1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 ** System Bus Adapter (SBA) I/O MMU manager
4 **
5 ** (c) Copyright 2000-2004 Grant Grundler <grundler @ parisc-linux x org>
6 ** (c) Copyright 2004 Naresh Kumar Inna <knaresh at india x hp x com>
7 ** (c) Copyright 2000-2004 Hewlett-Packard Company
8 **
9 ** Portions (c) 1999 Dave S. Miller (from sparc64 I/O MMU code)
10 **
11 **
12 **
13 ** This module initializes the IOC (I/O Controller) found on B1000/C3000/
14 ** J5000/J7000/N-class/L-class machines and their successors.
15 **
16 ** FIXME: add DMA hint support programming in both sba and lba modules.
17 */
18
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/spinlock.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24
25 #include <linux/mm.h>
26 #include <linux/string.h>
27 #include <linux/pci.h>
28 #include <linux/dma-map-ops.h>
29 #include <linux/scatterlist.h>
30 #include <linux/iommu-helper.h>
31 /*
32 * The semantics of 64 register access on 32bit systems can't be guaranteed
33 * by the C standard, we hope the _lo_hi() macros defining readq and writeq
34 * here will behave as expected.
35 */
36 #include <linux/io-64-nonatomic-lo-hi.h>
37
38 #include <asm/byteorder.h>
39 #include <asm/io.h>
40 #include <asm/dma.h> /* for DMA_CHUNK_SIZE */
41
42 #include <asm/hardware.h> /* for register_parisc_driver() stuff */
43
44 #include <linux/proc_fs.h>
45 #include <linux/seq_file.h>
46 #include <linux/module.h>
47
48 #include <asm/ropes.h>
49 #include <asm/page.h> /* for PAGE0 */
50 #include <asm/pdc.h> /* for PDC_MODEL_* */
51 #include <asm/pdcpat.h> /* for is_pdc_pat() */
52 #include <asm/parisc-device.h>
53
54 #include "iommu.h"
55
56 #define MODULE_NAME "SBA"
57
58 /*
59 ** The number of debug flags is a clue - this code is fragile.
60 ** Don't even think about messing with it unless you have
61 ** plenty of 710's to sacrifice to the computer gods. :^)
62 */
63 #undef DEBUG_SBA_INIT
64 #undef DEBUG_SBA_RUN
65 #undef DEBUG_SBA_RUN_SG
66 #undef DEBUG_SBA_RESOURCE
67 #undef ASSERT_PDIR_SANITY
68 #undef DEBUG_LARGE_SG_ENTRIES
69 #undef DEBUG_DMB_TRAP
70
71 #ifdef DEBUG_SBA_INIT
72 #define DBG_INIT(x...) printk(x)
73 #else
74 #define DBG_INIT(x...)
75 #endif
76
77 #ifdef DEBUG_SBA_RUN
78 #define DBG_RUN(x...) printk(x)
79 #else
80 #define DBG_RUN(x...)
81 #endif
82
83 #ifdef DEBUG_SBA_RUN_SG
84 #define DBG_RUN_SG(x...) printk(x)
85 #else
86 #define DBG_RUN_SG(x...)
87 #endif
88
89
90 #ifdef DEBUG_SBA_RESOURCE
91 #define DBG_RES(x...) printk(x)
92 #else
93 #define DBG_RES(x...)
94 #endif
95
96 #define DEFAULT_DMA_HINT_REG 0
97
98 struct sba_device *sba_list;
99 EXPORT_SYMBOL_GPL(sba_list);
100
101 static unsigned long ioc_needs_fdc = 0;
102
103 /* global count of IOMMUs in the system */
104 static unsigned int global_ioc_cnt = 0;
105
106 /* PA8700 (Piranha 2.2) bug workaround */
107 static unsigned long piranha_bad_128k = 0;
108
109 /* Looks nice and keeps the compiler happy */
110 #define SBA_DEV(d) ((struct sba_device *) (d))
111
112 #ifdef CONFIG_AGP_PARISC
113 #define SBA_AGP_SUPPORT
114 #endif /*CONFIG_AGP_PARISC*/
115
116 #ifdef SBA_AGP_SUPPORT
117 static int sba_reserve_agpgart = 1;
118 module_param(sba_reserve_agpgart, int, 0444);
119 MODULE_PARM_DESC(sba_reserve_agpgart, "Reserve half of IO pdir as AGPGART");
120 #endif
121
122 static struct proc_dir_entry *proc_runway_root __ro_after_init;
123 static struct proc_dir_entry *proc_mckinley_root __ro_after_init;
124
125 /************************************
126 ** SBA register read and write support
127 **
128 ** BE WARNED: register writes are posted.
129 ** (ie follow writes which must reach HW with a read)
130 **
131 ** Superdome (in particular, REO) allows only 64-bit CSR accesses.
132 */
133 #define READ_REG32(addr) readl(addr)
134 #define READ_REG64(addr) readq(addr)
135 #define WRITE_REG32(val, addr) writel((val), (addr))
136 #define WRITE_REG64(val, addr) writeq((val), (addr))
137
138 #ifdef CONFIG_64BIT
139 #define READ_REG(addr) READ_REG64(addr)
140 #define WRITE_REG(value, addr) WRITE_REG64(value, addr)
141 #else
142 #define READ_REG(addr) READ_REG32(addr)
143 #define WRITE_REG(value, addr) WRITE_REG32(value, addr)
144 #endif
145
146 #ifdef DEBUG_SBA_INIT
147
148 /* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */
149
150 /**
151 * sba_dump_ranges - debugging only - print ranges assigned to this IOA
152 * @hpa: base address of the sba
153 *
154 * Print the MMIO and IO Port address ranges forwarded by an Astro/Ike/RIO
155 * IO Adapter (aka Bus Converter).
156 */
157 static void
sba_dump_ranges(void __iomem * hpa)158 sba_dump_ranges(void __iomem *hpa)
159 {
160 DBG_INIT("SBA at 0x%p\n", hpa);
161 DBG_INIT("IOS_DIST_BASE : %Lx\n", READ_REG64(hpa+IOS_DIST_BASE));
162 DBG_INIT("IOS_DIST_MASK : %Lx\n", READ_REG64(hpa+IOS_DIST_MASK));
163 DBG_INIT("IOS_DIST_ROUTE : %Lx\n", READ_REG64(hpa+IOS_DIST_ROUTE));
164 DBG_INIT("\n");
165 DBG_INIT("IOS_DIRECT_BASE : %Lx\n", READ_REG64(hpa+IOS_DIRECT_BASE));
166 DBG_INIT("IOS_DIRECT_MASK : %Lx\n", READ_REG64(hpa+IOS_DIRECT_MASK));
167 DBG_INIT("IOS_DIRECT_ROUTE: %Lx\n", READ_REG64(hpa+IOS_DIRECT_ROUTE));
168 }
169
170 /**
171 * sba_dump_tlb - debugging only - print IOMMU operating parameters
172 * @hpa: base address of the IOMMU
173 *
174 * Print the size/location of the IO MMU PDIR.
175 */
sba_dump_tlb(void __iomem * hpa)176 static void sba_dump_tlb(void __iomem *hpa)
177 {
178 DBG_INIT("IO TLB at 0x%p\n", hpa);
179 DBG_INIT("IOC_IBASE : 0x%Lx\n", READ_REG64(hpa+IOC_IBASE));
180 DBG_INIT("IOC_IMASK : 0x%Lx\n", READ_REG64(hpa+IOC_IMASK));
181 DBG_INIT("IOC_TCNFG : 0x%Lx\n", READ_REG64(hpa+IOC_TCNFG));
182 DBG_INIT("IOC_PDIR_BASE: 0x%Lx\n", READ_REG64(hpa+IOC_PDIR_BASE));
183 DBG_INIT("\n");
184 }
185 #else
186 #define sba_dump_ranges(x)
187 #define sba_dump_tlb(x)
188 #endif /* DEBUG_SBA_INIT */
189
190
191 #ifdef ASSERT_PDIR_SANITY
192
193 /**
194 * sba_dump_pdir_entry - debugging only - print one IOMMU PDIR entry
195 * @ioc: IO MMU structure which owns the pdir we are interested in.
196 * @msg: text to print ont the output line.
197 * @pide: pdir index.
198 *
199 * Print one entry of the IO MMU PDIR in human readable form.
200 */
201 static void
sba_dump_pdir_entry(struct ioc * ioc,char * msg,uint pide)202 sba_dump_pdir_entry(struct ioc *ioc, char *msg, uint pide)
203 {
204 /* start printing from lowest pde in rval */
205 __le64 *ptr = &(ioc->pdir_base[pide & (~0U * BITS_PER_LONG)]);
206 unsigned long *rptr = (unsigned long *) &(ioc->res_map[(pide >>3) & ~(sizeof(unsigned long) - 1)]);
207 uint rcnt;
208
209 printk(KERN_DEBUG "SBA: %s rp %p bit %d rval 0x%lx\n",
210 msg,
211 rptr, pide & (BITS_PER_LONG - 1), *rptr);
212
213 rcnt = 0;
214 while (rcnt < BITS_PER_LONG) {
215 printk(KERN_DEBUG "%s %2d %p %016Lx\n",
216 (rcnt == (pide & (BITS_PER_LONG - 1)))
217 ? " -->" : " ",
218 rcnt, ptr, *ptr );
219 rcnt++;
220 ptr++;
221 }
222 printk(KERN_DEBUG "%s", msg);
223 }
224
225
226 /**
227 * sba_check_pdir - debugging only - consistency checker
228 * @ioc: IO MMU structure which owns the pdir we are interested in.
229 * @msg: text to print ont the output line.
230 *
231 * Verify the resource map and pdir state is consistent
232 */
233 static int
sba_check_pdir(struct ioc * ioc,char * msg)234 sba_check_pdir(struct ioc *ioc, char *msg)
235 {
236 u32 *rptr_end = (u32 *) &(ioc->res_map[ioc->res_size]);
237 u32 *rptr = (u32 *) ioc->res_map; /* resource map ptr */
238 u64 *pptr = ioc->pdir_base; /* pdir ptr */
239 uint pide = 0;
240
241 while (rptr < rptr_end) {
242 u32 rval = *rptr;
243 int rcnt = 32; /* number of bits we might check */
244
245 while (rcnt) {
246 /* Get last byte and highest bit from that */
247 u32 pde = ((u32) (((char *)pptr)[7])) << 24;
248 if ((rval ^ pde) & 0x80000000)
249 {
250 /*
251 ** BUMMER! -- res_map != pdir --
252 ** Dump rval and matching pdir entries
253 */
254 sba_dump_pdir_entry(ioc, msg, pide);
255 return(1);
256 }
257 rcnt--;
258 rval <<= 1; /* try the next bit */
259 pptr++;
260 pide++;
261 }
262 rptr++; /* look at next word of res_map */
263 }
264 /* It'd be nice if we always got here :^) */
265 return 0;
266 }
267
268
269 /**
270 * sba_dump_sg - debugging only - print Scatter-Gather list
271 * @ioc: IO MMU structure which owns the pdir we are interested in.
272 * @startsg: head of the SG list
273 * @nents: number of entries in SG list
274 *
275 * print the SG list so we can verify it's correct by hand.
276 */
277 static void
sba_dump_sg(struct ioc * ioc,struct scatterlist * startsg,int nents)278 sba_dump_sg( struct ioc *ioc, struct scatterlist *startsg, int nents)
279 {
280 while (nents-- > 0) {
281 printk(KERN_DEBUG " %d : %08lx/%05x %p/%05x\n",
282 nents,
283 (unsigned long) sg_dma_address(startsg),
284 sg_dma_len(startsg),
285 sg_virt(startsg), startsg->length);
286 startsg++;
287 }
288 }
289
290 #endif /* ASSERT_PDIR_SANITY */
291
292
293
294
295 /**************************************************************
296 *
297 * I/O Pdir Resource Management
298 *
299 * Bits set in the resource map are in use.
300 * Each bit can represent a number of pages.
301 * LSbs represent lower addresses (IOVA's).
302 *
303 ***************************************************************/
304 #define PAGES_PER_RANGE 1 /* could increase this to 4 or 8 if needed */
305
306 /* Convert from IOVP to IOVA and vice versa. */
307
308 #ifdef ZX1_SUPPORT
309 /* Pluto (aka ZX1) boxes need to set or clear the ibase bits appropriately */
310 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((ioc->ibase) | (iovp) | (offset))
311 #define SBA_IOVP(ioc,iova) ((iova) & (ioc)->iovp_mask)
312 #else
313 /* only support Astro and ancestors. Saves a few cycles in key places */
314 #define SBA_IOVA(ioc,iovp,offset,hint_reg) ((iovp) | (offset))
315 #define SBA_IOVP(ioc,iova) (iova)
316 #endif
317
318 #define PDIR_INDEX(iovp) ((iovp)>>IOVP_SHIFT)
319
320 #define RESMAP_MASK(n) (~0UL << (BITS_PER_LONG - (n)))
321 #define RESMAP_IDX_MASK (sizeof(unsigned long) - 1)
322
ptr_to_pide(struct ioc * ioc,unsigned long * res_ptr,unsigned int bitshiftcnt)323 static unsigned long ptr_to_pide(struct ioc *ioc, unsigned long *res_ptr,
324 unsigned int bitshiftcnt)
325 {
326 return (((unsigned long)res_ptr - (unsigned long)ioc->res_map) << 3)
327 + bitshiftcnt;
328 }
329
330 /**
331 * sba_search_bitmap - find free space in IO PDIR resource bitmap
332 * @ioc: IO MMU structure which owns the pdir we are interested in.
333 * @dev: device to query the bitmap for
334 * @bits_wanted: number of entries we need.
335 *
336 * Find consecutive free bits in resource bitmap.
337 * Each bit represents one entry in the IO Pdir.
338 * Cool perf optimization: search for log2(size) bits at a time.
339 */
340 static unsigned long
sba_search_bitmap(struct ioc * ioc,struct device * dev,unsigned long bits_wanted)341 sba_search_bitmap(struct ioc *ioc, struct device *dev,
342 unsigned long bits_wanted)
343 {
344 unsigned long *res_ptr = ioc->res_hint;
345 unsigned long *res_end = (unsigned long *) &(ioc->res_map[ioc->res_size]);
346 unsigned long pide = ~0UL, tpide;
347 unsigned long boundary_size;
348 unsigned long shift;
349 int ret;
350
351 boundary_size = dma_get_seg_boundary_nr_pages(dev, IOVP_SHIFT);
352
353 #if defined(ZX1_SUPPORT)
354 BUG_ON(ioc->ibase & ~IOVP_MASK);
355 shift = ioc->ibase >> IOVP_SHIFT;
356 #else
357 shift = 0;
358 #endif
359
360 if (bits_wanted > (BITS_PER_LONG/2)) {
361 /* Search word at a time - no mask needed */
362 for(; res_ptr < res_end; ++res_ptr) {
363 tpide = ptr_to_pide(ioc, res_ptr, 0);
364 ret = iommu_is_span_boundary(tpide, bits_wanted,
365 shift,
366 boundary_size);
367 if ((*res_ptr == 0) && !ret) {
368 *res_ptr = RESMAP_MASK(bits_wanted);
369 pide = tpide;
370 break;
371 }
372 }
373 /* point to the next word on next pass */
374 res_ptr++;
375 ioc->res_bitshift = 0;
376 } else {
377 /*
378 ** Search the resource bit map on well-aligned values.
379 ** "o" is the alignment.
380 ** We need the alignment to invalidate I/O TLB using
381 ** SBA HW features in the unmap path.
382 */
383 unsigned long o = 1 << get_order(bits_wanted << PAGE_SHIFT);
384 uint bitshiftcnt = ALIGN(ioc->res_bitshift, o);
385 unsigned long mask;
386
387 if (bitshiftcnt >= BITS_PER_LONG) {
388 bitshiftcnt = 0;
389 res_ptr++;
390 }
391 mask = RESMAP_MASK(bits_wanted) >> bitshiftcnt;
392
393 DBG_RES("%s() o %ld %p", __func__, o, res_ptr);
394 while(res_ptr < res_end)
395 {
396 DBG_RES(" %p %lx %lx\n", res_ptr, mask, *res_ptr);
397 WARN_ON(mask == 0);
398 tpide = ptr_to_pide(ioc, res_ptr, bitshiftcnt);
399 ret = iommu_is_span_boundary(tpide, bits_wanted,
400 shift,
401 boundary_size);
402 if ((((*res_ptr) & mask) == 0) && !ret) {
403 *res_ptr |= mask; /* mark resources busy! */
404 pide = tpide;
405 break;
406 }
407 mask >>= o;
408 bitshiftcnt += o;
409 if (mask == 0) {
410 mask = RESMAP_MASK(bits_wanted);
411 bitshiftcnt=0;
412 res_ptr++;
413 }
414 }
415 /* look in the same word on the next pass */
416 ioc->res_bitshift = bitshiftcnt + bits_wanted;
417 }
418
419 /* wrapped ? */
420 if (res_end <= res_ptr) {
421 ioc->res_hint = (unsigned long *) ioc->res_map;
422 ioc->res_bitshift = 0;
423 } else {
424 ioc->res_hint = res_ptr;
425 }
426 return (pide);
427 }
428
429
430 /**
431 * sba_alloc_range - find free bits and mark them in IO PDIR resource bitmap
432 * @ioc: IO MMU structure which owns the pdir we are interested in.
433 * @dev: device for which pages should be alloced
434 * @size: number of bytes to create a mapping for
435 *
436 * Given a size, find consecutive unmarked and then mark those bits in the
437 * resource bit map.
438 */
439 static int
sba_alloc_range(struct ioc * ioc,struct device * dev,size_t size)440 sba_alloc_range(struct ioc *ioc, struct device *dev, size_t size)
441 {
442 unsigned int pages_needed = size >> IOVP_SHIFT;
443 #ifdef SBA_COLLECT_STATS
444 unsigned long cr_start = mfctl(16);
445 #endif
446 unsigned long pide;
447
448 pide = sba_search_bitmap(ioc, dev, pages_needed);
449 if (pide >= (ioc->res_size << 3)) {
450 pide = sba_search_bitmap(ioc, dev, pages_needed);
451 if (pide >= (ioc->res_size << 3))
452 panic("%s: I/O MMU @ %p is out of mapping resources\n",
453 __FILE__, ioc->ioc_hpa);
454 }
455
456 #ifdef ASSERT_PDIR_SANITY
457 /* verify the first enable bit is clear */
458 if(0x00 != ((u8 *) ioc->pdir_base)[pide*sizeof(u64) + 7]) {
459 sba_dump_pdir_entry(ioc, "sba_search_bitmap() botched it?", pide);
460 }
461 #endif
462
463 DBG_RES("%s(%x) %d -> %lx hint %x/%x\n",
464 __func__, size, pages_needed, pide,
465 (uint) ((unsigned long) ioc->res_hint - (unsigned long) ioc->res_map),
466 ioc->res_bitshift );
467
468 #ifdef SBA_COLLECT_STATS
469 {
470 unsigned long cr_end = mfctl(16);
471 unsigned long tmp = cr_end - cr_start;
472 /* check for roll over */
473 cr_start = (cr_end < cr_start) ? -(tmp) : (tmp);
474 }
475 ioc->avg_search[ioc->avg_idx++] = cr_start;
476 ioc->avg_idx &= SBA_SEARCH_SAMPLE - 1;
477
478 ioc->used_pages += pages_needed;
479 #endif
480
481 return (pide);
482 }
483
484
485 /**
486 * sba_free_range - unmark bits in IO PDIR resource bitmap
487 * @ioc: IO MMU structure which owns the pdir we are interested in.
488 * @iova: IO virtual address which was previously allocated.
489 * @size: number of bytes to create a mapping for
490 *
491 * clear bits in the ioc's resource map
492 */
493 static void
sba_free_range(struct ioc * ioc,dma_addr_t iova,size_t size)494 sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size)
495 {
496 unsigned long iovp = SBA_IOVP(ioc, iova);
497 unsigned int pide = PDIR_INDEX(iovp);
498 unsigned int ridx = pide >> 3; /* convert bit to byte address */
499 unsigned long *res_ptr = (unsigned long *) &((ioc)->res_map[ridx & ~RESMAP_IDX_MASK]);
500
501 int bits_not_wanted = size >> IOVP_SHIFT;
502
503 /* 3-bits "bit" address plus 2 (or 3) bits for "byte" == bit in word */
504 unsigned long m = RESMAP_MASK(bits_not_wanted) >> (pide & (BITS_PER_LONG - 1));
505
506 DBG_RES("%s( ,%x,%x) %x/%lx %x %p %lx\n",
507 __func__, (uint) iova, size,
508 bits_not_wanted, m, pide, res_ptr, *res_ptr);
509
510 #ifdef SBA_COLLECT_STATS
511 ioc->used_pages -= bits_not_wanted;
512 #endif
513
514 *res_ptr &= ~m;
515 }
516
517
518 /**************************************************************
519 *
520 * "Dynamic DMA Mapping" support (aka "Coherent I/O")
521 *
522 ***************************************************************/
523
524 #ifdef SBA_HINT_SUPPORT
525 #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir)
526 #endif
527
528 typedef unsigned long space_t;
529 #define KERNEL_SPACE 0
530
531 /**
532 * sba_io_pdir_entry - fill in one IO PDIR entry
533 * @pdir_ptr: pointer to IO PDIR entry
534 * @sid: process Space ID - currently only support KERNEL_SPACE
535 * @pba: Physical address of buffer to map
536 * @hint: DMA hint set to use for this mapping
537 *
538 * SBA Mapping Routine
539 *
540 * Given a virtual address (vba, arg2) and space id, (sid, arg1)
541 * sba_io_pdir_entry() loads the I/O PDIR entry pointed to by
542 * pdir_ptr (arg0).
543 * Using the bass-ackwards HP bit numbering, Each IO Pdir entry
544 * for Astro/Ike looks like:
545 *
546 *
547 * 0 19 51 55 63
548 * +-+---------------------+----------------------------------+----+--------+
549 * |V| U | PPN[43:12] | U | VI |
550 * +-+---------------------+----------------------------------+----+--------+
551 *
552 * Pluto is basically identical, supports fewer physical address bits:
553 *
554 * 0 23 51 55 63
555 * +-+------------------------+-------------------------------+----+--------+
556 * |V| U | PPN[39:12] | U | VI |
557 * +-+------------------------+-------------------------------+----+--------+
558 *
559 * V == Valid Bit (Most Significant Bit is bit 0)
560 * U == Unused
561 * PPN == Physical Page Number
562 * VI == Virtual Index (aka Coherent Index)
563 *
564 * LPA instruction output is put into PPN field.
565 * LCI (Load Coherence Index) instruction provides the "VI" bits.
566 *
567 * We pre-swap the bytes since PCX-W is Big Endian and the
568 * IOMMU uses little endian for the pdir.
569 */
570
571 static void
sba_io_pdir_entry(__le64 * pdir_ptr,space_t sid,phys_addr_t pba,unsigned long hint)572 sba_io_pdir_entry(__le64 *pdir_ptr, space_t sid, phys_addr_t pba,
573 unsigned long hint)
574 {
575 register unsigned ci; /* coherent index */
576
577 asm("lci 0(%1), %0" : "=r" (ci) : "r" (phys_to_virt(pba)));
578 pba &= IOVP_MASK;
579 pba |= (ci >> PAGE_SHIFT) & 0xff; /* move CI (8 bits) into lowest byte */
580
581 pba |= SBA_PDIR_VALID_BIT; /* set "valid" bit */
582 *pdir_ptr = cpu_to_le64(pba); /* swap and store into I/O Pdir */
583
584 /*
585 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
586 * (bit #61, big endian), we have to flush and sync every time
587 * IO-PDIR is changed in Ike/Astro.
588 */
589 asm_io_fdc(pdir_ptr);
590 }
591
592
593 /**
594 * sba_mark_invalid - invalidate one or more IO PDIR entries
595 * @ioc: IO MMU structure which owns the pdir we are interested in.
596 * @iova: IO Virtual Address mapped earlier
597 * @byte_cnt: number of bytes this mapping covers.
598 *
599 * Marking the IO PDIR entry(ies) as Invalid and invalidate
600 * corresponding IO TLB entry. The Ike PCOM (Purge Command Register)
601 * is to purge stale entries in the IO TLB when unmapping entries.
602 *
603 * The PCOM register supports purging of multiple pages, with a minium
604 * of 1 page and a maximum of 2GB. Hardware requires the address be
605 * aligned to the size of the range being purged. The size of the range
606 * must be a power of 2. The "Cool perf optimization" in the
607 * allocation routine helps keep that true.
608 */
609 static void
sba_mark_invalid(struct ioc * ioc,dma_addr_t iova,size_t byte_cnt)610 sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt)
611 {
612 u32 iovp = (u32) SBA_IOVP(ioc,iova);
613 __le64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)];
614
615 #ifdef ASSERT_PDIR_SANITY
616 /* Assert first pdir entry is set.
617 **
618 ** Even though this is a big-endian machine, the entries
619 ** in the iopdir are little endian. That's why we look at
620 ** the byte at +7 instead of at +0.
621 */
622 if (0x80 != (((u8 *) pdir_ptr)[7])) {
623 sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp));
624 }
625 #endif
626
627 if (byte_cnt > IOVP_SIZE)
628 {
629 #if 0
630 unsigned long entries_per_cacheline = ioc_needs_fdc ?
631 L1_CACHE_ALIGN(((unsigned long) pdir_ptr))
632 - (unsigned long) pdir_ptr;
633 : 262144;
634 #endif
635
636 /* set "size" field for PCOM */
637 iovp |= get_order(byte_cnt) + PAGE_SHIFT;
638
639 do {
640 /* clear I/O Pdir entry "valid" bit first */
641 ((u8 *) pdir_ptr)[7] = 0;
642 asm_io_fdc(pdir_ptr);
643 if (ioc_needs_fdc) {
644 #if 0
645 entries_per_cacheline = L1_CACHE_SHIFT - 3;
646 #endif
647 }
648 pdir_ptr++;
649 byte_cnt -= IOVP_SIZE;
650 } while (byte_cnt > IOVP_SIZE);
651 } else
652 iovp |= IOVP_SHIFT; /* set "size" field for PCOM */
653
654 /*
655 ** clear I/O PDIR entry "valid" bit.
656 ** We have to R/M/W the cacheline regardless how much of the
657 ** pdir entry that we clobber.
658 ** The rest of the entry would be useful for debugging if we
659 ** could dump core on HPMC.
660 */
661 ((u8 *) pdir_ptr)[7] = 0;
662 asm_io_fdc(pdir_ptr);
663
664 WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM);
665 }
666
667 /**
668 * sba_dma_supported - PCI driver can query DMA support
669 * @dev: instance of PCI owned by the driver that's asking
670 * @mask: number of address bits this PCI device can handle
671 *
672 * See Documentation/core-api/dma-api-howto.rst
673 */
sba_dma_supported(struct device * dev,u64 mask)674 static int sba_dma_supported( struct device *dev, u64 mask)
675 {
676 struct ioc *ioc;
677
678 if (dev == NULL) {
679 printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n");
680 BUG();
681 return(0);
682 }
683
684 ioc = GET_IOC(dev);
685 if (!ioc)
686 return 0;
687
688 /*
689 * check if mask is >= than the current max IO Virt Address
690 * The max IO Virt address will *always* < 30 bits.
691 */
692 return((int)(mask >= (ioc->ibase - 1 +
693 (ioc->pdir_size / sizeof(u64) * IOVP_SIZE) )));
694 }
695
696
697 /**
698 * sba_map_single - map one buffer and return IOVA for DMA
699 * @dev: instance of PCI owned by the driver that's asking.
700 * @addr: driver buffer to map.
701 * @size: number of bytes to map in driver buffer.
702 * @direction: R/W or both.
703 *
704 * See Documentation/core-api/dma-api-howto.rst
705 */
706 static dma_addr_t
sba_map_single(struct device * dev,phys_addr_t addr,size_t size,enum dma_data_direction direction)707 sba_map_single(struct device *dev, phys_addr_t addr, size_t size,
708 enum dma_data_direction direction)
709 {
710 struct ioc *ioc;
711 unsigned long flags;
712 dma_addr_t iovp;
713 dma_addr_t offset;
714 __le64 *pdir_start;
715 int pide;
716
717 ioc = GET_IOC(dev);
718 if (!ioc)
719 return DMA_MAPPING_ERROR;
720
721 /* save offset bits */
722 offset = offset_in_page(addr);
723
724 /* round up to nearest IOVP_SIZE */
725 size = (size + offset + ~IOVP_MASK) & IOVP_MASK;
726
727 spin_lock_irqsave(&ioc->res_lock, flags);
728 #ifdef ASSERT_PDIR_SANITY
729 sba_check_pdir(ioc,"Check before sba_map_single()");
730 #endif
731
732 #ifdef SBA_COLLECT_STATS
733 ioc->msingle_calls++;
734 ioc->msingle_pages += size >> IOVP_SHIFT;
735 #endif
736 pide = sba_alloc_range(ioc, dev, size);
737 iovp = (dma_addr_t) pide << IOVP_SHIFT;
738
739 DBG_RUN("%s() 0x%pa -> 0x%lx\n",
740 __func__, &addr, (long) iovp | offset);
741
742 pdir_start = &(ioc->pdir_base[pide]);
743
744 while (size > 0) {
745 sba_io_pdir_entry(pdir_start, KERNEL_SPACE, addr, 0);
746
747 DBG_RUN(" pdir 0x%p %02x%02x%02x%02x%02x%02x%02x%02x\n",
748 pdir_start,
749 (u8) (((u8 *) pdir_start)[7]),
750 (u8) (((u8 *) pdir_start)[6]),
751 (u8) (((u8 *) pdir_start)[5]),
752 (u8) (((u8 *) pdir_start)[4]),
753 (u8) (((u8 *) pdir_start)[3]),
754 (u8) (((u8 *) pdir_start)[2]),
755 (u8) (((u8 *) pdir_start)[1]),
756 (u8) (((u8 *) pdir_start)[0])
757 );
758
759 addr += IOVP_SIZE;
760 size -= IOVP_SIZE;
761 pdir_start++;
762 }
763
764 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
765 asm_io_sync();
766
767 #ifdef ASSERT_PDIR_SANITY
768 sba_check_pdir(ioc,"Check after sba_map_single()");
769 #endif
770 spin_unlock_irqrestore(&ioc->res_lock, flags);
771
772 /* form complete address */
773 return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG);
774 }
775
776
777 static dma_addr_t
sba_map_phys(struct device * dev,phys_addr_t phys,size_t size,enum dma_data_direction direction,unsigned long attrs)778 sba_map_phys(struct device *dev, phys_addr_t phys, size_t size,
779 enum dma_data_direction direction, unsigned long attrs)
780 {
781 if (unlikely(attrs & DMA_ATTR_MMIO))
782 return DMA_MAPPING_ERROR;
783
784 return sba_map_single(dev, phys, size, direction);
785 }
786
787
788 /**
789 * sba_unmap_phys - unmap one IOVA and free resources
790 * @dev: instance of PCI owned by the driver that's asking.
791 * @iova: IOVA of driver buffer previously mapped.
792 * @size: number of bytes mapped in driver buffer.
793 * @direction: R/W or both.
794 * @attrs: attributes
795 *
796 * See Documentation/core-api/dma-api-howto.rst
797 */
798 static void
sba_unmap_phys(struct device * dev,dma_addr_t iova,size_t size,enum dma_data_direction direction,unsigned long attrs)799 sba_unmap_phys(struct device *dev, dma_addr_t iova, size_t size,
800 enum dma_data_direction direction, unsigned long attrs)
801 {
802 struct ioc *ioc;
803 #if DELAYED_RESOURCE_CNT > 0
804 struct sba_dma_pair *d;
805 #endif
806 unsigned long flags;
807 dma_addr_t offset;
808
809 DBG_RUN("%s() iovp 0x%lx/%x\n", __func__, (long) iova, size);
810
811 ioc = GET_IOC(dev);
812 if (!ioc) {
813 WARN_ON(!ioc);
814 return;
815 }
816 offset = iova & ~IOVP_MASK;
817 iova ^= offset; /* clear offset bits */
818 size += offset;
819 size = ALIGN(size, IOVP_SIZE);
820
821 spin_lock_irqsave(&ioc->res_lock, flags);
822
823 #ifdef SBA_COLLECT_STATS
824 ioc->usingle_calls++;
825 ioc->usingle_pages += size >> IOVP_SHIFT;
826 #endif
827
828 sba_mark_invalid(ioc, iova, size);
829
830 #if DELAYED_RESOURCE_CNT > 0
831 /* Delaying when we re-use a IO Pdir entry reduces the number
832 * of MMIO reads needed to flush writes to the PCOM register.
833 */
834 d = &(ioc->saved[ioc->saved_cnt]);
835 d->iova = iova;
836 d->size = size;
837 if (++(ioc->saved_cnt) >= DELAYED_RESOURCE_CNT) {
838 int cnt = ioc->saved_cnt;
839 while (cnt--) {
840 sba_free_range(ioc, d->iova, d->size);
841 d--;
842 }
843 ioc->saved_cnt = 0;
844
845 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
846 }
847 #else /* DELAYED_RESOURCE_CNT == 0 */
848 sba_free_range(ioc, iova, size);
849
850 /* If fdc's were issued, force fdc's to be visible now */
851 asm_io_sync();
852
853 READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */
854 #endif /* DELAYED_RESOURCE_CNT == 0 */
855
856 spin_unlock_irqrestore(&ioc->res_lock, flags);
857
858 /* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support.
859 ** For Astro based systems this isn't a big deal WRT performance.
860 ** As long as 2.4 kernels copyin/copyout data from/to userspace,
861 ** we don't need the syncdma. The issue here is I/O MMU cachelines
862 ** are *not* coherent in all cases. May be hwrev dependent.
863 ** Need to investigate more.
864 asm volatile("syncdma");
865 */
866 }
867
868
869 /**
870 * sba_alloc - allocate/map shared mem for DMA
871 * @hwdev: instance of PCI owned by the driver that's asking.
872 * @size: number of bytes mapped in driver buffer.
873 * @dma_handle: IOVA of new buffer.
874 * @gfp: allocation flags
875 * @attrs: attributes
876 *
877 * See Documentation/core-api/dma-api-howto.rst
878 */
sba_alloc(struct device * hwdev,size_t size,dma_addr_t * dma_handle,gfp_t gfp,unsigned long attrs)879 static void *sba_alloc(struct device *hwdev, size_t size, dma_addr_t *dma_handle,
880 gfp_t gfp, unsigned long attrs)
881 {
882 void *ret;
883
884 if (!hwdev) {
885 /* only support PCI */
886 *dma_handle = 0;
887 return NULL;
888 }
889
890 ret = (void *) __get_free_pages(gfp, get_order(size));
891
892 if (ret) {
893 memset(ret, 0, size);
894 *dma_handle = sba_map_single(hwdev, virt_to_phys(ret), size, 0);
895 }
896
897 return ret;
898 }
899
900
901 /**
902 * sba_free - free/unmap shared mem for DMA
903 * @hwdev: instance of PCI owned by the driver that's asking.
904 * @size: number of bytes mapped in driver buffer.
905 * @vaddr: virtual address IOVA of "consistent" buffer.
906 * @dma_handle: IO virtual address of "consistent" buffer.
907 * @attrs: attributes
908 *
909 * See Documentation/core-api/dma-api-howto.rst
910 */
911 static void
sba_free(struct device * hwdev,size_t size,void * vaddr,dma_addr_t dma_handle,unsigned long attrs)912 sba_free(struct device *hwdev, size_t size, void *vaddr,
913 dma_addr_t dma_handle, unsigned long attrs)
914 {
915 sba_unmap_phys(hwdev, dma_handle, size, 0, 0);
916 free_pages((unsigned long) vaddr, get_order(size));
917 }
918
919
920 /*
921 ** Since 0 is a valid pdir_base index value, can't use that
922 ** to determine if a value is valid or not. Use a flag to indicate
923 ** the SG list entry contains a valid pdir index.
924 */
925 #define PIDE_FLAG 0x80000000UL
926
927 #ifdef SBA_COLLECT_STATS
928 #define IOMMU_MAP_STATS
929 #endif
930 #include "iommu-helpers.h"
931
932 #ifdef DEBUG_LARGE_SG_ENTRIES
933 int dump_run_sg = 0;
934 #endif
935
936
937 /**
938 * sba_map_sg - map Scatter/Gather list
939 * @dev: instance of PCI owned by the driver that's asking.
940 * @sglist: array of buffer/length pairs
941 * @nents: number of entries in list
942 * @direction: R/W or both.
943 * @attrs: attributes
944 *
945 * See Documentation/core-api/dma-api-howto.rst
946 */
947 static int
sba_map_sg(struct device * dev,struct scatterlist * sglist,int nents,enum dma_data_direction direction,unsigned long attrs)948 sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents,
949 enum dma_data_direction direction, unsigned long attrs)
950 {
951 struct ioc *ioc;
952 int filled = 0;
953 unsigned long flags;
954
955 DBG_RUN_SG("%s() START %d entries\n", __func__, nents);
956
957 ioc = GET_IOC(dev);
958 if (!ioc)
959 return -EINVAL;
960
961 /* Fast path single entry scatterlists. */
962 if (nents == 1) {
963 sg_dma_address(sglist) = sba_map_single(dev, sg_phys(sglist),
964 sglist->length, direction);
965 sg_dma_len(sglist) = sglist->length;
966 return 1;
967 }
968
969 spin_lock_irqsave(&ioc->res_lock, flags);
970
971 #ifdef ASSERT_PDIR_SANITY
972 if (sba_check_pdir(ioc,"Check before sba_map_sg()"))
973 {
974 sba_dump_sg(ioc, sglist, nents);
975 panic("Check before sba_map_sg()");
976 }
977 #endif
978
979 #ifdef SBA_COLLECT_STATS
980 ioc->msg_calls++;
981 #endif
982
983 /*
984 ** First coalesce the chunks and allocate I/O pdir space
985 **
986 ** If this is one DMA stream, we can properly map using the
987 ** correct virtual address associated with each DMA page.
988 ** w/o this association, we wouldn't have coherent DMA!
989 ** Access to the virtual address is what forces a two pass algorithm.
990 */
991 iommu_coalesce_chunks(ioc, dev, sglist, nents, sba_alloc_range);
992
993 /*
994 ** Program the I/O Pdir
995 **
996 ** map the virtual addresses to the I/O Pdir
997 ** o dma_address will contain the pdir index
998 ** o dma_len will contain the number of bytes to map
999 ** o address contains the virtual address.
1000 */
1001 filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry);
1002
1003 /* force FDC ops in io_pdir_entry() to be visible to IOMMU */
1004 asm_io_sync();
1005
1006 #ifdef ASSERT_PDIR_SANITY
1007 if (sba_check_pdir(ioc,"Check after sba_map_sg()"))
1008 {
1009 sba_dump_sg(ioc, sglist, nents);
1010 panic("Check after sba_map_sg()\n");
1011 }
1012 #endif
1013
1014 spin_unlock_irqrestore(&ioc->res_lock, flags);
1015
1016 DBG_RUN_SG("%s() DONE %d mappings\n", __func__, filled);
1017
1018 return filled;
1019 }
1020
1021
1022 /**
1023 * sba_unmap_sg - unmap Scatter/Gather list
1024 * @dev: instance of PCI owned by the driver that's asking.
1025 * @sglist: array of buffer/length pairs
1026 * @nents: number of entries in list
1027 * @direction: R/W or both.
1028 * @attrs: attributes
1029 *
1030 * See Documentation/core-api/dma-api-howto.rst
1031 */
1032 static void
sba_unmap_sg(struct device * dev,struct scatterlist * sglist,int nents,enum dma_data_direction direction,unsigned long attrs)1033 sba_unmap_sg(struct device *dev, struct scatterlist *sglist, int nents,
1034 enum dma_data_direction direction, unsigned long attrs)
1035 {
1036 struct ioc *ioc;
1037 #ifdef ASSERT_PDIR_SANITY
1038 unsigned long flags;
1039 #endif
1040
1041 DBG_RUN_SG("%s() START %d entries, %p,%x\n",
1042 __func__, nents, sg_virt(sglist), sglist->length);
1043
1044 ioc = GET_IOC(dev);
1045 if (!ioc) {
1046 WARN_ON(!ioc);
1047 return;
1048 }
1049
1050 #ifdef SBA_COLLECT_STATS
1051 ioc->usg_calls++;
1052 #endif
1053
1054 #ifdef ASSERT_PDIR_SANITY
1055 spin_lock_irqsave(&ioc->res_lock, flags);
1056 sba_check_pdir(ioc,"Check before sba_unmap_sg()");
1057 spin_unlock_irqrestore(&ioc->res_lock, flags);
1058 #endif
1059
1060 while (nents && sg_dma_len(sglist)) {
1061
1062 sba_unmap_phys(dev, sg_dma_address(sglist), sg_dma_len(sglist),
1063 direction, 0);
1064 #ifdef SBA_COLLECT_STATS
1065 ioc->usg_pages += ((sg_dma_address(sglist) & ~IOVP_MASK) + sg_dma_len(sglist) + IOVP_SIZE - 1) >> PAGE_SHIFT;
1066 ioc->usingle_calls--; /* kluge since call is unmap_sg() */
1067 #endif
1068 ++sglist;
1069 nents--;
1070 }
1071
1072 DBG_RUN_SG("%s() DONE (nents %d)\n", __func__, nents);
1073
1074 #ifdef ASSERT_PDIR_SANITY
1075 spin_lock_irqsave(&ioc->res_lock, flags);
1076 sba_check_pdir(ioc,"Check after sba_unmap_sg()");
1077 spin_unlock_irqrestore(&ioc->res_lock, flags);
1078 #endif
1079
1080 }
1081
1082 static const struct dma_map_ops sba_ops = {
1083 .dma_supported = sba_dma_supported,
1084 .alloc = sba_alloc,
1085 .free = sba_free,
1086 .map_phys = sba_map_phys,
1087 .unmap_phys = sba_unmap_phys,
1088 .map_sg = sba_map_sg,
1089 .unmap_sg = sba_unmap_sg,
1090 .get_sgtable = dma_common_get_sgtable,
1091 .alloc_pages_op = dma_common_alloc_pages,
1092 .free_pages = dma_common_free_pages,
1093 };
1094
1095
1096 /**************************************************************************
1097 **
1098 ** SBA PAT PDC support
1099 **
1100 ** o call pdc_pat_cell_module()
1101 ** o store ranges in PCI "resource" structures
1102 **
1103 **************************************************************************/
1104
1105 static void
sba_get_pat_resources(struct sba_device * sba_dev)1106 sba_get_pat_resources(struct sba_device *sba_dev)
1107 {
1108 #if 0
1109 /*
1110 ** TODO/REVISIT/FIXME: support for directed ranges requires calls to
1111 ** PAT PDC to program the SBA/LBA directed range registers...this
1112 ** burden may fall on the LBA code since it directly supports the
1113 ** PCI subsystem. It's not clear yet. - ggg
1114 */
1115 PAT_MOD(mod)->mod_info.mod_pages = PAT_GET_MOD_PAGES(temp);
1116 FIXME : ???
1117 PAT_MOD(mod)->mod_info.dvi = PAT_GET_DVI(temp);
1118 Tells where the dvi bits are located in the address.
1119 PAT_MOD(mod)->mod_info.ioc = PAT_GET_IOC(temp);
1120 FIXME : ???
1121 #endif
1122 }
1123
1124
1125 /**************************************************************
1126 *
1127 * Initialization and claim
1128 *
1129 ***************************************************************/
1130 #define PIRANHA_ADDR_MASK 0x00160000UL /* bit 17,18,20 */
1131 #define PIRANHA_ADDR_VAL 0x00060000UL /* bit 17,18 on */
1132 static void *
sba_alloc_pdir(unsigned int pdir_size)1133 sba_alloc_pdir(unsigned int pdir_size)
1134 {
1135 unsigned long pdir_base;
1136 unsigned long pdir_order = get_order(pdir_size);
1137
1138 pdir_base = __get_free_pages(GFP_KERNEL, pdir_order);
1139 if (NULL == (void *) pdir_base) {
1140 panic("%s() could not allocate I/O Page Table\n",
1141 __func__);
1142 }
1143
1144 /* If this is not PA8700 (PCX-W2)
1145 ** OR newer than ver 2.2
1146 ** OR in a system that doesn't need VINDEX bits from SBA,
1147 **
1148 ** then we aren't exposed to the HW bug.
1149 */
1150 if ( ((boot_cpu_data.pdc.cpuid >> 5) & 0x7f) != 0x13
1151 || (boot_cpu_data.pdc.versions > 0x202)
1152 || (boot_cpu_data.pdc.capabilities & 0x08L) )
1153 return (void *) pdir_base;
1154
1155 /*
1156 * PA8700 (PCX-W2, aka piranha) silent data corruption fix
1157 *
1158 * An interaction between PA8700 CPU (Ver 2.2 or older) and
1159 * Ike/Astro can cause silent data corruption. This is only
1160 * a problem if the I/O PDIR is located in memory such that
1161 * (little-endian) bits 17 and 18 are on and bit 20 is off.
1162 *
1163 * Since the max IO Pdir size is 2MB, by cleverly allocating the
1164 * right physical address, we can either avoid (IOPDIR <= 1MB)
1165 * or minimize (2MB IO Pdir) the problem if we restrict the
1166 * IO Pdir to a maximum size of 2MB-128K (1902K).
1167 *
1168 * Because we always allocate 2^N sized IO pdirs, either of the
1169 * "bad" regions will be the last 128K if at all. That's easy
1170 * to test for.
1171 *
1172 */
1173 if (pdir_order <= (19-12)) {
1174 if (((virt_to_phys(pdir_base)+pdir_size-1) & PIRANHA_ADDR_MASK) == PIRANHA_ADDR_VAL) {
1175 /* allocate a new one on 512k alignment */
1176 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, (19-12));
1177 /* release original */
1178 free_pages(pdir_base, pdir_order);
1179
1180 pdir_base = new_pdir;
1181
1182 /* release excess */
1183 while (pdir_order < (19-12)) {
1184 new_pdir += pdir_size;
1185 free_pages(new_pdir, pdir_order);
1186 pdir_order +=1;
1187 pdir_size <<=1;
1188 }
1189 }
1190 } else {
1191 /*
1192 ** 1MB or 2MB Pdir
1193 ** Needs to be aligned on an "odd" 1MB boundary.
1194 */
1195 unsigned long new_pdir = __get_free_pages(GFP_KERNEL, pdir_order+1); /* 2 or 4MB */
1196
1197 /* release original */
1198 free_pages( pdir_base, pdir_order);
1199
1200 /* release first 1MB */
1201 free_pages(new_pdir, 20-12);
1202
1203 pdir_base = new_pdir + 1024*1024;
1204
1205 if (pdir_order > (20-12)) {
1206 /*
1207 ** 2MB Pdir.
1208 **
1209 ** Flag tells init_bitmap() to mark bad 128k as used
1210 ** and to reduce the size by 128k.
1211 */
1212 piranha_bad_128k = 1;
1213
1214 new_pdir += 3*1024*1024;
1215 /* release last 1MB */
1216 free_pages(new_pdir, 20-12);
1217
1218 /* release unusable 128KB */
1219 free_pages(new_pdir - 128*1024 , 17-12);
1220
1221 pdir_size -= 128*1024;
1222 }
1223 }
1224
1225 memset((void *) pdir_base, 0, pdir_size);
1226 return (void *) pdir_base;
1227 }
1228
1229 struct ibase_data_struct {
1230 struct ioc *ioc;
1231 int ioc_num;
1232 };
1233
setup_ibase_imask_callback(struct device * dev,void * data)1234 static int setup_ibase_imask_callback(struct device *dev, void *data)
1235 {
1236 struct parisc_device *lba = to_parisc_device(dev);
1237 struct ibase_data_struct *ibd = data;
1238 int rope_num = (lba->hpa.start >> 13) & 0xf;
1239 if (rope_num >> 3 == ibd->ioc_num)
1240 lba_set_iregs(lba, ibd->ioc->ibase, ibd->ioc->imask);
1241 return 0;
1242 }
1243
1244 /* setup Mercury or Elroy IBASE/IMASK registers. */
1245 static void
setup_ibase_imask(struct parisc_device * sba,struct ioc * ioc,int ioc_num)1246 setup_ibase_imask(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1247 {
1248 struct ibase_data_struct ibase_data = {
1249 .ioc = ioc,
1250 .ioc_num = ioc_num,
1251 };
1252
1253 device_for_each_child(&sba->dev, &ibase_data,
1254 setup_ibase_imask_callback);
1255 }
1256
1257 #ifdef SBA_AGP_SUPPORT
1258 static int
sba_ioc_find_quicksilver(struct device * dev,void * data)1259 sba_ioc_find_quicksilver(struct device *dev, void *data)
1260 {
1261 int *agp_found = data;
1262 struct parisc_device *lba = to_parisc_device(dev);
1263
1264 if (IS_QUICKSILVER(lba))
1265 *agp_found = 1;
1266 return 0;
1267 }
1268 #endif
1269
1270 static void
sba_ioc_init_pluto(struct parisc_device * sba,struct ioc * ioc,int ioc_num)1271 sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1272 {
1273 u32 iova_space_mask;
1274 u32 iova_space_size;
1275 int iov_order, tcnfg;
1276 #ifdef SBA_AGP_SUPPORT
1277 int agp_found = 0;
1278 #endif
1279 /*
1280 ** Firmware programs the base and size of a "safe IOVA space"
1281 ** (one that doesn't overlap memory or LMMIO space) in the
1282 ** IBASE and IMASK registers.
1283 */
1284 ioc->ibase = READ_REG(ioc->ioc_hpa + IOC_IBASE) & ~0x1fffffULL;
1285 iova_space_size = ~(READ_REG(ioc->ioc_hpa + IOC_IMASK) & 0xFFFFFFFFUL) + 1;
1286
1287 if ((ioc->ibase < 0xfed00000UL) && ((ioc->ibase + iova_space_size) > 0xfee00000UL)) {
1288 printk("WARNING: IOV space overlaps local config and interrupt message, truncating\n");
1289 iova_space_size /= 2;
1290 }
1291
1292 /*
1293 ** iov_order is always based on a 1GB IOVA space since we want to
1294 ** turn on the other half for AGP GART.
1295 */
1296 iov_order = get_order(iova_space_size >> (IOVP_SHIFT - PAGE_SHIFT));
1297 ioc->pdir_size = (iova_space_size / IOVP_SIZE) * sizeof(u64);
1298
1299 DBG_INIT("%s() hpa 0x%p IOV %dMB (%d bits)\n",
1300 __func__, ioc->ioc_hpa, iova_space_size >> 20,
1301 iov_order + PAGE_SHIFT);
1302
1303 ioc->pdir_base = (void *) __get_free_pages(GFP_KERNEL,
1304 get_order(ioc->pdir_size));
1305 if (!ioc->pdir_base)
1306 panic("Couldn't allocate I/O Page Table\n");
1307
1308 memset(ioc->pdir_base, 0, ioc->pdir_size);
1309
1310 DBG_INIT("%s() pdir %p size %x\n",
1311 __func__, ioc->pdir_base, ioc->pdir_size);
1312
1313 #ifdef SBA_HINT_SUPPORT
1314 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1315 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1316
1317 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1318 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1319 #endif
1320
1321 WARN_ON((((unsigned long) ioc->pdir_base) & PAGE_MASK) != (unsigned long) ioc->pdir_base);
1322 WRITE_REG(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1323
1324 /* build IMASK for IOC and Elroy */
1325 iova_space_mask = 0xffffffff;
1326 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1327 ioc->imask = iova_space_mask;
1328 #ifdef ZX1_SUPPORT
1329 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1330 #endif
1331 sba_dump_tlb(ioc->ioc_hpa);
1332
1333 setup_ibase_imask(sba, ioc, ioc_num);
1334
1335 WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK);
1336
1337 #ifdef CONFIG_64BIT
1338 /*
1339 ** Setting the upper bits makes checking for bypass addresses
1340 ** a little faster later on.
1341 */
1342 ioc->imask |= 0xFFFFFFFF00000000UL;
1343 #endif
1344
1345 /* Set I/O PDIR Page size to system page size */
1346 switch (PAGE_SHIFT) {
1347 case 12: tcnfg = 0; break; /* 4K */
1348 case 13: tcnfg = 1; break; /* 8K */
1349 case 14: tcnfg = 2; break; /* 16K */
1350 case 16: tcnfg = 3; break; /* 64K */
1351 default:
1352 panic(__FILE__ "Unsupported system page size %d",
1353 1 << PAGE_SHIFT);
1354 break;
1355 }
1356 WRITE_REG(tcnfg, ioc->ioc_hpa + IOC_TCNFG);
1357
1358 /*
1359 ** Program the IOC's ibase and enable IOVA translation
1360 ** Bit zero == enable bit.
1361 */
1362 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa + IOC_IBASE);
1363
1364 /*
1365 ** Clear I/O TLB of any possible entries.
1366 ** (Yes. This is a bit paranoid...but so what)
1367 */
1368 WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM);
1369
1370 #ifdef SBA_AGP_SUPPORT
1371
1372 /*
1373 ** If an AGP device is present, only use half of the IOV space
1374 ** for PCI DMA. Unfortunately we can't know ahead of time
1375 ** whether GART support will actually be used, for now we
1376 ** can just key on any AGP device found in the system.
1377 ** We program the next pdir index after we stop w/ a key for
1378 ** the GART code to handshake on.
1379 */
1380 device_for_each_child(&sba->dev, &agp_found, sba_ioc_find_quicksilver);
1381
1382 if (agp_found && sba_reserve_agpgart) {
1383 printk(KERN_INFO "%s: reserving %dMb of IOVA space for agpgart\n",
1384 __func__, (iova_space_size/2) >> 20);
1385 ioc->pdir_size /= 2;
1386 ioc->pdir_base[PDIR_INDEX(iova_space_size/2)] = SBA_AGPGART_COOKIE;
1387 }
1388 #endif /*SBA_AGP_SUPPORT*/
1389 }
1390
1391 static void
sba_ioc_init(struct parisc_device * sba,struct ioc * ioc,int ioc_num)1392 sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num)
1393 {
1394 u32 iova_space_size, iova_space_mask;
1395 unsigned int pdir_size, iov_order, tcnfg;
1396
1397 /*
1398 ** Determine IOVA Space size from memory size.
1399 **
1400 ** Ideally, PCI drivers would register the maximum number
1401 ** of DMA they can have outstanding for each device they
1402 ** own. Next best thing would be to guess how much DMA
1403 ** can be outstanding based on PCI Class/sub-class. Both
1404 ** methods still require some "extra" to support PCI
1405 ** Hot-Plug/Removal of PCI cards. (aka PCI OLARD).
1406 **
1407 ** While we have 32-bits "IOVA" space, top two 2 bits are used
1408 ** for DMA hints - ergo only 30 bits max.
1409 */
1410
1411 iova_space_size = (u32) (totalram_pages()/global_ioc_cnt);
1412
1413 /* limit IOVA space size to 1MB-1GB */
1414 if (iova_space_size < (1 << (20 - PAGE_SHIFT))) {
1415 iova_space_size = 1 << (20 - PAGE_SHIFT);
1416 }
1417 else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) {
1418 iova_space_size = 1 << (30 - PAGE_SHIFT);
1419 }
1420
1421 /*
1422 ** iova space must be log2() in size.
1423 ** thus, pdir/res_map will also be log2().
1424 ** PIRANHA BUG: Exception is when IO Pdir is 2MB (gets reduced)
1425 */
1426 iov_order = get_order(iova_space_size << PAGE_SHIFT);
1427
1428 /* iova_space_size is now bytes, not pages */
1429 iova_space_size = 1 << (iov_order + PAGE_SHIFT);
1430
1431 ioc->pdir_size = pdir_size = (iova_space_size/IOVP_SIZE) * sizeof(u64);
1432
1433 DBG_INIT("%s() hpa %px mem %ldMB IOV %dMB (%d bits)\n",
1434 __func__,
1435 ioc->ioc_hpa,
1436 (unsigned long) totalram_pages() >> (20 - PAGE_SHIFT),
1437 iova_space_size>>20,
1438 iov_order + PAGE_SHIFT);
1439
1440 ioc->pdir_base = sba_alloc_pdir(pdir_size);
1441
1442 DBG_INIT("%s() pdir %p size %x\n",
1443 __func__, ioc->pdir_base, pdir_size);
1444
1445 #ifdef SBA_HINT_SUPPORT
1446 /* FIXME : DMA HINTs not used */
1447 ioc->hint_shift_pdir = iov_order + PAGE_SHIFT;
1448 ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT));
1449
1450 DBG_INIT(" hint_shift_pdir %x hint_mask_pdir %lx\n",
1451 ioc->hint_shift_pdir, ioc->hint_mask_pdir);
1452 #endif
1453
1454 WRITE_REG64(virt_to_phys(ioc->pdir_base), ioc->ioc_hpa + IOC_PDIR_BASE);
1455
1456 /* build IMASK for IOC and Elroy */
1457 iova_space_mask = 0xffffffff;
1458 iova_space_mask <<= (iov_order + PAGE_SHIFT);
1459
1460 /*
1461 ** On C3000 w/512MB mem, HP-UX 10.20 reports:
1462 ** ibase=0, imask=0xFE000000, size=0x2000000.
1463 */
1464 ioc->ibase = 0;
1465 ioc->imask = iova_space_mask; /* save it */
1466 #ifdef ZX1_SUPPORT
1467 ioc->iovp_mask = ~(iova_space_mask + PAGE_SIZE - 1);
1468 #endif
1469
1470 DBG_INIT("%s() IOV base %#lx mask %#0lx\n",
1471 __func__, ioc->ibase, ioc->imask);
1472
1473 /*
1474 ** FIXME: Hint registers are programmed with default hint
1475 ** values during boot, so hints should be sane even if we
1476 ** can't reprogram them the way drivers want.
1477 */
1478
1479 setup_ibase_imask(sba, ioc, ioc_num);
1480
1481 /*
1482 ** Program the IOC's ibase and enable IOVA translation
1483 */
1484 WRITE_REG(ioc->ibase | 1, ioc->ioc_hpa+IOC_IBASE);
1485 WRITE_REG(ioc->imask, ioc->ioc_hpa+IOC_IMASK);
1486
1487 /* Set I/O PDIR Page size to system page size */
1488 switch (PAGE_SHIFT) {
1489 case 12: tcnfg = 0; break; /* 4K */
1490 case 13: tcnfg = 1; break; /* 8K */
1491 case 14: tcnfg = 2; break; /* 16K */
1492 case 16: tcnfg = 3; break; /* 64K */
1493 default:
1494 panic(__FILE__ "Unsupported system page size %d",
1495 1 << PAGE_SHIFT);
1496 break;
1497 }
1498 /* Set I/O PDIR Page size to PAGE_SIZE (4k/16k/...) */
1499 WRITE_REG(tcnfg, ioc->ioc_hpa+IOC_TCNFG);
1500
1501 /*
1502 ** Clear I/O TLB of any possible entries.
1503 ** (Yes. This is a bit paranoid...but so what)
1504 */
1505 WRITE_REG(0 | 31, ioc->ioc_hpa+IOC_PCOM);
1506
1507 ioc->ibase = 0; /* used by SBA_IOVA and related macros */
1508
1509 DBG_INIT("%s() DONE\n", __func__);
1510 }
1511
1512
1513
1514 /**************************************************************************
1515 **
1516 ** SBA initialization code (HW and SW)
1517 **
1518 ** o identify SBA chip itself
1519 ** o initialize SBA chip modes (HardFail)
1520 ** o initialize SBA chip modes (HardFail)
1521 ** o FIXME: initialize DMA hints for reasonable defaults
1522 **
1523 **************************************************************************/
1524
ioc_remap(struct sba_device * sba_dev,unsigned int offset)1525 static void __iomem *ioc_remap(struct sba_device *sba_dev, unsigned int offset)
1526 {
1527 return ioremap(sba_dev->dev->hpa.start + offset, SBA_FUNC_SIZE);
1528 }
1529
sba_hw_init(struct sba_device * sba_dev)1530 static void sba_hw_init(struct sba_device *sba_dev)
1531 {
1532 int i;
1533 int num_ioc;
1534 u64 ioc_ctl;
1535
1536 if (!is_pdc_pat()) {
1537 /* Shutdown the USB controller on Astro-based workstations.
1538 ** Once we reprogram the IOMMU, the next DMA performed by
1539 ** USB will HPMC the box. USB is only enabled if a
1540 ** keyboard is present and found.
1541 **
1542 ** With serial console, j6k v5.0 firmware says:
1543 ** mem_kbd hpa 0xfee003f8 sba 0x0 pad 0x0 cl_class 0x7
1544 **
1545 ** FIXME: Using GFX+USB console at power up but direct
1546 ** linux to serial console is still broken.
1547 ** USB could generate DMA so we must reset USB.
1548 ** The proper sequence would be:
1549 ** o block console output
1550 ** o reset USB device
1551 ** o reprogram serial port
1552 ** o unblock console output
1553 */
1554 if (PAGE0->mem_kbd.cl_class == CL_KEYBD) {
1555 pdc_io_reset_devices();
1556 }
1557
1558 }
1559
1560
1561 #if 0
1562 printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
1563 PAGE0->mem_boot.spa, PAGE0->mem_boot.pad, PAGE0->mem_boot.cl_class);
1564
1565 /*
1566 ** Need to deal with DMA from LAN.
1567 ** Maybe use page zero boot device as a handle to talk
1568 ** to PDC about which device to shutdown.
1569 **
1570 ** Netbooting, j6k v5.0 firmware says:
1571 ** mem_boot hpa 0xf4008000 sba 0x0 pad 0x0 cl_class 0x1002
1572 ** ARGH! invalid class.
1573 */
1574 if ((PAGE0->mem_boot.cl_class != CL_RANDOM)
1575 && (PAGE0->mem_boot.cl_class != CL_SEQU)) {
1576 pdc_io_reset();
1577 }
1578 #endif
1579
1580 if (!IS_PLUTO(sba_dev->dev)) {
1581 ioc_ctl = READ_REG(sba_dev->sba_hpa+IOC_CTRL);
1582 DBG_INIT("%s() hpa %px ioc_ctl 0x%Lx ->",
1583 __func__, sba_dev->sba_hpa, ioc_ctl);
1584 ioc_ctl &= ~(IOC_CTRL_RM | IOC_CTRL_NC | IOC_CTRL_CE);
1585 ioc_ctl |= IOC_CTRL_DD | IOC_CTRL_D4 | IOC_CTRL_TC;
1586 /* j6700 v1.6 firmware sets 0x294f */
1587 /* A500 firmware sets 0x4d */
1588
1589 WRITE_REG(ioc_ctl, sba_dev->sba_hpa+IOC_CTRL);
1590
1591 #ifdef DEBUG_SBA_INIT
1592 ioc_ctl = READ_REG64(sba_dev->sba_hpa+IOC_CTRL);
1593 DBG_INIT(" 0x%Lx\n", ioc_ctl);
1594 #endif
1595 } /* if !PLUTO */
1596
1597 if (IS_ASTRO(sba_dev->dev)) {
1598 int err;
1599 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, ASTRO_IOC_OFFSET);
1600 num_ioc = 1;
1601
1602 sba_dev->chip_resv.name = "Astro Intr Ack";
1603 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfef00000UL;
1604 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff000000UL - 1) ;
1605 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1606 BUG_ON(err < 0);
1607
1608 } else if (IS_PLUTO(sba_dev->dev)) {
1609 int err;
1610
1611 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, PLUTO_IOC_OFFSET);
1612 num_ioc = 1;
1613
1614 sba_dev->chip_resv.name = "Pluto Intr/PIOP/VGA";
1615 sba_dev->chip_resv.start = PCI_F_EXTEND | 0xfee00000UL;
1616 sba_dev->chip_resv.end = PCI_F_EXTEND | (0xff200000UL - 1);
1617 err = request_resource(&iomem_resource, &(sba_dev->chip_resv));
1618 WARN_ON(err < 0);
1619
1620 sba_dev->iommu_resv.name = "IOVA Space";
1621 sba_dev->iommu_resv.start = 0x40000000UL;
1622 sba_dev->iommu_resv.end = 0x50000000UL - 1;
1623 err = request_resource(&iomem_resource, &(sba_dev->iommu_resv));
1624 WARN_ON(err < 0);
1625 } else {
1626 /* IKE, REO */
1627 sba_dev->ioc[0].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(0));
1628 sba_dev->ioc[1].ioc_hpa = ioc_remap(sba_dev, IKE_IOC_OFFSET(1));
1629 num_ioc = 2;
1630
1631 /* TODO - LOOKUP Ike/Stretch chipset mem map */
1632 }
1633 /* XXX: What about Reo Grande? */
1634
1635 sba_dev->num_ioc = num_ioc;
1636 for (i = 0; i < num_ioc; i++) {
1637 void __iomem *ioc_hpa = sba_dev->ioc[i].ioc_hpa;
1638 unsigned int j;
1639
1640 for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) {
1641
1642 /*
1643 * Clear ROPE(N)_CONFIG AO bit.
1644 * Disables "NT Ordering" (~= !"Relaxed Ordering")
1645 * Overrides bit 1 in DMA Hint Sets.
1646 * Improves netperf UDP_STREAM by ~10% for bcm5701.
1647 */
1648 if (IS_PLUTO(sba_dev->dev)) {
1649 void __iomem *rope_cfg;
1650 unsigned long cfg_val;
1651
1652 rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j;
1653 cfg_val = READ_REG(rope_cfg);
1654 cfg_val &= ~IOC_ROPE_AO;
1655 WRITE_REG(cfg_val, rope_cfg);
1656 }
1657
1658 /*
1659 ** Make sure the box crashes on rope errors.
1660 */
1661 WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j);
1662 }
1663
1664 /* flush out the last writes */
1665 READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
1666
1667 DBG_INIT(" ioc[%d] ROPE_CFG %#lx ROPE_DBG %lx\n",
1668 i,
1669 (unsigned long) READ_REG(sba_dev->ioc[i].ioc_hpa + 0x40),
1670 (unsigned long) READ_REG(sba_dev->ioc[i].ioc_hpa + 0x50)
1671 );
1672 DBG_INIT(" STATUS_CONTROL %#lx FLUSH_CTRL %#lx\n",
1673 (unsigned long) READ_REG(sba_dev->ioc[i].ioc_hpa + 0x108),
1674 (unsigned long) READ_REG(sba_dev->ioc[i].ioc_hpa + 0x400)
1675 );
1676
1677 if (IS_PLUTO(sba_dev->dev)) {
1678 sba_ioc_init_pluto(sba_dev->dev, &(sba_dev->ioc[i]), i);
1679 } else {
1680 sba_ioc_init(sba_dev->dev, &(sba_dev->ioc[i]), i);
1681 }
1682 }
1683 }
1684
1685 static void
sba_common_init(struct sba_device * sba_dev)1686 sba_common_init(struct sba_device *sba_dev)
1687 {
1688 int i;
1689
1690 /* add this one to the head of the list (order doesn't matter)
1691 ** This will be useful for debugging - especially if we get coredumps
1692 */
1693 sba_dev->next = sba_list;
1694 sba_list = sba_dev;
1695
1696 for(i=0; i< sba_dev->num_ioc; i++) {
1697 int res_size;
1698 #ifdef DEBUG_DMB_TRAP
1699 extern void iterate_pages(unsigned long , unsigned long ,
1700 void (*)(pte_t * , unsigned long),
1701 unsigned long );
1702 void set_data_memory_break(pte_t * , unsigned long);
1703 #endif
1704 /* resource map size dictated by pdir_size */
1705 res_size = sba_dev->ioc[i].pdir_size/sizeof(u64); /* entries */
1706
1707 /* Second part of PIRANHA BUG */
1708 if (piranha_bad_128k) {
1709 res_size -= (128*1024)/sizeof(u64);
1710 }
1711
1712 res_size >>= 3; /* convert bit count to byte count */
1713 DBG_INIT("%s() res_size 0x%x\n",
1714 __func__, res_size);
1715
1716 sba_dev->ioc[i].res_size = res_size;
1717 sba_dev->ioc[i].res_map = (char *) __get_free_pages(GFP_KERNEL, get_order(res_size));
1718
1719 #ifdef DEBUG_DMB_TRAP
1720 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1721 set_data_memory_break, 0);
1722 #endif
1723
1724 if (NULL == sba_dev->ioc[i].res_map)
1725 {
1726 panic("%s:%s() could not allocate resource map\n",
1727 __FILE__, __func__ );
1728 }
1729
1730 memset(sba_dev->ioc[i].res_map, 0, res_size);
1731 /* next available IOVP - circular search */
1732 sba_dev->ioc[i].res_hint = (unsigned long *)
1733 &(sba_dev->ioc[i].res_map[L1_CACHE_BYTES]);
1734
1735 #ifdef ASSERT_PDIR_SANITY
1736 /* Mark first bit busy - ie no IOVA 0 */
1737 sba_dev->ioc[i].res_map[0] = 0x80;
1738 sba_dev->ioc[i].pdir_base[0] = (__force __le64) 0xeeffc0addbba0080ULL;
1739 #endif
1740
1741 /* Third (and last) part of PIRANHA BUG */
1742 if (piranha_bad_128k) {
1743 /* region from +1408K to +1536 is un-usable. */
1744
1745 int idx_start = (1408*1024/sizeof(u64)) >> 3;
1746 int idx_end = (1536*1024/sizeof(u64)) >> 3;
1747 long *p_start = (long *) &(sba_dev->ioc[i].res_map[idx_start]);
1748 long *p_end = (long *) &(sba_dev->ioc[i].res_map[idx_end]);
1749
1750 /* mark that part of the io pdir busy */
1751 while (p_start < p_end)
1752 *p_start++ = -1;
1753
1754 }
1755
1756 #ifdef DEBUG_DMB_TRAP
1757 iterate_pages( sba_dev->ioc[i].res_map, res_size,
1758 set_data_memory_break, 0);
1759 iterate_pages( sba_dev->ioc[i].pdir_base, sba_dev->ioc[i].pdir_size,
1760 set_data_memory_break, 0);
1761 #endif
1762
1763 DBG_INIT("%s() %d res_map %x %p\n",
1764 __func__, i, res_size, sba_dev->ioc[i].res_map);
1765 }
1766
1767 spin_lock_init(&sba_dev->sba_lock);
1768 ioc_needs_fdc = boot_cpu_data.pdc.capabilities & PDC_MODEL_IOPDIR_FDC;
1769
1770 #ifdef DEBUG_SBA_INIT
1771 /*
1772 * If the PDC_MODEL capabilities has Non-coherent IO-PDIR bit set
1773 * (bit #61, big endian), we have to flush and sync every time
1774 * IO-PDIR is changed in Ike/Astro.
1775 */
1776 if (ioc_needs_fdc) {
1777 printk(KERN_INFO MODULE_NAME " FDC/SYNC required.\n");
1778 } else {
1779 printk(KERN_INFO MODULE_NAME " IOC has cache coherent PDIR.\n");
1780 }
1781 #endif
1782 }
1783
1784 #ifdef CONFIG_PROC_FS
sba_proc_info(struct seq_file * m,void * p)1785 static int sba_proc_info(struct seq_file *m, void *p)
1786 {
1787 struct sba_device *sba_dev = sba_list;
1788 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
1789 int total_pages = (int) (ioc->res_size << 3); /* 8 bits per byte */
1790 #ifdef SBA_COLLECT_STATS
1791 unsigned long avg = 0, min, max;
1792 #endif
1793 int i;
1794
1795 seq_printf(m, "%s rev %d.%d\n",
1796 sba_dev->name,
1797 (sba_dev->hw_rev & 0x7) + 1,
1798 (sba_dev->hw_rev & 0x18) >> 3);
1799 seq_printf(m, "IO PDIR size : %d bytes (%d entries)\n",
1800 (int)((ioc->res_size << 3) * sizeof(u64)), /* 8 bits/byte */
1801 total_pages);
1802
1803 seq_printf(m, "Resource bitmap : %d bytes (%d pages)\n",
1804 ioc->res_size, ioc->res_size << 3); /* 8 bits per byte */
1805
1806 seq_printf(m, "LMMIO_BASE/MASK/ROUTE %08x %08x %08x\n",
1807 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_BASE),
1808 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_MASK),
1809 READ_REG32(sba_dev->sba_hpa + LMMIO_DIST_ROUTE));
1810
1811 for (i=0; i<4; i++)
1812 seq_printf(m, "DIR%d_BASE/MASK/ROUTE %08x %08x %08x\n",
1813 i,
1814 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_BASE + i*0x18),
1815 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_MASK + i*0x18),
1816 READ_REG32(sba_dev->sba_hpa + LMMIO_DIRECT0_ROUTE + i*0x18));
1817
1818 #ifdef SBA_COLLECT_STATS
1819 seq_printf(m, "IO PDIR entries : %ld free %ld used (%d%%)\n",
1820 total_pages - ioc->used_pages, ioc->used_pages,
1821 (int)(ioc->used_pages * 100 / total_pages));
1822
1823 min = max = ioc->avg_search[0];
1824 for (i = 0; i < SBA_SEARCH_SAMPLE; i++) {
1825 avg += ioc->avg_search[i];
1826 if (ioc->avg_search[i] > max) max = ioc->avg_search[i];
1827 if (ioc->avg_search[i] < min) min = ioc->avg_search[i];
1828 }
1829 avg /= SBA_SEARCH_SAMPLE;
1830 seq_printf(m, " Bitmap search : %ld/%ld/%ld (min/avg/max CPU Cycles)\n",
1831 min, avg, max);
1832
1833 seq_printf(m, "pci_map_single(): %12ld calls %12ld pages (avg %d/1000)\n",
1834 ioc->msingle_calls, ioc->msingle_pages,
1835 (int)((ioc->msingle_pages * 1000)/ioc->msingle_calls));
1836
1837 /* KLUGE - unmap_sg calls unmap_single for each mapped page */
1838 min = ioc->usingle_calls;
1839 max = ioc->usingle_pages - ioc->usg_pages;
1840 seq_printf(m, "pci_unmap_single: %12ld calls %12ld pages (avg %d/1000)\n",
1841 min, max, (int)((max * 1000)/min));
1842
1843 seq_printf(m, "pci_map_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1844 ioc->msg_calls, ioc->msg_pages,
1845 (int)((ioc->msg_pages * 1000)/ioc->msg_calls));
1846
1847 seq_printf(m, "pci_unmap_sg() : %12ld calls %12ld pages (avg %d/1000)\n",
1848 ioc->usg_calls, ioc->usg_pages,
1849 (int)((ioc->usg_pages * 1000)/ioc->usg_calls));
1850 #endif
1851
1852 return 0;
1853 }
1854
1855 static int
sba_proc_bitmap_info(struct seq_file * m,void * p)1856 sba_proc_bitmap_info(struct seq_file *m, void *p)
1857 {
1858 struct sba_device *sba_dev = sba_list;
1859 struct ioc *ioc = &sba_dev->ioc[0]; /* FIXME: Multi-IOC support! */
1860
1861 seq_hex_dump(m, " ", DUMP_PREFIX_NONE, 32, 4, ioc->res_map,
1862 ioc->res_size, false);
1863 seq_putc(m, '\n');
1864
1865 return 0;
1866 }
1867 #endif /* CONFIG_PROC_FS */
1868
1869 static const struct parisc_device_id sba_tbl[] __initconst = {
1870 { HPHW_IOA, HVERSION_REV_ANY_ID, ASTRO_RUNWAY_PORT, 0xb },
1871 { HPHW_BCPORT, HVERSION_REV_ANY_ID, IKE_MERCED_PORT, 0xc },
1872 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REO_MERCED_PORT, 0xc },
1873 { HPHW_BCPORT, HVERSION_REV_ANY_ID, REOG_MERCED_PORT, 0xc },
1874 { HPHW_IOA, HVERSION_REV_ANY_ID, PLUTO_MCKINLEY_PORT, 0xc },
1875 { 0, }
1876 };
1877
1878 static int sba_driver_callback(struct parisc_device *);
1879
1880 static struct parisc_driver sba_driver __refdata = {
1881 .name = MODULE_NAME,
1882 .id_table = sba_tbl,
1883 .probe = sba_driver_callback,
1884 };
1885
1886 /*
1887 ** Determine if sba should claim this chip (return 0) or not (return 1).
1888 ** If so, initialize the chip and tell other partners in crime they
1889 ** have work to do.
1890 */
sba_driver_callback(struct parisc_device * dev)1891 static int __init sba_driver_callback(struct parisc_device *dev)
1892 {
1893 struct sba_device *sba_dev;
1894 u32 func_class;
1895 int i;
1896 char *version;
1897 void __iomem *sba_addr = ioremap(dev->hpa.start, SBA_FUNC_SIZE);
1898 struct proc_dir_entry *root __maybe_unused;
1899
1900 sba_dump_ranges(sba_addr);
1901
1902 /* Read HW Rev First */
1903 func_class = READ_REG(sba_addr + SBA_FCLASS);
1904
1905 if (IS_ASTRO(dev)) {
1906 unsigned long fclass;
1907 static char astro_rev[]="Astro ?.?";
1908
1909 /* Astro is broken...Read HW Rev First */
1910 fclass = READ_REG(sba_addr);
1911
1912 astro_rev[6] = '1' + (char) (fclass & 0x7);
1913 astro_rev[8] = '0' + (char) ((fclass & 0x18) >> 3);
1914 version = astro_rev;
1915
1916 } else if (IS_IKE(dev)) {
1917 static char ike_rev[] = "Ike rev ?";
1918 ike_rev[8] = '0' + (char) (func_class & 0xff);
1919 version = ike_rev;
1920 } else if (IS_PLUTO(dev)) {
1921 static char pluto_rev[]="Pluto ?.?";
1922 pluto_rev[6] = '0' + (char) ((func_class & 0xf0) >> 4);
1923 pluto_rev[8] = '0' + (char) (func_class & 0x0f);
1924 version = pluto_rev;
1925 } else {
1926 static char reo_rev[] = "REO rev ?";
1927 reo_rev[8] = '0' + (char) (func_class & 0xff);
1928 version = reo_rev;
1929 }
1930
1931 if (!global_ioc_cnt) {
1932 global_ioc_cnt = count_parisc_driver(&sba_driver);
1933
1934 /* Astro and Pluto have one IOC per SBA */
1935 if ((!IS_ASTRO(dev)) || (!IS_PLUTO(dev)))
1936 global_ioc_cnt *= 2;
1937 }
1938
1939 printk(KERN_INFO "%s found %s at 0x%llx\n",
1940 MODULE_NAME, version, (unsigned long long)dev->hpa.start);
1941
1942 sba_dev = kzalloc(sizeof(struct sba_device), GFP_KERNEL);
1943 if (!sba_dev) {
1944 printk(KERN_ERR MODULE_NAME " - couldn't alloc sba_device\n");
1945 return -ENOMEM;
1946 }
1947
1948 parisc_set_drvdata(dev, sba_dev);
1949
1950 for(i=0; i<MAX_IOC; i++)
1951 spin_lock_init(&(sba_dev->ioc[i].res_lock));
1952
1953 sba_dev->dev = dev;
1954 sba_dev->hw_rev = func_class;
1955 sba_dev->name = dev->name;
1956 sba_dev->sba_hpa = sba_addr;
1957
1958 sba_get_pat_resources(sba_dev);
1959 sba_hw_init(sba_dev);
1960 sba_common_init(sba_dev);
1961
1962 hppa_dma_ops = &sba_ops;
1963
1964 switch (dev->id.hversion) {
1965 case PLUTO_MCKINLEY_PORT:
1966 if (!proc_mckinley_root)
1967 proc_mckinley_root = proc_mkdir("bus/mckinley", NULL);
1968 root = proc_mckinley_root;
1969 break;
1970 case ASTRO_RUNWAY_PORT:
1971 case IKE_MERCED_PORT:
1972 default:
1973 if (!proc_runway_root)
1974 proc_runway_root = proc_mkdir("bus/runway", NULL);
1975 root = proc_runway_root;
1976 break;
1977 }
1978
1979 proc_create_single("sba_iommu", 0, root, sba_proc_info);
1980 proc_create_single("sba_iommu-bitmap", 0, root, sba_proc_bitmap_info);
1981 return 0;
1982 }
1983
1984 /*
1985 ** One time initialization to let the world know the SBA was found.
1986 ** This is the only routine which is NOT static.
1987 ** Must be called exactly once before pci_init().
1988 */
sba_init(void)1989 static int __init sba_init(void)
1990 {
1991 return register_parisc_driver(&sba_driver);
1992 }
1993 arch_initcall(sba_init);
1994
1995
1996 /**
1997 * sba_get_iommu - Assign the iommu pointer for the pci bus controller.
1998 * @pci_hba: The parisc device.
1999 *
2000 * Returns the appropriate IOMMU data for the given parisc PCI controller.
2001 * This is cached and used later for PCI DMA Mapping.
2002 */
sba_get_iommu(struct parisc_device * pci_hba)2003 void * sba_get_iommu(struct parisc_device *pci_hba)
2004 {
2005 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2006 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2007 char t = sba_dev->id.hw_type;
2008 int iocnum = (pci_hba->hw_path >> 3); /* IOC # */
2009
2010 WARN_ON((t != HPHW_IOA) && (t != HPHW_BCPORT));
2011
2012 return &(sba->ioc[iocnum]);
2013 }
2014
2015
2016 /**
2017 * sba_directed_lmmio - return first directed LMMIO range routed to rope
2018 * @pci_hba: The parisc device.
2019 * @r: resource PCI host controller wants start/end fields assigned.
2020 *
2021 * For the given parisc PCI controller, determine if any direct ranges
2022 * are routed down the corresponding rope.
2023 */
sba_directed_lmmio(struct parisc_device * pci_hba,struct resource * r)2024 void sba_directed_lmmio(struct parisc_device *pci_hba, struct resource *r)
2025 {
2026 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2027 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2028 char t = sba_dev->id.hw_type;
2029 int i;
2030 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2031
2032 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2033
2034 r->start = r->end = 0;
2035
2036 /* Astro has 4 directed ranges. Not sure about Ike/Pluto/et al */
2037 for (i=0; i<4; i++) {
2038 int base, size;
2039 void __iomem *reg = sba->sba_hpa + i*0x18;
2040
2041 base = READ_REG32(reg + LMMIO_DIRECT0_BASE);
2042 if ((base & 1) == 0)
2043 continue; /* not enabled */
2044
2045 size = READ_REG32(reg + LMMIO_DIRECT0_ROUTE);
2046
2047 if ((size & (ROPES_PER_IOC-1)) != rope)
2048 continue; /* directed down different rope */
2049
2050 r->start = (base & ~1UL) | PCI_F_EXTEND;
2051 size = ~ READ_REG32(reg + LMMIO_DIRECT0_MASK);
2052 r->end = r->start + size;
2053 r->flags = IORESOURCE_MEM;
2054 }
2055 }
2056
2057
2058 /**
2059 * sba_distributed_lmmio - return portion of distributed LMMIO range
2060 * @pci_hba: The parisc device.
2061 * @r: resource PCI host controller wants start/end fields assigned.
2062 *
2063 * For the given parisc PCI controller, return portion of distributed LMMIO
2064 * range. The distributed LMMIO is always present and it's just a question
2065 * of the base address and size of the range.
2066 */
sba_distributed_lmmio(struct parisc_device * pci_hba,struct resource * r)2067 void sba_distributed_lmmio(struct parisc_device *pci_hba, struct resource *r )
2068 {
2069 struct parisc_device *sba_dev = parisc_parent(pci_hba);
2070 struct sba_device *sba = dev_get_drvdata(&sba_dev->dev);
2071 char t = sba_dev->id.hw_type;
2072 int base, size;
2073 int rope = (pci_hba->hw_path & (ROPES_PER_IOC-1)); /* rope # */
2074
2075 BUG_ON((t!=HPHW_IOA) && (t!=HPHW_BCPORT));
2076
2077 r->start = r->end = 0;
2078
2079 base = READ_REG32(sba->sba_hpa + LMMIO_DIST_BASE);
2080 if ((base & 1) == 0) {
2081 BUG(); /* Gah! Distr Range wasn't enabled! */
2082 return;
2083 }
2084
2085 r->start = (base & ~1UL) | PCI_F_EXTEND;
2086
2087 size = (~READ_REG32(sba->sba_hpa + LMMIO_DIST_MASK)) / ROPES_PER_IOC;
2088 r->start += rope * (size + 1); /* adjust base for this rope */
2089 r->end = r->start + size;
2090 r->flags = IORESOURCE_MEM;
2091 }
2092