "r" (pa));
 587	asm volatile ("extrd,u %1,31,16,%0" : "+r" (pa) : "r" (pa));
 588	asm volatile ("depd  %1,35,4,%0" : "+r" (pa) : "r" (ci));
 589#else
 590	pa = 0;
 591#endif
 592	/*
 593	** get CPU coherency index bits
 594	** Grab virtual index [0:11]
 595	** Deposit virt_idx bits into I/O PDIR word
 596	*/
 597	asm volatile ("lci %%r0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
 598	asm volatile ("extru %1,19,12,%0" : "+r" (ci) : "r" (ci));
 599	asm volatile ("depw  %1,15,12,%0" : "+r" (pa) : "r" (ci));
 600
 601	((u32 *)pdir_ptr)[0] = (u32) pa;
 602
 603
 604	/* FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360)
 605	**        PCX-U/U+ do. (eg C200/C240)
 606	**        PCX-T'? Don't know. (eg C110 or similar K-class)
 607	**
 608	** See PDC_MODEL/option 0/SW_CAP word for "Non-coherent IO-PDIR bit".
 609	** Hopefully we can patch (NOP) these out at boot time somehow.
 610	**
 611	** "Since PCX-U employs an offset hash that is incompatible with
 612	** the real mode coherence index generation of U2, the PDIR entry
 613	** must be flushed to memory to retain coherence."
 614	*/
 615	asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
 616	asm volatile("sync");
 617}
 618
 619/**
 620 * ccio_clear_io_tlb - Remove stale entries from the I/O TLB.
 621 * @ioc: The I/O Controller.
 622 * @iovp: The I/O Virtual Page.
 623 * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir.
 624 *
 625 * Purge invalid I/O PDIR entries from the I/O TLB.
 626 *
 627 * FIXME: Can we change the byte_cnt to pages_mapped?
 628 */
 629static CCIO_INLINE void
 630ccio_clear_io_tlb(struct ioc *ioc, dma_addr_t iovp, size_t byte_cnt)
 631{
 632	u32 chain_size = 1 << ioc->chainid_shift;
 633
 634	iovp &= IOVP_MASK;	/* clear offset bits, just want pagenum */
 635	byte_cnt += chain_size;
 636
 637	while(byte_cnt > chain_size) {
 638		WRITE_U32(CMD_TLB_PURGE | iovp, &ioc->ioc_regs->io_command);
 639		iovp += chain_size;
 640		byte_cnt -= chain_size;
 641	}
 642}
 643
 644/**
 645 * ccio_mark_invalid - Mark the I/O Pdir entries invalid.
 646 * @ioc: The I/O Controller.
 647 * @iova: The I/O Virtual Address.
 648 * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir.
 649 *
 650 * Mark the I/O Pdir entries invalid and blow away the corresponding I/O
 651 * TLB entries.
 652 *
 653 * FIXME: at some threshold it might be "cheaper" to just blow
 654 *        away the entire I/O TLB instead of individual entries.
 655 *
 656 * FIXME: Uturn has 256 TLB entries. We don't need to purge every
 657 *        PDIR entry - just once for each possible TLB entry.
 658 *        (We do need to maker I/O PDIR entries invalid regardless).
 659 *
 660 * FIXME: Can we change byte_cnt to pages_mapped?
 661 */ 
 662static CCIO_INLINE void
 663ccio_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
 664{
 665	u32 iovp = (u32)CCIO_IOVP(iova);
 666	size_t saved_byte_cnt;
 667
 668	/* round up to nearest page size */
 669	saved_byte_cnt = byte_cnt = ALIGN(byte_cnt, IOVP_SIZE);
 670
 671	while(byte_cnt > 0) {
 672		/* invalidate one page at a time */
 673		unsigned int idx = PDIR_INDEX(iovp);
 674		char *pdir_ptr = (char *) &(ioc->pdir_base[idx]);
 675
 676		BUG_ON(idx >= (ioc->pdir_size / sizeof(u64)));
 677		pdir_ptr[7] = 0;	/* clear only VALID bit */ 
 678		/*
 679		** FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360)
 680		**   PCX-U/U+ do. (eg C200/C240)
 681		** See PDC_MODEL/option 0/SW_CAP for "Non-coherent IO-PDIR bit".
 682		**
 683		** Hopefully someone figures out how to patch (NOP) the
 684		** FDC/SYNC out at boot time.
 685		*/
 686		asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr[7]));
 687
 688		iovp     += IOVP_SIZE;
 689		byte_cnt -= IOVP_SIZE;
 690	}
 691
 692	asm volatile("sync");
 693	ccio_clear_io_tlb(ioc, CCIO_IOVP(iova), saved_byte_cnt);
 694}
 695
 696/****************************************************************
 697**
 698**          CCIO dma_ops
 699**
 700*****************************************************************/
 701
 702/**
 703 * ccio_dma_supported - Verify the IOMMU supports the DMA address range.
 704 * @dev: The PCI device.
 705 * @mask: A bit mask describing the DMA address range of the device.
 706 *
 707 * This function implements the pci_dma_supported function.
 708 */
 709static int 
 710ccio_dma_supported(struct device *dev, u64 mask)
 711{
 712	if(dev == NULL) {
 713		printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
 714		BUG();
 715		return 0;
 716	}
 717
 718	/* only support 32-bit devices (ie PCI/GSC) */
 719	return (int)(mask == 0xffffffffUL);
 720}
 721
 722/**
 723 * ccio_map_single - Map an address range into the IOMMU.
 724 * @dev: The PCI device.
 725 * @addr: The start address of the DMA region.
 726 * @size: The length of the DMA region.
 727 * @direction: The direction of the DMA transaction (to/from device).
 728 *
 729 * This function implements the pci_map_single function.
 730 */
 731static dma_addr_t 
 732ccio_map_single(struct device *dev, void *addr, size_t size,
 733		enum dma_data_direction direction)
 734{
 735	int idx;
 736	struct ioc *ioc;
 737	unsigned long flags;
 738	dma_addr_t iovp;
 739	dma_addr_t offset;
 740	u64 *pdir_start;
 741	unsigned long hint = hint_lookup[(int)direction];
 742
 743	BUG_ON(!dev);
 744	ioc = GET_IOC(dev);
 
 
 745
 746	BUG_ON(size <= 0);
 747
 748	/* save offset bits */
 749	offset = ((unsigned long) addr) & ~IOVP_MASK;
 750
 751	/* round up to nearest IOVP_SIZE */
 752	size = ALIGN(size + offset, IOVP_SIZE);
 753	spin_lock_irqsave(&ioc->res_lock, flags);
 754
 755#ifdef CCIO_COLLECT_STATS
 756	ioc->msingle_calls++;
 757	ioc->msingle_pages += size >> IOVP_SHIFT;
 758#endif
 759
 760	idx = ccio_alloc_range(ioc, dev, size);
 761	iovp = (dma_addr_t)MKIOVP(idx);
 762
 763	pdir_start = &(ioc->pdir_base[idx]);
 764
 765	DBG_RUN("%s() 0x%p -> 0x%lx size: %0x%x\n",
 766		__func__, addr, (long)iovp | offset, size);
 767
 768	/* If not cacheline aligned, force SAFE_DMA on the whole mess */
 769	if((size % L1_CACHE_BYTES) || ((unsigned long)addr % L1_CACHE_BYTES))
 770		hint |= HINT_SAFE_DMA;
 771
 772	while(size > 0) {
 773		ccio_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long)addr, hint);
 774
 775		DBG_RUN(" pdir %p %08x%08x\n",
 776			pdir_start,
 777			(u32) (((u32 *) pdir_start)[0]),
 778			(u32) (((u32 *) pdir_start)[1]));
 779		++pdir_start;
 780		addr += IOVP_SIZE;
 781		size -= IOVP_SIZE;
 782	}
 783
 784	spin_unlock_irqrestore(&ioc->res_lock, flags);
 785
 786	/* form complete address */
 787	return CCIO_IOVA(iovp, offset);
 788}
 789
 
 
 
 
 
 
 
 
 
 
 
 790/**
 791 * ccio_unmap_single - Unmap an address range from the IOMMU.
 792 * @dev: The PCI device.
 793 * @addr: The start address of the DMA region.
 794 * @size: The length of the DMA region.
 795 * @direction: The direction of the DMA transaction (to/from device).
 796 *
 797 * This function implements the pci_unmap_single function.
 798 */
 799static void 
 800ccio_unmap_single(struct device *dev, dma_addr_t iova, size_t size, 
 801		  enum dma_data_direction direction)
 802{
 803	struct ioc *ioc;
 804	unsigned long flags; 
 805	dma_addr_t offset = iova & ~IOVP_MASK;
 806	
 807	BUG_ON(!dev);
 808	ioc = GET_IOC(dev);
 
 
 
 
 809
 810	DBG_RUN("%s() iovp 0x%lx/%x\n",
 811		__func__, (long)iova, size);
 812
 813	iova ^= offset;        /* clear offset bits */
 814	size += offset;
 815	size = ALIGN(size, IOVP_SIZE);
 816
 817	spin_lock_irqsave(&ioc->res_lock, flags);
 818
 819#ifdef CCIO_COLLECT_STATS
 820	ioc->usingle_calls++;
 821	ioc->usingle_pages += size >> IOVP_SHIFT;
 822#endif
 823
 824	ccio_mark_invalid(ioc, iova, size);
 825	ccio_free_range(ioc, iova, (size >> IOVP_SHIFT));
 826	spin_unlock_irqrestore(&ioc->res_lock, flags);
 827}
 828
 829/**
 830 * ccio_alloc_consistent - Allocate a consistent DMA mapping.
 831 * @dev: The PCI device.
 832 * @size: The length of the DMA region.
 833 * @dma_handle: The DMA address handed back to the device (not the cpu).
 834 *
 835 * This function implements the pci_alloc_consistent function.
 836 */
 837static void * 
 838ccio_alloc_consistent(struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag)
 
 839{
 840      void *ret;
 841#if 0
 842/* GRANT Need to establish hierarchy for non-PCI devs as well
 843** and then provide matching gsc_map_xxx() functions for them as well.
 844*/
 845	if(!hwdev) {
 846		/* only support PCI */
 847		*dma_handle = 0;
 848		return 0;
 849	}
 850#endif
 851        ret = (void *) __get_free_pages(flag, get_order(size));
 852
 853	if (ret) {
 854		memset(ret, 0, size);
 855		*dma_handle = ccio_map_single(dev, ret, size, PCI_DMA_BIDIRECTIONAL);
 856	}
 857
 858	return ret;
 859}
 860
 861/**
 862 * ccio_free_consistent - Free a consistent DMA mapping.
 863 * @dev: The PCI device.
 864 * @size: The length of the DMA region.
 865 * @cpu_addr: The cpu address returned from the ccio_alloc_consistent.
 866 * @dma_handle: The device address returned from the ccio_alloc_consistent.
 867 *
 868 * This function implements the pci_free_consistent function.
 869 */
 870static void 
 871ccio_free_consistent(struct device *dev, size_t size, void *cpu_addr, 
 872		     dma_addr_t dma_handle)
 873{
 874	ccio_unmap_single(dev, dma_handle, size, 0);
 875	free_pages((unsigned long)cpu_addr, get_order(size));
 876}
 877
 878/*
 879** Since 0 is a valid pdir_base index value, can't use that
 880** to determine if a value is valid or not. Use a flag to indicate
 881** the SG list entry contains a valid pdir index.
 882*/
 883#define PIDE_FLAG 0x80000000UL
 884
 885#ifdef CCIO_COLLECT_STATS
 886#define IOMMU_MAP_STATS
 887#endif
 888#include "iommu-helpers.h"
 889
 890/**
 891 * ccio_map_sg - Map the scatter/gather list into the IOMMU.
 892 * @dev: The PCI device.
 893 * @sglist: The scatter/gather list to be mapped in the IOMMU.
 894 * @nents: The number of entries in the scatter/gather list.
 895 * @direction: The direction of the DMA transaction (to/from device).
 896 *
 897 * This function implements the pci_map_sg function.
 898 */
 899static int
 900ccio_map_sg(struct device *dev, struct scatterlist *sglist, int nents, 
 901	    enum dma_data_direction direction)
 902{
 903	struct ioc *ioc;
 904	int coalesced, filled = 0;
 905	unsigned long flags;
 906	unsigned long hint = hint_lookup[(int)direction];
 907	unsigned long prev_len = 0, current_len = 0;
 908	int i;
 909	
 910	BUG_ON(!dev);
 911	ioc = GET_IOC(dev);
 
 
 912	
 913	DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
 914
 915	/* Fast path single entry scatterlists. */
 916	if (nents == 1) {
 917		sg_dma_address(sglist) = ccio_map_single(dev,
 918				(void *)sg_virt_addr(sglist), sglist->length,
 919				direction);
 920		sg_dma_len(sglist) = sglist->length;
 921		return 1;
 922	}
 923
 924	for(i = 0; i < nents; i++)
 925		prev_len += sglist[i].length;
 926	
 927	spin_lock_irqsave(&ioc->res_lock, flags);
 928
 929#ifdef CCIO_COLLECT_STATS
 930	ioc->msg_calls++;
 931#endif
 932
 933	/*
 934	** First coalesce the chunks and allocate I/O pdir space
 935	**
 936	** If this is one DMA stream, we can properly map using the
 937	** correct virtual address associated with each DMA page.
 938	** w/o this association, we wouldn't have coherent DMA!
 939	** Access to the virtual address is what forces a two pass algorithm.
 940	*/
 941	coalesced = iommu_coalesce_chunks(ioc, dev, sglist, nents, ccio_alloc_range);
 942
 943	/*
 944	** Program the I/O Pdir
 945	**
 946	** map the virtual addresses to the I/O Pdir
 947	** o dma_address will contain the pdir index
 948	** o dma_len will contain the number of bytes to map 
 949	** o page/offset contain the virtual address.
 950	*/
 951	filled = iommu_fill_pdir(ioc, sglist, nents, hint, ccio_io_pdir_entry);
 952
 953	spin_unlock_irqrestore(&ioc->res_lock, flags);
 954
 955	BUG_ON(coalesced != filled);
 956
 957	DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
 958
 959	for (i = 0; i < filled; i++)
 960		current_len += sg_dma_len(sglist + i);
 961
 962	BUG_ON(current_len != prev_len);
 963
 964	return filled;
 965}
 966
 967/**
 968 * ccio_unmap_sg - Unmap the scatter/gather list from the IOMMU.
 969 * @dev: The PCI device.
 970 * @sglist: The scatter/gather list to be unmapped from the IOMMU.
 971 * @nents: The number of entries in the scatter/gather list.
 972 * @direction: The direction of the DMA transaction (to/from device).
 973 *
 974 * This function implements the pci_unmap_sg function.
 975 */
 976static void 
 977ccio_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents, 
 978	      enum dma_data_direction direction)
 979{
 980	struct ioc *ioc;
 981
 982	BUG_ON(!dev);
 983	ioc = GET_IOC(dev);
 
 
 
 
 984
 985	DBG_RUN_SG("%s() START %d entries,  %08lx,%x\n",
 986		__func__, nents, sg_virt_addr(sglist), sglist->length);
 987
 988#ifdef CCIO_COLLECT_STATS
 989	ioc->usg_calls++;
 990#endif
 991
 992	while(sg_dma_len(sglist) && nents--) {
 993
 994#ifdef CCIO_COLLECT_STATS
 995		ioc->usg_pages += sg_dma_len(sglist) >> PAGE_SHIFT;
 996#endif
 997		ccio_unmap_single(dev, sg_dma_address(sglist),
 998				  sg_dma_len(sglist), direction);
 999		++sglist;
1000	}
1001
1002	DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents);
1003}
1004
1005static struct hppa_dma_ops ccio_ops = {
 
 
 
 
 
1006	.dma_supported =	ccio_dma_supported,
1007	.alloc_consistent =	ccio_alloc_consistent,
1008	.alloc_noncoherent =	ccio_alloc_consistent,
1009	.free_consistent =	ccio_free_consistent,
1010	.map_single =		ccio_map_single,
1011	.unmap_single =		ccio_unmap_single,
1012	.map_sg = 		ccio_map_sg,
1013	.unmap_sg = 		ccio_unmap_sg,
1014	.dma_sync_single_for_cpu =	NULL,	/* NOP for U2/Uturn */
1015	.dma_sync_single_for_device =	NULL,	/* NOP for U2/Uturn */
1016	.dma_sync_sg_for_cpu =		NULL,	/* ditto */
1017	.dma_sync_sg_for_device =		NULL,	/* ditto */
1018};
1019
1020#ifdef CONFIG_PROC_FS
1021static int ccio_proc_info(struct seq_file *m, void *p)
1022{
1023	int len = 0;
1024	struct ioc *ioc = ioc_list;
1025
1026	while (ioc != NULL) {
1027		unsigned int total_pages = ioc->res_size << 3;
1028#ifdef CCIO_COLLECT_STATS
1029		unsigned long avg = 0, min, max;
1030		int j;
1031#endif
1032
1033		len += seq_printf(m, "%s\n", ioc->name);
1034		
1035		len += seq_printf(m, "Cujo 2.0 bug    : %s\n",
1036				  (ioc->cujo20_bug ? "yes" : "no"));
1037		
1038		len += seq_printf(m, "IO PDIR size    : %d bytes (%d entries)\n",
1039			       total_pages * 8, total_pages);
1040
1041#ifdef CCIO_COLLECT_STATS
1042		len += seq_printf(m, "IO PDIR entries : %ld free  %ld used (%d%%)\n",
1043				  total_pages - ioc->used_pages, ioc->used_pages,
1044				  (int)(ioc->used_pages * 100 / total_pages));
1045#endif
1046
1047		len += seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n", 
1048				  ioc->res_size, total_pages);
1049
1050#ifdef CCIO_COLLECT_STATS
1051		min = max = ioc->avg_search[0];
1052		for(j = 0; j < CCIO_SEARCH_SAMPLE; ++j) {
1053			avg += ioc->avg_search[j];
1054			if(ioc->avg_search[j] > max) 
1055				max = ioc->avg_search[j];
1056			if(ioc->avg_search[j] < min) 
1057				min = ioc->avg_search[j];
1058		}
1059		avg /= CCIO_SEARCH_SAMPLE;
1060		len += seq_printf(m, "  Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1061				  min, avg, max);
1062
1063		len += seq_printf(m, "pci_map_single(): %8ld calls  %8ld pages (avg %d/1000)\n",
1064				  ioc->msingle_calls, ioc->msingle_pages,
1065				  (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1066
1067		/* KLUGE - unmap_sg calls unmap_single for each mapped page */
1068		min = ioc->usingle_calls - ioc->usg_calls;
1069		max = ioc->usingle_pages - ioc->usg_pages;
1070		len += seq_printf(m, "pci_unmap_single: %8ld calls  %8ld pages (avg %d/1000)\n",
1071				  min, max, (int)((max * 1000)/min));
1072 
1073		len += seq_printf(m, "pci_map_sg()    : %8ld calls  %8ld pages (avg %d/1000)\n",
1074				  ioc->msg_calls, ioc->msg_pages,
1075				  (int)((ioc->msg_pages * 1000)/ioc->msg_calls));
1076
1077		len += seq_printf(m, "pci_unmap_sg()  : %8ld calls  %8ld pages (avg %d/1000)\n\n\n",
1078				  ioc->usg_calls, ioc->usg_pages,
1079				  (int)((ioc->usg_pages * 1000)/ioc->usg_calls));
1080#endif	/* CCIO_COLLECT_STATS */
1081
1082		ioc = ioc->next;
1083	}
1084
1085	return 0;
1086}
1087
1088static int ccio_proc_info_open(struct inode *inode, struct file *file)
1089{
1090	return single_open(file, &ccio_proc_info, NULL);
1091}
1092
1093static const struct file_operations ccio_proc_info_fops = {
1094	.owner = THIS_MODULE,
1095	.open = ccio_proc_info_open,
1096	.read = seq_read,
1097	.llseek = seq_lseek,
1098	.release = single_release,
1099};
1100
1101static int ccio_proc_bitmap_info(struct seq_file *m, void *p)
1102{
1103	int len = 0;
1104	struct ioc *ioc = ioc_list;
1105
1106	while (ioc != NULL) {
1107		u32 *res_ptr = (u32 *)ioc->res_map;
1108		int j;
1109
1110		for (j = 0; j < (ioc->res_size / sizeof(u32)); j++) {
1111			if ((j & 7) == 0)
1112				len += seq_puts(m, "\n   ");
1113			len += seq_printf(m, "%08x", *res_ptr);
1114			res_ptr++;
1115		}
1116		len += seq_puts(m, "\n\n");
1117		ioc = ioc->next;
1118		break; /* XXX - remove me */
1119	}
1120
1121	return 0;
1122}
1123
1124static int ccio_proc_bitmap_open(struct inode *inode, struct file *file)
1125{
1126	return single_open(file, &ccio_proc_bitmap_info, NULL);
1127}
1128
1129static const struct file_operations ccio_proc_bitmap_fops = {
1130	.owner = THIS_MODULE,
1131	.open = ccio_proc_bitmap_open,
1132	.read = seq_read,
1133	.llseek = seq_lseek,
1134	.release = single_release,
1135};
1136#endif /* CONFIG_PROC_FS */
1137
1138/**
1139 * ccio_find_ioc - Find the ioc in the ioc_list
1140 * @hw_path: The hardware path of the ioc.
1141 *
1142 * This function searches the ioc_list for an ioc that matches
1143 * the provide hardware path.
1144 */
1145static struct ioc * ccio_find_ioc(int hw_path)
1146{
1147	int i;
1148	struct ioc *ioc;
1149
1150	ioc = ioc_list;
1151	for (i = 0; i < ioc_count; i++) {
1152		if (ioc->hw_path == hw_path)
1153			return ioc;
1154
1155		ioc = ioc->next;
1156	}
1157
1158	return NULL;
1159}
1160
1161/**
1162 * ccio_get_iommu - Find the iommu which controls this device
1163 * @dev: The parisc device.
1164 *
1165 * This function searches through the registered IOMMU's and returns
1166 * the appropriate IOMMU for the device based on its hardware path.
1167 */
1168void * ccio_get_iommu(const struct parisc_device *dev)
1169{
1170	dev = find_pa_parent_type(dev, HPHW_IOA);
1171	if (!dev)
1172		return NULL;
1173
1174	return ccio_find_ioc(dev->hw_path);
1175}
1176
1177#define CUJO_20_STEP       0x10000000	/* inc upper nibble */
1178
1179/* Cujo 2.0 has a bug which will silently corrupt data being transferred
1180 * to/from certain pages.  To avoid this happening, we mark these pages
1181 * as `used', and ensure that nothing will try to allocate from them.
1182 */
1183void ccio_cujo20_fixup(struct parisc_device *cujo, u32 iovp)
1184{
1185	unsigned int idx;
1186	struct parisc_device *dev = parisc_parent(cujo);
1187	struct ioc *ioc = ccio_get_iommu(dev);
1188	u8 *res_ptr;
1189
1190	ioc->cujo20_bug = 1;
1191	res_ptr = ioc->res_map;
1192	idx = PDIR_INDEX(iovp) >> 3;
1193
1194	while (idx < ioc->res_size) {
1195 		res_ptr[idx] |= 0xff;
1196		idx += PDIR_INDEX(CUJO_20_STEP) >> 3;
1197	}
1198}
1199
1200#if 0
1201/* GRANT -  is this needed for U2 or not? */
1202
1203/*
1204** Get the size of the I/O TLB for this I/O MMU.
1205**
1206** If spa_shift is non-zero (ie probably U2),
1207** then calculate the I/O TLB size using spa_shift.
1208**
1209** Otherwise we are supposed to get the IODC entry point ENTRY TLB
1210** and execute it. However, both U2 and Uturn firmware supplies spa_shift.
1211** I think only Java (K/D/R-class too?) systems don't do this.
1212*/
1213static int
1214ccio_get_iotlb_size(struct parisc_device *dev)
1215{
1216	if (dev->spa_shift == 0) {
1217		panic("%s() : Can't determine I/O TLB size.\n", __func__);
1218	}
1219	return (1 << dev->spa_shift);
1220}
1221#else
1222
1223/* Uturn supports 256 TLB entries */
1224#define CCIO_CHAINID_SHIFT	8
1225#define CCIO_CHAINID_MASK	0xff
1226#endif /* 0 */
1227
1228/* We *can't* support JAVA (T600). Venture there at your own risk. */
1229static const struct parisc_device_id ccio_tbl[] = {
1230	{ HPHW_IOA, HVERSION_REV_ANY_ID, U2_IOA_RUNWAY, 0xb }, /* U2 */
1231	{ HPHW_IOA, HVERSION_REV_ANY_ID, UTURN_IOA_RUNWAY, 0xb }, /* UTurn */
1232	{ 0, }
1233};
1234
1235static int ccio_probe(struct parisc_device *dev);
1236
1237static struct parisc_driver ccio_driver = {
1238	.name =		"ccio",
1239	.id_table =	ccio_tbl,
1240	.probe =	ccio_probe,
1241};
1242
1243/**
1244 * ccio_ioc_init - Initialize the I/O Controller
1245 * @ioc: The I/O Controller.
1246 *
1247 * Initialize the I/O Controller which includes setting up the
1248 * I/O Page Directory, the resource map, and initalizing the
1249 * U2/Uturn chip into virtual mode.
1250 */
1251static void
1252ccio_ioc_init(struct ioc *ioc)
1253{
1254	int i;
1255	unsigned int iov_order;
1256	u32 iova_space_size;
1257
1258	/*
1259	** Determine IOVA Space size from memory size.
1260	**
1261	** Ideally, PCI drivers would register the maximum number
1262	** of DMA they can have outstanding for each device they
1263	** own.  Next best thing would be to guess how much DMA
1264	** can be outstanding based on PCI Class/sub-class. Both
1265	** methods still require some "extra" to support PCI
1266	** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1267	*/
1268
1269	iova_space_size = (u32) (totalram_pages / count_parisc_driver(&ccio_driver));
1270
1271	/* limit IOVA space size to 1MB-1GB */
1272
1273	if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1274		iova_space_size =  1 << (20 - PAGE_SHIFT);
1275#ifdef __LP64__
1276	} else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1277		iova_space_size =  1 << (30 - PAGE_SHIFT);
1278#endif
1279	}
1280
1281	/*
1282	** iova space must be log2() in size.
1283	** thus, pdir/res_map will also be log2().
1284	*/
1285
1286	/* We could use larger page sizes in order to *decrease* the number
1287	** of mappings needed.  (ie 8k pages means 1/2 the mappings).
1288	**
1289	** Note: Grant Grunder says "Using 8k I/O pages isn't trivial either
1290	**   since the pages must also be physically contiguous - typically
1291	**   this is the case under linux."
1292	*/
1293
1294	iov_order = get_order(iova_space_size << PAGE_SHIFT);
1295
1296	/* iova_space_size is now bytes, not pages */
1297	iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1298
1299	ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1300
1301	BUG_ON(ioc->pdir_size > 8 * 1024 * 1024);   /* max pdir size <= 8MB */
1302
1303	/* Verify it's a power of two */
1304	BUG_ON((1 << get_order(ioc->pdir_size)) != (ioc->pdir_size >> PAGE_SHIFT));
1305
1306	DBG_INIT("%s() hpa 0x%p mem %luMB IOV %dMB (%d bits)\n",
1307			__func__, ioc->ioc_regs,
1308			(unsigned long) totalram_pages >> (20 - PAGE_SHIFT),
1309			iova_space_size>>20,
1310			iov_order + PAGE_SHIFT);
1311
1312	ioc->pdir_base = (u64 *)__get_free_pages(GFP_KERNEL, 
1313						 get_order(ioc->pdir_size));
1314	if(NULL == ioc->pdir_base) {
1315		panic("%s() could not allocate I/O Page Table\n", __func__);
1316	}
1317	memset(ioc->pdir_base, 0, ioc->pdir_size);
1318
1319	BUG_ON((((unsigned long)ioc->pdir_base) & PAGE_MASK) != (unsigned long)ioc->pdir_base);
1320	DBG_INIT(" base %p\n", ioc->pdir_base);
1321
1322	/* resource map size dictated by pdir_size */
1323 	ioc->res_size = (ioc->pdir_size / sizeof(u64)) >> 3;
1324	DBG_INIT("%s() res_size 0x%x\n", __func__, ioc->res_size);
1325	
1326	ioc->res_map = (u8 *)__get_free_pages(GFP_KERNEL, 
1327					      get_order(ioc->res_size));
1328	if(NULL == ioc->res_map) {
1329		panic("%s() could not allocate resource map\n", __func__);
1330	}
1331	memset(ioc->res_map, 0, ioc->res_size);
1332
1333	/* Initialize the res_hint to 16 */
1334	ioc->res_hint = 16;
1335
1336	/* Initialize the spinlock */
1337	spin_lock_init(&ioc->res_lock);
1338
1339	/*
1340	** Chainid is the upper most bits of an IOVP used to determine
1341	** which TLB entry an IOVP will use.
1342	*/
1343	ioc->chainid_shift = get_order(iova_space_size) + PAGE_SHIFT - CCIO_CHAINID_SHIFT;
1344	DBG_INIT(" chainid_shift 0x%x\n", ioc->chainid_shift);
1345
1346	/*
1347	** Initialize IOA hardware
1348	*/
1349	WRITE_U32(CCIO_CHAINID_MASK << ioc->chainid_shift, 
1350		  &ioc->ioc_regs->io_chain_id_mask);
1351
1352	WRITE_U32(virt_to_phys(ioc->pdir_base), 
1353		  &ioc->ioc_regs->io_pdir_base);
1354
1355	/*
1356	** Go to "Virtual Mode"
1357	*/
1358	WRITE_U32(IOA_NORMAL_MODE, &ioc->ioc_regs->io_control);
1359
1360	/*
1361	** Initialize all I/O TLB entries to 0 (Valid bit off).
1362	*/
1363	WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_m);
1364	WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_l);
1365
1366	for(i = 1 << CCIO_CHAINID_SHIFT; i ; i--) {
1367		WRITE_U32((CMD_TLB_DIRECT_WRITE | (i << ioc->chainid_shift)),
1368			  &ioc->ioc_regs->io_command);
1369	}
1370}
1371
1372static void __init
1373ccio_init_resource(struct resource *res, char *name, void __iomem *ioaddr)
1374{
1375	int result;
1376
1377	res->parent = NULL;
1378	res->flags = IORESOURCE_MEM;
1379	/*
1380	 * bracing ((signed) ...) are required for 64bit kernel because
1381	 * we only want to sign extend the lower 16 bits of the register.
1382	 * The upper 16-bits of range registers are hardcoded to 0xffff.
1383	 */
1384	res->start = (unsigned long)((signed) READ_U32(ioaddr) << 16);
1385	res->end = (unsigned long)((signed) (READ_U32(ioaddr + 4) << 16) - 1);
1386	res->name = name;
1387	/*
1388	 * Check if this MMIO range is disable
1389	 */
1390	if (res->end + 1 == res->start)
1391		return;
1392
1393	/* On some platforms (e.g. K-Class), we have already registered
1394	 * resources for devices reported by firmware. Some are children
1395	 * of ccio.
1396	 * "insert" ccio ranges in the mmio hierarchy (/proc/iomem).
1397	 */
1398	result = insert_resource(&iomem_resource, res);
1399	if (result < 0) {
1400		printk(KERN_ERR "%s() failed to claim CCIO bus address space (%08lx,%08lx)\n", 
1401			__func__, (unsigned long)res->start, (unsigned long)res->end);
1402	}
1403}
1404
1405static void __init ccio_init_resources(struct ioc *ioc)
1406{
1407	struct resource *res = ioc->mmio_region;
1408	char *name = kmalloc(14, GFP_KERNEL);
1409
1410	snprintf(name, 14, "GSC Bus [%d/]", ioc->hw_path);
1411
1412	ccio_init_resource(res, name, &ioc->ioc_regs->io_io_low);
1413	ccio_init_resource(res + 1, name, &ioc->ioc_regs->io_io_low_hv);
1414}
1415
1416static int new_ioc_area(struct resource *res, unsigned long size,
1417		unsigned long min, unsigned long max, unsigned long align)
1418{
1419	if (max <= min)
1420		return -EBUSY;
1421
1422	res->start = (max - size + 1) &~ (align - 1);
1423	res->end = res->start + size;
1424	
1425	/* We might be trying to expand the MMIO range to include
1426	 * a child device that has already registered it's MMIO space.
1427	 * Use "insert" instead of request_resource().
1428	 */
1429	if (!insert_resource(&iomem_resource, res))
1430		return 0;
1431
1432	return new_ioc_area(res, size, min, max - size, align);
1433}
1434
1435static int expand_ioc_area(struct resource *res, unsigned long size,
1436		unsigned long min, unsigned long max, unsigned long align)
1437{
1438	unsigned long start, len;
1439
1440	if (!res->parent)
1441		return new_ioc_area(res, size, min, max, align);
1442
1443	start = (res->start - size) &~ (align - 1);
1444	len = res->end - start + 1;
1445	if (start >= min) {
1446		if (!adjust_resource(res, start, len))
1447			return 0;
1448	}
1449
1450	start = res->start;
1451	len = ((size + res->end + align) &~ (align - 1)) - start;
1452	if (start + len <= max) {
1453		if (!adjust_resource(res, start, len))
1454			return 0;
1455	}
1456
1457	return -EBUSY;
1458}
1459
1460/*
1461 * Dino calls this function.  Beware that we may get called on systems
1462 * which have no IOC (725, B180, C160L, etc) but do have a Dino.
1463 * So it's legal to find no parent IOC.
1464 *
1465 * Some other issues: one of the resources in the ioc may be unassigned.
1466 */
1467int ccio_allocate_resource(const struct parisc_device *dev,
1468		struct resource *res, unsigned long size,
1469		unsigned long min, unsigned long max, unsigned long align)
1470{
1471	struct resource *parent = &iomem_resource;
1472	struct ioc *ioc = ccio_get_iommu(dev);
1473	if (!ioc)
1474		goto out;
1475
1476	parent = ioc->mmio_region;
1477	if (parent->parent &&
1478	    !allocate_resource(parent, res, size, min, max, align, NULL, NULL))
1479		return 0;
1480
1481	if ((parent + 1)->parent &&
1482	    !allocate_resource(parent + 1, res, size, min, max, align,
1483				NULL, NULL))
1484		return 0;
1485
1486	if (!expand_ioc_area(parent, size, min, max, align)) {
1487		__raw_writel(((parent->start)>>16) | 0xffff0000,
1488			     &ioc->ioc_regs->io_io_low);
1489		__raw_writel(((parent->end)>>16) | 0xffff0000,
1490			     &ioc->ioc_regs->io_io_high);
1491	} else if (!expand_ioc_area(parent + 1, size, min, max, align)) {
1492		parent++;
1493		__raw_writel(((parent->start)>>16) | 0xffff0000,
1494			     &ioc->ioc_regs->io_io_low_hv);
1495		__raw_writel(((parent->end)>>16) | 0xffff0000,
1496			     &ioc->ioc_regs->io_io_high_hv);
1497	} else {
1498		return -EBUSY;
1499	}
1500
1501 out:
1502	return allocate_resource(parent, res, size, min, max, align, NULL,NULL);
1503}
1504
1505int ccio_request_resource(const struct parisc_device *dev,
1506		struct resource *res)
1507{
1508	struct resource *parent;
1509	struct ioc *ioc = ccio_get_iommu(dev);
1510
1511	if (!ioc) {
1512		parent = &iomem_resource;
1513	} else if ((ioc->mmio_region->start <= res->start) &&
1514			(res->end <= ioc->mmio_region->end)) {
1515		parent = ioc->mmio_region;
1516	} else if (((ioc->mmio_region + 1)->start <= res->start) &&
1517			(res->end <= (ioc->mmio_region + 1)->end)) {
1518		parent = ioc->mmio_region + 1;
1519	} else {
1520		return -EBUSY;
1521	}
1522
1523	/* "transparent" bus bridges need to register MMIO resources
1524	 * firmware assigned them. e.g. children of hppb.c (e.g. K-class)
1525	 * registered their resources in the PDC "bus walk" (See
1526	 * arch/parisc/kernel/inventory.c).
1527	 */
1528	return insert_resource(parent, res);
1529}
1530
1531/**
1532 * ccio_probe - Determine if ccio should claim this device.
1533 * @dev: The device which has been found
1534 *
1535 * Determine if ccio should claim this chip (return 0) or not (return 1).
1536 * If so, initialize the chip and tell other partners in crime they
1537 * have work to do.
1538 */
1539static int __init ccio_probe(struct parisc_device *dev)
1540{
1541	int i;
1542	struct ioc *ioc, **ioc_p = &ioc_list;
1543
1544	ioc = kzalloc(sizeof(struct ioc), GFP_KERNEL);
1545	if (ioc == NULL) {
1546		printk(KERN_ERR MODULE_NAME ": memory allocation failure\n");
1547		return 1;
1548	}
1549
1550	ioc->name = dev->id.hversion == U2_IOA_RUNWAY ? "U2" : "UTurn";
1551
1552	printk(KERN_INFO "Found %s at 0x%lx\n", ioc->name,
1553		(unsigned long)dev->hpa.start);
1554
1555	for (i = 0; i < ioc_count; i++) {
1556		ioc_p = &(*ioc_p)->next;
1557	}
1558	*ioc_p = ioc;
1559
1560	ioc->hw_path = dev->hw_path;
1561	ioc->ioc_regs = ioremap_nocache(dev->hpa.start, 4096);
 
 
 
 
1562	ccio_ioc_init(ioc);
1563	ccio_init_resources(ioc);
1564	hppa_dma_ops = &ccio_ops;
1565	dev->dev.platform_data = kzalloc(sizeof(struct pci_hba_data), GFP_KERNEL);
1566
1567	/* if this fails, no I/O cards will work, so may as well bug */
1568	BUG_ON(dev->dev.platform_data == NULL);
1569	HBA_DATA(dev->dev.platform_data)->iommu = ioc;
1570
1571#ifdef CONFIG_PROC_FS
1572	if (ioc_count == 0) {
1573		proc_create(MODULE_NAME, 0, proc_runway_root,
1574			    &ccio_proc_info_fops);
1575		proc_create(MODULE_NAME"-bitmap", 0, proc_runway_root,
1576			    &ccio_proc_bitmap_fops);
1577	}
1578#endif
1579	ioc_count++;
1580
1581	parisc_has_iommu();
1582	return 0;
1583}
1584
1585/**
1586 * ccio_init - ccio initialization procedure.
1587 *
1588 * Register this driver.
1589 */
1590void __init ccio_init(void)
1591{
1592	register_parisc_driver(&ccio_driver);
1593}
1594