Merge rsync://rsync.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / arch / x86_64 / kernel / pci-calgary.c
1 /*
2  * Derived from arch/powerpc/kernel/iommu.c
3  *
4  * Copyright (C) IBM Corporation, 2006
5  *
6  * Author: Jon Mason <jdmason@us.ibm.com>
7  * Author: Muli Ben-Yehuda <muli@il.ibm.com>
8
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation; either version 2 of the License, or
12  * (at your option) any later version.
13  *
14  * This program is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  * GNU General Public License for more details.
18  *
19  * You should have received a copy of the GNU General Public License
20  * along with this program; if not, write to the Free Software
21  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
22  */
23
24 #include <linux/config.h>
25 #include <linux/kernel.h>
26 #include <linux/init.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/mm.h>
30 #include <linux/spinlock.h>
31 #include <linux/string.h>
32 #include <linux/dma-mapping.h>
33 #include <linux/init.h>
34 #include <linux/bitops.h>
35 #include <linux/pci_ids.h>
36 #include <linux/pci.h>
37 #include <linux/delay.h>
38 #include <asm/proto.h>
39 #include <asm/calgary.h>
40 #include <asm/tce.h>
41 #include <asm/pci-direct.h>
42 #include <asm/system.h>
43 #include <asm/dma.h>
44
45 #define PCI_DEVICE_ID_IBM_CALGARY 0x02a1
46 #define PCI_VENDOR_DEVICE_ID_CALGARY \
47         (PCI_VENDOR_ID_IBM | PCI_DEVICE_ID_IBM_CALGARY << 16)
48
49 /* we need these for register space address calculation */
50 #define START_ADDRESS           0xfe000000
51 #define CHASSIS_BASE            0
52 #define ONE_BASED_CHASSIS_NUM   1
53
54 /* register offsets inside the host bridge space */
55 #define PHB_CSR_OFFSET          0x0110
56 #define PHB_PLSSR_OFFSET        0x0120
57 #define PHB_CONFIG_RW_OFFSET    0x0160
58 #define PHB_IOBASE_BAR_LOW      0x0170
59 #define PHB_IOBASE_BAR_HIGH     0x0180
60 #define PHB_MEM_1_LOW           0x0190
61 #define PHB_MEM_1_HIGH          0x01A0
62 #define PHB_IO_ADDR_SIZE        0x01B0
63 #define PHB_MEM_1_SIZE          0x01C0
64 #define PHB_MEM_ST_OFFSET       0x01D0
65 #define PHB_AER_OFFSET          0x0200
66 #define PHB_CONFIG_0_HIGH       0x0220
67 #define PHB_CONFIG_0_LOW        0x0230
68 #define PHB_CONFIG_0_END        0x0240
69 #define PHB_MEM_2_LOW           0x02B0
70 #define PHB_MEM_2_HIGH          0x02C0
71 #define PHB_MEM_2_SIZE_HIGH     0x02D0
72 #define PHB_MEM_2_SIZE_LOW      0x02E0
73 #define PHB_DOSHOLE_OFFSET      0x08E0
74
75 /* PHB_CONFIG_RW */
76 #define PHB_TCE_ENABLE          0x20000000
77 #define PHB_SLOT_DISABLE        0x1C000000
78 #define PHB_DAC_DISABLE         0x01000000
79 #define PHB_MEM2_ENABLE         0x00400000
80 #define PHB_MCSR_ENABLE         0x00100000
81 /* TAR (Table Address Register) */
82 #define TAR_SW_BITS             0x0000ffffffff800fUL
83 #define TAR_VALID               0x0000000000000008UL
84 /* CSR (Channel/DMA Status Register) */
85 #define CSR_AGENT_MASK          0xffe0ffff
86
87 #define MAX_NUM_OF_PHBS         8 /* how many PHBs in total? */
88 #define MAX_NUM_CHASSIS         8 /* max number of chassis */
89 #define MAX_PHB_BUS_NUM         (MAX_NUM_OF_PHBS * MAX_NUM_CHASSIS * 2) /* max dev->bus->number */
90 #define PHBS_PER_CALGARY        4
91
92 /* register offsets in Calgary's internal register space */
93 static const unsigned long tar_offsets[] = {
94         0x0580 /* TAR0 */,
95         0x0588 /* TAR1 */,
96         0x0590 /* TAR2 */,
97         0x0598 /* TAR3 */
98 };
99
100 static const unsigned long split_queue_offsets[] = {
101         0x4870 /* SPLIT QUEUE 0 */,
102         0x5870 /* SPLIT QUEUE 1 */,
103         0x6870 /* SPLIT QUEUE 2 */,
104         0x7870 /* SPLIT QUEUE 3 */
105 };
106
107 static const unsigned long phb_offsets[] = {
108         0x8000 /* PHB0 */,
109         0x9000 /* PHB1 */,
110         0xA000 /* PHB2 */,
111         0xB000 /* PHB3 */
112 };
113
114 static char bus_to_phb[MAX_PHB_BUS_NUM];
115 void* tce_table_kva[MAX_PHB_BUS_NUM];
116 unsigned int specified_table_size = TCE_TABLE_SIZE_UNSPECIFIED;
117 static int translate_empty_slots __read_mostly = 0;
118 static int calgary_detected __read_mostly = 0;
119
120 /*
121  * the bitmap of PHBs the user requested that we disable
122  * translation on.
123  */
124 static DECLARE_BITMAP(translation_disabled, MAX_PHB_BUS_NUM);
125
126 static void tce_cache_blast(struct iommu_table *tbl);
127
128 /* enable this to stress test the chip's TCE cache */
129 #ifdef CONFIG_IOMMU_DEBUG
130 static inline void tce_cache_blast_stress(struct iommu_table *tbl)
131 {
132         tce_cache_blast(tbl);
133 }
134 #else
135 static inline void tce_cache_blast_stress(struct iommu_table *tbl)
136 {
137 }
138 #endif /* BLAST_TCE_CACHE_ON_UNMAP */
139
140 static inline unsigned int num_dma_pages(unsigned long dma, unsigned int dmalen)
141 {
142         unsigned int npages;
143
144         npages = PAGE_ALIGN(dma + dmalen) - (dma & PAGE_MASK);
145         npages >>= PAGE_SHIFT;
146
147         return npages;
148 }
149
150 static inline int translate_phb(struct pci_dev* dev)
151 {
152         int disabled = test_bit(dev->bus->number, translation_disabled);
153         return !disabled;
154 }
155
156 static void iommu_range_reserve(struct iommu_table *tbl,
157         unsigned long start_addr, unsigned int npages)
158 {
159         unsigned long index;
160         unsigned long end;
161
162         index = start_addr >> PAGE_SHIFT;
163
164         /* bail out if we're asked to reserve a region we don't cover */
165         if (index >= tbl->it_size)
166                 return;
167
168         end = index + npages;
169         if (end > tbl->it_size) /* don't go off the table */
170                 end = tbl->it_size;
171
172         while (index < end) {
173                 if (test_bit(index, tbl->it_map))
174                         printk(KERN_ERR "Calgary: entry already allocated at "
175                                "0x%lx tbl %p dma 0x%lx npages %u\n",
176                                index, tbl, start_addr, npages);
177                 ++index;
178         }
179         set_bit_string(tbl->it_map, start_addr >> PAGE_SHIFT, npages);
180 }
181
182 static unsigned long iommu_range_alloc(struct iommu_table *tbl,
183         unsigned int npages)
184 {
185         unsigned long offset;
186
187         BUG_ON(npages == 0);
188
189         offset = find_next_zero_string(tbl->it_map, tbl->it_hint,
190                                        tbl->it_size, npages);
191         if (offset == ~0UL) {
192                 tce_cache_blast(tbl);
193                 offset = find_next_zero_string(tbl->it_map, 0,
194                                                tbl->it_size, npages);
195                 if (offset == ~0UL) {
196                         printk(KERN_WARNING "Calgary: IOMMU full.\n");
197                         if (panic_on_overflow)
198                                 panic("Calgary: fix the allocator.\n");
199                         else
200                                 return bad_dma_address;
201                 }
202         }
203
204         set_bit_string(tbl->it_map, offset, npages);
205         tbl->it_hint = offset + npages;
206         BUG_ON(tbl->it_hint > tbl->it_size);
207
208         return offset;
209 }
210
211 static dma_addr_t iommu_alloc(struct iommu_table *tbl, void *vaddr,
212         unsigned int npages, int direction)
213 {
214         unsigned long entry, flags;
215         dma_addr_t ret = bad_dma_address;
216
217         spin_lock_irqsave(&tbl->it_lock, flags);
218
219         entry = iommu_range_alloc(tbl, npages);
220
221         if (unlikely(entry == bad_dma_address))
222                 goto error;
223
224         /* set the return dma address */
225         ret = (entry << PAGE_SHIFT) | ((unsigned long)vaddr & ~PAGE_MASK);
226
227         /* put the TCEs in the HW table */
228         tce_build(tbl, entry, npages, (unsigned long)vaddr & PAGE_MASK,
229                   direction);
230
231         spin_unlock_irqrestore(&tbl->it_lock, flags);
232
233         return ret;
234
235 error:
236         spin_unlock_irqrestore(&tbl->it_lock, flags);
237         printk(KERN_WARNING "Calgary: failed to allocate %u pages in "
238                "iommu %p\n", npages, tbl);
239         return bad_dma_address;
240 }
241
242 static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
243         unsigned int npages)
244 {
245         unsigned long entry;
246         unsigned long i;
247
248         entry = dma_addr >> PAGE_SHIFT;
249
250         BUG_ON(entry + npages > tbl->it_size);
251
252         tce_free(tbl, entry, npages);
253
254         for (i = 0; i < npages; ++i) {
255                 if (!test_bit(entry + i, tbl->it_map))
256                         printk(KERN_ERR "Calgary: bit is off at 0x%lx "
257                                "tbl %p dma 0x%Lx entry 0x%lx npages %u\n",
258                                entry + i, tbl, dma_addr, entry, npages);
259         }
260
261         __clear_bit_string(tbl->it_map, entry, npages);
262
263         tce_cache_blast_stress(tbl);
264 }
265
266 static void iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
267         unsigned int npages)
268 {
269         unsigned long flags;
270
271         spin_lock_irqsave(&tbl->it_lock, flags);
272
273         __iommu_free(tbl, dma_addr, npages);
274
275         spin_unlock_irqrestore(&tbl->it_lock, flags);
276 }
277
278 static void __calgary_unmap_sg(struct iommu_table *tbl,
279         struct scatterlist *sglist, int nelems, int direction)
280 {
281         while (nelems--) {
282                 unsigned int npages;
283                 dma_addr_t dma = sglist->dma_address;
284                 unsigned int dmalen = sglist->dma_length;
285
286                 if (dmalen == 0)
287                         break;
288
289                 npages = num_dma_pages(dma, dmalen);
290                 __iommu_free(tbl, dma, npages);
291                 sglist++;
292         }
293 }
294
295 void calgary_unmap_sg(struct device *dev, struct scatterlist *sglist,
296                       int nelems, int direction)
297 {
298         unsigned long flags;
299         struct iommu_table *tbl = to_pci_dev(dev)->bus->self->sysdata;
300
301         if (!translate_phb(to_pci_dev(dev)))
302                 return;
303
304         spin_lock_irqsave(&tbl->it_lock, flags);
305
306         __calgary_unmap_sg(tbl, sglist, nelems, direction);
307
308         spin_unlock_irqrestore(&tbl->it_lock, flags);
309 }
310
311 static int calgary_nontranslate_map_sg(struct device* dev,
312         struct scatterlist *sg, int nelems, int direction)
313 {
314         int i;
315
316         for (i = 0; i < nelems; i++ ) {
317                 struct scatterlist *s = &sg[i];
318                 BUG_ON(!s->page);
319                 s->dma_address = virt_to_bus(page_address(s->page) +s->offset);
320                 s->dma_length = s->length;
321         }
322         return nelems;
323 }
324
325 int calgary_map_sg(struct device *dev, struct scatterlist *sg,
326         int nelems, int direction)
327 {
328         struct iommu_table *tbl = to_pci_dev(dev)->bus->self->sysdata;
329         unsigned long flags;
330         unsigned long vaddr;
331         unsigned int npages;
332         unsigned long entry;
333         int i;
334
335         if (!translate_phb(to_pci_dev(dev)))
336                 return calgary_nontranslate_map_sg(dev, sg, nelems, direction);
337
338         spin_lock_irqsave(&tbl->it_lock, flags);
339
340         for (i = 0; i < nelems; i++ ) {
341                 struct scatterlist *s = &sg[i];
342                 BUG_ON(!s->page);
343
344                 vaddr = (unsigned long)page_address(s->page) + s->offset;
345                 npages = num_dma_pages(vaddr, s->length);
346
347                 entry = iommu_range_alloc(tbl, npages);
348                 if (entry == bad_dma_address) {
349                         /* makes sure unmap knows to stop */
350                         s->dma_length = 0;
351                         goto error;
352                 }
353
354                 s->dma_address = (entry << PAGE_SHIFT) | s->offset;
355
356                 /* insert into HW table */
357                 tce_build(tbl, entry, npages, vaddr & PAGE_MASK,
358                           direction);
359
360                 s->dma_length = s->length;
361         }
362
363         spin_unlock_irqrestore(&tbl->it_lock, flags);
364
365         return nelems;
366 error:
367         __calgary_unmap_sg(tbl, sg, nelems, direction);
368         for (i = 0; i < nelems; i++) {
369                 sg[i].dma_address = bad_dma_address;
370                 sg[i].dma_length = 0;
371         }
372         spin_unlock_irqrestore(&tbl->it_lock, flags);
373         return 0;
374 }
375
376 dma_addr_t calgary_map_single(struct device *dev, void *vaddr,
377         size_t size, int direction)
378 {
379         dma_addr_t dma_handle = bad_dma_address;
380         unsigned long uaddr;
381         unsigned int npages;
382         struct iommu_table *tbl = to_pci_dev(dev)->bus->self->sysdata;
383
384         uaddr = (unsigned long)vaddr;
385         npages = num_dma_pages(uaddr, size);
386
387         if (translate_phb(to_pci_dev(dev)))
388                 dma_handle = iommu_alloc(tbl, vaddr, npages, direction);
389         else
390                 dma_handle = virt_to_bus(vaddr);
391
392         return dma_handle;
393 }
394
395 void calgary_unmap_single(struct device *dev, dma_addr_t dma_handle,
396         size_t size, int direction)
397 {
398         struct iommu_table *tbl = to_pci_dev(dev)->bus->self->sysdata;
399         unsigned int npages;
400
401         if (!translate_phb(to_pci_dev(dev)))
402                 return;
403
404         npages = num_dma_pages(dma_handle, size);
405         iommu_free(tbl, dma_handle, npages);
406 }
407
408 void* calgary_alloc_coherent(struct device *dev, size_t size,
409         dma_addr_t *dma_handle, gfp_t flag)
410 {
411         void *ret = NULL;
412         dma_addr_t mapping;
413         unsigned int npages, order;
414         struct iommu_table *tbl;
415
416         tbl = to_pci_dev(dev)->bus->self->sysdata;
417
418         size = PAGE_ALIGN(size); /* size rounded up to full pages */
419         npages = size >> PAGE_SHIFT;
420         order = get_order(size);
421
422         /* alloc enough pages (and possibly more) */
423         ret = (void *)__get_free_pages(flag, order);
424         if (!ret)
425                 goto error;
426         memset(ret, 0, size);
427
428         if (translate_phb(to_pci_dev(dev))) {
429                 /* set up tces to cover the allocated range */
430                 mapping = iommu_alloc(tbl, ret, npages, DMA_BIDIRECTIONAL);
431                 if (mapping == bad_dma_address)
432                         goto free;
433
434                 *dma_handle = mapping;
435         } else /* non translated slot */
436                 *dma_handle = virt_to_bus(ret);
437
438         return ret;
439
440 free:
441         free_pages((unsigned long)ret, get_order(size));
442         ret = NULL;
443 error:
444         return ret;
445 }
446
447 static struct dma_mapping_ops calgary_dma_ops = {
448         .alloc_coherent = calgary_alloc_coherent,
449         .map_single = calgary_map_single,
450         .unmap_single = calgary_unmap_single,
451         .map_sg = calgary_map_sg,
452         .unmap_sg = calgary_unmap_sg,
453 };
454
455 static inline int busno_to_phbid(unsigned char num)
456 {
457         return bus_to_phb[num];
458 }
459
460 static inline unsigned long split_queue_offset(unsigned char num)
461 {
462         size_t idx = busno_to_phbid(num);
463
464         return split_queue_offsets[idx];
465 }
466
467 static inline unsigned long tar_offset(unsigned char num)
468 {
469         size_t idx = busno_to_phbid(num);
470
471         return tar_offsets[idx];
472 }
473
474 static inline unsigned long phb_offset(unsigned char num)
475 {
476         size_t idx = busno_to_phbid(num);
477
478         return phb_offsets[idx];
479 }
480
481 static inline void __iomem* calgary_reg(void __iomem *bar, unsigned long offset)
482 {
483         unsigned long target = ((unsigned long)bar) | offset;
484         return (void __iomem*)target;
485 }
486
487 static void tce_cache_blast(struct iommu_table *tbl)
488 {
489         u64 val;
490         u32 aer;
491         int i = 0;
492         void __iomem *bbar = tbl->bbar;
493         void __iomem *target;
494
495         /* disable arbitration on the bus */
496         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_AER_OFFSET);
497         aer = readl(target);
498         writel(0, target);
499
500         /* read plssr to ensure it got there */
501         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_PLSSR_OFFSET);
502         val = readl(target);
503
504         /* poll split queues until all DMA activity is done */
505         target = calgary_reg(bbar, split_queue_offset(tbl->it_busno));
506         do {
507                 val = readq(target);
508                 i++;
509         } while ((val & 0xff) != 0xff && i < 100);
510         if (i == 100)
511                 printk(KERN_WARNING "Calgary: PCI bus not quiesced, "
512                        "continuing anyway\n");
513
514         /* invalidate TCE cache */
515         target = calgary_reg(bbar, tar_offset(tbl->it_busno));
516         writeq(tbl->tar_val, target);
517
518         /* enable arbitration */
519         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_AER_OFFSET);
520         writel(aer, target);
521         (void)readl(target); /* flush */
522 }
523
524 static void __init calgary_reserve_mem_region(struct pci_dev *dev, u64 start,
525         u64 limit)
526 {
527         unsigned int numpages;
528
529         limit = limit | 0xfffff;
530         limit++;
531
532         numpages = ((limit - start) >> PAGE_SHIFT);
533         iommu_range_reserve(dev->sysdata, start, numpages);
534 }
535
536 static void __init calgary_reserve_peripheral_mem_1(struct pci_dev *dev)
537 {
538         void __iomem *target;
539         u64 low, high, sizelow;
540         u64 start, limit;
541         struct iommu_table *tbl = dev->sysdata;
542         unsigned char busnum = dev->bus->number;
543         void __iomem *bbar = tbl->bbar;
544
545         /* peripheral MEM_1 region */
546         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_LOW);
547         low = be32_to_cpu(readl(target));
548         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_HIGH);
549         high = be32_to_cpu(readl(target));
550         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_1_SIZE);
551         sizelow = be32_to_cpu(readl(target));
552
553         start = (high << 32) | low;
554         limit = sizelow;
555
556         calgary_reserve_mem_region(dev, start, limit);
557 }
558
559 static void __init calgary_reserve_peripheral_mem_2(struct pci_dev *dev)
560 {
561         void __iomem *target;
562         u32 val32;
563         u64 low, high, sizelow, sizehigh;
564         u64 start, limit;
565         struct iommu_table *tbl = dev->sysdata;
566         unsigned char busnum = dev->bus->number;
567         void __iomem *bbar = tbl->bbar;
568
569         /* is it enabled? */
570         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
571         val32 = be32_to_cpu(readl(target));
572         if (!(val32 & PHB_MEM2_ENABLE))
573                 return;
574
575         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_LOW);
576         low = be32_to_cpu(readl(target));
577         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_HIGH);
578         high = be32_to_cpu(readl(target));
579         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_SIZE_LOW);
580         sizelow = be32_to_cpu(readl(target));
581         target = calgary_reg(bbar, phb_offset(busnum) | PHB_MEM_2_SIZE_HIGH);
582         sizehigh = be32_to_cpu(readl(target));
583
584         start = (high << 32) | low;
585         limit = (sizehigh << 32) | sizelow;
586
587         calgary_reserve_mem_region(dev, start, limit);
588 }
589
590 /*
591  * some regions of the IO address space do not get translated, so we
592  * must not give devices IO addresses in those regions. The regions
593  * are the 640KB-1MB region and the two PCI peripheral memory holes.
594  * Reserve all of them in the IOMMU bitmap to avoid giving them out
595  * later.
596  */
597 static void __init calgary_reserve_regions(struct pci_dev *dev)
598 {
599         unsigned int npages;
600         void __iomem *bbar;
601         unsigned char busnum;
602         u64 start;
603         struct iommu_table *tbl = dev->sysdata;
604
605         bbar = tbl->bbar;
606         busnum = dev->bus->number;
607
608         /* reserve bad_dma_address in case it's a legal address */
609         iommu_range_reserve(tbl, bad_dma_address, 1);
610
611         /* avoid the BIOS/VGA first 640KB-1MB region */
612         start = (640 * 1024);
613         npages = ((1024 - 640) * 1024) >> PAGE_SHIFT;
614         iommu_range_reserve(tbl, start, npages);
615
616         /* reserve the two PCI peripheral memory regions in IO space */
617         calgary_reserve_peripheral_mem_1(dev);
618         calgary_reserve_peripheral_mem_2(dev);
619 }
620
621 static int __init calgary_setup_tar(struct pci_dev *dev, void __iomem *bbar)
622 {
623         u64 val64;
624         u64 table_phys;
625         void __iomem *target;
626         int ret;
627         struct iommu_table *tbl;
628
629         /* build TCE tables for each PHB */
630         ret = build_tce_table(dev, bbar);
631         if (ret)
632                 return ret;
633
634         calgary_reserve_regions(dev);
635
636         /* set TARs for each PHB */
637         target = calgary_reg(bbar, tar_offset(dev->bus->number));
638         val64 = be64_to_cpu(readq(target));
639
640         /* zero out all TAR bits under sw control */
641         val64 &= ~TAR_SW_BITS;
642
643         tbl = dev->sysdata;
644         table_phys = (u64)__pa(tbl->it_base);
645         val64 |= table_phys;
646
647         BUG_ON(specified_table_size > TCE_TABLE_SIZE_8M);
648         val64 |= (u64) specified_table_size;
649
650         tbl->tar_val = cpu_to_be64(val64);
651         writeq(tbl->tar_val, target);
652         readq(target); /* flush */
653
654         return 0;
655 }
656
657 static void __init calgary_free_tar(struct pci_dev *dev)
658 {
659         u64 val64;
660         struct iommu_table *tbl = dev->sysdata;
661         void __iomem *target;
662
663         target = calgary_reg(tbl->bbar, tar_offset(dev->bus->number));
664         val64 = be64_to_cpu(readq(target));
665         val64 &= ~TAR_SW_BITS;
666         writeq(cpu_to_be64(val64), target);
667         readq(target); /* flush */
668
669         kfree(tbl);
670         dev->sysdata = NULL;
671 }
672
673 static void calgary_watchdog(unsigned long data)
674 {
675         struct pci_dev *dev = (struct pci_dev *)data;
676         struct iommu_table *tbl = dev->sysdata;
677         void __iomem *bbar = tbl->bbar;
678         u32 val32;
679         void __iomem *target;
680
681         target = calgary_reg(bbar, phb_offset(tbl->it_busno) | PHB_CSR_OFFSET);
682         val32 = be32_to_cpu(readl(target));
683
684         /* If no error, the agent ID in the CSR is not valid */
685         if (val32 & CSR_AGENT_MASK) {
686                 printk(KERN_EMERG "calgary_watchdog: DMA error on bus %d, "
687                                   "CSR = %#x\n", dev->bus->number, val32);
688                 writel(0, target);
689
690                 /* Disable bus that caused the error */
691                 target = calgary_reg(bbar, phb_offset(tbl->it_busno) |
692                                            PHB_CONFIG_RW_OFFSET);
693                 val32 = be32_to_cpu(readl(target));
694                 val32 |= PHB_SLOT_DISABLE;
695                 writel(cpu_to_be32(val32), target);
696                 readl(target); /* flush */
697         } else {
698                 /* Reset the timer */
699                 mod_timer(&tbl->watchdog_timer, jiffies + 2 * HZ);
700         }
701 }
702
703 static void __init calgary_enable_translation(struct pci_dev *dev)
704 {
705         u32 val32;
706         unsigned char busnum;
707         void __iomem *target;
708         void __iomem *bbar;
709         struct iommu_table *tbl;
710
711         busnum = dev->bus->number;
712         tbl = dev->sysdata;
713         bbar = tbl->bbar;
714
715         /* enable TCE in PHB Config Register */
716         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
717         val32 = be32_to_cpu(readl(target));
718         val32 |= PHB_TCE_ENABLE | PHB_DAC_DISABLE | PHB_MCSR_ENABLE;
719
720         printk(KERN_INFO "Calgary: enabling translation on PHB %d\n", busnum);
721         printk(KERN_INFO "Calgary: errant DMAs will now be prevented on this "
722                "bus.\n");
723
724         writel(cpu_to_be32(val32), target);
725         readl(target); /* flush */
726
727         init_timer(&tbl->watchdog_timer);
728         tbl->watchdog_timer.function = &calgary_watchdog;
729         tbl->watchdog_timer.data = (unsigned long)dev;
730         mod_timer(&tbl->watchdog_timer, jiffies);
731 }
732
733 static void __init calgary_disable_translation(struct pci_dev *dev)
734 {
735         u32 val32;
736         unsigned char busnum;
737         void __iomem *target;
738         void __iomem *bbar;
739         struct iommu_table *tbl;
740
741         busnum = dev->bus->number;
742         tbl = dev->sysdata;
743         bbar = tbl->bbar;
744
745         /* disable TCE in PHB Config Register */
746         target = calgary_reg(bbar, phb_offset(busnum) | PHB_CONFIG_RW_OFFSET);
747         val32 = be32_to_cpu(readl(target));
748         val32 &= ~(PHB_TCE_ENABLE | PHB_DAC_DISABLE | PHB_MCSR_ENABLE);
749
750         printk(KERN_INFO "Calgary: disabling translation on PHB %d!\n", busnum);
751         writel(cpu_to_be32(val32), target);
752         readl(target); /* flush */
753
754         del_timer_sync(&tbl->watchdog_timer);
755 }
756
757 static inline unsigned int __init locate_register_space(struct pci_dev *dev)
758 {
759         int rionodeid;
760         u32 address;
761
762         rionodeid = (dev->bus->number % 15 > 4) ? 3 : 2;
763         /*
764          * register space address calculation as follows:
765          * FE0MB-8MB*OneBasedChassisNumber+1MB*(RioNodeId-ChassisBase)
766          * ChassisBase is always zero for x366/x260/x460
767          * RioNodeId is 2 for first Calgary, 3 for second Calgary
768          */
769         address = START_ADDRESS -
770                 (0x800000 * (ONE_BASED_CHASSIS_NUM + dev->bus->number / 15)) +
771                 (0x100000) * (rionodeid - CHASSIS_BASE);
772         return address;
773 }
774
775 static int __init calgary_init_one_nontraslated(struct pci_dev *dev)
776 {
777         dev->sysdata = NULL;
778         dev->bus->self = dev;
779
780         return 0;
781 }
782
783 static int __init calgary_init_one(struct pci_dev *dev)
784 {
785         u32 address;
786         void __iomem *bbar;
787         int ret;
788
789         address = locate_register_space(dev);
790         /* map entire 1MB of Calgary config space */
791         bbar = ioremap_nocache(address, 1024 * 1024);
792         if (!bbar) {
793                 ret = -ENODATA;
794                 goto done;
795         }
796
797         ret = calgary_setup_tar(dev, bbar);
798         if (ret)
799                 goto iounmap;
800
801         dev->bus->self = dev;
802         calgary_enable_translation(dev);
803
804         return 0;
805
806 iounmap:
807         iounmap(bbar);
808 done:
809         return ret;
810 }
811
812 static int __init calgary_init(void)
813 {
814         int i, ret = -ENODEV;
815         struct pci_dev *dev = NULL;
816
817         for (i = 0; i < MAX_PHB_BUS_NUM; i++) {
818                 dev = pci_get_device(PCI_VENDOR_ID_IBM,
819                                      PCI_DEVICE_ID_IBM_CALGARY,
820                                      dev);
821                 if (!dev)
822                         break;
823                 if (!translate_phb(dev)) {
824                         calgary_init_one_nontraslated(dev);
825                         continue;
826                 }
827                 if (!tce_table_kva[dev->bus->number] && !translate_empty_slots) {
828                         pci_dev_put(dev);
829                         continue;
830                 }
831                 ret = calgary_init_one(dev);
832                 if (ret)
833                         goto error;
834         }
835
836         return ret;
837
838 error:
839         for (i--; i >= 0; i--) {
840                 dev = pci_find_device_reverse(PCI_VENDOR_ID_IBM,
841                                               PCI_DEVICE_ID_IBM_CALGARY,
842                                               dev);
843                 if (!translate_phb(dev)) {
844                         pci_dev_put(dev);
845                         continue;
846                 }
847                 if (!tce_table_kva[dev->bus->number] && !translate_empty_slots)
848                         continue;
849                 calgary_disable_translation(dev);
850                 calgary_free_tar(dev);
851                 pci_dev_put(dev);
852         }
853
854         return ret;
855 }
856
857 static inline int __init determine_tce_table_size(u64 ram)
858 {
859         int ret;
860
861         if (specified_table_size != TCE_TABLE_SIZE_UNSPECIFIED)
862                 return specified_table_size;
863
864         /*
865          * Table sizes are from 0 to 7 (TCE_TABLE_SIZE_64K to
866          * TCE_TABLE_SIZE_8M). Table size 0 has 8K entries and each
867          * larger table size has twice as many entries, so shift the
868          * max ram address by 13 to divide by 8K and then look at the
869          * order of the result to choose between 0-7.
870          */
871         ret = get_order(ram >> 13);
872         if (ret > TCE_TABLE_SIZE_8M)
873                 ret = TCE_TABLE_SIZE_8M;
874
875         return ret;
876 }
877
878 void __init detect_calgary(void)
879 {
880         u32 val;
881         int bus;
882         void *tbl;
883         int calgary_found = 0;
884         int phb = -1;
885
886         /*
887          * if the user specified iommu=off or iommu=soft or we found
888          * another HW IOMMU already, bail out.
889          */
890         if (swiotlb || no_iommu || iommu_detected)
891                 return;
892
893         specified_table_size = determine_tce_table_size(end_pfn * PAGE_SIZE);
894
895         for (bus = 0; bus < MAX_PHB_BUS_NUM; bus++) {
896                 int dev;
897
898                 tce_table_kva[bus] = NULL;
899                 bus_to_phb[bus] = -1;
900
901                 if (read_pci_config(bus, 0, 0, 0) != PCI_VENDOR_DEVICE_ID_CALGARY)
902                         continue;
903
904                 /*
905                  * There are 4 PHBs per Calgary chip.  Set phb to which phb (0-3)
906                  * it is connected to releative to the clagary chip.
907                  */
908                 phb = (phb + 1) % PHBS_PER_CALGARY;
909
910                 if (test_bit(bus, translation_disabled)) {
911                         printk(KERN_INFO "Calgary: translation is disabled for "
912                                "PHB 0x%x\n", bus);
913                         /* skip this phb, don't allocate a tbl for it */
914                         continue;
915                 }
916                 /*
917                  * Scan the slots of the PCI bus to see if there is a device present.
918                  * The parent bus will be the zero-ith device, so start at 1.
919                  */
920                 for (dev = 1; dev < 8; dev++) {
921                         val = read_pci_config(bus, dev, 0, 0);
922                         if (val != 0xffffffff || translate_empty_slots) {
923                                 tbl = alloc_tce_table();
924                                 if (!tbl)
925                                         goto cleanup;
926                                 tce_table_kva[bus] = tbl;
927                                 bus_to_phb[bus] = phb;
928                                 calgary_found = 1;
929                                 break;
930                         }
931                 }
932         }
933
934         if (calgary_found) {
935                 iommu_detected = 1;
936                 calgary_detected = 1;
937                 printk(KERN_INFO "PCI-DMA: Calgary IOMMU detected. "
938                        "TCE table spec is %d.\n", specified_table_size);
939         }
940         return;
941
942 cleanup:
943         for (--bus; bus >= 0; --bus)
944                 if (tce_table_kva[bus])
945                         free_tce_table(tce_table_kva[bus]);
946 }
947
948 int __init calgary_iommu_init(void)
949 {
950         int ret;
951
952         if (no_iommu || swiotlb)
953                 return -ENODEV;
954
955         if (!calgary_detected)
956                 return -ENODEV;
957
958         /* ok, we're trying to use Calgary - let's roll */
959         printk(KERN_INFO "PCI-DMA: Using Calgary IOMMU\n");
960
961         ret = calgary_init();
962         if (ret) {
963                 printk(KERN_ERR "PCI-DMA: Calgary init failed %d, "
964                        "falling back to no_iommu\n", ret);
965                 if (end_pfn > MAX_DMA32_PFN)
966                         printk(KERN_ERR "WARNING more than 4GB of memory, "
967                                         "32bit PCI may malfunction.\n");
968                 return ret;
969         }
970
971         force_iommu = 1;
972         dma_ops = &calgary_dma_ops;
973
974         return 0;
975 }
976
977 static int __init calgary_parse_options(char *p)
978 {
979         unsigned int bridge;
980         size_t len;
981         char* endp;
982
983         while (*p) {
984                 if (!strncmp(p, "64k", 3))
985                         specified_table_size = TCE_TABLE_SIZE_64K;
986                 else if (!strncmp(p, "128k", 4))
987                         specified_table_size = TCE_TABLE_SIZE_128K;
988                 else if (!strncmp(p, "256k", 4))
989                         specified_table_size = TCE_TABLE_SIZE_256K;
990                 else if (!strncmp(p, "512k", 4))
991                         specified_table_size = TCE_TABLE_SIZE_512K;
992                 else if (!strncmp(p, "1M", 2))
993                         specified_table_size = TCE_TABLE_SIZE_1M;
994                 else if (!strncmp(p, "2M", 2))
995                         specified_table_size = TCE_TABLE_SIZE_2M;
996                 else if (!strncmp(p, "4M", 2))
997                         specified_table_size = TCE_TABLE_SIZE_4M;
998                 else if (!strncmp(p, "8M", 2))
999                         specified_table_size = TCE_TABLE_SIZE_8M;
1000
1001                 len = strlen("translate_empty_slots");
1002                 if (!strncmp(p, "translate_empty_slots", len))
1003                         translate_empty_slots = 1;
1004
1005                 len = strlen("disable");
1006                 if (!strncmp(p, "disable", len)) {
1007                         p += len;
1008                         if (*p == '=')
1009                                 ++p;
1010                         if (*p == '\0')
1011                                 break;
1012                         bridge = simple_strtol(p, &endp, 0);
1013                         if (p == endp)
1014                                 break;
1015
1016                         if (bridge < MAX_PHB_BUS_NUM) {
1017                                 printk(KERN_INFO "Calgary: disabling "
1018                                        "translation for PHB 0x%x\n", bridge);
1019                                 set_bit(bridge, translation_disabled);
1020                         }
1021                 }
1022
1023                 p = strpbrk(p, ",");
1024                 if (!p)
1025                         break;
1026
1027                 p++; /* skip ',' */
1028         }
1029         return 1;
1030 }
1031 __setup("calgary=", calgary_parse_options);