xref: /openbmc/linux/drivers/char/agp/amd-k7-agp.c (revision 384740dc)
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 #ifndef CONFIG_X86
45 	SetPageReserved(virt_to_page(page_map->real));
46 	global_cache_flush();
47 	page_map->remapped = ioremap_nocache(virt_to_gart(page_map->real),
48 					    PAGE_SIZE);
49 	if (page_map->remapped == NULL) {
50 		ClearPageReserved(virt_to_page(page_map->real));
51 		free_page((unsigned long) page_map->real);
52 		page_map->real = NULL;
53 		return -ENOMEM;
54 	}
55 	global_cache_flush();
56 #else
57 	set_memory_uc((unsigned long)page_map->real, 1);
58 	page_map->remapped = page_map->real;
59 #endif
60 
61 	for (i = 0; i < PAGE_SIZE / sizeof(unsigned long); i++) {
62 		writel(agp_bridge->scratch_page, page_map->remapped+i);
63 		readl(page_map->remapped+i);	/* PCI Posting. */
64 	}
65 
66 	return 0;
67 }
68 
69 static void amd_free_page_map(struct amd_page_map *page_map)
70 {
71 #ifndef CONFIG_X86
72 	iounmap(page_map->remapped);
73 	ClearPageReserved(virt_to_page(page_map->real));
74 #else
75 	set_memory_wb((unsigned long)page_map->real, 1);
76 #endif
77 	free_page((unsigned long) page_map->real);
78 }
79 
80 static void amd_free_gatt_pages(void)
81 {
82 	int i;
83 	struct amd_page_map **tables;
84 	struct amd_page_map *entry;
85 
86 	tables = amd_irongate_private.gatt_pages;
87 	for (i = 0; i < amd_irongate_private.num_tables; i++) {
88 		entry = tables[i];
89 		if (entry != NULL) {
90 			if (entry->real != NULL)
91 				amd_free_page_map(entry);
92 			kfree(entry);
93 		}
94 	}
95 	kfree(tables);
96 	amd_irongate_private.gatt_pages = NULL;
97 }
98 
99 static int amd_create_gatt_pages(int nr_tables)
100 {
101 	struct amd_page_map **tables;
102 	struct amd_page_map *entry;
103 	int retval = 0;
104 	int i;
105 
106 	tables = kzalloc((nr_tables + 1) * sizeof(struct amd_page_map *),GFP_KERNEL);
107 	if (tables == NULL)
108 		return -ENOMEM;
109 
110 	for (i = 0; i < nr_tables; i++) {
111 		entry = kzalloc(sizeof(struct amd_page_map), GFP_KERNEL);
112 		tables[i] = entry;
113 		if (entry == NULL) {
114 			retval = -ENOMEM;
115 			break;
116 		}
117 		retval = amd_create_page_map(entry);
118 		if (retval != 0)
119 			break;
120 	}
121 	amd_irongate_private.num_tables = i;
122 	amd_irongate_private.gatt_pages = tables;
123 
124 	if (retval != 0)
125 		amd_free_gatt_pages();
126 
127 	return retval;
128 }
129 
130 /* Since we don't need contiguous memory we just try
131  * to get the gatt table once
132  */
133 
134 #define GET_PAGE_DIR_OFF(addr) (addr >> 22)
135 #define GET_PAGE_DIR_IDX(addr) (GET_PAGE_DIR_OFF(addr) - \
136 	GET_PAGE_DIR_OFF(agp_bridge->gart_bus_addr))
137 #define GET_GATT_OFF(addr) ((addr & 0x003ff000) >> 12)
138 #define GET_GATT(addr) (amd_irongate_private.gatt_pages[\
139 	GET_PAGE_DIR_IDX(addr)]->remapped)
140 
141 static int amd_create_gatt_table(struct agp_bridge_data *bridge)
142 {
143 	struct aper_size_info_lvl2 *value;
144 	struct amd_page_map page_dir;
145 	unsigned long addr;
146 	int retval;
147 	u32 temp;
148 	int i;
149 
150 	value = A_SIZE_LVL2(agp_bridge->current_size);
151 	retval = amd_create_page_map(&page_dir);
152 	if (retval != 0)
153 		return retval;
154 
155 	retval = amd_create_gatt_pages(value->num_entries / 1024);
156 	if (retval != 0) {
157 		amd_free_page_map(&page_dir);
158 		return retval;
159 	}
160 
161 	agp_bridge->gatt_table_real = (u32 *)page_dir.real;
162 	agp_bridge->gatt_table = (u32 __iomem *)page_dir.remapped;
163 	agp_bridge->gatt_bus_addr = virt_to_gart(page_dir.real);
164 
165 	/* Get the address for the gart region.
166 	 * This is a bus address even on the alpha, b/c its
167 	 * used to program the agp master not the cpu
168 	 */
169 
170 	pci_read_config_dword(agp_bridge->dev, AGP_APBASE, &temp);
171 	addr = (temp & PCI_BASE_ADDRESS_MEM_MASK);
172 	agp_bridge->gart_bus_addr = addr;
173 
174 	/* Calculate the agp offset */
175 	for (i = 0; i < value->num_entries / 1024; i++, addr += 0x00400000) {
176 		writel(virt_to_gart(amd_irongate_private.gatt_pages[i]->real) | 1,
177 			page_dir.remapped+GET_PAGE_DIR_OFF(addr));
178 		readl(page_dir.remapped+GET_PAGE_DIR_OFF(addr));	/* PCI Posting. */
179 	}
180 
181 	return 0;
182 }
183 
184 static int amd_free_gatt_table(struct agp_bridge_data *bridge)
185 {
186 	struct amd_page_map page_dir;
187 
188 	page_dir.real = (unsigned long *)agp_bridge->gatt_table_real;
189 	page_dir.remapped = (unsigned long __iomem *)agp_bridge->gatt_table;
190 
191 	amd_free_gatt_pages();
192 	amd_free_page_map(&page_dir);
193 	return 0;
194 }
195 
196 static int amd_irongate_fetch_size(void)
197 {
198 	int i;
199 	u32 temp;
200 	struct aper_size_info_lvl2 *values;
201 
202 	pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
203 	temp = (temp & 0x0000000e);
204 	values = A_SIZE_LVL2(agp_bridge->driver->aperture_sizes);
205 	for (i = 0; i < agp_bridge->driver->num_aperture_sizes; i++) {
206 		if (temp == values[i].size_value) {
207 			agp_bridge->previous_size =
208 			    agp_bridge->current_size = (void *) (values + i);
209 
210 			agp_bridge->aperture_size_idx = i;
211 			return values[i].size;
212 		}
213 	}
214 
215 	return 0;
216 }
217 
218 static int amd_irongate_configure(void)
219 {
220 	struct aper_size_info_lvl2 *current_size;
221 	u32 temp;
222 	u16 enable_reg;
223 
224 	current_size = A_SIZE_LVL2(agp_bridge->current_size);
225 
226 	/* Get the memory mapped registers */
227 	pci_read_config_dword(agp_bridge->dev, AMD_MMBASE, &temp);
228 	temp = (temp & PCI_BASE_ADDRESS_MEM_MASK);
229 	amd_irongate_private.registers = (volatile u8 __iomem *) ioremap(temp, 4096);
230 	if (!amd_irongate_private.registers)
231 		return -ENOMEM;
232 
233 	/* Write out the address of the gatt table */
234 	writel(agp_bridge->gatt_bus_addr, amd_irongate_private.registers+AMD_ATTBASE);
235 	readl(amd_irongate_private.registers+AMD_ATTBASE);	/* PCI Posting. */
236 
237 	/* Write the Sync register */
238 	pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL, 0x80);
239 
240 	/* Set indexing mode */
241 	pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL2, 0x00);
242 
243 	/* Write the enable register */
244 	enable_reg = readw(amd_irongate_private.registers+AMD_GARTENABLE);
245 	enable_reg = (enable_reg | 0x0004);
246 	writew(enable_reg, amd_irongate_private.registers+AMD_GARTENABLE);
247 	readw(amd_irongate_private.registers+AMD_GARTENABLE);	/* PCI Posting. */
248 
249 	/* Write out the size register */
250 	pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
251 	temp = (((temp & ~(0x0000000e)) | current_size->size_value) | 1);
252 	pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);
253 
254 	/* Flush the tlb */
255 	writel(1, amd_irongate_private.registers+AMD_TLBFLUSH);
256 	readl(amd_irongate_private.registers+AMD_TLBFLUSH);	/* PCI Posting.*/
257 	return 0;
258 }
259 
260 static void amd_irongate_cleanup(void)
261 {
262 	struct aper_size_info_lvl2 *previous_size;
263 	u32 temp;
264 	u16 enable_reg;
265 
266 	previous_size = A_SIZE_LVL2(agp_bridge->previous_size);
267 
268 	enable_reg = readw(amd_irongate_private.registers+AMD_GARTENABLE);
269 	enable_reg = (enable_reg & ~(0x0004));
270 	writew(enable_reg, amd_irongate_private.registers+AMD_GARTENABLE);
271 	readw(amd_irongate_private.registers+AMD_GARTENABLE);	/* PCI Posting. */
272 
273 	/* Write back the previous size and disable gart translation */
274 	pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
275 	temp = ((temp & ~(0x0000000f)) | previous_size->size_value);
276 	pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);
277 	iounmap((void __iomem *) amd_irongate_private.registers);
278 }
279 
280 /*
281  * This routine could be implemented by taking the addresses
282  * written to the GATT, and flushing them individually.  However
283  * currently it just flushes the whole table.  Which is probably
284  * more efficent, since agp_memory blocks can be a large number of
285  * entries.
286  */
287 
288 static void amd_irongate_tlbflush(struct agp_memory *temp)
289 {
290 	writel(1, amd_irongate_private.registers+AMD_TLBFLUSH);
291 	readl(amd_irongate_private.registers+AMD_TLBFLUSH);	/* PCI Posting. */
292 }
293 
294 static int amd_insert_memory(struct agp_memory *mem, off_t pg_start, int type)
295 {
296 	int i, j, num_entries;
297 	unsigned long __iomem *cur_gatt;
298 	unsigned long addr;
299 
300 	num_entries = A_SIZE_LVL2(agp_bridge->current_size)->num_entries;
301 
302 	if (type != 0 || mem->type != 0)
303 		return -EINVAL;
304 
305 	if ((pg_start + mem->page_count) > num_entries)
306 		return -EINVAL;
307 
308 	j = pg_start;
309 	while (j < (pg_start + mem->page_count)) {
310 		addr = (j * PAGE_SIZE) + agp_bridge->gart_bus_addr;
311 		cur_gatt = GET_GATT(addr);
312 		if (!PGE_EMPTY(agp_bridge, readl(cur_gatt+GET_GATT_OFF(addr))))
313 			return -EBUSY;
314 		j++;
315 	}
316 
317 	if (!mem->is_flushed) {
318 		global_cache_flush();
319 		mem->is_flushed = true;
320 	}
321 
322 	for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
323 		addr = (j * PAGE_SIZE) + agp_bridge->gart_bus_addr;
324 		cur_gatt = GET_GATT(addr);
325 		writel(agp_generic_mask_memory(agp_bridge,
326 			mem->memory[i], mem->type), cur_gatt+GET_GATT_OFF(addr));
327 		readl(cur_gatt+GET_GATT_OFF(addr));	/* PCI Posting. */
328 	}
329 	amd_irongate_tlbflush(mem);
330 	return 0;
331 }
332 
333 static int amd_remove_memory(struct agp_memory *mem, off_t pg_start, int type)
334 {
335 	int i;
336 	unsigned long __iomem *cur_gatt;
337 	unsigned long addr;
338 
339 	if (type != 0 || mem->type != 0)
340 		return -EINVAL;
341 
342 	for (i = pg_start; i < (mem->page_count + pg_start); i++) {
343 		addr = (i * PAGE_SIZE) + agp_bridge->gart_bus_addr;
344 		cur_gatt = GET_GATT(addr);
345 		writel(agp_bridge->scratch_page, cur_gatt+GET_GATT_OFF(addr));
346 		readl(cur_gatt+GET_GATT_OFF(addr));	/* PCI Posting. */
347 	}
348 
349 	amd_irongate_tlbflush(mem);
350 	return 0;
351 }
352 
353 static const struct aper_size_info_lvl2 amd_irongate_sizes[7] =
354 {
355 	{2048, 524288, 0x0000000c},
356 	{1024, 262144, 0x0000000a},
357 	{512, 131072, 0x00000008},
358 	{256, 65536, 0x00000006},
359 	{128, 32768, 0x00000004},
360 	{64, 16384, 0x00000002},
361 	{32, 8192, 0x00000000}
362 };
363 
364 static const struct gatt_mask amd_irongate_masks[] =
365 {
366 	{.mask = 1, .type = 0}
367 };
368 
369 static const struct agp_bridge_driver amd_irongate_driver = {
370 	.owner			= THIS_MODULE,
371 	.aperture_sizes		= amd_irongate_sizes,
372 	.size_type		= LVL2_APER_SIZE,
373 	.num_aperture_sizes	= 7,
374 	.configure		= amd_irongate_configure,
375 	.fetch_size		= amd_irongate_fetch_size,
376 	.cleanup		= amd_irongate_cleanup,
377 	.tlb_flush		= amd_irongate_tlbflush,
378 	.mask_memory		= agp_generic_mask_memory,
379 	.masks			= amd_irongate_masks,
380 	.agp_enable		= agp_generic_enable,
381 	.cache_flush		= global_cache_flush,
382 	.create_gatt_table	= amd_create_gatt_table,
383 	.free_gatt_table	= amd_free_gatt_table,
384 	.insert_memory		= amd_insert_memory,
385 	.remove_memory		= amd_remove_memory,
386 	.alloc_by_type		= agp_generic_alloc_by_type,
387 	.free_by_type		= agp_generic_free_by_type,
388 	.agp_alloc_page		= agp_generic_alloc_page,
389 	.agp_destroy_page	= agp_generic_destroy_page,
390 	.agp_type_to_mask_type  = agp_generic_type_to_mask_type,
391 };
392 
393 static struct agp_device_ids amd_agp_device_ids[] __devinitdata =
394 {
395 	{
396 		.device_id	= PCI_DEVICE_ID_AMD_FE_GATE_7006,
397 		.chipset_name	= "Irongate",
398 	},
399 	{
400 		.device_id	= PCI_DEVICE_ID_AMD_FE_GATE_700E,
401 		.chipset_name	= "761",
402 	},
403 	{
404 		.device_id	= PCI_DEVICE_ID_AMD_FE_GATE_700C,
405 		.chipset_name	= "760MP",
406 	},
407 	{ }, /* dummy final entry, always present */
408 };
409 
410 static int __devinit agp_amdk7_probe(struct pci_dev *pdev,
411 				     const struct pci_device_id *ent)
412 {
413 	struct agp_bridge_data *bridge;
414 	u8 cap_ptr;
415 	int j;
416 
417 	cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP);
418 	if (!cap_ptr)
419 		return -ENODEV;
420 
421 	j = ent - agp_amdk7_pci_table;
422 	dev_info(&pdev->dev, "AMD %s chipset\n",
423 		 amd_agp_device_ids[j].chipset_name);
424 
425 	bridge = agp_alloc_bridge();
426 	if (!bridge)
427 		return -ENOMEM;
428 
429 	bridge->driver = &amd_irongate_driver;
430 	bridge->dev_private_data = &amd_irongate_private,
431 	bridge->dev = pdev;
432 	bridge->capndx = cap_ptr;
433 
434 	/* 751 Errata (22564_B-1.PDF)
435 	   erratum 20: strobe glitch with Nvidia NV10 GeForce cards.
436 	   system controller may experience noise due to strong drive strengths
437 	 */
438 	if (agp_bridge->dev->device == PCI_DEVICE_ID_AMD_FE_GATE_7006) {
439 		struct pci_dev *gfxcard=NULL;
440 
441 		cap_ptr = 0;
442 		while (!cap_ptr) {
443 			gfxcard = pci_get_class(PCI_CLASS_DISPLAY_VGA<<8, gfxcard);
444 			if (!gfxcard) {
445 				dev_info(&pdev->dev, "no AGP VGA controller\n");
446 				return -ENODEV;
447 			}
448 			cap_ptr = pci_find_capability(gfxcard, PCI_CAP_ID_AGP);
449 		}
450 
451 		/* With so many variants of NVidia cards, it's simpler just
452 		   to blacklist them all, and then whitelist them as needed
453 		   (if necessary at all). */
454 		if (gfxcard->vendor == PCI_VENDOR_ID_NVIDIA) {
455 			agp_bridge->flags |= AGP_ERRATA_1X;
456 			dev_info(&pdev->dev, "AMD 751 chipset with NVidia GeForce; forcing 1X due to errata\n");
457 		}
458 		pci_dev_put(gfxcard);
459 	}
460 
461 	/* 761 Errata (23613_F.pdf)
462 	 * Revisions B0/B1 were a disaster.
463 	 * erratum 44: SYSCLK/AGPCLK skew causes 2X failures -- Force mode to 1X
464 	 * erratum 45: Timing problem prevents fast writes -- Disable fast write.
465 	 * erratum 46: Setup violation on AGP SBA pins - Disable side band addressing.
466 	 * With this lot disabled, we should prevent lockups. */
467 	if (agp_bridge->dev->device == PCI_DEVICE_ID_AMD_FE_GATE_700E) {
468 		if (pdev->revision == 0x10 || pdev->revision == 0x11) {
469 			agp_bridge->flags = AGP_ERRATA_FASTWRITES;
470 			agp_bridge->flags |= AGP_ERRATA_SBA;
471 			agp_bridge->flags |= AGP_ERRATA_1X;
472 			dev_info(&pdev->dev, "AMD 761 chipset with errata; disabling AGP fast writes & SBA and forcing to 1X\n");
473 		}
474 	}
475 
476 	/* Fill in the mode register */
477 	pci_read_config_dword(pdev,
478 			bridge->capndx+PCI_AGP_STATUS,
479 			&bridge->mode);
480 
481 	pci_set_drvdata(pdev, bridge);
482 	return agp_add_bridge(bridge);
483 }
484 
485 static void __devexit agp_amdk7_remove(struct pci_dev *pdev)
486 {
487 	struct agp_bridge_data *bridge = pci_get_drvdata(pdev);
488 
489 	agp_remove_bridge(bridge);
490 	agp_put_bridge(bridge);
491 }
492 
493 /* must be the same order as name table above */
494 static struct pci_device_id agp_amdk7_pci_table[] = {
495 	{
496 	.class		= (PCI_CLASS_BRIDGE_HOST << 8),
497 	.class_mask	= ~0,
498 	.vendor		= PCI_VENDOR_ID_AMD,
499 	.device		= PCI_DEVICE_ID_AMD_FE_GATE_7006,
500 	.subvendor	= PCI_ANY_ID,
501 	.subdevice	= PCI_ANY_ID,
502 	},
503 	{
504 	.class		= (PCI_CLASS_BRIDGE_HOST << 8),
505 	.class_mask	= ~0,
506 	.vendor		= PCI_VENDOR_ID_AMD,
507 	.device		= PCI_DEVICE_ID_AMD_FE_GATE_700E,
508 	.subvendor	= PCI_ANY_ID,
509 	.subdevice	= PCI_ANY_ID,
510 	},
511 	{
512 	.class		= (PCI_CLASS_BRIDGE_HOST << 8),
513 	.class_mask	= ~0,
514 	.vendor		= PCI_VENDOR_ID_AMD,
515 	.device		= PCI_DEVICE_ID_AMD_FE_GATE_700C,
516 	.subvendor	= PCI_ANY_ID,
517 	.subdevice	= PCI_ANY_ID,
518 	},
519 	{ }
520 };
521 
522 MODULE_DEVICE_TABLE(pci, agp_amdk7_pci_table);
523 
524 static struct pci_driver agp_amdk7_pci_driver = {
525 	.name		= "agpgart-amdk7",
526 	.id_table	= agp_amdk7_pci_table,
527 	.probe		= agp_amdk7_probe,
528 	.remove		= agp_amdk7_remove,
529 };
530 
531 static int __init agp_amdk7_init(void)
532 {
533 	if (agp_off)
534 		return -EINVAL;
535 	return pci_register_driver(&agp_amdk7_pci_driver);
536 }
537 
538 static void __exit agp_amdk7_cleanup(void)
539 {
540 	pci_unregister_driver(&agp_amdk7_pci_driver);
541 }
542 
543 module_init(agp_amdk7_init);
544 module_exit(agp_amdk7_cleanup);
545 
546 MODULE_LICENSE("GPL and additional rights");
547