xref: /openbmc/linux/drivers/parisc/ccio-dma.c (revision feac8c8b)
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 <linux/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 
113 #define CCIO_MAPPING_ERROR    (~(dma_addr_t)0)
114 
115 struct ioa_registers {
116         /* Runway Supervisory Set */
117         int32_t    unused1[12];
118         uint32_t   io_command;             /* Offset 12 */
119         uint32_t   io_status;              /* Offset 13 */
120         uint32_t   io_control;             /* Offset 14 */
121         int32_t    unused2[1];
122 
123         /* Runway Auxiliary Register Set */
124         uint32_t   io_err_resp;            /* Offset  0 */
125         uint32_t   io_err_info;            /* Offset  1 */
126         uint32_t   io_err_req;             /* Offset  2 */
127         uint32_t   io_err_resp_hi;         /* Offset  3 */
128         uint32_t   io_tlb_entry_m;         /* Offset  4 */
129         uint32_t   io_tlb_entry_l;         /* Offset  5 */
130         uint32_t   unused3[1];
131         uint32_t   io_pdir_base;           /* Offset  7 */
132         uint32_t   io_io_low_hv;           /* Offset  8 */
133         uint32_t   io_io_high_hv;          /* Offset  9 */
134         uint32_t   unused4[1];
135         uint32_t   io_chain_id_mask;       /* Offset 11 */
136         uint32_t   unused5[2];
137         uint32_t   io_io_low;              /* Offset 14 */
138         uint32_t   io_io_high;             /* Offset 15 */
139 };
140 
141 /*
142 ** IOA Registers
143 ** -------------
144 **
145 ** Runway IO_CONTROL Register (+0x38)
146 **
147 ** The Runway IO_CONTROL register controls the forwarding of transactions.
148 **
149 ** | 0  ...  13  |  14 15 | 16 ... 21 | 22 | 23 24 |  25 ... 31 |
150 ** |    HV       |   TLB  |  reserved | HV | mode  |  reserved  |
151 **
152 ** o mode field indicates the address translation of transactions
153 **   forwarded from Runway to GSC+:
154 **       Mode Name     Value        Definition
155 **       Off (default)   0          Opaque to matching addresses.
156 **       Include         1          Transparent for matching addresses.
157 **       Peek            3          Map matching addresses.
158 **
159 **       + "Off" mode: Runway transactions which match the I/O range
160 **         specified by the IO_IO_LOW/IO_IO_HIGH registers will be ignored.
161 **       + "Include" mode: all addresses within the I/O range specified
162 **         by the IO_IO_LOW and IO_IO_HIGH registers are transparently
163 **         forwarded. This is the I/O Adapter's normal operating mode.
164 **       + "Peek" mode: used during system configuration to initialize the
165 **         GSC+ bus. Runway Write_Shorts in the address range specified by
166 **         IO_IO_LOW and IO_IO_HIGH are forwarded through the I/O Adapter
167 **         *AND* the GSC+ address is remapped to the Broadcast Physical
168 **         Address space by setting the 14 high order address bits of the
169 **         32 bit GSC+ address to ones.
170 **
171 ** o TLB field affects transactions which are forwarded from GSC+ to Runway.
172 **   "Real" mode is the poweron default.
173 **
174 **   TLB Mode  Value  Description
175 **   Real        0    No TLB translation. Address is directly mapped and the
176 **                    virtual address is composed of selected physical bits.
177 **   Error       1    Software fills the TLB manually.
178 **   Normal      2    IOA fetches IO TLB misses from IO PDIR (in host memory).
179 **
180 **
181 ** IO_IO_LOW_HV	  +0x60 (HV dependent)
182 ** IO_IO_HIGH_HV  +0x64 (HV dependent)
183 ** IO_IO_LOW      +0x78	(Architected register)
184 ** IO_IO_HIGH     +0x7c	(Architected register)
185 **
186 ** IO_IO_LOW and IO_IO_HIGH set the lower and upper bounds of the
187 ** I/O Adapter address space, respectively.
188 **
189 ** 0  ... 7 | 8 ... 15 |  16   ...   31 |
190 ** 11111111 | 11111111 |      address   |
191 **
192 ** Each LOW/HIGH pair describes a disjoint address space region.
193 ** (2 per GSC+ port). Each incoming Runway transaction address is compared
194 ** with both sets of LOW/HIGH registers. If the address is in the range
195 ** greater than or equal to IO_IO_LOW and less than IO_IO_HIGH the transaction
196 ** for forwarded to the respective GSC+ bus.
197 ** Specify IO_IO_LOW equal to or greater than IO_IO_HIGH to avoid specifying
198 ** an address space region.
199 **
200 ** In order for a Runway address to reside within GSC+ extended address space:
201 **	Runway Address [0:7]    must identically compare to 8'b11111111
202 **	Runway Address [8:11]   must be equal to IO_IO_LOW(_HV)[16:19]
203 ** 	Runway Address [12:23]  must be greater than or equal to
204 **	           IO_IO_LOW(_HV)[20:31] and less than IO_IO_HIGH(_HV)[20:31].
205 **	Runway Address [24:39]  is not used in the comparison.
206 **
207 ** When the Runway transaction is forwarded to GSC+, the GSC+ address is
208 ** as follows:
209 **	GSC+ Address[0:3]	4'b1111
210 **	GSC+ Address[4:29]	Runway Address[12:37]
211 **	GSC+ Address[30:31]	2'b00
212 **
213 ** All 4 Low/High registers must be initialized (by PDC) once the lower bus
214 ** is interrogated and address space is defined. The operating system will
215 ** modify the architectural IO_IO_LOW and IO_IO_HIGH registers following
216 ** the PDC initialization.  However, the hardware version dependent IO_IO_LOW
217 ** and IO_IO_HIGH registers should not be subsequently altered by the OS.
218 **
219 ** Writes to both sets of registers will take effect immediately, bypassing
220 ** the queues, which ensures that subsequent Runway transactions are checked
221 ** against the updated bounds values. However reads are queued, introducing
222 ** the possibility of a read being bypassed by a subsequent write to the same
223 ** register. This sequence can be avoided by having software wait for read
224 ** returns before issuing subsequent writes.
225 */
226 
227 struct ioc {
228 	struct ioa_registers __iomem *ioc_regs;  /* I/O MMU base address */
229 	u8  *res_map;	                /* resource map, bit == pdir entry */
230 	u64 *pdir_base;	                /* physical base address */
231 	u32 pdir_size; 			/* bytes, function of IOV Space size */
232 	u32 res_hint;	                /* next available IOVP -
233 					   circular search */
234 	u32 res_size;		    	/* size of resource map in bytes */
235 	spinlock_t res_lock;
236 
237 #ifdef CCIO_COLLECT_STATS
238 #define CCIO_SEARCH_SAMPLE 0x100
239 	unsigned long avg_search[CCIO_SEARCH_SAMPLE];
240 	unsigned long avg_idx;		  /* current index into avg_search */
241 	unsigned long used_pages;
242 	unsigned long msingle_calls;
243 	unsigned long msingle_pages;
244 	unsigned long msg_calls;
245 	unsigned long msg_pages;
246 	unsigned long usingle_calls;
247 	unsigned long usingle_pages;
248 	unsigned long usg_calls;
249 	unsigned long usg_pages;
250 #endif
251 	unsigned short cujo20_bug;
252 
253 	/* STUFF We don't need in performance path */
254 	u32 chainid_shift; 		/* specify bit location of chain_id */
255 	struct ioc *next;		/* Linked list of discovered iocs */
256 	const char *name;		/* device name from firmware */
257 	unsigned int hw_path;           /* the hardware path this ioc is associatd with */
258 	struct pci_dev *fake_pci_dev;   /* the fake pci_dev for non-pci devs */
259 	struct resource mmio_region[2]; /* The "routed" MMIO regions */
260 };
261 
262 static struct ioc *ioc_list;
263 static int ioc_count;
264 
265 /**************************************************************
266 *
267 *   I/O Pdir Resource Management
268 *
269 *   Bits set in the resource map are in use.
270 *   Each bit can represent a number of pages.
271 *   LSbs represent lower addresses (IOVA's).
272 *
273 *   This was was copied from sba_iommu.c. Don't try to unify
274 *   the two resource managers unless a way to have different
275 *   allocation policies is also adjusted. We'd like to avoid
276 *   I/O TLB thrashing by having resource allocation policy
277 *   match the I/O TLB replacement policy.
278 *
279 ***************************************************************/
280 #define IOVP_SIZE PAGE_SIZE
281 #define IOVP_SHIFT PAGE_SHIFT
282 #define IOVP_MASK PAGE_MASK
283 
284 /* Convert from IOVP to IOVA and vice versa. */
285 #define CCIO_IOVA(iovp,offset) ((iovp) | (offset))
286 #define CCIO_IOVP(iova) ((iova) & IOVP_MASK)
287 
288 #define PDIR_INDEX(iovp)    ((iovp)>>IOVP_SHIFT)
289 #define MKIOVP(pdir_idx)    ((long)(pdir_idx) << IOVP_SHIFT)
290 #define MKIOVA(iovp,offset) (dma_addr_t)((long)iovp | (long)offset)
291 
292 /*
293 ** Don't worry about the 150% average search length on a miss.
294 ** If the search wraps around, and passes the res_hint, it will
295 ** cause the kernel to panic anyhow.
296 */
297 #define CCIO_SEARCH_LOOP(ioc, res_idx, mask, size)  \
298        for(; res_ptr < res_end; ++res_ptr) { \
299 		int ret;\
300 		unsigned int idx;\
301 		idx = (unsigned int)((unsigned long)res_ptr - (unsigned long)ioc->res_map); \
302 		ret = iommu_is_span_boundary(idx << 3, pages_needed, 0, boundary_size);\
303 		if ((0 == (*res_ptr & mask)) && !ret) { \
304 			*res_ptr |= mask; \
305 			res_idx = idx;\
306 			ioc->res_hint = res_idx + (size >> 3); \
307 			goto resource_found; \
308 		} \
309 	}
310 
311 #define CCIO_FIND_FREE_MAPPING(ioa, res_idx, mask, size) \
312        u##size *res_ptr = (u##size *)&((ioc)->res_map[ioa->res_hint & ~((size >> 3) - 1)]); \
313        u##size *res_end = (u##size *)&(ioc)->res_map[ioa->res_size]; \
314        CCIO_SEARCH_LOOP(ioc, res_idx, mask, size); \
315        res_ptr = (u##size *)&(ioc)->res_map[0]; \
316        CCIO_SEARCH_LOOP(ioa, res_idx, mask, size);
317 
318 /*
319 ** Find available bit in this ioa's resource map.
320 ** Use a "circular" search:
321 **   o Most IOVA's are "temporary" - avg search time should be small.
322 ** o keep a history of what happened for debugging
323 ** o KISS.
324 **
325 ** Perf optimizations:
326 ** o search for log2(size) bits at a time.
327 ** o search for available resource bits using byte/word/whatever.
328 ** o use different search for "large" (eg > 4 pages) or "very large"
329 **   (eg > 16 pages) mappings.
330 */
331 
332 /**
333  * ccio_alloc_range - Allocate pages in the ioc's resource map.
334  * @ioc: The I/O Controller.
335  * @pages_needed: The requested number of pages to be mapped into the
336  * I/O Pdir...
337  *
338  * This function searches the resource map of the ioc to locate a range
339  * of available pages for the requested size.
340  */
341 static int
342 ccio_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
343 {
344 	unsigned int pages_needed = size >> IOVP_SHIFT;
345 	unsigned int res_idx;
346 	unsigned long boundary_size;
347 #ifdef CCIO_COLLECT_STATS
348 	unsigned long cr_start = mfctl(16);
349 #endif
350 
351 	BUG_ON(pages_needed == 0);
352 	BUG_ON((pages_needed * IOVP_SIZE) > DMA_CHUNK_SIZE);
353 
354 	DBG_RES("%s() size: %d pages_needed %d\n",
355 		__func__, size, pages_needed);
356 
357 	/*
358 	** "seek and ye shall find"...praying never hurts either...
359 	** ggg sacrifices another 710 to the computer gods.
360 	*/
361 
362 	boundary_size = ALIGN((unsigned long long)dma_get_seg_boundary(dev) + 1,
363 			      1ULL << IOVP_SHIFT) >> IOVP_SHIFT;
364 
365 	if (pages_needed <= 8) {
366 		/*
367 		 * LAN traffic will not thrash the TLB IFF the same NIC
368 		 * uses 8 adjacent pages to map separate payload data.
369 		 * ie the same byte in the resource bit map.
370 		 */
371 #if 0
372 		/* FIXME: bit search should shift it's way through
373 		 * an unsigned long - not byte at a time. As it is now,
374 		 * we effectively allocate this byte to this mapping.
375 		 */
376 		unsigned long mask = ~(~0UL >> pages_needed);
377 		CCIO_FIND_FREE_MAPPING(ioc, res_idx, mask, 8);
378 #else
379 		CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xff, 8);
380 #endif
381 	} else if (pages_needed <= 16) {
382 		CCIO_FIND_FREE_MAPPING(ioc, res_idx, 0xffff, 16);
383 	} else if (pages_needed <= 32) {
384 		CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~(unsigned int)0, 32);
385 #ifdef __LP64__
386 	} else if (pages_needed <= 64) {
387 		CCIO_FIND_FREE_MAPPING(ioc, res_idx, ~0UL, 64);
388 #endif
389 	} else {
390 		panic("%s: %s() Too many pages to map. pages_needed: %u\n",
391 		       __FILE__,  __func__, pages_needed);
392 	}
393 
394 	panic("%s: %s() I/O MMU is out of mapping resources.\n", __FILE__,
395 	      __func__);
396 
397 resource_found:
398 
399 	DBG_RES("%s() res_idx %d res_hint: %d\n",
400 		__func__, res_idx, ioc->res_hint);
401 
402 #ifdef CCIO_COLLECT_STATS
403 	{
404 		unsigned long cr_end = mfctl(16);
405 		unsigned long tmp = cr_end - cr_start;
406 		/* check for roll over */
407 		cr_start = (cr_end < cr_start) ?  -(tmp) : (tmp);
408 	}
409 	ioc->avg_search[ioc->avg_idx++] = cr_start;
410 	ioc->avg_idx &= CCIO_SEARCH_SAMPLE - 1;
411 	ioc->used_pages += pages_needed;
412 #endif
413 	/*
414 	** return the bit address.
415 	*/
416 	return res_idx << 3;
417 }
418 
419 #define CCIO_FREE_MAPPINGS(ioc, res_idx, mask, size) \
420         u##size *res_ptr = (u##size *)&((ioc)->res_map[res_idx]); \
421         BUG_ON((*res_ptr & mask) != mask); \
422         *res_ptr &= ~(mask);
423 
424 /**
425  * ccio_free_range - Free pages from the ioc's resource map.
426  * @ioc: The I/O Controller.
427  * @iova: The I/O Virtual Address.
428  * @pages_mapped: The requested number of pages to be freed from the
429  * I/O Pdir.
430  *
431  * This function frees the resouces allocated for the iova.
432  */
433 static void
434 ccio_free_range(struct ioc *ioc, dma_addr_t iova, unsigned long pages_mapped)
435 {
436 	unsigned long iovp = CCIO_IOVP(iova);
437 	unsigned int res_idx = PDIR_INDEX(iovp) >> 3;
438 
439 	BUG_ON(pages_mapped == 0);
440 	BUG_ON((pages_mapped * IOVP_SIZE) > DMA_CHUNK_SIZE);
441 	BUG_ON(pages_mapped > BITS_PER_LONG);
442 
443 	DBG_RES("%s():  res_idx: %d pages_mapped %d\n",
444 		__func__, res_idx, pages_mapped);
445 
446 #ifdef CCIO_COLLECT_STATS
447 	ioc->used_pages -= pages_mapped;
448 #endif
449 
450 	if(pages_mapped <= 8) {
451 #if 0
452 		/* see matching comments in alloc_range */
453 		unsigned long mask = ~(~0UL >> pages_mapped);
454 		CCIO_FREE_MAPPINGS(ioc, res_idx, mask, 8);
455 #else
456 		CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffUL, 8);
457 #endif
458 	} else if(pages_mapped <= 16) {
459 		CCIO_FREE_MAPPINGS(ioc, res_idx, 0xffffUL, 16);
460 	} else if(pages_mapped <= 32) {
461 		CCIO_FREE_MAPPINGS(ioc, res_idx, ~(unsigned int)0, 32);
462 #ifdef __LP64__
463 	} else if(pages_mapped <= 64) {
464 		CCIO_FREE_MAPPINGS(ioc, res_idx, ~0UL, 64);
465 #endif
466 	} else {
467 		panic("%s:%s() Too many pages to unmap.\n", __FILE__,
468 		      __func__);
469 	}
470 }
471 
472 /****************************************************************
473 **
474 **          CCIO dma_ops support routines
475 **
476 *****************************************************************/
477 
478 typedef unsigned long space_t;
479 #define KERNEL_SPACE 0
480 
481 /*
482 ** DMA "Page Type" and Hints
483 ** o if SAFE_DMA isn't set, mapping is for FAST_DMA. SAFE_DMA should be
484 **   set for subcacheline DMA transfers since we don't want to damage the
485 **   other part of a cacheline.
486 ** o SAFE_DMA must be set for "memory" allocated via pci_alloc_consistent().
487 **   This bit tells U2 to do R/M/W for partial cachelines. "Streaming"
488 **   data can avoid this if the mapping covers full cache lines.
489 ** o STOP_MOST is needed for atomicity across cachelines.
490 **   Apparently only "some EISA devices" need this.
491 **   Using CONFIG_ISA is hack. Only the IOA with EISA under it needs
492 **   to use this hint iff the EISA devices needs this feature.
493 **   According to the U2 ERS, STOP_MOST enabled pages hurt performance.
494 ** o PREFETCH should *not* be set for cases like Multiple PCI devices
495 **   behind GSCtoPCI (dino) bus converter. Only one cacheline per GSC
496 **   device can be fetched and multiply DMA streams will thrash the
497 **   prefetch buffer and burn memory bandwidth. See 6.7.3 "Prefetch Rules
498 **   and Invalidation of Prefetch Entries".
499 **
500 ** FIXME: the default hints need to be per GSC device - not global.
501 **
502 ** HP-UX dorks: linux device driver programming model is totally different
503 **    than HP-UX's. HP-UX always sets HINT_PREFETCH since it's drivers
504 **    do special things to work on non-coherent platforms...linux has to
505 **    be much more careful with this.
506 */
507 #define IOPDIR_VALID    0x01UL
508 #define HINT_SAFE_DMA   0x02UL	/* used for pci_alloc_consistent() pages */
509 #ifdef CONFIG_EISA
510 #define HINT_STOP_MOST  0x04UL	/* LSL support */
511 #else
512 #define HINT_STOP_MOST  0x00UL	/* only needed for "some EISA devices" */
513 #endif
514 #define HINT_UDPATE_ENB 0x08UL  /* not used/supported by U2 */
515 #define HINT_PREFETCH   0x10UL	/* for outbound pages which are not SAFE */
516 
517 
518 /*
519 ** Use direction (ie PCI_DMA_TODEVICE) to pick hint.
520 ** ccio_alloc_consistent() depends on this to get SAFE_DMA
521 ** when it passes in BIDIRECTIONAL flag.
522 */
523 static u32 hint_lookup[] = {
524 	[PCI_DMA_BIDIRECTIONAL]	= HINT_STOP_MOST | HINT_SAFE_DMA | IOPDIR_VALID,
525 	[PCI_DMA_TODEVICE]	= HINT_STOP_MOST | HINT_PREFETCH | IOPDIR_VALID,
526 	[PCI_DMA_FROMDEVICE]	= HINT_STOP_MOST | IOPDIR_VALID,
527 };
528 
529 /**
530  * ccio_io_pdir_entry - Initialize an I/O Pdir.
531  * @pdir_ptr: A pointer into I/O Pdir.
532  * @sid: The Space Identifier.
533  * @vba: The virtual address.
534  * @hints: The DMA Hint.
535  *
536  * Given a virtual address (vba, arg2) and space id, (sid, arg1),
537  * load the I/O PDIR entry pointed to by pdir_ptr (arg0). Each IO Pdir
538  * entry consists of 8 bytes as shown below (MSB == bit 0):
539  *
540  *
541  * WORD 0:
542  * +------+----------------+-----------------------------------------------+
543  * | Phys | Virtual Index  |               Phys                            |
544  * | 0:3  |     0:11       |               4:19                            |
545  * |4 bits|   12 bits      |              16 bits                          |
546  * +------+----------------+-----------------------------------------------+
547  * WORD 1:
548  * +-----------------------+-----------------------------------------------+
549  * |      Phys    |  Rsvd  | Prefetch |Update |Rsvd  |Lock  |Safe  |Valid  |
550  * |     20:39    |        | Enable   |Enable |      |Enable|DMA   |       |
551  * |    20 bits   | 5 bits | 1 bit    |1 bit  |2 bits|1 bit |1 bit |1 bit  |
552  * +-----------------------+-----------------------------------------------+
553  *
554  * The virtual index field is filled with the results of the LCI
555  * (Load Coherence Index) instruction.  The 8 bits used for the virtual
556  * index are bits 12:19 of the value returned by LCI.
557  */
558 static void CCIO_INLINE
559 ccio_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba,
560 		   unsigned long hints)
561 {
562 	register unsigned long pa;
563 	register unsigned long ci; /* coherent index */
564 
565 	/* We currently only support kernel addresses */
566 	BUG_ON(sid != KERNEL_SPACE);
567 
568 	mtsp(sid,1);
569 
570 	/*
571 	** WORD 1 - low order word
572 	** "hints" parm includes the VALID bit!
573 	** "dep" clobbers the physical address offset bits as well.
574 	*/
575 	pa = virt_to_phys(vba);
576 	asm volatile("depw  %1,31,12,%0" : "+r" (pa) : "r" (hints));
577 	((u32 *)pdir_ptr)[1] = (u32) pa;
578 
579 	/*
580 	** WORD 0 - high order word
581 	*/
582 
583 #ifdef __LP64__
584 	/*
585 	** get bits 12:15 of physical address
586 	** shift bits 16:31 of physical address
587 	** and deposit them
588 	*/
589 	asm volatile ("extrd,u %1,15,4,%0" : "=r" (ci) : "r" (pa));
590 	asm volatile ("extrd,u %1,31,16,%0" : "+r" (pa) : "r" (pa));
591 	asm volatile ("depd  %1,35,4,%0" : "+r" (pa) : "r" (ci));
592 #else
593 	pa = 0;
594 #endif
595 	/*
596 	** get CPU coherency index bits
597 	** Grab virtual index [0:11]
598 	** Deposit virt_idx bits into I/O PDIR word
599 	*/
600 	asm volatile ("lci %%r0(%%sr1, %1), %0" : "=r" (ci) : "r" (vba));
601 	asm volatile ("extru %1,19,12,%0" : "+r" (ci) : "r" (ci));
602 	asm volatile ("depw  %1,15,12,%0" : "+r" (pa) : "r" (ci));
603 
604 	((u32 *)pdir_ptr)[0] = (u32) pa;
605 
606 
607 	/* FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360)
608 	**        PCX-U/U+ do. (eg C200/C240)
609 	**        PCX-T'? Don't know. (eg C110 or similar K-class)
610 	**
611 	** See PDC_MODEL/option 0/SW_CAP word for "Non-coherent IO-PDIR bit".
612 	** Hopefully we can patch (NOP) these out at boot time somehow.
613 	**
614 	** "Since PCX-U employs an offset hash that is incompatible with
615 	** the real mode coherence index generation of U2, the PDIR entry
616 	** must be flushed to memory to retain coherence."
617 	*/
618 	asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr));
619 	asm volatile("sync");
620 }
621 
622 /**
623  * ccio_clear_io_tlb - Remove stale entries from the I/O TLB.
624  * @ioc: The I/O Controller.
625  * @iovp: The I/O Virtual Page.
626  * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir.
627  *
628  * Purge invalid I/O PDIR entries from the I/O TLB.
629  *
630  * FIXME: Can we change the byte_cnt to pages_mapped?
631  */
632 static CCIO_INLINE void
633 ccio_clear_io_tlb(struct ioc *ioc, dma_addr_t iovp, size_t byte_cnt)
634 {
635 	u32 chain_size = 1 << ioc->chainid_shift;
636 
637 	iovp &= IOVP_MASK;	/* clear offset bits, just want pagenum */
638 	byte_cnt += chain_size;
639 
640 	while(byte_cnt > chain_size) {
641 		WRITE_U32(CMD_TLB_PURGE | iovp, &ioc->ioc_regs->io_command);
642 		iovp += chain_size;
643 		byte_cnt -= chain_size;
644 	}
645 }
646 
647 /**
648  * ccio_mark_invalid - Mark the I/O Pdir entries invalid.
649  * @ioc: The I/O Controller.
650  * @iova: The I/O Virtual Address.
651  * @byte_cnt: The requested number of bytes to be freed from the I/O Pdir.
652  *
653  * Mark the I/O Pdir entries invalid and blow away the corresponding I/O
654  * TLB entries.
655  *
656  * FIXME: at some threshold it might be "cheaper" to just blow
657  *        away the entire I/O TLB instead of individual entries.
658  *
659  * FIXME: Uturn has 256 TLB entries. We don't need to purge every
660  *        PDIR entry - just once for each possible TLB entry.
661  *        (We do need to maker I/O PDIR entries invalid regardless).
662  *
663  * FIXME: Can we change byte_cnt to pages_mapped?
664  */
665 static CCIO_INLINE void
666 ccio_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
667 {
668 	u32 iovp = (u32)CCIO_IOVP(iova);
669 	size_t saved_byte_cnt;
670 
671 	/* round up to nearest page size */
672 	saved_byte_cnt = byte_cnt = ALIGN(byte_cnt, IOVP_SIZE);
673 
674 	while(byte_cnt > 0) {
675 		/* invalidate one page at a time */
676 		unsigned int idx = PDIR_INDEX(iovp);
677 		char *pdir_ptr = (char *) &(ioc->pdir_base[idx]);
678 
679 		BUG_ON(idx >= (ioc->pdir_size / sizeof(u64)));
680 		pdir_ptr[7] = 0;	/* clear only VALID bit */
681 		/*
682 		** FIXME: PCX_W platforms don't need FDC/SYNC. (eg C360)
683 		**   PCX-U/U+ do. (eg C200/C240)
684 		** See PDC_MODEL/option 0/SW_CAP for "Non-coherent IO-PDIR bit".
685 		**
686 		** Hopefully someone figures out how to patch (NOP) the
687 		** FDC/SYNC out at boot time.
688 		*/
689 		asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr[7]));
690 
691 		iovp     += IOVP_SIZE;
692 		byte_cnt -= IOVP_SIZE;
693 	}
694 
695 	asm volatile("sync");
696 	ccio_clear_io_tlb(ioc, CCIO_IOVP(iova), saved_byte_cnt);
697 }
698 
699 /****************************************************************
700 **
701 **          CCIO dma_ops
702 **
703 *****************************************************************/
704 
705 /**
706  * ccio_dma_supported - Verify the IOMMU supports the DMA address range.
707  * @dev: The PCI device.
708  * @mask: A bit mask describing the DMA address range of the device.
709  */
710 static int
711 ccio_dma_supported(struct device *dev, u64 mask)
712 {
713 	if(dev == NULL) {
714 		printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
715 		BUG();
716 		return 0;
717 	}
718 
719 	/* only support 32-bit devices (ie PCI/GSC) */
720 	return (int)(mask == 0xffffffffUL);
721 }
722 
723 /**
724  * ccio_map_single - Map an address range into the IOMMU.
725  * @dev: The PCI device.
726  * @addr: The start address of the DMA region.
727  * @size: The length of the DMA region.
728  * @direction: The direction of the DMA transaction (to/from device).
729  *
730  * This function implements the pci_map_single function.
731  */
732 static dma_addr_t
733 ccio_map_single(struct device *dev, void *addr, size_t size,
734 		enum dma_data_direction direction)
735 {
736 	int idx;
737 	struct ioc *ioc;
738 	unsigned long flags;
739 	dma_addr_t iovp;
740 	dma_addr_t offset;
741 	u64 *pdir_start;
742 	unsigned long hint = hint_lookup[(int)direction];
743 
744 	BUG_ON(!dev);
745 	ioc = GET_IOC(dev);
746 	if (!ioc)
747 		return CCIO_MAPPING_ERROR;
748 
749 	BUG_ON(size <= 0);
750 
751 	/* save offset bits */
752 	offset = ((unsigned long) addr) & ~IOVP_MASK;
753 
754 	/* round up to nearest IOVP_SIZE */
755 	size = ALIGN(size + offset, IOVP_SIZE);
756 	spin_lock_irqsave(&ioc->res_lock, flags);
757 
758 #ifdef CCIO_COLLECT_STATS
759 	ioc->msingle_calls++;
760 	ioc->msingle_pages += size >> IOVP_SHIFT;
761 #endif
762 
763 	idx = ccio_alloc_range(ioc, dev, size);
764 	iovp = (dma_addr_t)MKIOVP(idx);
765 
766 	pdir_start = &(ioc->pdir_base[idx]);
767 
768 	DBG_RUN("%s() 0x%p -> 0x%lx size: %0x%x\n",
769 		__func__, addr, (long)iovp | offset, size);
770 
771 	/* If not cacheline aligned, force SAFE_DMA on the whole mess */
772 	if((size % L1_CACHE_BYTES) || ((unsigned long)addr % L1_CACHE_BYTES))
773 		hint |= HINT_SAFE_DMA;
774 
775 	while(size > 0) {
776 		ccio_io_pdir_entry(pdir_start, KERNEL_SPACE, (unsigned long)addr, hint);
777 
778 		DBG_RUN(" pdir %p %08x%08x\n",
779 			pdir_start,
780 			(u32) (((u32 *) pdir_start)[0]),
781 			(u32) (((u32 *) pdir_start)[1]));
782 		++pdir_start;
783 		addr += IOVP_SIZE;
784 		size -= IOVP_SIZE;
785 	}
786 
787 	spin_unlock_irqrestore(&ioc->res_lock, flags);
788 
789 	/* form complete address */
790 	return CCIO_IOVA(iovp, offset);
791 }
792 
793 
794 static dma_addr_t
795 ccio_map_page(struct device *dev, struct page *page, unsigned long offset,
796 		size_t size, enum dma_data_direction direction,
797 		unsigned long attrs)
798 {
799 	return ccio_map_single(dev, page_address(page) + offset, size,
800 			direction);
801 }
802 
803 
804 /**
805  * ccio_unmap_page - Unmap an address range from the IOMMU.
806  * @dev: The PCI device.
807  * @addr: The start address of the DMA region.
808  * @size: The length of the DMA region.
809  * @direction: The direction of the DMA transaction (to/from device).
810  */
811 static void
812 ccio_unmap_page(struct device *dev, dma_addr_t iova, size_t size,
813 		enum dma_data_direction direction, unsigned long attrs)
814 {
815 	struct ioc *ioc;
816 	unsigned long flags;
817 	dma_addr_t offset = iova & ~IOVP_MASK;
818 
819 	BUG_ON(!dev);
820 	ioc = GET_IOC(dev);
821 	if (!ioc) {
822 		WARN_ON(!ioc);
823 		return;
824 	}
825 
826 	DBG_RUN("%s() iovp 0x%lx/%x\n",
827 		__func__, (long)iova, size);
828 
829 	iova ^= offset;        /* clear offset bits */
830 	size += offset;
831 	size = ALIGN(size, IOVP_SIZE);
832 
833 	spin_lock_irqsave(&ioc->res_lock, flags);
834 
835 #ifdef CCIO_COLLECT_STATS
836 	ioc->usingle_calls++;
837 	ioc->usingle_pages += size >> IOVP_SHIFT;
838 #endif
839 
840 	ccio_mark_invalid(ioc, iova, size);
841 	ccio_free_range(ioc, iova, (size >> IOVP_SHIFT));
842 	spin_unlock_irqrestore(&ioc->res_lock, flags);
843 }
844 
845 /**
846  * ccio_alloc - Allocate a consistent DMA mapping.
847  * @dev: The PCI device.
848  * @size: The length of the DMA region.
849  * @dma_handle: The DMA address handed back to the device (not the cpu).
850  *
851  * This function implements the pci_alloc_consistent function.
852  */
853 static void *
854 ccio_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag,
855 		unsigned long attrs)
856 {
857       void *ret;
858 #if 0
859 /* GRANT Need to establish hierarchy for non-PCI devs as well
860 ** and then provide matching gsc_map_xxx() functions for them as well.
861 */
862 	if(!hwdev) {
863 		/* only support PCI */
864 		*dma_handle = 0;
865 		return 0;
866 	}
867 #endif
868         ret = (void *) __get_free_pages(flag, get_order(size));
869 
870 	if (ret) {
871 		memset(ret, 0, size);
872 		*dma_handle = ccio_map_single(dev, ret, size, PCI_DMA_BIDIRECTIONAL);
873 	}
874 
875 	return ret;
876 }
877 
878 /**
879  * ccio_free - Free a consistent DMA mapping.
880  * @dev: The PCI device.
881  * @size: The length of the DMA region.
882  * @cpu_addr: The cpu address returned from the ccio_alloc_consistent.
883  * @dma_handle: The device address returned from the ccio_alloc_consistent.
884  *
885  * This function implements the pci_free_consistent function.
886  */
887 static void
888 ccio_free(struct device *dev, size_t size, void *cpu_addr,
889 		dma_addr_t dma_handle, unsigned long attrs)
890 {
891 	ccio_unmap_page(dev, dma_handle, size, 0, 0);
892 	free_pages((unsigned long)cpu_addr, get_order(size));
893 }
894 
895 /*
896 ** Since 0 is a valid pdir_base index value, can't use that
897 ** to determine if a value is valid or not. Use a flag to indicate
898 ** the SG list entry contains a valid pdir index.
899 */
900 #define PIDE_FLAG 0x80000000UL
901 
902 #ifdef CCIO_COLLECT_STATS
903 #define IOMMU_MAP_STATS
904 #endif
905 #include "iommu-helpers.h"
906 
907 /**
908  * ccio_map_sg - Map the scatter/gather list into the IOMMU.
909  * @dev: The PCI device.
910  * @sglist: The scatter/gather list to be mapped in the IOMMU.
911  * @nents: The number of entries in the scatter/gather list.
912  * @direction: The direction of the DMA transaction (to/from device).
913  *
914  * This function implements the pci_map_sg function.
915  */
916 static int
917 ccio_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
918 	    enum dma_data_direction direction, unsigned long attrs)
919 {
920 	struct ioc *ioc;
921 	int coalesced, filled = 0;
922 	unsigned long flags;
923 	unsigned long hint = hint_lookup[(int)direction];
924 	unsigned long prev_len = 0, current_len = 0;
925 	int i;
926 
927 	BUG_ON(!dev);
928 	ioc = GET_IOC(dev);
929 	if (!ioc)
930 		return 0;
931 
932 	DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
933 
934 	/* Fast path single entry scatterlists. */
935 	if (nents == 1) {
936 		sg_dma_address(sglist) = ccio_map_single(dev,
937 				sg_virt(sglist), sglist->length,
938 				direction);
939 		sg_dma_len(sglist) = sglist->length;
940 		return 1;
941 	}
942 
943 	for(i = 0; i < nents; i++)
944 		prev_len += sglist[i].length;
945 
946 	spin_lock_irqsave(&ioc->res_lock, flags);
947 
948 #ifdef CCIO_COLLECT_STATS
949 	ioc->msg_calls++;
950 #endif
951 
952 	/*
953 	** First coalesce the chunks and allocate I/O pdir space
954 	**
955 	** If this is one DMA stream, we can properly map using the
956 	** correct virtual address associated with each DMA page.
957 	** w/o this association, we wouldn't have coherent DMA!
958 	** Access to the virtual address is what forces a two pass algorithm.
959 	*/
960 	coalesced = iommu_coalesce_chunks(ioc, dev, sglist, nents, ccio_alloc_range);
961 
962 	/*
963 	** Program the I/O Pdir
964 	**
965 	** map the virtual addresses to the I/O Pdir
966 	** o dma_address will contain the pdir index
967 	** o dma_len will contain the number of bytes to map
968 	** o page/offset contain the virtual address.
969 	*/
970 	filled = iommu_fill_pdir(ioc, sglist, nents, hint, ccio_io_pdir_entry);
971 
972 	spin_unlock_irqrestore(&ioc->res_lock, flags);
973 
974 	BUG_ON(coalesced != filled);
975 
976 	DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
977 
978 	for (i = 0; i < filled; i++)
979 		current_len += sg_dma_len(sglist + i);
980 
981 	BUG_ON(current_len != prev_len);
982 
983 	return filled;
984 }
985 
986 /**
987  * ccio_unmap_sg - Unmap the scatter/gather list from the IOMMU.
988  * @dev: The PCI device.
989  * @sglist: The scatter/gather list to be unmapped from the IOMMU.
990  * @nents: The number of entries in the scatter/gather list.
991  * @direction: The direction of the DMA transaction (to/from device).
992  *
993  * This function implements the pci_unmap_sg function.
994  */
995 static void
996 ccio_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
997 	      enum dma_data_direction direction, unsigned long attrs)
998 {
999 	struct ioc *ioc;
1000 
1001 	BUG_ON(!dev);
1002 	ioc = GET_IOC(dev);
1003 	if (!ioc) {
1004 		WARN_ON(!ioc);
1005 		return;
1006 	}
1007 
1008 	DBG_RUN_SG("%s() START %d entries, %p,%x\n",
1009 		__func__, nents, sg_virt(sglist), sglist->length);
1010 
1011 #ifdef CCIO_COLLECT_STATS
1012 	ioc->usg_calls++;
1013 #endif
1014 
1015 	while(sg_dma_len(sglist) && nents--) {
1016 
1017 #ifdef CCIO_COLLECT_STATS
1018 		ioc->usg_pages += sg_dma_len(sglist) >> PAGE_SHIFT;
1019 #endif
1020 		ccio_unmap_page(dev, sg_dma_address(sglist),
1021 				  sg_dma_len(sglist), direction, 0);
1022 		++sglist;
1023 	}
1024 
1025 	DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents);
1026 }
1027 
1028 static int ccio_mapping_error(struct device *dev, dma_addr_t dma_addr)
1029 {
1030 	return dma_addr == CCIO_MAPPING_ERROR;
1031 }
1032 
1033 static const struct dma_map_ops ccio_ops = {
1034 	.dma_supported =	ccio_dma_supported,
1035 	.alloc =		ccio_alloc,
1036 	.free =			ccio_free,
1037 	.map_page =		ccio_map_page,
1038 	.unmap_page =		ccio_unmap_page,
1039 	.map_sg = 		ccio_map_sg,
1040 	.unmap_sg = 		ccio_unmap_sg,
1041 	.mapping_error =	ccio_mapping_error,
1042 };
1043 
1044 #ifdef CONFIG_PROC_FS
1045 static int ccio_proc_info(struct seq_file *m, void *p)
1046 {
1047 	struct ioc *ioc = ioc_list;
1048 
1049 	while (ioc != NULL) {
1050 		unsigned int total_pages = ioc->res_size << 3;
1051 #ifdef CCIO_COLLECT_STATS
1052 		unsigned long avg = 0, min, max;
1053 		int j;
1054 #endif
1055 
1056 		seq_printf(m, "%s\n", ioc->name);
1057 
1058 		seq_printf(m, "Cujo 2.0 bug    : %s\n",
1059 			   (ioc->cujo20_bug ? "yes" : "no"));
1060 
1061 		seq_printf(m, "IO PDIR size    : %d bytes (%d entries)\n",
1062 			   total_pages * 8, total_pages);
1063 
1064 #ifdef CCIO_COLLECT_STATS
1065 		seq_printf(m, "IO PDIR entries : %ld free  %ld used (%d%%)\n",
1066 			   total_pages - ioc->used_pages, ioc->used_pages,
1067 			   (int)(ioc->used_pages * 100 / total_pages));
1068 #endif
1069 
1070 		seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1071 			   ioc->res_size, total_pages);
1072 
1073 #ifdef CCIO_COLLECT_STATS
1074 		min = max = ioc->avg_search[0];
1075 		for(j = 0; j < CCIO_SEARCH_SAMPLE; ++j) {
1076 			avg += ioc->avg_search[j];
1077 			if(ioc->avg_search[j] > max)
1078 				max = ioc->avg_search[j];
1079 			if(ioc->avg_search[j] < min)
1080 				min = ioc->avg_search[j];
1081 		}
1082 		avg /= CCIO_SEARCH_SAMPLE;
1083 		seq_printf(m, "  Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1084 			   min, avg, max);
1085 
1086 		seq_printf(m, "pci_map_single(): %8ld calls  %8ld pages (avg %d/1000)\n",
1087 			   ioc->msingle_calls, ioc->msingle_pages,
1088 			   (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1089 
1090 		/* KLUGE - unmap_sg calls unmap_page for each mapped page */
1091 		min = ioc->usingle_calls - ioc->usg_calls;
1092 		max = ioc->usingle_pages - ioc->usg_pages;
1093 		seq_printf(m, "pci_unmap_single: %8ld calls  %8ld pages (avg %d/1000)\n",
1094 			   min, max, (int)((max * 1000)/min));
1095 
1096 		seq_printf(m, "pci_map_sg()    : %8ld calls  %8ld pages (avg %d/1000)\n",
1097 			   ioc->msg_calls, ioc->msg_pages,
1098 			   (int)((ioc->msg_pages * 1000)/ioc->msg_calls));
1099 
1100 		seq_printf(m, "pci_unmap_sg()  : %8ld calls  %8ld pages (avg %d/1000)\n\n\n",
1101 			   ioc->usg_calls, ioc->usg_pages,
1102 			   (int)((ioc->usg_pages * 1000)/ioc->usg_calls));
1103 #endif	/* CCIO_COLLECT_STATS */
1104 
1105 		ioc = ioc->next;
1106 	}
1107 
1108 	return 0;
1109 }
1110 
1111 static int ccio_proc_info_open(struct inode *inode, struct file *file)
1112 {
1113 	return single_open(file, &ccio_proc_info, NULL);
1114 }
1115 
1116 static const struct file_operations ccio_proc_info_fops = {
1117 	.owner = THIS_MODULE,
1118 	.open = ccio_proc_info_open,
1119 	.read = seq_read,
1120 	.llseek = seq_lseek,
1121 	.release = single_release,
1122 };
1123 
1124 static int ccio_proc_bitmap_info(struct seq_file *m, void *p)
1125 {
1126 	struct ioc *ioc = ioc_list;
1127 
1128 	while (ioc != NULL) {
1129 		seq_hex_dump(m, "   ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map,
1130 			     ioc->res_size, false);
1131 		seq_putc(m, '\n');
1132 		ioc = ioc->next;
1133 		break; /* XXX - remove me */
1134 	}
1135 
1136 	return 0;
1137 }
1138 
1139 static int ccio_proc_bitmap_open(struct inode *inode, struct file *file)
1140 {
1141 	return single_open(file, &ccio_proc_bitmap_info, NULL);
1142 }
1143 
1144 static const struct file_operations ccio_proc_bitmap_fops = {
1145 	.owner = THIS_MODULE,
1146 	.open = ccio_proc_bitmap_open,
1147 	.read = seq_read,
1148 	.llseek = seq_lseek,
1149 	.release = single_release,
1150 };
1151 #endif /* CONFIG_PROC_FS */
1152 
1153 /**
1154  * ccio_find_ioc - Find the ioc in the ioc_list
1155  * @hw_path: The hardware path of the ioc.
1156  *
1157  * This function searches the ioc_list for an ioc that matches
1158  * the provide hardware path.
1159  */
1160 static struct ioc * ccio_find_ioc(int hw_path)
1161 {
1162 	int i;
1163 	struct ioc *ioc;
1164 
1165 	ioc = ioc_list;
1166 	for (i = 0; i < ioc_count; i++) {
1167 		if (ioc->hw_path == hw_path)
1168 			return ioc;
1169 
1170 		ioc = ioc->next;
1171 	}
1172 
1173 	return NULL;
1174 }
1175 
1176 /**
1177  * ccio_get_iommu - Find the iommu which controls this device
1178  * @dev: The parisc device.
1179  *
1180  * This function searches through the registered IOMMU's and returns
1181  * the appropriate IOMMU for the device based on its hardware path.
1182  */
1183 void * ccio_get_iommu(const struct parisc_device *dev)
1184 {
1185 	dev = find_pa_parent_type(dev, HPHW_IOA);
1186 	if (!dev)
1187 		return NULL;
1188 
1189 	return ccio_find_ioc(dev->hw_path);
1190 }
1191 
1192 #define CUJO_20_STEP       0x10000000	/* inc upper nibble */
1193 
1194 /* Cujo 2.0 has a bug which will silently corrupt data being transferred
1195  * to/from certain pages.  To avoid this happening, we mark these pages
1196  * as `used', and ensure that nothing will try to allocate from them.
1197  */
1198 void ccio_cujo20_fixup(struct parisc_device *cujo, u32 iovp)
1199 {
1200 	unsigned int idx;
1201 	struct parisc_device *dev = parisc_parent(cujo);
1202 	struct ioc *ioc = ccio_get_iommu(dev);
1203 	u8 *res_ptr;
1204 
1205 	ioc->cujo20_bug = 1;
1206 	res_ptr = ioc->res_map;
1207 	idx = PDIR_INDEX(iovp) >> 3;
1208 
1209 	while (idx < ioc->res_size) {
1210  		res_ptr[idx] |= 0xff;
1211 		idx += PDIR_INDEX(CUJO_20_STEP) >> 3;
1212 	}
1213 }
1214 
1215 #if 0
1216 /* GRANT -  is this needed for U2 or not? */
1217 
1218 /*
1219 ** Get the size of the I/O TLB for this I/O MMU.
1220 **
1221 ** If spa_shift is non-zero (ie probably U2),
1222 ** then calculate the I/O TLB size using spa_shift.
1223 **
1224 ** Otherwise we are supposed to get the IODC entry point ENTRY TLB
1225 ** and execute it. However, both U2 and Uturn firmware supplies spa_shift.
1226 ** I think only Java (K/D/R-class too?) systems don't do this.
1227 */
1228 static int
1229 ccio_get_iotlb_size(struct parisc_device *dev)
1230 {
1231 	if (dev->spa_shift == 0) {
1232 		panic("%s() : Can't determine I/O TLB size.\n", __func__);
1233 	}
1234 	return (1 << dev->spa_shift);
1235 }
1236 #else
1237 
1238 /* Uturn supports 256 TLB entries */
1239 #define CCIO_CHAINID_SHIFT	8
1240 #define CCIO_CHAINID_MASK	0xff
1241 #endif /* 0 */
1242 
1243 /* We *can't* support JAVA (T600). Venture there at your own risk. */
1244 static const struct parisc_device_id ccio_tbl[] __initconst = {
1245 	{ HPHW_IOA, HVERSION_REV_ANY_ID, U2_IOA_RUNWAY, 0xb }, /* U2 */
1246 	{ HPHW_IOA, HVERSION_REV_ANY_ID, UTURN_IOA_RUNWAY, 0xb }, /* UTurn */
1247 	{ 0, }
1248 };
1249 
1250 static int ccio_probe(struct parisc_device *dev);
1251 
1252 static struct parisc_driver ccio_driver __refdata = {
1253 	.name =		"ccio",
1254 	.id_table =	ccio_tbl,
1255 	.probe =	ccio_probe,
1256 };
1257 
1258 /**
1259  * ccio_ioc_init - Initialize the I/O Controller
1260  * @ioc: The I/O Controller.
1261  *
1262  * Initialize the I/O Controller which includes setting up the
1263  * I/O Page Directory, the resource map, and initalizing the
1264  * U2/Uturn chip into virtual mode.
1265  */
1266 static void
1267 ccio_ioc_init(struct ioc *ioc)
1268 {
1269 	int i;
1270 	unsigned int iov_order;
1271 	u32 iova_space_size;
1272 
1273 	/*
1274 	** Determine IOVA Space size from memory size.
1275 	**
1276 	** Ideally, PCI drivers would register the maximum number
1277 	** of DMA they can have outstanding for each device they
1278 	** own.  Next best thing would be to guess how much DMA
1279 	** can be outstanding based on PCI Class/sub-class. Both
1280 	** methods still require some "extra" to support PCI
1281 	** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1282 	*/
1283 
1284 	iova_space_size = (u32) (totalram_pages / count_parisc_driver(&ccio_driver));
1285 
1286 	/* limit IOVA space size to 1MB-1GB */
1287 
1288 	if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1289 		iova_space_size =  1 << (20 - PAGE_SHIFT);
1290 #ifdef __LP64__
1291 	} else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1292 		iova_space_size =  1 << (30 - PAGE_SHIFT);
1293 #endif
1294 	}
1295 
1296 	/*
1297 	** iova space must be log2() in size.
1298 	** thus, pdir/res_map will also be log2().
1299 	*/
1300 
1301 	/* We could use larger page sizes in order to *decrease* the number
1302 	** of mappings needed.  (ie 8k pages means 1/2 the mappings).
1303 	**
1304 	** Note: Grant Grunder says "Using 8k I/O pages isn't trivial either
1305 	**   since the pages must also be physically contiguous - typically
1306 	**   this is the case under linux."
1307 	*/
1308 
1309 	iov_order = get_order(iova_space_size << PAGE_SHIFT);
1310 
1311 	/* iova_space_size is now bytes, not pages */
1312 	iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1313 
1314 	ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1315 
1316 	BUG_ON(ioc->pdir_size > 8 * 1024 * 1024);   /* max pdir size <= 8MB */
1317 
1318 	/* Verify it's a power of two */
1319 	BUG_ON((1 << get_order(ioc->pdir_size)) != (ioc->pdir_size >> PAGE_SHIFT));
1320 
1321 	DBG_INIT("%s() hpa 0x%p mem %luMB IOV %dMB (%d bits)\n",
1322 			__func__, ioc->ioc_regs,
1323 			(unsigned long) totalram_pages >> (20 - PAGE_SHIFT),
1324 			iova_space_size>>20,
1325 			iov_order + PAGE_SHIFT);
1326 
1327 	ioc->pdir_base = (u64 *)__get_free_pages(GFP_KERNEL,
1328 						 get_order(ioc->pdir_size));
1329 	if(NULL == ioc->pdir_base) {
1330 		panic("%s() could not allocate I/O Page Table\n", __func__);
1331 	}
1332 	memset(ioc->pdir_base, 0, ioc->pdir_size);
1333 
1334 	BUG_ON((((unsigned long)ioc->pdir_base) & PAGE_MASK) != (unsigned long)ioc->pdir_base);
1335 	DBG_INIT(" base %p\n", ioc->pdir_base);
1336 
1337 	/* resource map size dictated by pdir_size */
1338  	ioc->res_size = (ioc->pdir_size / sizeof(u64)) >> 3;
1339 	DBG_INIT("%s() res_size 0x%x\n", __func__, ioc->res_size);
1340 
1341 	ioc->res_map = (u8 *)__get_free_pages(GFP_KERNEL,
1342 					      get_order(ioc->res_size));
1343 	if(NULL == ioc->res_map) {
1344 		panic("%s() could not allocate resource map\n", __func__);
1345 	}
1346 	memset(ioc->res_map, 0, ioc->res_size);
1347 
1348 	/* Initialize the res_hint to 16 */
1349 	ioc->res_hint = 16;
1350 
1351 	/* Initialize the spinlock */
1352 	spin_lock_init(&ioc->res_lock);
1353 
1354 	/*
1355 	** Chainid is the upper most bits of an IOVP used to determine
1356 	** which TLB entry an IOVP will use.
1357 	*/
1358 	ioc->chainid_shift = get_order(iova_space_size) + PAGE_SHIFT - CCIO_CHAINID_SHIFT;
1359 	DBG_INIT(" chainid_shift 0x%x\n", ioc->chainid_shift);
1360 
1361 	/*
1362 	** Initialize IOA hardware
1363 	*/
1364 	WRITE_U32(CCIO_CHAINID_MASK << ioc->chainid_shift,
1365 		  &ioc->ioc_regs->io_chain_id_mask);
1366 
1367 	WRITE_U32(virt_to_phys(ioc->pdir_base),
1368 		  &ioc->ioc_regs->io_pdir_base);
1369 
1370 	/*
1371 	** Go to "Virtual Mode"
1372 	*/
1373 	WRITE_U32(IOA_NORMAL_MODE, &ioc->ioc_regs->io_control);
1374 
1375 	/*
1376 	** Initialize all I/O TLB entries to 0 (Valid bit off).
1377 	*/
1378 	WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_m);
1379 	WRITE_U32(0, &ioc->ioc_regs->io_tlb_entry_l);
1380 
1381 	for(i = 1 << CCIO_CHAINID_SHIFT; i ; i--) {
1382 		WRITE_U32((CMD_TLB_DIRECT_WRITE | (i << ioc->chainid_shift)),
1383 			  &ioc->ioc_regs->io_command);
1384 	}
1385 }
1386 
1387 static void __init
1388 ccio_init_resource(struct resource *res, char *name, void __iomem *ioaddr)
1389 {
1390 	int result;
1391 
1392 	res->parent = NULL;
1393 	res->flags = IORESOURCE_MEM;
1394 	/*
1395 	 * bracing ((signed) ...) are required for 64bit kernel because
1396 	 * we only want to sign extend the lower 16 bits of the register.
1397 	 * The upper 16-bits of range registers are hardcoded to 0xffff.
1398 	 */
1399 	res->start = (unsigned long)((signed) READ_U32(ioaddr) << 16);
1400 	res->end = (unsigned long)((signed) (READ_U32(ioaddr + 4) << 16) - 1);
1401 	res->name = name;
1402 	/*
1403 	 * Check if this MMIO range is disable
1404 	 */
1405 	if (res->end + 1 == res->start)
1406 		return;
1407 
1408 	/* On some platforms (e.g. K-Class), we have already registered
1409 	 * resources for devices reported by firmware. Some are children
1410 	 * of ccio.
1411 	 * "insert" ccio ranges in the mmio hierarchy (/proc/iomem).
1412 	 */
1413 	result = insert_resource(&iomem_resource, res);
1414 	if (result < 0) {
1415 		printk(KERN_ERR "%s() failed to claim CCIO bus address space (%08lx,%08lx)\n",
1416 			__func__, (unsigned long)res->start, (unsigned long)res->end);
1417 	}
1418 }
1419 
1420 static void __init ccio_init_resources(struct ioc *ioc)
1421 {
1422 	struct resource *res = ioc->mmio_region;
1423 	char *name = kmalloc(14, GFP_KERNEL);
1424 
1425 	snprintf(name, 14, "GSC Bus [%d/]", ioc->hw_path);
1426 
1427 	ccio_init_resource(res, name, &ioc->ioc_regs->io_io_low);
1428 	ccio_init_resource(res + 1, name, &ioc->ioc_regs->io_io_low_hv);
1429 }
1430 
1431 static int new_ioc_area(struct resource *res, unsigned long size,
1432 		unsigned long min, unsigned long max, unsigned long align)
1433 {
1434 	if (max <= min)
1435 		return -EBUSY;
1436 
1437 	res->start = (max - size + 1) &~ (align - 1);
1438 	res->end = res->start + size;
1439 
1440 	/* We might be trying to expand the MMIO range to include
1441 	 * a child device that has already registered it's MMIO space.
1442 	 * Use "insert" instead of request_resource().
1443 	 */
1444 	if (!insert_resource(&iomem_resource, res))
1445 		return 0;
1446 
1447 	return new_ioc_area(res, size, min, max - size, align);
1448 }
1449 
1450 static int expand_ioc_area(struct resource *res, unsigned long size,
1451 		unsigned long min, unsigned long max, unsigned long align)
1452 {
1453 	unsigned long start, len;
1454 
1455 	if (!res->parent)
1456 		return new_ioc_area(res, size, min, max, align);
1457 
1458 	start = (res->start - size) &~ (align - 1);
1459 	len = res->end - start + 1;
1460 	if (start >= min) {
1461 		if (!adjust_resource(res, start, len))
1462 			return 0;
1463 	}
1464 
1465 	start = res->start;
1466 	len = ((size + res->end + align) &~ (align - 1)) - start;
1467 	if (start + len <= max) {
1468 		if (!adjust_resource(res, start, len))
1469 			return 0;
1470 	}
1471 
1472 	return -EBUSY;
1473 }
1474 
1475 /*
1476  * Dino calls this function.  Beware that we may get called on systems
1477  * which have no IOC (725, B180, C160L, etc) but do have a Dino.
1478  * So it's legal to find no parent IOC.
1479  *
1480  * Some other issues: one of the resources in the ioc may be unassigned.
1481  */
1482 int ccio_allocate_resource(const struct parisc_device *dev,
1483 		struct resource *res, unsigned long size,
1484 		unsigned long min, unsigned long max, unsigned long align)
1485 {
1486 	struct resource *parent = &iomem_resource;
1487 	struct ioc *ioc = ccio_get_iommu(dev);
1488 	if (!ioc)
1489 		goto out;
1490 
1491 	parent = ioc->mmio_region;
1492 	if (parent->parent &&
1493 	    !allocate_resource(parent, res, size, min, max, align, NULL, NULL))
1494 		return 0;
1495 
1496 	if ((parent + 1)->parent &&
1497 	    !allocate_resource(parent + 1, res, size, min, max, align,
1498 				NULL, NULL))
1499 		return 0;
1500 
1501 	if (!expand_ioc_area(parent, size, min, max, align)) {
1502 		__raw_writel(((parent->start)>>16) | 0xffff0000,
1503 			     &ioc->ioc_regs->io_io_low);
1504 		__raw_writel(((parent->end)>>16) | 0xffff0000,
1505 			     &ioc->ioc_regs->io_io_high);
1506 	} else if (!expand_ioc_area(parent + 1, size, min, max, align)) {
1507 		parent++;
1508 		__raw_writel(((parent->start)>>16) | 0xffff0000,
1509 			     &ioc->ioc_regs->io_io_low_hv);
1510 		__raw_writel(((parent->end)>>16) | 0xffff0000,
1511 			     &ioc->ioc_regs->io_io_high_hv);
1512 	} else {
1513 		return -EBUSY;
1514 	}
1515 
1516  out:
1517 	return allocate_resource(parent, res, size, min, max, align, NULL,NULL);
1518 }
1519 
1520 int ccio_request_resource(const struct parisc_device *dev,
1521 		struct resource *res)
1522 {
1523 	struct resource *parent;
1524 	struct ioc *ioc = ccio_get_iommu(dev);
1525 
1526 	if (!ioc) {
1527 		parent = &iomem_resource;
1528 	} else if ((ioc->mmio_region->start <= res->start) &&
1529 			(res->end <= ioc->mmio_region->end)) {
1530 		parent = ioc->mmio_region;
1531 	} else if (((ioc->mmio_region + 1)->start <= res->start) &&
1532 			(res->end <= (ioc->mmio_region + 1)->end)) {
1533 		parent = ioc->mmio_region + 1;
1534 	} else {
1535 		return -EBUSY;
1536 	}
1537 
1538 	/* "transparent" bus bridges need to register MMIO resources
1539 	 * firmware assigned them. e.g. children of hppb.c (e.g. K-class)
1540 	 * registered their resources in the PDC "bus walk" (See
1541 	 * arch/parisc/kernel/inventory.c).
1542 	 */
1543 	return insert_resource(parent, res);
1544 }
1545 
1546 /**
1547  * ccio_probe - Determine if ccio should claim this device.
1548  * @dev: The device which has been found
1549  *
1550  * Determine if ccio should claim this chip (return 0) or not (return 1).
1551  * If so, initialize the chip and tell other partners in crime they
1552  * have work to do.
1553  */
1554 static int __init ccio_probe(struct parisc_device *dev)
1555 {
1556 	int i;
1557 	struct ioc *ioc, **ioc_p = &ioc_list;
1558 
1559 	ioc = kzalloc(sizeof(struct ioc), GFP_KERNEL);
1560 	if (ioc == NULL) {
1561 		printk(KERN_ERR MODULE_NAME ": memory allocation failure\n");
1562 		return -ENOMEM;
1563 	}
1564 
1565 	ioc->name = dev->id.hversion == U2_IOA_RUNWAY ? "U2" : "UTurn";
1566 
1567 	printk(KERN_INFO "Found %s at 0x%lx\n", ioc->name,
1568 		(unsigned long)dev->hpa.start);
1569 
1570 	for (i = 0; i < ioc_count; i++) {
1571 		ioc_p = &(*ioc_p)->next;
1572 	}
1573 	*ioc_p = ioc;
1574 
1575 	ioc->hw_path = dev->hw_path;
1576 	ioc->ioc_regs = ioremap_nocache(dev->hpa.start, 4096);
1577 	if (!ioc->ioc_regs) {
1578 		kfree(ioc);
1579 		return -ENOMEM;
1580 	}
1581 	ccio_ioc_init(ioc);
1582 	ccio_init_resources(ioc);
1583 	hppa_dma_ops = &ccio_ops;
1584 	dev->dev.platform_data = kzalloc(sizeof(struct pci_hba_data), GFP_KERNEL);
1585 
1586 	/* if this fails, no I/O cards will work, so may as well bug */
1587 	BUG_ON(dev->dev.platform_data == NULL);
1588 	HBA_DATA(dev->dev.platform_data)->iommu = ioc;
1589 
1590 #ifdef CONFIG_PROC_FS
1591 	if (ioc_count == 0) {
1592 		proc_create(MODULE_NAME, 0, proc_runway_root,
1593 			    &ccio_proc_info_fops);
1594 		proc_create(MODULE_NAME"-bitmap", 0, proc_runway_root,
1595 			    &ccio_proc_bitmap_fops);
1596 	}
1597 #endif
1598 	ioc_count++;
1599 
1600 	parisc_has_iommu();
1601 	return 0;
1602 }
1603 
1604 /**
1605  * ccio_init - ccio initialization procedure.
1606  *
1607  * Register this driver.
1608  */
1609 void __init ccio_init(void)
1610 {
1611 	register_parisc_driver(&ccio_driver);
1612 }
1613 
1614