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