2 ** System Bus Adapter (SBA) I/O MMU manager
4 ** (c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org>
5 ** (c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com>
6 ** (c) Copyright 2000-2004 Hewlett-Packard Company
8 ** Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
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.
16 ** This module initializes the IOC (I/O Controller) found on B1000/C3000/
17 ** J5000/J7000/N-class/L-class machines and their successors.
19 ** FIXME: add DMA hint support programming in both sba and lba modules.
22 #include <linux/types.h>
23 #include <linux/kernel.h>
24 #include <linux/spinlock.h>
25 #include <linux/slab.h>
26 #include <linux/init.h>
29 #include <linux/string.h>
30 #include <linux/pci.h>
32 #include <asm/byteorder.h>
34 #include <asm/dma.h> /* for DMA_CHUNK_SIZE */
36 #include <asm/hardware.h> /* for register_parisc_driver() stuff */
38 #include <linux/proc_fs.h>
39 #include <linux/seq_file.h>
41 #include <asm/mckinley.h> /* for proc_mckinley_root */
42 #include <asm/runway.h> /* for proc_runway_root */
43 #include <asm/pdc.h> /* for PDC_MODEL_* */
44 #include <asm/pdcpat.h> /* for is_pdc_pat() */
45 #include <asm/parisc-device.h>
47 #define MODULE_NAME "SBA"
50 /* depends on proc fs support. But costs CPU performance */
51 #undef SBA_COLLECT_STATS
55 ** The number of debug flags is a clue - this code is fragile.
56 ** Don't even think about messing with it unless you have
57 ** plenty of 710's to sacrifice to the computer gods. :^)
61 #undef DEBUG_SBA_RUN_SG
62 #undef DEBUG_SBA_RESOURCE
63 #undef ASSERT_PDIR_SANITY
64 #undef DEBUG_LARGE_SG_ENTRIES
68 #define DBG_INIT(x...) printk(x)
70 #define DBG_INIT(x...)
74 #define DBG_RUN(x...) printk(x)
79 #ifdef DEBUG_SBA_RUN_SG
80 #define DBG_RUN_SG(x...) printk(x)
82 #define DBG_RUN_SG(x...)
86 #ifdef DEBUG_SBA_RESOURCE
87 #define DBG_RES(x...) printk(x)
92 #if defined(CONFIG_64BIT)
93 /* "low end" PA8800 machines use ZX1 chipset: PAT PDC and only run 64-bit */
97 #define SBA_INLINE __inline__
101 ** The number of pdir entries to "free" before issueing
102 ** a read to PCOM register to flush out PCOM writes.
103 ** Interacts with allocation granularity (ie 4 or 8 entries
104 ** allocated and free'd/purged at a time might make this
105 ** less interesting).
107 #define DELAYED_RESOURCE_CNT 16
109 #define DEFAULT_DMA_HINT_REG 0
111 #define ASTRO_RUNWAY_PORT 0x582
112 #define IKE_MERCED_PORT 0x803
113 #define REO_MERCED_PORT 0x804
114 #define REOG_MERCED_PORT 0x805
115 #define PLUTO_MCKINLEY_PORT 0x880
117 #define SBA_FUNC_ID 0x0000 /* function id */
118 #define SBA_FCLASS 0x0008 /* function class, bist, header, rev... */
120 #define IS_ASTRO(id) ((id)->hversion == ASTRO_RUNWAY_PORT)
121 #define IS_IKE(id) ((id)->hversion == IKE_MERCED_PORT)
122 #define IS_PLUTO(id) ((id)->hversion == PLUTO_MCKINLEY_PORT)
124 #define SBA_FUNC_SIZE 4096 /* SBA configuration function reg set */
126 #define ASTRO_IOC_OFFSET (32 * SBA_FUNC_SIZE)
127 #define PLUTO_IOC_OFFSET (1 * SBA_FUNC_SIZE)
128 /* Ike's IOC's occupy functions 2 and 3 */
129 #define IKE_IOC_OFFSET(p) ((p+2) * SBA_FUNC_SIZE)
131 #define IOC_CTRL 0x8 /* IOC_CTRL offset */
132 #define IOC_CTRL_TC (1 << 0) /* TOC Enable */
133 #define IOC_CTRL_CE (1 << 1) /* Coalesce Enable */
134 #define IOC_CTRL_DE (1 << 2) /* Dillon Enable */
135 #define IOC_CTRL_RM (1 << 8) /* Real Mode */
136 #define IOC_CTRL_NC (1 << 9) /* Non Coherent Mode */
137 #define IOC_CTRL_D4 (1 << 11) /* Disable 4-byte coalescing */
138 #define IOC_CTRL_DD (1 << 13) /* Disable distr. LMMIO range coalescing */
140 #define MAX_IOC 2 /* per Ike. Pluto/Astro only have 1. */
142 #define ROPES_PER_IOC 8 /* per Ike half or Pluto/Astro */
146 ** Offsets into MBIB (Function 0 on Ike and hopefully Astro)
147 ** Firmware programs this stuff. Don't touch it.
149 #define LMMIO_DIRECT0_BASE 0x300
150 #define LMMIO_DIRECT0_MASK 0x308
151 #define LMMIO_DIRECT0_ROUTE 0x310
153 #define LMMIO_DIST_BASE 0x360
154 #define LMMIO_DIST_MASK 0x368
155 #define LMMIO_DIST_ROUTE 0x370
157 #define IOS_DIST_BASE 0x390
158 #define IOS_DIST_MASK 0x398
159 #define IOS_DIST_ROUTE 0x3A0
161 #define IOS_DIRECT_BASE 0x3C0
162 #define IOS_DIRECT_MASK 0x3C8
163 #define IOS_DIRECT_ROUTE 0x3D0
166 ** Offsets into I/O TLB (Function 2 and 3 on Ike)
168 #define ROPE0_CTL 0x200 /* "regbus pci0" */
169 #define ROPE1_CTL 0x208
170 #define ROPE2_CTL 0x210
171 #define ROPE3_CTL 0x218
172 #define ROPE4_CTL 0x220
173 #define ROPE5_CTL 0x228
174 #define ROPE6_CTL 0x230
175 #define ROPE7_CTL 0x238
177 #define IOC_ROPE0_CFG 0x500 /* pluto only */
178 #define IOC_ROPE_AO 0x10 /* Allow "Relaxed Ordering" */
182 #define HF_ENABLE 0x40
185 #define IOC_IBASE 0x300 /* IO TLB */
186 #define IOC_IMASK 0x308
187 #define IOC_PCOM 0x310
188 #define IOC_TCNFG 0x318
189 #define IOC_PDIR_BASE 0x320
191 /* AGP GART driver looks for this */
192 #define SBA_IOMMU_COOKIE 0x0000badbadc0ffeeUL
196 ** IOC supports 4/8/16/64KB page sizes (see TCNFG register)
197 ** It's safer (avoid memory corruption) to keep DMA page mappings
198 ** equivalently sized to VM PAGE_SIZE.
200 ** We really can't avoid generating a new mapping for each
201 ** page since the Virtual Coherence Index has to be generated
202 ** and updated for each page.
204 ** PAGE_SIZE could be greater than IOVP_SIZE. But not the inverse.
206 #define IOVP_SIZE PAGE_SIZE
207 #define IOVP_SHIFT PAGE_SHIFT
208 #define IOVP_MASK PAGE_MASK
210 #define SBA_PERF_CFG 0x708 /* Performance Counter stuff */
211 #define SBA_PERF_MASK1 0x718
212 #define SBA_PERF_MASK2 0x730
216 ** Offsets into PCI Performance Counters (functions 12 and 13)
217 ** Controlled by PERF registers in function 2 & 3 respectively.
219 #define SBA_PERF_CNT1 0x200
220 #define SBA_PERF_CNT2 0x208
221 #define SBA_PERF_CNT3 0x210
225 void __iomem
*ioc_hpa
; /* I/O MMU base address */
226 char *res_map
; /* resource map, bit == pdir entry */
227 u64
*pdir_base
; /* physical base address */
228 unsigned long ibase
; /* pdir IOV Space base - shared w/lba_pci */
229 unsigned long imask
; /* pdir IOV Space mask - shared w/lba_pci */
231 unsigned long iovp_mask
; /* help convert IOVA to IOVP */
233 unsigned long *res_hint
; /* next avail IOVP - circular search */
235 unsigned int res_bitshift
; /* from the LEFT! */
236 unsigned int res_size
; /* size of resource map in bytes */
237 #ifdef SBA_HINT_SUPPORT
238 /* FIXME : DMA HINTs not used */
239 unsigned long hint_mask_pdir
; /* bits used for DMA hints */
240 unsigned int hint_shift_pdir
;
242 #if DELAYED_RESOURCE_CNT > 0
244 struct sba_dma_pair
{
247 } saved
[DELAYED_RESOURCE_CNT
];
250 #ifdef SBA_COLLECT_STATS
251 #define SBA_SEARCH_SAMPLE 0x100
252 unsigned long avg_search
[SBA_SEARCH_SAMPLE
];
253 unsigned long avg_idx
; /* current index into avg_search */
254 unsigned long used_pages
;
255 unsigned long msingle_calls
;
256 unsigned long msingle_pages
;
257 unsigned long msg_calls
;
258 unsigned long msg_pages
;
259 unsigned long usingle_calls
;
260 unsigned long usingle_pages
;
261 unsigned long usg_calls
;
262 unsigned long usg_pages
;
265 /* STUFF We don't need in performance path */
266 unsigned int pdir_size
; /* in bytes, determined by IOV Space size */
270 struct sba_device
*next
; /* list of SBA's in system */
271 struct parisc_device
*dev
; /* dev found in bus walk */
272 struct parisc_device_id
*iodc
; /* data about dev from firmware */
274 void __iomem
*sba_hpa
; /* base address */
276 unsigned int flags
; /* state/functionality enabled */
277 unsigned int hw_rev
; /* HW revision of chip */
279 struct resource chip_resv
; /* MMIO reserved for chip */
280 struct resource iommu_resv
; /* MMIO reserved for iommu */
282 unsigned int num_ioc
; /* number of on-board IOC's */
283 struct ioc ioc
[MAX_IOC
];
287 static struct sba_device
*sba_list
;
289 static unsigned long ioc_needs_fdc
= 0;
291 /* global count of IOMMUs in the system */
292 static unsigned int global_ioc_cnt
= 0;
294 /* PA8700 (Piranha 2.2) bug workaround */
295 static unsigned long piranha_bad_128k
= 0;
297 /* Looks nice and keeps the compiler happy */
298 #define SBA_DEV(d) ((struct sba_device *) (d))
300 #ifdef SBA_AGP_SUPPORT
301 static int reserve_sba_gart
= 1;
304 #define ROUNDUP(x,y) ((x + ((y)-1)) & ~((y)-1))
307 /************************************
308 ** SBA register read and write support
310 ** BE WARNED: register writes are posted.
311 ** (ie follow writes which must reach HW with a read)
313 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
315 #define READ_REG32(addr) readl(addr)
316 #define READ_REG64(addr) readq(addr)
317 #define WRITE_REG32(val, addr) writel((val), (addr))
318 #define WRITE_REG64(val, addr) writeq((val), (addr))
321 #define READ_REG(addr) READ_REG64(addr)
322 #define WRITE_REG(value, addr) WRITE_REG64(value, addr)
324 #define READ_REG(addr) READ_REG32(addr)
325 #define WRITE_REG(value, addr) WRITE_REG32(value, addr)
328 #ifdef DEBUG_SBA_INIT
330 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
333 * sba_dump_ranges - debugging only - print ranges assigned to this IOA
334 * @hpa: base address of the sba
336 * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
337 * IO Adapter (aka Bus Converter).
340 sba_dump_ranges(void __iomem
*hpa
)
342 DBG_INIT("SBA at 0x%p\n", hpa
);
343 DBG_INIT("IOS_DIST_BASE : %Lx\n", READ_REG64(hpa
+IOS_DIST_BASE
));
344 DBG_INIT("IOS_DIST_MASK : %Lx\n", READ_REG64(hpa
+IOS_DIST_MASK
));
345 DBG_INIT("IOS_DIST_ROUTE : %Lx\n", READ_REG64(hpa
+IOS_DIST_ROUTE
));
347 DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa
+IOS_DIRECT_BASE
));
348 DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa
+IOS_DIRECT_MASK
));
349 DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa
+IOS_DIRECT_ROUTE
));
353 * sba_dump_tlb - debugging only - print IOMMU operating parameters
354 * @hpa: base address of the IOMMU
356 * Print the size/location of the IO MMU PDIR.
358 static void sba_dump_tlb(void __iomem
*hpa
)
360 DBG_INIT("IO TLB at 0x%p\n", hpa
);
361 DBG_INIT("IOC_IBASE : 0x%Lx\n", READ_REG64(hpa
+IOC_IBASE
));
362 DBG_INIT("IOC_IMASK : 0x%Lx\n", READ_REG64(hpa
+IOC_IMASK
));
363 DBG_INIT("IOC_TCNFG : 0x%Lx\n", READ_REG64(hpa
+IOC_TCNFG
));
364 DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa
+IOC_PDIR_BASE
));
368 #define sba_dump_ranges(x)
369 #define sba_dump_tlb(x)
370 #endif /* DEBUG_SBA_INIT */
373 #ifdef ASSERT_PDIR_SANITY
376 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
377 * @ioc: IO MMU structure which owns the pdir we are interested in.
378 * @msg: text to print ont the output line.
381 * Print one entry of the IO MMU PDIR in human readable form.
384 sba_dump_pdir_entry(struct ioc
*ioc
, char *msg
, uint pide
)
386 /* start printing from lowest pde in rval */
387 u64
*ptr
= &(ioc
->pdir_base
[pide
& (~0U * BITS_PER_LONG
)]);
388 unsigned long *rptr
= (unsigned long *) &(ioc
->res_map
[(pide
>>3) & ~(sizeof(unsigned long) - 1)]);
391 printk(KERN_DEBUG
"SBA: %s rp %p bit %d rval 0x%lx\n",
393 rptr
, pide
& (BITS_PER_LONG
- 1), *rptr
);
396 while (rcnt
< BITS_PER_LONG
) {
397 printk(KERN_DEBUG
"%s %2d %p %016Lx\n",
398 (rcnt
== (pide
& (BITS_PER_LONG
- 1)))
404 printk(KERN_DEBUG
"%s", msg
);
409 * sba_check_pdir - debugging only - consistency checker
410 * @ioc: IO MMU structure which owns the pdir we are interested in.
411 * @msg: text to print ont the output line.
413 * Verify the resource map and pdir state is consistent
416 sba_check_pdir(struct ioc
*ioc
, char *msg
)
418 u32
*rptr_end
= (u32
*) &(ioc
->res_map
[ioc
->res_size
]);
419 u32
*rptr
= (u32
*) ioc
->res_map
; /* resource map ptr */
420 u64
*pptr
= ioc
->pdir_base
; /* pdir ptr */
423 while (rptr
< rptr_end
) {
425 int rcnt
= 32; /* number of bits we might check */
428 /* Get last byte and highest bit from that */
429 u32 pde
= ((u32
) (((char *)pptr
)[7])) << 24;
430 if ((rval
^ pde
) & 0x80000000)
433 ** BUMMER! -- res_map != pdir --
434 ** Dump rval and matching pdir entries
436 sba_dump_pdir_entry(ioc
, msg
, pide
);
440 rval
<<= 1; /* try the next bit */
444 rptr
++; /* look at next word of res_map */
446 /* It'd be nice if we always got here :^) */
452 * sba_dump_sg - debugging only - print Scatter-Gather list
453 * @ioc: IO MMU structure which owns the pdir we are interested in.
454 * @startsg: head of the SG list
455 * @nents: number of entries in SG list
457 * print the SG list so we can verify it's correct by hand.
460 sba_dump_sg( struct ioc
*ioc
, struct scatterlist
*startsg
, int nents
)
462 while (nents
-- > 0) {
463 printk(KERN_DEBUG
" %d : %08lx/%05x %p/%05x\n",
465 (unsigned long) sg_dma_address(startsg
),
467 sg_virt_addr(startsg
), startsg
->length
);
472 #endif /* ASSERT_PDIR_SANITY */
477 /**************************************************************
479 * I/O Pdir Resource Management
481 * Bits set in the resource map are in use.
482 * Each bit can represent a number of pages.
483 * LSbs represent lower addresses (IOVA's).
485 ***************************************************************/
486 #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */
488 /* Convert from IOVP to IOVA and vice versa. */
491 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
492 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
493 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
495 /* only support Astro and ancestors. Saves a few cycles in key places */
496 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
497 #define SBA_IOVP(ioc,iova) (iova)
500 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT)
502 #define RESMAP_MASK(n) (~0UL << (BITS_PER_LONG - (n)))
503 #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1)
507 * sba_search_bitmap - find free space in IO PDIR resource bitmap
508 * @ioc: IO MMU structure which owns the pdir we are interested in.
509 * @bits_wanted: number of entries we need.
511 * Find consecutive free bits in resource bitmap.
512 * Each bit represents one entry in the IO Pdir.
513 * Cool perf optimization: search for log2(size) bits at a time.
515 static SBA_INLINE
unsigned long
516 sba_search_bitmap(struct ioc
*ioc
, unsigned long bits_wanted
)
518 unsigned long *res_ptr
= ioc
->res_hint
;
519 unsigned long *res_end
= (unsigned long *) &(ioc
->res_map
[ioc
->res_size
]);
520 unsigned long pide
= ~0UL;
522 if (bits_wanted
> (BITS_PER_LONG
/2)) {
523 /* Search word at a time - no mask needed */
524 for(; res_ptr
< res_end
; ++res_ptr
) {
526 *res_ptr
= RESMAP_MASK(bits_wanted
);
527 pide
= ((unsigned long)res_ptr
- (unsigned long)ioc
->res_map
);
528 pide
<<= 3; /* convert to bit address */
532 /* point to the next word on next pass */
534 ioc
->res_bitshift
= 0;
537 ** Search the resource bit map on well-aligned values.
538 ** "o" is the alignment.
539 ** We need the alignment to invalidate I/O TLB using
540 ** SBA HW features in the unmap path.
542 unsigned long o
= 1 << get_order(bits_wanted
<< PAGE_SHIFT
);
543 uint bitshiftcnt
= ROUNDUP(ioc
->res_bitshift
, o
);
546 if (bitshiftcnt
>= BITS_PER_LONG
) {
550 mask
= RESMAP_MASK(bits_wanted
) >> bitshiftcnt
;
552 DBG_RES("%s() o %ld %p", __FUNCTION__
, o
, res_ptr
);
553 while(res_ptr
< res_end
)
555 DBG_RES(" %p %lx %lx\n", res_ptr
, mask
, *res_ptr
);
557 if(((*res_ptr
) & mask
) == 0) {
558 *res_ptr
|= mask
; /* mark resources busy! */
559 pide
= ((unsigned long)res_ptr
- (unsigned long)ioc
->res_map
);
560 pide
<<= 3; /* convert to bit address */
567 mask
= RESMAP_MASK(bits_wanted
);
572 /* look in the same word on the next pass */
573 ioc
->res_bitshift
= bitshiftcnt
+ bits_wanted
;
577 if (res_end
<= res_ptr
) {
578 ioc
->res_hint
= (unsigned long *) ioc
->res_map
;
579 ioc
->res_bitshift
= 0;
581 ioc
->res_hint
= res_ptr
;
588 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
589 * @ioc: IO MMU structure which owns the pdir we are interested in.
590 * @size: number of bytes to create a mapping for
592 * Given a size, find consecutive unmarked and then mark those bits in the
596 sba_alloc_range(struct ioc
*ioc
, size_t size
)
598 unsigned int pages_needed
= size
>> IOVP_SHIFT
;
599 #ifdef SBA_COLLECT_STATS
600 unsigned long cr_start
= mfctl(16);
604 pide
= sba_search_bitmap(ioc
, pages_needed
);
605 if (pide
>= (ioc
->res_size
<< 3)) {
606 pide
= sba_search_bitmap(ioc
, pages_needed
);
607 if (pide
>= (ioc
->res_size
<< 3))
608 panic("%s: I/O MMU @ %p is out of mapping resources\n",
609 __FILE__
, ioc
->ioc_hpa
);
612 #ifdef ASSERT_PDIR_SANITY
613 /* verify the first enable bit is clear */
614 if(0x00 != ((u8
*) ioc
->pdir_base
)[pide
*sizeof(u64
) + 7]) {
615 sba_dump_pdir_entry(ioc
, "sba_search_bitmap() botched it?", pide
);
619 DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
620 __FUNCTION__
, size
, pages_needed
, pide
,
621 (uint
) ((unsigned long) ioc
->res_hint
- (unsigned long) ioc
->res_map
),
624 #ifdef SBA_COLLECT_STATS
626 unsigned long cr_end
= mfctl(16);
627 unsigned long tmp
= cr_end
- cr_start
;
628 /* check for roll over */
629 cr_start
= (cr_end
< cr_start
) ? -(tmp
) : (tmp
);
631 ioc
->avg_search
[ioc
->avg_idx
++] = cr_start
;
632 ioc
->avg_idx
&= SBA_SEARCH_SAMPLE
- 1;
634 ioc
->used_pages
+= pages_needed
;
642 * sba_free_range - unmark bits in IO PDIR resource bitmap
643 * @ioc: IO MMU structure which owns the pdir we are interested in.
644 * @iova: IO virtual address which was previously allocated.
645 * @size: number of bytes to create a mapping for
647 * clear bits in the ioc's resource map
649 static SBA_INLINE
void
650 sba_free_range(struct ioc
*ioc
, dma_addr_t iova
, size_t size
)
652 unsigned long iovp
= SBA_IOVP(ioc
, iova
);
653 unsigned int pide
= PDIR_INDEX(iovp
);
654 unsigned int ridx
= pide
>> 3; /* convert bit to byte address */
655 unsigned long *res_ptr
= (unsigned long *) &((ioc
)->res_map
[ridx
& ~RESMAP_IDX_MASK
]);
657 int bits_not_wanted
= size
>> IOVP_SHIFT
;
659 /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
660 unsigned long m
= RESMAP_MASK(bits_not_wanted
) >> (pide
& (BITS_PER_LONG
- 1));
662 DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
663 __FUNCTION__
, (uint
) iova
, size
,
664 bits_not_wanted
, m
, pide
, res_ptr
, *res_ptr
);
666 #ifdef SBA_COLLECT_STATS
667 ioc
->used_pages
-= bits_not_wanted
;
674 /**************************************************************
676 * "Dynamic DMA Mapping" support (aka "Coherent I/O")
678 ***************************************************************/
680 #ifdef SBA_HINT_SUPPORT
681 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
684 typedef unsigned long space_t
;
685 #define KERNEL_SPACE 0
688 * sba_io_pdir_entry - fill in one IO PDIR entry
689 * @pdir_ptr: pointer to IO PDIR entry
690 * @sid: process Space ID - currently only support KERNEL_SPACE
691 * @vba: Virtual CPU address of buffer to map
692 * @hint: DMA hint set to use for this mapping
694 * SBA Mapping Routine
696 * Given a virtual address (vba, arg2) and space id, (sid, arg1)
697 * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
699 * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
700 * for Astro/Ike looks like:
704 * +-+---------------------+----------------------------------+----+--------+
705 * |V| U | PPN[43:12] | U | VI |
706 * +-+---------------------+----------------------------------+----+--------+
708 * Pluto is basically identical, supports fewer physical address bits:
711 * +-+------------------------+-------------------------------+----+--------+
712 * |V| U | PPN[39:12] | U | VI |
713 * +-+------------------------+-------------------------------+----+--------+
715 * V == Valid Bit (Most Significant Bit is bit 0)
717 * PPN == Physical Page Number
718 * VI == Virtual Index (aka Coherent Index)
720 * LPA instruction output is put into PPN field.
721 * LCI (Load Coherence Index) instruction provides the "VI" bits.
723 * We pre-swap the bytes since PCX-W is Big Endian and the
724 * IOMMU uses little endian for the pdir.
728 sba_io_pdir_entry(u64
*pdir_ptr
, space_t sid
, unsigned long vba
,
731 u64 pa
; /* physical address */
732 register unsigned ci
; /* coherent index */
734 pa
= virt_to_phys(vba
);
738 asm("lci 0(%%sr1, %1), %0" : "=r" (ci
) : "r" (vba
));
739 pa
|= (ci
>> 12) & 0xff; /* move CI (8 bits) into lowest byte */
741 pa
|= 0x8000000000000000ULL
; /* set "valid" bit */
742 *pdir_ptr
= cpu_to_le64(pa
); /* swap and store into I/O Pdir */
745 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
746 * (bit #61, big endian), we have to flush and sync every time
747 * IO-PDIR is changed in Ike/Astro.
750 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr
));
755 * sba_mark_invalid - invalidate one or more IO PDIR entries
756 * @ioc: IO MMU structure which owns the pdir we are interested in.
757 * @iova: IO Virtual Address mapped earlier
758 * @byte_cnt: number of bytes this mapping covers.
760 * Marking the IO PDIR entry(ies) as Invalid and invalidate
761 * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
762 * is to purge stale entries in the IO TLB when unmapping entries.
764 * The PCOM register supports purging of multiple pages, with a minium
765 * of 1 page and a maximum of 2GB. Hardware requires the address be
766 * aligned to the size of the range being purged. The size of the range
767 * must be a power of 2. The "Cool perf optimization" in the
768 * allocation routine helps keep that true.
770 static SBA_INLINE
void
771 sba_mark_invalid(struct ioc
*ioc
, dma_addr_t iova
, size_t byte_cnt
)
773 u32 iovp
= (u32
) SBA_IOVP(ioc
,iova
);
774 u64
*pdir_ptr
= &ioc
->pdir_base
[PDIR_INDEX(iovp
)];
776 #ifdef ASSERT_PDIR_SANITY
777 /* Assert first pdir entry is set.
779 ** Even though this is a big-endian machine, the entries
780 ** in the iopdir are little endian. That's why we look at
781 ** the byte at +7 instead of at +0.
783 if (0x80 != (((u8
*) pdir_ptr
)[7])) {
784 sba_dump_pdir_entry(ioc
,"sba_mark_invalid()", PDIR_INDEX(iovp
));
788 if (byte_cnt
> IOVP_SIZE
)
791 unsigned long entries_per_cacheline
= ioc_needs_fdc
?
792 L1_CACHE_ALIGN(((unsigned long) pdir_ptr
))
793 - (unsigned long) pdir_ptr
;
797 /* set "size" field for PCOM */
798 iovp
|= get_order(byte_cnt
) + PAGE_SHIFT
;
801 /* clear I/O Pdir entry "valid" bit first */
802 ((u8
*) pdir_ptr
)[7] = 0;
804 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr
));
806 entries_per_cacheline
= L1_CACHE_SHIFT
- 3;
810 byte_cnt
-= IOVP_SIZE
;
811 } while (byte_cnt
> IOVP_SIZE
);
813 iovp
|= IOVP_SHIFT
; /* set "size" field for PCOM */
816 ** clear I/O PDIR entry "valid" bit.
817 ** We have to R/M/W the cacheline regardless how much of the
818 ** pdir entry that we clobber.
819 ** The rest of the entry would be useful for debugging if we
820 ** could dump core on HPMC.
822 ((u8
*) pdir_ptr
)[7] = 0;
824 asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr
));
826 WRITE_REG( SBA_IOVA(ioc
, iovp
, 0, 0), ioc
->ioc_hpa
+IOC_PCOM
);
830 * sba_dma_supported - PCI driver can query DMA support
831 * @dev: instance of PCI owned by the driver that's asking
832 * @mask: number of address bits this PCI device can handle
834 * See Documentation/DMA-mapping.txt
836 static int sba_dma_supported( struct device
*dev
, u64 mask
)
841 printk(KERN_ERR MODULE_NAME
": EISA/ISA/et al not supported\n");
846 /* Documentation/DMA-mapping.txt tells drivers to try 64-bit first,
847 * then fall back to 32-bit if that fails.
848 * We are just "encouraging" 32-bit DMA masks here since we can
849 * never allow IOMMU bypass unless we add special support for ZX1.
857 * check if mask is >= than the current max IO Virt Address
858 * The max IO Virt address will *always* < 30 bits.
860 return((int)(mask
>= (ioc
->ibase
- 1 +
861 (ioc
->pdir_size
/ sizeof(u64
) * IOVP_SIZE
) )));
866 * sba_map_single - map one buffer and return IOVA for DMA
867 * @dev: instance of PCI owned by the driver that's asking.
868 * @addr: driver buffer to map.
869 * @size: number of bytes to map in driver buffer.
870 * @direction: R/W or both.
872 * See Documentation/DMA-mapping.txt
875 sba_map_single(struct device
*dev
, void *addr
, size_t size
,
876 enum dma_data_direction direction
)
887 /* save offset bits */
888 offset
= ((dma_addr_t
) (long) addr
) & ~IOVP_MASK
;
890 /* round up to nearest IOVP_SIZE */
891 size
= (size
+ offset
+ ~IOVP_MASK
) & IOVP_MASK
;
893 spin_lock_irqsave(&ioc
->res_lock
, flags
);
894 #ifdef ASSERT_PDIR_SANITY
895 sba_check_pdir(ioc
,"Check before sba_map_single()");
898 #ifdef SBA_COLLECT_STATS
899 ioc
->msingle_calls
++;
900 ioc
->msingle_pages
+= size
>> IOVP_SHIFT
;
902 pide
= sba_alloc_range(ioc
, size
);
903 iovp
= (dma_addr_t
) pide
<< IOVP_SHIFT
;
905 DBG_RUN("%s() 0x%p -> 0x%lx\n",
906 __FUNCTION__
, addr
, (long) iovp
| offset
);
908 pdir_start
= &(ioc
->pdir_base
[pide
]);
911 sba_io_pdir_entry(pdir_start
, KERNEL_SPACE
, (unsigned long) addr
, 0);
913 DBG_RUN(" pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
915 (u8
) (((u8
*) pdir_start
)[7]),
916 (u8
) (((u8
*) pdir_start
)[6]),
917 (u8
) (((u8
*) pdir_start
)[5]),
918 (u8
) (((u8
*) pdir_start
)[4]),
919 (u8
) (((u8
*) pdir_start
)[3]),
920 (u8
) (((u8
*) pdir_start
)[2]),
921 (u8
) (((u8
*) pdir_start
)[1]),
922 (u8
) (((u8
*) pdir_start
)[0])
930 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
932 asm volatile("sync" : : );
934 #ifdef ASSERT_PDIR_SANITY
935 sba_check_pdir(ioc
,"Check after sba_map_single()");
937 spin_unlock_irqrestore(&ioc
->res_lock
, flags
);
939 /* form complete address */
940 return SBA_IOVA(ioc
, iovp
, offset
, DEFAULT_DMA_HINT_REG
);
945 * sba_unmap_single - unmap one IOVA and free resources
946 * @dev: instance of PCI owned by the driver that's asking.
947 * @iova: IOVA of driver buffer previously mapped.
948 * @size: number of bytes mapped in driver buffer.
949 * @direction: R/W or both.
951 * See Documentation/DMA-mapping.txt
954 sba_unmap_single(struct device
*dev
, dma_addr_t iova
, size_t size
,
955 enum dma_data_direction direction
)
958 #if DELAYED_RESOURCE_CNT > 0
959 struct sba_dma_pair
*d
;
964 DBG_RUN("%s() iovp 0x%lx/%x\n", __FUNCTION__
, (long) iova
, size
);
967 offset
= iova
& ~IOVP_MASK
;
968 iova
^= offset
; /* clear offset bits */
970 size
= ROUNDUP(size
, IOVP_SIZE
);
972 spin_lock_irqsave(&ioc
->res_lock
, flags
);
974 #ifdef SBA_COLLECT_STATS
975 ioc
->usingle_calls
++;
976 ioc
->usingle_pages
+= size
>> IOVP_SHIFT
;
979 sba_mark_invalid(ioc
, iova
, size
);
981 #if DELAYED_RESOURCE_CNT > 0
982 /* Delaying when we re-use a IO Pdir entry reduces the number
983 * of MMIO reads needed to flush writes to the PCOM register.
985 d
= &(ioc
->saved
[ioc
->saved_cnt
]);
988 if (++(ioc
->saved_cnt
) >= DELAYED_RESOURCE_CNT
) {
989 int cnt
= ioc
->saved_cnt
;
991 sba_free_range(ioc
, d
->iova
, d
->size
);
996 READ_REG(ioc
->ioc_hpa
+IOC_PCOM
); /* flush purges */
998 #else /* DELAYED_RESOURCE_CNT == 0 */
999 sba_free_range(ioc
, iova
, size
);
1001 /* If fdc's were issued, force fdc's to be visible now */
1003 asm volatile("sync" : : );
1005 READ_REG(ioc
->ioc_hpa
+IOC_PCOM
); /* flush purges */
1006 #endif /* DELAYED_RESOURCE_CNT == 0 */
1008 spin_unlock_irqrestore(&ioc
->res_lock
, flags
);
1010 /* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
1011 ** For Astro based systems this isn't a big deal WRT performance.
1012 ** As long as 2.4 kernels copyin/copyout data from/to userspace,
1013 ** we don't need the syncdma. The issue here is I/O MMU cachelines
1014 ** are *not* coherent in all cases. May be hwrev dependent.
1015 ** Need to investigate more.
1016 asm volatile("syncdma");
1022 * sba_alloc_consistent - allocate/map shared mem for DMA
1023 * @hwdev: instance of PCI owned by the driver that's asking.
1024 * @size: number of bytes mapped in driver buffer.
1025 * @dma_handle: IOVA of new buffer.
1027 * See Documentation/DMA-mapping.txt
1029 static void *sba_alloc_consistent(struct device
*hwdev
, size_t size
,
1030 dma_addr_t
*dma_handle
, gfp_t gfp
)
1035 /* only support PCI */
1040 ret
= (void *) __get_free_pages(gfp
, get_order(size
));
1043 memset(ret
, 0, size
);
1044 *dma_handle
= sba_map_single(hwdev
, ret
, size
, 0);
1052 * sba_free_consistent - free/unmap shared mem for DMA
1053 * @hwdev: instance of PCI owned by the driver that's asking.
1054 * @size: number of bytes mapped in driver buffer.
1055 * @vaddr: virtual address IOVA of "consistent" buffer.
1056 * @dma_handler: IO virtual address of "consistent" buffer.
1058 * See Documentation/DMA-mapping.txt
1061 sba_free_consistent(struct device
*hwdev
, size_t size
, void *vaddr
,
1062 dma_addr_t dma_handle
)
1064 sba_unmap_single(hwdev
, dma_handle
, size
, 0);
1065 free_pages((unsigned long) vaddr
, get_order(size
));
1070 ** Since 0 is a valid pdir_base index value, can't use that
1071 ** to determine if a value is valid or not. Use a flag to indicate
1072 ** the SG list entry contains a valid pdir index.
1074 #define PIDE_FLAG 0x80000000UL
1076 #ifdef SBA_COLLECT_STATS
1077 #define IOMMU_MAP_STATS
1079 #include "iommu-helpers.h"
1081 #ifdef DEBUG_LARGE_SG_ENTRIES
1082 int dump_run_sg
= 0;
1087 * sba_map_sg - map Scatter/Gather list
1088 * @dev: instance of PCI owned by the driver that's asking.
1089 * @sglist: array of buffer/length pairs
1090 * @nents: number of entries in list
1091 * @direction: R/W or both.
1093 * See Documentation/DMA-mapping.txt
1096 sba_map_sg(struct device
*dev
, struct scatterlist
*sglist
, int nents
,
1097 enum dma_data_direction direction
)
1100 int coalesced
, filled
= 0;
1101 unsigned long flags
;
1103 DBG_RUN_SG("%s() START %d entries\n", __FUNCTION__
, nents
);
1107 /* Fast path single entry scatterlists. */
1109 sg_dma_address(sglist
) = sba_map_single(dev
,
1110 (void *)sg_virt_addr(sglist
),
1111 sglist
->length
, direction
);
1112 sg_dma_len(sglist
) = sglist
->length
;
1116 spin_lock_irqsave(&ioc
->res_lock
, flags
);
1118 #ifdef ASSERT_PDIR_SANITY
1119 if (sba_check_pdir(ioc
,"Check before sba_map_sg()"))
1121 sba_dump_sg(ioc
, sglist
, nents
);
1122 panic("Check before sba_map_sg()");
1126 #ifdef SBA_COLLECT_STATS
1131 ** First coalesce the chunks and allocate I/O pdir space
1133 ** If this is one DMA stream, we can properly map using the
1134 ** correct virtual address associated with each DMA page.
1135 ** w/o this association, we wouldn't have coherent DMA!
1136 ** Access to the virtual address is what forces a two pass algorithm.
1138 coalesced
= iommu_coalesce_chunks(ioc
, sglist
, nents
, sba_alloc_range
);
1141 ** Program the I/O Pdir
1143 ** map the virtual addresses to the I/O Pdir
1144 ** o dma_address will contain the pdir index
1145 ** o dma_len will contain the number of bytes to map
1146 ** o address contains the virtual address.
1148 filled
= iommu_fill_pdir(ioc
, sglist
, nents
, 0, sba_io_pdir_entry
);
1150 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
1152 asm volatile("sync" : : );
1154 #ifdef ASSERT_PDIR_SANITY
1155 if (sba_check_pdir(ioc
,"Check after sba_map_sg()"))
1157 sba_dump_sg(ioc
, sglist
, nents
);
1158 panic("Check after sba_map_sg()\n");
1162 spin_unlock_irqrestore(&ioc
->res_lock
, flags
);
1164 DBG_RUN_SG("%s() DONE %d mappings\n", __FUNCTION__
, filled
);
1171 * sba_unmap_sg - unmap Scatter/Gather list
1172 * @dev: instance of PCI owned by the driver that's asking.
1173 * @sglist: array of buffer/length pairs
1174 * @nents: number of entries in list
1175 * @direction: R/W or both.
1177 * See Documentation/DMA-mapping.txt
1180 sba_unmap_sg(struct device
*dev
, struct scatterlist
*sglist
, int nents
,
1181 enum dma_data_direction direction
)
1184 #ifdef ASSERT_PDIR_SANITY
1185 unsigned long flags
;
1188 DBG_RUN_SG("%s() START %d entries, %p,%x\n",
1189 __FUNCTION__
, nents
, sg_virt_addr(sglist
), sglist
->length
);
1193 #ifdef SBA_COLLECT_STATS
1197 #ifdef ASSERT_PDIR_SANITY
1198 spin_lock_irqsave(&ioc
->res_lock
, flags
);
1199 sba_check_pdir(ioc
,"Check before sba_unmap_sg()");
1200 spin_unlock_irqrestore(&ioc
->res_lock
, flags
);
1203 while (sg_dma_len(sglist
) && nents
--) {
1205 sba_unmap_single(dev
, sg_dma_address(sglist
), sg_dma_len(sglist
), direction
);
1206 #ifdef SBA_COLLECT_STATS
1207 ioc
->usg_pages
+= ((sg_dma_address(sglist
) & ~IOVP_MASK
) + sg_dma_len(sglist
) + IOVP_SIZE
- 1) >> PAGE_SHIFT
;
1208 ioc
->usingle_calls
--; /* kluge since call is unmap_sg() */
1213 DBG_RUN_SG("%s() DONE (nents %d)\n", __FUNCTION__
, nents
);
1215 #ifdef ASSERT_PDIR_SANITY
1216 spin_lock_irqsave(&ioc
->res_lock
, flags
);
1217 sba_check_pdir(ioc
,"Check after sba_unmap_sg()");
1218 spin_unlock_irqrestore(&ioc
->res_lock
, flags
);
1223 static struct hppa_dma_ops sba_ops
= {
1224 .dma_supported
= sba_dma_supported
,
1225 .alloc_consistent
= sba_alloc_consistent
,
1226 .alloc_noncoherent
= sba_alloc_consistent
,
1227 .free_consistent
= sba_free_consistent
,
1228 .map_single
= sba_map_single
,
1229 .unmap_single
= sba_unmap_single
,
1230 .map_sg
= sba_map_sg
,
1231 .unmap_sg
= sba_unmap_sg
,
1232 .dma_sync_single_for_cpu
= NULL
,
1233 .dma_sync_single_for_device
= NULL
,
1234 .dma_sync_sg_for_cpu
= NULL
,
1235 .dma_sync_sg_for_device
= NULL
,
1239 /**************************************************************************
1241 ** SBA PAT PDC support
1243 ** o call pdc_pat_cell_module()
1244 ** o store ranges in PCI "resource" structures
1246 **************************************************************************/
1249 sba_get_pat_resources(struct sba_device
*sba_dev
)
1253 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1254 ** PAT PDC to program the SBA/LBA directed range registers...this
1255 ** burden may fall on the LBA code since it directly supports the
1256 ** PCI subsystem. It's not clear yet. - ggg
1258 PAT_MOD(mod
)->mod_info
.mod_pages
= PAT_GET_MOD_PAGES(temp
);
1260 PAT_MOD(mod
)->mod_info
.dvi
= PAT_GET_DVI(temp
);
1261 Tells where the dvi bits are located in the address
.
1262 PAT_MOD(mod
)->mod_info
.ioc
= PAT_GET_IOC(temp
);
1268 /**************************************************************
1270 * Initialization and claim
1272 ***************************************************************/
1273 #define PIRANHA_ADDR_MASK 0x00160000UL /* bit 17,18,20 */
1274 #define PIRANHA_ADDR_VAL 0x00060000UL /* bit 17,18 on */
1276 sba_alloc_pdir(unsigned int pdir_size
)
1278 unsigned long pdir_base
;
1279 unsigned long pdir_order
= get_order(pdir_size
);
1281 pdir_base
= __get_free_pages(GFP_KERNEL
, pdir_order
);
1282 if (NULL
== (void *) pdir_base
) {
1283 panic("%s() could not allocate I/O Page Table\n",
1287 /* If this is not PA8700 (PCX-W2)
1288 ** OR newer than ver 2.2
1289 ** OR in a system that doesn't need VINDEX bits from SBA,
1291 ** then we aren't exposed to the HW bug.
1293 if ( ((boot_cpu_data
.pdc
.cpuid
>> 5) & 0x7f) != 0x13
1294 || (boot_cpu_data
.pdc
.versions
> 0x202)
1295 || (boot_cpu_data
.pdc
.capabilities
& 0x08L
) )
1296 return (void *) pdir_base
;
1299 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1301 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1302 * Ike/Astro can cause silent data corruption. This is only
1303 * a problem if the I/O PDIR is located in memory such that
1304 * (little-endian) bits 17 and 18 are on and bit 20 is off.
1306 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1307 * right physical address, we can either avoid (IOPDIR <= 1MB)
1308 * or minimize (2MB IO Pdir) the problem if we restrict the
1309 * IO Pdir to a maximum size of 2MB-128K (1902K).
1311 * Because we always allocate 2^N sized IO pdirs, either of the
1312 * "bad" regions will be the last 128K if at all. That's easy
1316 if (pdir_order
<= (19-12)) {
1317 if (((virt_to_phys(pdir_base
)+pdir_size
-1) & PIRANHA_ADDR_MASK
) == PIRANHA_ADDR_VAL
) {
1318 /* allocate a new one on 512k alignment */
1319 unsigned long new_pdir
= __get_free_pages(GFP_KERNEL
, (19-12));
1320 /* release original */
1321 free_pages(pdir_base
, pdir_order
);
1323 pdir_base
= new_pdir
;
1325 /* release excess */
1326 while (pdir_order
< (19-12)) {
1327 new_pdir
+= pdir_size
;
1328 free_pages(new_pdir
, pdir_order
);
1336 ** Needs to be aligned on an "odd" 1MB boundary.
1338 unsigned long new_pdir
= __get_free_pages(GFP_KERNEL
, pdir_order
+1); /* 2 or 4MB */
1340 /* release original */
1341 free_pages( pdir_base
, pdir_order
);
1343 /* release first 1MB */
1344 free_pages(new_pdir
, 20-12);
1346 pdir_base
= new_pdir
+ 1024*1024;
1348 if (pdir_order
> (20-12)) {
1352 ** Flag tells init_bitmap() to mark bad 128k as used
1353 ** and to reduce the size by 128k.
1355 piranha_bad_128k
= 1;
1357 new_pdir
+= 3*1024*1024;
1358 /* release last 1MB */
1359 free_pages(new_pdir
, 20-12);
1361 /* release unusable 128KB */
1362 free_pages(new_pdir
- 128*1024 , 17-12);
1364 pdir_size
-= 128*1024;
1368 memset((void *) pdir_base
, 0, pdir_size
);
1369 return (void *) pdir_base
;
1372 static struct device
*next_device(struct klist_iter
*i
)
1374 struct klist_node
* n
= klist_next(i
);
1375 return n
? container_of(n
, struct device
, knode_parent
) : NULL
;
1378 /* setup Mercury or Elroy IBASE/IMASK registers. */
1380 setup_ibase_imask(struct parisc_device
*sba
, struct ioc
*ioc
, int ioc_num
)
1382 /* lba_set_iregs() is in drivers/parisc/lba_pci.c */
1383 extern void lba_set_iregs(struct parisc_device
*, u32
, u32
);
1385 struct klist_iter i
;
1387 klist_iter_init(&sba
->dev
.klist_children
, &i
);
1388 while ((dev
= next_device(&i
))) {
1389 struct parisc_device
*lba
= to_parisc_device(dev
);
1390 int rope_num
= (lba
->hpa
.start
>> 13) & 0xf;
1391 if (rope_num
>> 3 == ioc_num
)
1392 lba_set_iregs(lba
, ioc
->ibase
, ioc
->imask
);
1394 klist_iter_exit(&i
);
1398 sba_ioc_init_pluto(struct parisc_device
*sba
, struct ioc
*ioc
, int ioc_num
)
1400 u32 iova_space_mask
;
1401 u32 iova_space_size
;
1402 int iov_order
, tcnfg
;
1403 #ifdef SBA_AGP_SUPPORT
1407 ** Firmware programs the base and size of a "safe IOVA space"
1408 ** (one that doesn't overlap memory or LMMIO space) in the
1409 ** IBASE and IMASK registers.
1411 ioc
->ibase
= READ_REG(ioc
->ioc_hpa
+ IOC_IBASE
);
1412 iova_space_size
= ~(READ_REG(ioc
->ioc_hpa
+ IOC_IMASK
) & 0xFFFFFFFFUL
) + 1;
1414 if ((ioc
->ibase
< 0xfed00000UL
) && ((ioc
->ibase
+ iova_space_size
) > 0xfee00000UL
)) {
1415 printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1416 iova_space_size
/= 2;
1420 ** iov_order is always based on a 1GB IOVA space since we want to
1421 ** turn on the other half for AGP GART.
1423 iov_order
= get_order(iova_space_size
>> (IOVP_SHIFT
- PAGE_SHIFT
));
1424 ioc
->pdir_size
= (iova_space_size
/ IOVP_SIZE
) * sizeof(u64
);
1426 DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n",
1427 __FUNCTION__
, ioc
->ioc_hpa
, iova_space_size
>> 20,
1428 iov_order
+ PAGE_SHIFT
);
1430 ioc
->pdir_base
= (void *) __get_free_pages(GFP_KERNEL
,
1431 get_order(ioc
->pdir_size
));
1432 if (!ioc
->pdir_base
)
1433 panic("Couldn't allocate I/O Page Table\n");
1435 memset(ioc
->pdir_base
, 0, ioc
->pdir_size
);
1437 DBG_INIT("%s() pdir %p size %x\n",
1438 __FUNCTION__
, ioc
->pdir_base
, ioc
->pdir_size
);
1440 #ifdef SBA_HINT_SUPPORT
1441 ioc
->hint_shift_pdir
= iov_order
+ PAGE_SHIFT
;
1442 ioc
->hint_mask_pdir
= ~(0x3 << (iov_order
+ PAGE_SHIFT
));
1444 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1445 ioc
->hint_shift_pdir
, ioc
->hint_mask_pdir
);
1448 WARN_ON((((unsigned long) ioc
->pdir_base
) & PAGE_MASK
) != (unsigned long) ioc
->pdir_base
);
1449 WRITE_REG(virt_to_phys(ioc
->pdir_base
), ioc
->ioc_hpa
+ IOC_PDIR_BASE
);
1451 /* build IMASK for IOC and Elroy */
1452 iova_space_mask
= 0xffffffff;
1453 iova_space_mask
<<= (iov_order
+ PAGE_SHIFT
);
1454 ioc
->imask
= iova_space_mask
;
1456 ioc
->iovp_mask
= ~(iova_space_mask
+ PAGE_SIZE
- 1);
1458 sba_dump_tlb(ioc
->ioc_hpa
);
1460 setup_ibase_imask(sba
, ioc
, ioc_num
);
1462 WRITE_REG(ioc
->imask
, ioc
->ioc_hpa
+ IOC_IMASK
);
1466 ** Setting the upper bits makes checking for bypass addresses
1467 ** a little faster later on.
1469 ioc
->imask
|= 0xFFFFFFFF00000000UL
;
1472 /* Set I/O PDIR Page size to system page size */
1473 switch (PAGE_SHIFT
) {
1474 case 12: tcnfg
= 0; break; /* 4K */
1475 case 13: tcnfg
= 1; break; /* 8K */
1476 case 14: tcnfg
= 2; break; /* 16K */
1477 case 16: tcnfg
= 3; break; /* 64K */
1479 panic(__FILE__
"Unsupported system page size %d",
1483 WRITE_REG(tcnfg
, ioc
->ioc_hpa
+ IOC_TCNFG
);
1486 ** Program the IOC's ibase and enable IOVA translation
1487 ** Bit zero == enable bit.
1489 WRITE_REG(ioc
->ibase
| 1, ioc
->ioc_hpa
+ IOC_IBASE
);
1492 ** Clear I/O TLB of any possible entries.
1493 ** (Yes. This is a bit paranoid...but so what)
1495 WRITE_REG(ioc
->ibase
| 31, ioc
->ioc_hpa
+ IOC_PCOM
);
1497 #ifdef SBA_AGP_SUPPORT
1499 ** If an AGP device is present, only use half of the IOV space
1500 ** for PCI DMA. Unfortunately we can't know ahead of time
1501 ** whether GART support will actually be used, for now we
1502 ** can just key on any AGP device found in the system.
1503 ** We program the next pdir index after we stop w/ a key for
1504 ** the GART code to handshake on.
1507 for (lba
= sba
->child
; lba
; lba
= lba
->sibling
) {
1508 if (IS_QUICKSILVER(lba
))
1513 DBG_INIT("%s: Reserving half of IOVA space for AGP GART support\n", __FUNCTION__
);
1514 ioc
->pdir_size
/= 2;
1515 ((u64
*)ioc
->pdir_base
)[PDIR_INDEX(iova_space_size
/2)] = SBA_IOMMU_COOKIE
;
1517 DBG_INIT("%s: No GART needed - no AGP controller found\n", __FUNCTION__
);
1524 sba_ioc_init(struct parisc_device
*sba
, struct ioc
*ioc
, int ioc_num
)
1526 u32 iova_space_size
, iova_space_mask
;
1527 unsigned int pdir_size
, iov_order
;
1530 ** Determine IOVA Space size from memory size.
1532 ** Ideally, PCI drivers would register the maximum number
1533 ** of DMA they can have outstanding for each device they
1534 ** own. Next best thing would be to guess how much DMA
1535 ** can be outstanding based on PCI Class/sub-class. Both
1536 ** methods still require some "extra" to support PCI
1537 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1539 ** While we have 32-bits "IOVA" space, top two 2 bits are used
1540 ** for DMA hints - ergo only 30 bits max.
1543 iova_space_size
= (u32
) (num_physpages
/global_ioc_cnt
);
1545 /* limit IOVA space size to 1MB-1GB */
1546 if (iova_space_size
< (1 << (20 - PAGE_SHIFT
))) {
1547 iova_space_size
= 1 << (20 - PAGE_SHIFT
);
1549 else if (iova_space_size
> (1 << (30 - PAGE_SHIFT
))) {
1550 iova_space_size
= 1 << (30 - PAGE_SHIFT
);
1554 ** iova space must be log2() in size.
1555 ** thus, pdir/res_map will also be log2().
1556 ** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1558 iov_order
= get_order(iova_space_size
<< PAGE_SHIFT
);
1560 /* iova_space_size is now bytes, not pages */
1561 iova_space_size
= 1 << (iov_order
+ PAGE_SHIFT
);
1563 ioc
->pdir_size
= pdir_size
= (iova_space_size
/IOVP_SIZE
) * sizeof(u64
);
1565 DBG_INIT("%s() hpa 0x%lx mem %ldMB IOV %dMB (%d bits)\n",
1568 (unsigned long) num_physpages
>> (20 - PAGE_SHIFT
),
1569 iova_space_size
>>20,
1570 iov_order
+ PAGE_SHIFT
);
1572 ioc
->pdir_base
= sba_alloc_pdir(pdir_size
);
1574 DBG_INIT("%s() pdir %p size %x\n",
1575 __FUNCTION__
, ioc
->pdir_base
, pdir_size
);
1577 #ifdef SBA_HINT_SUPPORT
1578 /* FIXME : DMA HINTs not used */
1579 ioc
->hint_shift_pdir
= iov_order
+ PAGE_SHIFT
;
1580 ioc
->hint_mask_pdir
= ~(0x3 << (iov_order
+ PAGE_SHIFT
));
1582 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1583 ioc
->hint_shift_pdir
, ioc
->hint_mask_pdir
);
1586 WRITE_REG64(virt_to_phys(ioc
->pdir_base
), ioc
->ioc_hpa
+ IOC_PDIR_BASE
);
1588 /* build IMASK for IOC and Elroy */
1589 iova_space_mask
= 0xffffffff;
1590 iova_space_mask
<<= (iov_order
+ PAGE_SHIFT
);
1593 ** On C3000 w/512MB mem, HP-UX 10.20 reports:
1594 ** ibase=0, imask=0xFE000000, size=0x2000000.
1597 ioc
->imask
= iova_space_mask
; /* save it */
1599 ioc
->iovp_mask
= ~(iova_space_mask
+ PAGE_SIZE
- 1);
1602 DBG_INIT("%s() IOV base 0x%lx mask 0x%0lx\n",
1603 __FUNCTION__
, ioc
->ibase
, ioc
->imask
);
1606 ** FIXME: Hint registers are programmed with default hint
1607 ** values during boot, so hints should be sane even if we
1608 ** can't reprogram them the way drivers want.
1611 setup_ibase_imask(sba
, ioc
, ioc_num
);
1614 ** Program the IOC's ibase and enable IOVA translation
1616 WRITE_REG(ioc
->ibase
| 1, ioc
->ioc_hpa
+IOC_IBASE
);
1617 WRITE_REG(ioc
->imask
, ioc
->ioc_hpa
+IOC_IMASK
);
1619 /* Set I/O PDIR Page size to 4K */
1620 WRITE_REG(0, ioc
->ioc_hpa
+IOC_TCNFG
);
1623 ** Clear I/O TLB of any possible entries.
1624 ** (Yes. This is a bit paranoid...but so what)
1626 WRITE_REG(0 | 31, ioc
->ioc_hpa
+IOC_PCOM
);
1628 ioc
->ibase
= 0; /* used by SBA_IOVA and related macros */
1630 DBG_INIT("%s() DONE\n", __FUNCTION__
);
1635 /**************************************************************************
1637 ** SBA initialization code (HW and SW)
1639 ** o identify SBA chip itself
1640 ** o initialize SBA chip modes (HardFail)
1641 ** o initialize SBA chip modes (HardFail)
1642 ** o FIXME: initialize DMA hints for reasonable defaults
1644 **************************************************************************/
1646 static void __iomem
*ioc_remap(struct sba_device
*sba_dev
, unsigned int offset
)
1648 return ioremap_nocache(sba_dev
->dev
->hpa
.start
+ offset
, SBA_FUNC_SIZE
);
1651 static void sba_hw_init(struct sba_device
*sba_dev
)
1657 if (!is_pdc_pat()) {
1658 /* Shutdown the USB controller on Astro-based workstations.
1659 ** Once we reprogram the IOMMU, the next DMA performed by
1660 ** USB will HPMC the box. USB is only enabled if a
1661 ** keyboard is present and found.
1663 ** With serial console, j6k v5.0 firmware says:
1664 ** mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1666 ** FIXME: Using GFX+USB console at power up but direct
1667 ** linux to serial console is still broken.
1668 ** USB could generate DMA so we must reset USB.
1669 ** The proper sequence would be:
1670 ** o block console output
1671 ** o reset USB device
1672 ** o reprogram serial port
1673 ** o unblock console output
1675 if (PAGE0
->mem_kbd
.cl_class
== CL_KEYBD
) {
1676 pdc_io_reset_devices();
1683 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0
->mem_boot
.hpa
,
1684 PAGE0
->mem_boot
.spa
, PAGE0
->mem_boot
.pad
, PAGE0
->mem_boot
.cl_class
);
1687 ** Need to deal with DMA from LAN.
1688 ** Maybe use page zero boot device as a handle to talk
1689 ** to PDC about which device to shutdown.
1691 ** Netbooting, j6k v5.0 firmware says:
1692 ** mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1693 ** ARGH! invalid class.
1695 if ((PAGE0
->mem_boot
.cl_class
!= CL_RANDOM
)
1696 && (PAGE0
->mem_boot
.cl_class
!= CL_SEQU
)) {
1701 if (!IS_PLUTO(sba_dev
->iodc
)) {
1702 ioc_ctl
= READ_REG(sba_dev
->sba_hpa
+IOC_CTRL
);
1703 DBG_INIT("%s() hpa 0x%lx ioc_ctl 0x%Lx ->",
1704 __FUNCTION__
, sba_dev
->sba_hpa
, ioc_ctl
);
1705 ioc_ctl
&= ~(IOC_CTRL_RM
| IOC_CTRL_NC
| IOC_CTRL_CE
);
1706 ioc_ctl
|= IOC_CTRL_DD
| IOC_CTRL_D4
| IOC_CTRL_TC
;
1707 /* j6700 v1.6 firmware sets 0x294f */
1708 /* A500 firmware sets 0x4d */
1710 WRITE_REG(ioc_ctl
, sba_dev
->sba_hpa
+IOC_CTRL
);
1712 #ifdef DEBUG_SBA_INIT
1713 ioc_ctl
= READ_REG64(sba_dev
->sba_hpa
+IOC_CTRL
);
1714 DBG_INIT(" 0x%Lx\n", ioc_ctl
);
1718 if (IS_ASTRO(sba_dev
->iodc
)) {
1720 /* PAT_PDC (L-class) also reports the same goofy base */
1721 sba_dev
->ioc
[0].ioc_hpa
= ioc_remap(sba_dev
, ASTRO_IOC_OFFSET
);
1724 sba_dev
->chip_resv
.name
= "Astro Intr Ack";
1725 sba_dev
->chip_resv
.start
= PCI_F_EXTEND
| 0xfef00000UL
;
1726 sba_dev
->chip_resv
.end
= PCI_F_EXTEND
| (0xff000000UL
- 1) ;
1727 err
= request_resource(&iomem_resource
, &(sba_dev
->chip_resv
));
1730 } else if (IS_PLUTO(sba_dev
->iodc
)) {
1733 /* We use a negative value for IOC HPA so it gets
1734 * corrected when we add it with IKE's IOC offset.
1735 * Doesnt look clean, but fewer code.
1737 sba_dev
->ioc
[0].ioc_hpa
= ioc_remap(sba_dev
, PLUTO_IOC_OFFSET
);
1740 sba_dev
->chip_resv
.name
= "Pluto Intr/PIOP/VGA";
1741 sba_dev
->chip_resv
.start
= PCI_F_EXTEND
| 0xfee00000UL
;
1742 sba_dev
->chip_resv
.end
= PCI_F_EXTEND
| (0xff200000UL
- 1);
1743 err
= request_resource(&iomem_resource
, &(sba_dev
->chip_resv
));
1746 sba_dev
->iommu_resv
.name
= "IOVA Space";
1747 sba_dev
->iommu_resv
.start
= 0x40000000UL
;
1748 sba_dev
->iommu_resv
.end
= 0x50000000UL
- 1;
1749 err
= request_resource(&iomem_resource
, &(sba_dev
->iommu_resv
));
1752 /* IS_IKE (ie N-class, L3000, L1500) */
1753 sba_dev
->ioc
[0].ioc_hpa
= ioc_remap(sba_dev
, IKE_IOC_OFFSET(0));
1754 sba_dev
->ioc
[1].ioc_hpa
= ioc_remap(sba_dev
, IKE_IOC_OFFSET(1));
1757 /* TODO - LOOKUP Ike/Stretch chipset mem map */
1759 /* XXX: What about Reo? */
1761 sba_dev
->num_ioc
= num_ioc
;
1762 for (i
= 0; i
< num_ioc
; i
++) {
1763 void __iomem
*ioc_hpa
= sba_dev
->ioc
[i
].ioc_hpa
;
1766 for (j
=0; j
< sizeof(u64
) * ROPES_PER_IOC
; j
+=sizeof(u64
)) {
1769 * Clear ROPE(N)_CONFIG AO bit.
1770 * Disables "NT Ordering" (~= !"Relaxed Ordering")
1771 * Overrides bit 1 in DMA Hint Sets.
1772 * Improves netperf UDP_STREAM by ~10% for bcm5701.
1774 if (IS_PLUTO(sba_dev
->iodc
)) {
1775 void __iomem
*rope_cfg
;
1776 unsigned long cfg_val
;
1778 rope_cfg
= ioc_hpa
+ IOC_ROPE0_CFG
+ j
;
1779 cfg_val
= READ_REG(rope_cfg
);
1780 cfg_val
&= ~IOC_ROPE_AO
;
1781 WRITE_REG(cfg_val
, rope_cfg
);
1785 ** Make sure the box crashes on rope errors.
1787 WRITE_REG(HF_ENABLE
, ioc_hpa
+ ROPE0_CTL
+ j
);
1790 /* flush out the last writes */
1791 READ_REG(sba_dev
->ioc
[i
].ioc_hpa
+ ROPE7_CTL
);
1793 DBG_INIT(" ioc[%d] ROPE_CFG 0x%Lx ROPE_DBG 0x%Lx\n",
1795 READ_REG(sba_dev
->ioc
[i
].ioc_hpa
+ 0x40),
1796 READ_REG(sba_dev
->ioc
[i
].ioc_hpa
+ 0x50)
1798 DBG_INIT(" STATUS_CONTROL 0x%Lx FLUSH_CTRL 0x%Lx\n",
1799 READ_REG(sba_dev
->ioc
[i
].ioc_hpa
+ 0x108),
1800 READ_REG(sba_dev
->ioc
[i
].ioc_hpa
+ 0x400)
1803 if (IS_PLUTO(sba_dev
->iodc
)) {
1804 sba_ioc_init_pluto(sba_dev
->dev
, &(sba_dev
->ioc
[i
]), i
);
1806 sba_ioc_init(sba_dev
->dev
, &(sba_dev
->ioc
[i
]), i
);
1812 sba_common_init(struct sba_device
*sba_dev
)
1816 /* add this one to the head of the list (order doesn't matter)
1817 ** This will be useful for debugging - especially if we get coredumps
1819 sba_dev
->next
= sba_list
;
1822 for(i
=0; i
< sba_dev
->num_ioc
; i
++) {
1824 #ifdef DEBUG_DMB_TRAP
1825 extern void iterate_pages(unsigned long , unsigned long ,
1826 void (*)(pte_t
* , unsigned long),
1828 void set_data_memory_break(pte_t
* , unsigned long);
1830 /* resource map size dictated by pdir_size */
1831 res_size
= sba_dev
->ioc
[i
].pdir_size
/sizeof(u64
); /* entries */
1833 /* Second part of PIRANHA BUG */
1834 if (piranha_bad_128k
) {
1835 res_size
-= (128*1024)/sizeof(u64
);
1838 res_size
>>= 3; /* convert bit count to byte count */
1839 DBG_INIT("%s() res_size 0x%x\n",
1840 __FUNCTION__
, res_size
);
1842 sba_dev
->ioc
[i
].res_size
= res_size
;
1843 sba_dev
->ioc
[i
].res_map
= (char *) __get_free_pages(GFP_KERNEL
, get_order(res_size
));
1845 #ifdef DEBUG_DMB_TRAP
1846 iterate_pages( sba_dev
->ioc
[i
].res_map
, res_size
,
1847 set_data_memory_break
, 0);
1850 if (NULL
== sba_dev
->ioc
[i
].res_map
)
1852 panic("%s:%s() could not allocate resource map\n",
1853 __FILE__
, __FUNCTION__
);
1856 memset(sba_dev
->ioc
[i
].res_map
, 0, res_size
);
1857 /* next available IOVP - circular search */
1858 sba_dev
->ioc
[i
].res_hint
= (unsigned long *)
1859 &(sba_dev
->ioc
[i
].res_map
[L1_CACHE_BYTES
]);
1861 #ifdef ASSERT_PDIR_SANITY
1862 /* Mark first bit busy - ie no IOVA 0 */
1863 sba_dev
->ioc
[i
].res_map
[0] = 0x80;
1864 sba_dev
->ioc
[i
].pdir_base
[0] = 0xeeffc0addbba0080ULL
;
1867 /* Third (and last) part of PIRANHA BUG */
1868 if (piranha_bad_128k
) {
1869 /* region from +1408K to +1536 is un-usable. */
1871 int idx_start
= (1408*1024/sizeof(u64
)) >> 3;
1872 int idx_end
= (1536*1024/sizeof(u64
)) >> 3;
1873 long *p_start
= (long *) &(sba_dev
->ioc
[i
].res_map
[idx_start
]);
1874 long *p_end
= (long *) &(sba_dev
->ioc
[i
].res_map
[idx_end
]);
1876 /* mark that part of the io pdir busy */
1877 while (p_start
< p_end
)
1882 #ifdef DEBUG_DMB_TRAP
1883 iterate_pages( sba_dev
->ioc
[i
].res_map
, res_size
,
1884 set_data_memory_break
, 0);
1885 iterate_pages( sba_dev
->ioc
[i
].pdir_base
, sba_dev
->ioc
[i
].pdir_size
,
1886 set_data_memory_break
, 0);
1889 DBG_INIT("%s() %d res_map %x %p\n",
1890 __FUNCTION__
, i
, res_size
, sba_dev
->ioc
[i
].res_map
);
1893 spin_lock_init(&sba_dev
->sba_lock
);
1894 ioc_needs_fdc
= boot_cpu_data
.pdc
.capabilities
& PDC_MODEL_IOPDIR_FDC
;
1896 #ifdef DEBUG_SBA_INIT
1898 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1899 * (bit #61, big endian), we have to flush and sync every time
1900 * IO-PDIR is changed in Ike/Astro.
1902 if (ioc_needs_fdc
) {
1903 printk(KERN_INFO MODULE_NAME
" FDC/SYNC required.\n");
1905 printk(KERN_INFO MODULE_NAME
" IOC has cache coherent PDIR.\n");
1910 #ifdef CONFIG_PROC_FS
1911 static int sba_proc_info(struct seq_file
*m
, void *p
)
1913 struct sba_device
*sba_dev
= sba_list
;
1914 struct ioc
*ioc
= &sba_dev
->ioc
[0]; /* FIXME: Multi-IOC support! */
1915 int total_pages
= (int) (ioc
->res_size
<< 3); /* 8 bits per byte */
1916 #ifdef SBA_COLLECT_STATS
1917 unsigned long avg
= 0, min
, max
;
1921 len
+= seq_printf(m
, "%s rev %d.%d\n",
1923 (sba_dev
->hw_rev
& 0x7) + 1,
1924 (sba_dev
->hw_rev
& 0x18) >> 3
1926 len
+= seq_printf(m
, "IO PDIR size : %d bytes (%d entries)\n",
1927 (int) ((ioc
->res_size
<< 3) * sizeof(u64
)), /* 8 bits/byte */
1930 len
+= seq_printf(m
, "Resource bitmap : %d bytes (%d pages)\n",
1931 ioc
->res_size
, ioc
->res_size
<< 3); /* 8 bits per byte */
1933 len
+= seq_printf(m
, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1934 READ_REG32(sba_dev
->sba_hpa
+ LMMIO_DIST_BASE
),
1935 READ_REG32(sba_dev
->sba_hpa
+ LMMIO_DIST_MASK
),
1936 READ_REG32(sba_dev
->sba_hpa
+ LMMIO_DIST_ROUTE
)
1940 len
+= seq_printf(m
, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n", i
,
1941 READ_REG32(sba_dev
->sba_hpa
+ LMMIO_DIRECT0_BASE
+ i
*0x18),
1942 READ_REG32(sba_dev
->sba_hpa
+ LMMIO_DIRECT0_MASK
+ i
*0x18),
1943 READ_REG32(sba_dev
->sba_hpa
+ LMMIO_DIRECT0_ROUTE
+ i
*0x18)
1946 #ifdef SBA_COLLECT_STATS
1947 len
+= seq_printf(m
, "IO PDIR entries : %ld free %ld used (%d%%)\n",
1948 total_pages
- ioc
->used_pages
, ioc
->used_pages
,
1949 (int) (ioc
->used_pages
* 100 / total_pages
));
1951 min
= max
= ioc
->avg_search
[0];
1952 for (i
= 0; i
< SBA_SEARCH_SAMPLE
; i
++) {
1953 avg
+= ioc
->avg_search
[i
];
1954 if (ioc
->avg_search
[i
] > max
) max
= ioc
->avg_search
[i
];
1955 if (ioc
->avg_search
[i
] < min
) min
= ioc
->avg_search
[i
];
1957 avg
/= SBA_SEARCH_SAMPLE
;
1958 len
+= seq_printf(m
, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1961 len
+= seq_printf(m
, "pci_map_single(): %12ld calls %12ld pages (avg %d/1000)\n",
1962 ioc
->msingle_calls
, ioc
->msingle_pages
,
1963 (int) ((ioc
->msingle_pages
* 1000)/ioc
->msingle_calls
));
1965 /* KLUGE - unmap_sg calls unmap_single for each mapped page */
1966 min
= ioc
->usingle_calls
;
1967 max
= ioc
->usingle_pages
- ioc
->usg_pages
;
1968 len
+= seq_printf(m
, "pci_unmap_single: %12ld calls %12ld pages (avg %d/1000)\n",
1969 min
, max
, (int) ((max
* 1000)/min
));
1971 len
+= seq_printf(m
, "pci_map_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1972 ioc
->msg_calls
, ioc
->msg_pages
,
1973 (int) ((ioc
->msg_pages
* 1000)/ioc
->msg_calls
));
1975 len
+= seq_printf(m
, "pci_unmap_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1976 ioc
->usg_calls
, ioc
->usg_pages
,
1977 (int) ((ioc
->usg_pages
* 1000)/ioc
->usg_calls
));
1984 sba_proc_open(struct inode
*i
, struct file
*f
)
1986 return single_open(f
, &sba_proc_info
, NULL
);
1989 static struct file_operations sba_proc_fops
= {
1990 .owner
= THIS_MODULE
,
1991 .open
= sba_proc_open
,
1993 .llseek
= seq_lseek
,
1994 .release
= single_release
,
1998 sba_proc_bitmap_info(struct seq_file
*m
, void *p
)
2000 struct sba_device
*sba_dev
= sba_list
;
2001 struct ioc
*ioc
= &sba_dev
->ioc
[0]; /* FIXME: Multi-IOC support! */
2002 unsigned int *res_ptr
= (unsigned int *)ioc
->res_map
;
2005 for (i
= 0; i
< (ioc
->res_size
/sizeof(unsigned int)); ++i
, ++res_ptr
) {
2007 len
+= seq_printf(m
, "\n ");
2008 len
+= seq_printf(m
, " %08x", *res_ptr
);
2010 len
+= seq_printf(m
, "\n");
2016 sba_proc_bitmap_open(struct inode
*i
, struct file
*f
)
2018 return single_open(f
, &sba_proc_bitmap_info
, NULL
);
2021 static struct file_operations sba_proc_bitmap_fops
= {
2022 .owner
= THIS_MODULE
,
2023 .open
= sba_proc_bitmap_open
,
2025 .llseek
= seq_lseek
,
2026 .release
= single_release
,
2028 #endif /* CONFIG_PROC_FS */
2030 static struct parisc_device_id sba_tbl
[] = {
2031 { HPHW_IOA
, HVERSION_REV_ANY_ID
, ASTRO_RUNWAY_PORT
, 0xb },
2032 { HPHW_BCPORT
, HVERSION_REV_ANY_ID
, IKE_MERCED_PORT
, 0xc },
2033 { HPHW_BCPORT
, HVERSION_REV_ANY_ID
, REO_MERCED_PORT
, 0xc },
2034 { HPHW_BCPORT
, HVERSION_REV_ANY_ID
, REOG_MERCED_PORT
, 0xc },
2035 { HPHW_IOA
, HVERSION_REV_ANY_ID
, PLUTO_MCKINLEY_PORT
, 0xc },
2039 int sba_driver_callback(struct parisc_device
*);
2041 static struct parisc_driver sba_driver
= {
2042 .name
= MODULE_NAME
,
2043 .id_table
= sba_tbl
,
2044 .probe
= sba_driver_callback
,
2048 ** Determine if sba should claim this chip (return 0) or not (return 1).
2049 ** If so, initialize the chip and tell other partners in crime they
2053 sba_driver_callback(struct parisc_device
*dev
)
2055 struct sba_device
*sba_dev
;
2059 void __iomem
*sba_addr
= ioremap_nocache(dev
->hpa
.start
, SBA_FUNC_SIZE
);
2060 struct proc_dir_entry
*info_entry
, *bitmap_entry
, *root
;
2062 sba_dump_ranges(sba_addr
);
2064 /* Read HW Rev First */
2065 func_class
= READ_REG(sba_addr
+ SBA_FCLASS
);
2067 if (IS_ASTRO(&dev
->id
)) {
2068 unsigned long fclass
;
2069 static char astro_rev
[]="Astro ?.?";
2071 /* Astro is broken...Read HW Rev First */
2072 fclass
= READ_REG(sba_addr
);
2074 astro_rev
[6] = '1' + (char) (fclass
& 0x7);
2075 astro_rev
[8] = '0' + (char) ((fclass
& 0x18) >> 3);
2076 version
= astro_rev
;
2078 } else if (IS_IKE(&dev
->id
)) {
2079 static char ike_rev
[] = "Ike rev ?";
2080 ike_rev
[8] = '0' + (char) (func_class
& 0xff);
2082 } else if (IS_PLUTO(&dev
->id
)) {
2083 static char pluto_rev
[]="Pluto ?.?";
2084 pluto_rev
[6] = '0' + (char) ((func_class
& 0xf0) >> 4);
2085 pluto_rev
[8] = '0' + (char) (func_class
& 0x0f);
2086 version
= pluto_rev
;
2088 static char reo_rev
[] = "REO rev ?";
2089 reo_rev
[8] = '0' + (char) (func_class
& 0xff);
2093 if (!global_ioc_cnt
) {
2094 global_ioc_cnt
= count_parisc_driver(&sba_driver
);
2096 /* Astro and Pluto have one IOC per SBA */
2097 if ((!IS_ASTRO(&dev
->id
)) || (!IS_PLUTO(&dev
->id
)))
2098 global_ioc_cnt
*= 2;
2101 printk(KERN_INFO
"%s found %s at 0x%lx\n",
2102 MODULE_NAME
, version
, dev
->hpa
.start
);
2104 sba_dev
= kzalloc(sizeof(struct sba_device
), GFP_KERNEL
);
2106 printk(KERN_ERR MODULE_NAME
" - couldn't alloc sba_device\n");
2110 parisc_set_drvdata(dev
, sba_dev
);
2112 for(i
=0; i
<MAX_IOC
; i
++)
2113 spin_lock_init(&(sba_dev
->ioc
[i
].res_lock
));
2116 sba_dev
->hw_rev
= func_class
;
2117 sba_dev
->iodc
= &dev
->id
;
2118 sba_dev
->name
= dev
->name
;
2119 sba_dev
->sba_hpa
= sba_addr
;
2121 sba_get_pat_resources(sba_dev
);
2122 sba_hw_init(sba_dev
);
2123 sba_common_init(sba_dev
);
2125 hppa_dma_ops
= &sba_ops
;
2127 #ifdef CONFIG_PROC_FS
2128 switch (dev
->id
.hversion
) {
2129 case PLUTO_MCKINLEY_PORT
:
2130 root
= proc_mckinley_root
;
2132 case ASTRO_RUNWAY_PORT
:
2133 case IKE_MERCED_PORT
:
2135 root
= proc_runway_root
;
2139 info_entry
= create_proc_entry("sba_iommu", 0, root
);
2140 bitmap_entry
= create_proc_entry("sba_iommu-bitmap", 0, root
);
2143 info_entry
->proc_fops
= &sba_proc_fops
;
2146 bitmap_entry
->proc_fops
= &sba_proc_bitmap_fops
;
2149 parisc_vmerge_boundary
= IOVP_SIZE
;
2150 parisc_vmerge_max_size
= IOVP_SIZE
* BITS_PER_LONG
;
2156 ** One time initialization to let the world know the SBA was found.
2157 ** This is the only routine which is NOT static.
2158 ** Must be called exactly once before pci_init().
2160 void __init
sba_init(void)
2162 register_parisc_driver(&sba_driver
);
2167 * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
2168 * @dev: The parisc device.
2170 * Returns the appropriate IOMMU data for the given parisc PCI controller.
2171 * This is cached and used later for PCI DMA Mapping.
2173 void * sba_get_iommu(struct parisc_device
*pci_hba
)
2175 struct parisc_device
*sba_dev
= parisc_parent(pci_hba
);
2176 struct sba_device
*sba
= sba_dev
->dev
.driver_data
;
2177 char t
= sba_dev
->id
.hw_type
;
2178 int iocnum
= (pci_hba
->hw_path
>> 3); /* rope # */
2180 WARN_ON((t
!= HPHW_IOA
) && (t
!= HPHW_BCPORT
));
2182 return &(sba
->ioc
[iocnum
]);
2187 * sba_directed_lmmio - return first directed LMMIO range routed to rope
2188 * @pa_dev: The parisc device.
2189 * @r: resource PCI host controller wants start/end fields assigned.
2191 * For the given parisc PCI controller, determine if any direct ranges
2192 * are routed down the corresponding rope.
2194 void sba_directed_lmmio(struct parisc_device
*pci_hba
, struct resource
*r
)
2196 struct parisc_device
*sba_dev
= parisc_parent(pci_hba
);
2197 struct sba_device
*sba
= sba_dev
->dev
.driver_data
;
2198 char t
= sba_dev
->id
.hw_type
;
2200 int rope
= (pci_hba
->hw_path
& (ROPES_PER_IOC
-1)); /* rope # */
2202 BUG_ON((t
!=HPHW_IOA
) && (t
!=HPHW_BCPORT
));
2204 r
->start
= r
->end
= 0;
2206 /* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2207 for (i
=0; i
<4; i
++) {
2209 void __iomem
*reg
= sba
->sba_hpa
+ i
*0x18;
2211 base
= READ_REG32(reg
+ LMMIO_DIRECT0_BASE
);
2212 if ((base
& 1) == 0)
2213 continue; /* not enabled */
2215 size
= READ_REG32(reg
+ LMMIO_DIRECT0_ROUTE
);
2217 if ((size
& (ROPES_PER_IOC
-1)) != rope
)
2218 continue; /* directed down different rope */
2220 r
->start
= (base
& ~1UL) | PCI_F_EXTEND
;
2221 size
= ~ READ_REG32(reg
+ LMMIO_DIRECT0_MASK
);
2222 r
->end
= r
->start
+ size
;
2228 * sba_distributed_lmmio - return portion of distributed LMMIO range
2229 * @pa_dev: The parisc device.
2230 * @r: resource PCI host controller wants start/end fields assigned.
2232 * For the given parisc PCI controller, return portion of distributed LMMIO
2233 * range. The distributed LMMIO is always present and it's just a question
2234 * of the base address and size of the range.
2236 void sba_distributed_lmmio(struct parisc_device
*pci_hba
, struct resource
*r
)
2238 struct parisc_device
*sba_dev
= parisc_parent(pci_hba
);
2239 struct sba_device
*sba
= sba_dev
->dev
.driver_data
;
2240 char t
= sba_dev
->id
.hw_type
;
2242 int rope
= (pci_hba
->hw_path
& (ROPES_PER_IOC
-1)); /* rope # */
2244 BUG_ON((t
!=HPHW_IOA
) && (t
!=HPHW_BCPORT
));
2246 r
->start
= r
->end
= 0;
2248 base
= READ_REG32(sba
->sba_hpa
+ LMMIO_DIST_BASE
);
2249 if ((base
& 1) == 0) {
2250 BUG(); /* Gah! Distr Range wasn't enabled! */
2254 r
->start
= (base
& ~1UL) | PCI_F_EXTEND
;
2256 size
= (~READ_REG32(sba
->sba_hpa
+ LMMIO_DIST_MASK
)) / ROPES_PER_IOC
;
2257 r
->start
+= rope
* (size
+ 1); /* adjust base for this rope */
2258 r
->end
= r
->start
+ size
;