Merge with rsync://rsync.kernel.org/pub/scm/linux/kernel/git/sfrench/cifs-2.6.git
[linux-2.6] / drivers / char / agp / amd-k7-agp.c
1 /*
2  * AMD K7 AGPGART routines.
3  */
4
5 #include <linux/module.h>
6 #include <linux/pci.h>
7 #include <linux/init.h>
8 #include <linux/agp_backend.h>
9 #include <linux/gfp.h>
10 #include <linux/page-flags.h>
11 #include <linux/mm.h>
12 #include "agp.h"
13
14 #define AMD_MMBASE      0x14
15 #define AMD_APSIZE      0xac
16 #define AMD_MODECNTL    0xb0
17 #define AMD_MODECNTL2   0xb2
18 #define AMD_GARTENABLE  0x02    /* In mmio region (16-bit register) */
19 #define AMD_ATTBASE     0x04    /* In mmio region (32-bit register) */
20 #define AMD_TLBFLUSH    0x0c    /* In mmio region (32-bit register) */
21 #define AMD_CACHEENTRY  0x10    /* In mmio region (32-bit register) */
22
23 static struct pci_device_id agp_amdk7_pci_table[];
24
25 struct amd_page_map {
26         unsigned long *real;
27         unsigned long __iomem *remapped;
28 };
29
30 static struct _amd_irongate_private {
31         volatile u8 __iomem *registers;
32         struct amd_page_map **gatt_pages;
33         int num_tables;
34 } amd_irongate_private;
35
36 static int amd_create_page_map(struct amd_page_map *page_map)
37 {
38         int i;
39
40         page_map->real = (unsigned long *) __get_free_page(GFP_KERNEL);
41         if (page_map->real == NULL)
42                 return -ENOMEM;
43
44         SetPageReserved(virt_to_page(page_map->real));
45         global_cache_flush();
46         page_map->remapped = ioremap_nocache(virt_to_gart(page_map->real),
47                                             PAGE_SIZE);
48         if (page_map->remapped == NULL) {
49                 ClearPageReserved(virt_to_page(page_map->real));
50                 free_page((unsigned long) page_map->real);
51                 page_map->real = NULL;
52                 return -ENOMEM;
53         }
54         global_cache_flush();
55
56         for (i = 0; i < PAGE_SIZE / sizeof(unsigned long); i++) {
57                 writel(agp_bridge->scratch_page, page_map->remapped+i);
58                 readl(page_map->remapped+i);    /* PCI Posting. */
59         }
60
61         return 0;
62 }
63
64 static void amd_free_page_map(struct amd_page_map *page_map)
65 {
66         iounmap(page_map->remapped);
67         ClearPageReserved(virt_to_page(page_map->real));
68         free_page((unsigned long) page_map->real);
69 }
70
71 static void amd_free_gatt_pages(void)
72 {
73         int i;
74         struct amd_page_map **tables;
75         struct amd_page_map *entry;
76
77         tables = amd_irongate_private.gatt_pages;
78         for (i = 0; i < amd_irongate_private.num_tables; i++) {
79                 entry = tables[i];
80                 if (entry != NULL) {
81                         if (entry->real != NULL)
82                                 amd_free_page_map(entry);
83                         kfree(entry);
84                 }
85         }
86         kfree(tables);
87         amd_irongate_private.gatt_pages = NULL;
88 }
89
90 static int amd_create_gatt_pages(int nr_tables)
91 {
92         struct amd_page_map **tables;
93         struct amd_page_map *entry;
94         int retval = 0;
95         int i;
96
97         tables = kmalloc((nr_tables + 1) * sizeof(struct amd_page_map *),
98                          GFP_KERNEL);
99         if (tables == NULL)
100                 return -ENOMEM;
101
102         memset (tables, 0, sizeof(struct amd_page_map *) * (nr_tables + 1));
103         for (i = 0; i < nr_tables; i++) {
104                 entry = kmalloc(sizeof(struct amd_page_map), GFP_KERNEL);
105                 if (entry == NULL) {
106                         retval = -ENOMEM;
107                         break;
108                 }
109                 memset (entry, 0, sizeof(struct amd_page_map));
110                 tables[i] = entry;
111                 retval = amd_create_page_map(entry);
112                 if (retval != 0)
113                         break;
114         }
115         amd_irongate_private.num_tables = nr_tables;
116         amd_irongate_private.gatt_pages = tables;
117
118         if (retval != 0)
119                 amd_free_gatt_pages();
120
121         return retval;
122 }
123
124 /* Since we don't need contigious memory we just try
125  * to get the gatt table once
126  */
127
128 #define GET_PAGE_DIR_OFF(addr) (addr >> 22)
129 #define GET_PAGE_DIR_IDX(addr) (GET_PAGE_DIR_OFF(addr) - \
130         GET_PAGE_DIR_OFF(agp_bridge->gart_bus_addr))
131 #define GET_GATT_OFF(addr) ((addr & 0x003ff000) >> 12)
132 #define GET_GATT(addr) (amd_irongate_private.gatt_pages[\
133         GET_PAGE_DIR_IDX(addr)]->remapped)
134
135 static int amd_create_gatt_table(struct agp_bridge_data *bridge)
136 {
137         struct aper_size_info_lvl2 *value;
138         struct amd_page_map page_dir;
139         unsigned long addr;
140         int retval;
141         u32 temp;
142         int i;
143
144         value = A_SIZE_LVL2(agp_bridge->current_size);
145         retval = amd_create_page_map(&page_dir);
146         if (retval != 0)
147                 return retval;
148
149         retval = amd_create_gatt_pages(value->num_entries / 1024);
150         if (retval != 0) {
151                 amd_free_page_map(&page_dir);
152                 return retval;
153         }
154
155         agp_bridge->gatt_table_real = (u32 *)page_dir.real;
156         agp_bridge->gatt_table = (u32 __iomem *)page_dir.remapped;
157         agp_bridge->gatt_bus_addr = virt_to_gart(page_dir.real);
158
159         /* Get the address for the gart region.
160          * This is a bus address even on the alpha, b/c its
161          * used to program the agp master not the cpu
162          */
163
164         pci_read_config_dword(agp_bridge->dev, AGP_APBASE, &temp);
165         addr = (temp & PCI_BASE_ADDRESS_MEM_MASK);
166         agp_bridge->gart_bus_addr = addr;
167
168         /* Calculate the agp offset */
169         for (i = 0; i < value->num_entries / 1024; i++, addr += 0x00400000) {
170                 writel(virt_to_gart(amd_irongate_private.gatt_pages[i]->real) | 1,
171                         page_dir.remapped+GET_PAGE_DIR_OFF(addr));
172                 readl(page_dir.remapped+GET_PAGE_DIR_OFF(addr));        /* PCI Posting. */
173         }
174
175         return 0;
176 }
177
178 static int amd_free_gatt_table(struct agp_bridge_data *bridge)
179 {
180         struct amd_page_map page_dir;
181
182         page_dir.real = (unsigned long *)agp_bridge->gatt_table_real;
183         page_dir.remapped = (unsigned long __iomem *)agp_bridge->gatt_table;
184
185         amd_free_gatt_pages();
186         amd_free_page_map(&page_dir);
187         return 0;
188 }
189
190 static int amd_irongate_fetch_size(void)
191 {
192         int i;
193         u32 temp;
194         struct aper_size_info_lvl2 *values;
195
196         pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
197         temp = (temp & 0x0000000e);
198         values = A_SIZE_LVL2(agp_bridge->driver->aperture_sizes);
199         for (i = 0; i < agp_bridge->driver->num_aperture_sizes; i++) {
200                 if (temp == values[i].size_value) {
201                         agp_bridge->previous_size =
202                             agp_bridge->current_size = (void *) (values + i);
203
204                         agp_bridge->aperture_size_idx = i;
205                         return values[i].size;
206                 }
207         }
208
209         return 0;
210 }
211
212 static int amd_irongate_configure(void)
213 {
214         struct aper_size_info_lvl2 *current_size;
215         u32 temp;
216         u16 enable_reg;
217
218         current_size = A_SIZE_LVL2(agp_bridge->current_size);
219
220         /* Get the memory mapped registers */
221         pci_read_config_dword(agp_bridge->dev, AMD_MMBASE, &temp);
222         temp = (temp & PCI_BASE_ADDRESS_MEM_MASK);
223         amd_irongate_private.registers = (volatile u8 __iomem *) ioremap(temp, 4096);
224
225         /* Write out the address of the gatt table */
226         writel(agp_bridge->gatt_bus_addr, amd_irongate_private.registers+AMD_ATTBASE);
227         readl(amd_irongate_private.registers+AMD_ATTBASE);      /* PCI Posting. */
228
229         /* Write the Sync register */
230         pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL, 0x80);
231
232         /* Set indexing mode */
233         pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL2, 0x00);
234
235         /* Write the enable register */
236         enable_reg = readw(amd_irongate_private.registers+AMD_GARTENABLE);
237         enable_reg = (enable_reg | 0x0004);
238         writew(enable_reg, amd_irongate_private.registers+AMD_GARTENABLE);
239         readw(amd_irongate_private.registers+AMD_GARTENABLE);   /* PCI Posting. */
240
241         /* Write out the size register */
242         pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
243         temp = (((temp & ~(0x0000000e)) | current_size->size_value) | 1);
244         pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);
245
246         /* Flush the tlb */
247         writel(1, amd_irongate_private.registers+AMD_TLBFLUSH);
248         readl(amd_irongate_private.registers+AMD_TLBFLUSH);     /* PCI Posting.*/
249         return 0;
250 }
251
252 static void amd_irongate_cleanup(void)
253 {
254         struct aper_size_info_lvl2 *previous_size;
255         u32 temp;
256         u16 enable_reg;
257
258         previous_size = A_SIZE_LVL2(agp_bridge->previous_size);
259
260         enable_reg = readw(amd_irongate_private.registers+AMD_GARTENABLE);
261         enable_reg = (enable_reg & ~(0x0004));
262         writew(enable_reg, amd_irongate_private.registers+AMD_GARTENABLE);
263         readw(amd_irongate_private.registers+AMD_GARTENABLE);   /* PCI Posting. */
264
265         /* Write back the previous size and disable gart translation */
266         pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
267         temp = ((temp & ~(0x0000000f)) | previous_size->size_value);
268         pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);
269         iounmap((void __iomem *) amd_irongate_private.registers);
270 }
271
272 /*
273  * This routine could be implemented by taking the addresses
274  * written to the GATT, and flushing them individually.  However
275  * currently it just flushes the whole table.  Which is probably
276  * more efficent, since agp_memory blocks can be a large number of
277  * entries.
278  */
279
280 static void amd_irongate_tlbflush(struct agp_memory *temp)
281 {
282         writel(1, amd_irongate_private.registers+AMD_TLBFLUSH);
283         readl(amd_irongate_private.registers+AMD_TLBFLUSH);     /* PCI Posting. */
284 }
285
286 static int amd_insert_memory(struct agp_memory *mem, off_t pg_start, int type)
287 {
288         int i, j, num_entries;
289         unsigned long __iomem *cur_gatt;
290         unsigned long addr;
291
292         num_entries = A_SIZE_LVL2(agp_bridge->current_size)->num_entries;
293
294         if (type != 0 || mem->type != 0)
295                 return -EINVAL;
296
297         if ((pg_start + mem->page_count) > num_entries)
298                 return -EINVAL;
299
300         j = pg_start;
301         while (j < (pg_start + mem->page_count)) {
302                 addr = (j * PAGE_SIZE) + agp_bridge->gart_bus_addr;
303                 cur_gatt = GET_GATT(addr);
304                 if (!PGE_EMPTY(agp_bridge, readl(cur_gatt+GET_GATT_OFF(addr))))
305                         return -EBUSY;
306                 j++;
307         }
308
309         if (mem->is_flushed == FALSE) {
310                 global_cache_flush();
311                 mem->is_flushed = TRUE;
312         }
313
314         for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
315                 addr = (j * PAGE_SIZE) + agp_bridge->gart_bus_addr;
316                 cur_gatt = GET_GATT(addr);
317                 writel(agp_generic_mask_memory(agp_bridge,
318                         mem->memory[i], mem->type), cur_gatt+GET_GATT_OFF(addr));
319                 readl(cur_gatt+GET_GATT_OFF(addr));     /* PCI Posting. */
320         }
321         amd_irongate_tlbflush(mem);
322         return 0;
323 }
324
325 static int amd_remove_memory(struct agp_memory *mem, off_t pg_start, int type)
326 {
327         int i;
328         unsigned long __iomem *cur_gatt;
329         unsigned long addr;
330
331         if (type != 0 || mem->type != 0)
332                 return -EINVAL;
333
334         for (i = pg_start; i < (mem->page_count + pg_start); i++) {
335                 addr = (i * PAGE_SIZE) + agp_bridge->gart_bus_addr;
336                 cur_gatt = GET_GATT(addr);
337                 writel(agp_bridge->scratch_page, cur_gatt+GET_GATT_OFF(addr));
338                 readl(cur_gatt+GET_GATT_OFF(addr));     /* PCI Posting. */
339         }
340
341         amd_irongate_tlbflush(mem);
342         return 0;
343 }
344
345 static struct aper_size_info_lvl2 amd_irongate_sizes[7] =
346 {
347         {2048, 524288, 0x0000000c},
348         {1024, 262144, 0x0000000a},
349         {512, 131072, 0x00000008},
350         {256, 65536, 0x00000006},
351         {128, 32768, 0x00000004},
352         {64, 16384, 0x00000002},
353         {32, 8192, 0x00000000}
354 };
355
356 static struct gatt_mask amd_irongate_masks[] =
357 {
358         {.mask = 1, .type = 0}
359 };
360
361 static struct agp_bridge_driver amd_irongate_driver = {
362         .owner                  = THIS_MODULE,
363         .aperture_sizes         = amd_irongate_sizes,
364         .size_type              = LVL2_APER_SIZE,
365         .num_aperture_sizes     = 7,
366         .configure              = amd_irongate_configure,
367         .fetch_size             = amd_irongate_fetch_size,
368         .cleanup                = amd_irongate_cleanup,
369         .tlb_flush              = amd_irongate_tlbflush,
370         .mask_memory            = agp_generic_mask_memory,
371         .masks                  = amd_irongate_masks,
372         .agp_enable             = agp_generic_enable,
373         .cache_flush            = global_cache_flush,
374         .create_gatt_table      = amd_create_gatt_table,
375         .free_gatt_table        = amd_free_gatt_table,
376         .insert_memory          = amd_insert_memory,
377         .remove_memory          = amd_remove_memory,
378         .alloc_by_type          = agp_generic_alloc_by_type,
379         .free_by_type           = agp_generic_free_by_type,
380         .agp_alloc_page         = agp_generic_alloc_page,
381         .agp_destroy_page       = agp_generic_destroy_page,
382 };
383
384 static struct agp_device_ids amd_agp_device_ids[] __devinitdata =
385 {
386         {
387                 .device_id      = PCI_DEVICE_ID_AMD_FE_GATE_7006,
388                 .chipset_name   = "Irongate",
389         },
390         {
391                 .device_id      = PCI_DEVICE_ID_AMD_FE_GATE_700E,
392                 .chipset_name   = "761",
393         },
394         {
395                 .device_id      = PCI_DEVICE_ID_AMD_FE_GATE_700C,
396                 .chipset_name   = "760MP",
397         },
398         { }, /* dummy final entry, always present */
399 };
400
401 static int __devinit agp_amdk7_probe(struct pci_dev *pdev,
402                                      const struct pci_device_id *ent)
403 {
404         struct agp_bridge_data *bridge;
405         u8 cap_ptr;
406         int j;
407
408         cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP);
409         if (!cap_ptr)
410                 return -ENODEV;
411
412         j = ent - agp_amdk7_pci_table;
413         printk(KERN_INFO PFX "Detected AMD %s chipset\n",
414                amd_agp_device_ids[j].chipset_name);
415
416         bridge = agp_alloc_bridge();
417         if (!bridge)
418                 return -ENOMEM;
419
420         bridge->driver = &amd_irongate_driver;
421         bridge->dev_private_data = &amd_irongate_private,
422         bridge->dev = pdev;
423         bridge->capndx = cap_ptr;
424
425         /* 751 Errata (22564_B-1.PDF)
426            erratum 20: strobe glitch with Nvidia NV10 GeForce cards.
427            system controller may experience noise due to strong drive strengths
428          */
429         if (agp_bridge->dev->device == PCI_DEVICE_ID_AMD_FE_GATE_7006) {
430                 u8 cap_ptr=0;
431                 struct pci_dev *gfxcard=NULL;
432                 while (!cap_ptr) {
433                         gfxcard = pci_get_class(PCI_CLASS_DISPLAY_VGA<<8, gfxcard);
434                         if (!gfxcard) {
435                                 printk (KERN_INFO PFX "Couldn't find an AGP VGA controller.\n");
436                                 return -ENODEV;
437                         }
438                         cap_ptr = pci_find_capability(gfxcard, PCI_CAP_ID_AGP);
439                         if (!cap_ptr) {
440                                 pci_dev_put(gfxcard);
441                                 continue;
442                         }
443                 }
444
445                 /* With so many variants of NVidia cards, it's simpler just
446                    to blacklist them all, and then whitelist them as needed
447                    (if necessary at all). */
448                 if (gfxcard->vendor == PCI_VENDOR_ID_NVIDIA) {
449                         agp_bridge->flags |= AGP_ERRATA_1X;
450                         printk (KERN_INFO PFX "AMD 751 chipset with NVidia GeForce detected. Forcing to 1X due to errata.\n");
451                 }
452                 pci_dev_put(gfxcard);
453         }
454
455         /* 761 Errata (23613_F.pdf)
456          * Revisions B0/B1 were a disaster.
457          * erratum 44: SYSCLK/AGPCLK skew causes 2X failures -- Force mode to 1X
458          * erratum 45: Timing problem prevents fast writes -- Disable fast write.
459          * erratum 46: Setup violation on AGP SBA pins - Disable side band addressing.
460          * With this lot disabled, we should prevent lockups. */
461         if (agp_bridge->dev->device == PCI_DEVICE_ID_AMD_FE_GATE_700E) {
462                 u8 revision=0;
463                 pci_read_config_byte(pdev, PCI_REVISION_ID, &revision);
464                 if (revision == 0x10 || revision == 0x11) {
465                         agp_bridge->flags = AGP_ERRATA_FASTWRITES;
466                         agp_bridge->flags |= AGP_ERRATA_SBA;
467                         agp_bridge->flags |= AGP_ERRATA_1X;
468                         printk (KERN_INFO PFX "AMD 761 chipset with errata detected - disabling AGP fast writes & SBA and forcing to 1X.\n");
469                 }
470         }
471
472         /* Fill in the mode register */
473         pci_read_config_dword(pdev,
474                         bridge->capndx+PCI_AGP_STATUS,
475                         &bridge->mode);
476
477         pci_set_drvdata(pdev, bridge);
478         return agp_add_bridge(bridge);
479 }
480
481 static void __devexit agp_amdk7_remove(struct pci_dev *pdev)
482 {
483         struct agp_bridge_data *bridge = pci_get_drvdata(pdev);
484
485         agp_remove_bridge(bridge);
486         agp_put_bridge(bridge);
487 }
488
489 /* must be the same order as name table above */
490 static struct pci_device_id agp_amdk7_pci_table[] = {
491         {
492         .class          = (PCI_CLASS_BRIDGE_HOST << 8),
493         .class_mask     = ~0,
494         .vendor         = PCI_VENDOR_ID_AMD,
495         .device         = PCI_DEVICE_ID_AMD_FE_GATE_7006,
496         .subvendor      = PCI_ANY_ID,
497         .subdevice      = PCI_ANY_ID,
498         },
499         {
500         .class          = (PCI_CLASS_BRIDGE_HOST << 8),
501         .class_mask     = ~0,
502         .vendor         = PCI_VENDOR_ID_AMD,
503         .device         = PCI_DEVICE_ID_AMD_FE_GATE_700E,
504         .subvendor      = PCI_ANY_ID,
505         .subdevice      = PCI_ANY_ID,
506         },
507         {
508         .class          = (PCI_CLASS_BRIDGE_HOST << 8),
509         .class_mask     = ~0,
510         .vendor         = PCI_VENDOR_ID_AMD,
511         .device         = PCI_DEVICE_ID_AMD_FE_GATE_700C,
512         .subvendor      = PCI_ANY_ID,
513         .subdevice      = PCI_ANY_ID,
514         },
515         { }
516 };
517
518 MODULE_DEVICE_TABLE(pci, agp_amdk7_pci_table);
519
520 static struct pci_driver agp_amdk7_pci_driver = {
521         .name           = "agpgart-amdk7",
522         .id_table       = agp_amdk7_pci_table,
523         .probe          = agp_amdk7_probe,
524         .remove         = agp_amdk7_remove,
525 };
526
527 static int __init agp_amdk7_init(void)
528 {
529         if (agp_off)
530                 return -EINVAL;
531         return pci_register_driver(&agp_amdk7_pci_driver);
532 }
533
534 static void __exit agp_amdk7_cleanup(void)
535 {
536         pci_unregister_driver(&agp_amdk7_pci_driver);
537 }
538
539 module_init(agp_amdk7_init);
540 module_exit(agp_amdk7_cleanup);
541
542 MODULE_LICENSE("GPL and additional rights");