2 * Copyright 2001-2003 SuSE Labs.
3 * Distributed under the GNU public license, v2.
5 * This is a GART driver for the AMD Opteron/Athlon64 on-CPU northbridge.
6 * It also includes support for the AMD 8151 AGP bridge,
7 * although it doesn't actually do much, as all the real
8 * work is done in the northbridge(s).
11 #include <linux/module.h>
12 #include <linux/pci.h>
13 #include <linux/init.h>
14 #include <linux/agp_backend.h>
15 #include <linux/mmzone.h>
16 #include <asm/page.h> /* PAGE_SIZE */
22 /* NVIDIA K8 registers */
23 #define NVIDIA_X86_64_0_APBASE 0x10
24 #define NVIDIA_X86_64_1_APBASE1 0x50
25 #define NVIDIA_X86_64_1_APLIMIT1 0x54
26 #define NVIDIA_X86_64_1_APSIZE 0xa8
27 #define NVIDIA_X86_64_1_APBASE2 0xd8
28 #define NVIDIA_X86_64_1_APLIMIT2 0xdc
30 /* ULi K8 registers */
31 #define ULI_X86_64_BASE_ADDR 0x10
32 #define ULI_X86_64_HTT_FEA_REG 0x50
33 #define ULI_X86_64_ENU_SCR_REG 0x54
35 static struct resource *aperture_resource;
36 static int __initdata agp_try_unsupported = 1;
37 static int agp_bridges_found;
39 static void amd64_tlbflush(struct agp_memory *temp)
44 static int amd64_insert_memory(struct agp_memory *mem, off_t pg_start, int type)
46 int i, j, num_entries;
49 struct agp_bridge_data *bridge = mem->bridge;
52 num_entries = agp_num_entries();
54 if (type != mem->type)
56 mask_type = bridge->driver->agp_type_to_mask_type(bridge, type);
61 /* Make sure we can fit the range in the gatt table. */
62 /* FIXME: could wrap */
63 if (((unsigned long)pg_start + mem->page_count) > num_entries)
68 /* gatt table should be empty. */
69 while (j < (pg_start + mem->page_count)) {
70 if (!PGE_EMPTY(agp_bridge, readl(agp_bridge->gatt_table+j)))
75 if (!mem->is_flushed) {
77 mem->is_flushed = true;
80 for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
81 tmp = agp_bridge->driver->mask_memory(agp_bridge,
82 mem->memory[i], mask_type);
84 BUG_ON(tmp & 0xffffff0000000ffcULL);
85 pte = (tmp & 0x000000ff00000000ULL) >> 28;
86 pte |=(tmp & 0x00000000fffff000ULL);
87 pte |= GPTE_VALID | GPTE_COHERENT;
89 writel(pte, agp_bridge->gatt_table+j);
90 readl(agp_bridge->gatt_table+j); /* PCI Posting. */
97 * This hack alters the order element according
98 * to the size of a long. It sucks. I totally disown this, even
99 * though it does appear to work for the most part.
101 static struct aper_size_info_32 amd64_aperture_sizes[7] =
103 {32, 8192, 3+(sizeof(long)/8), 0 },
104 {64, 16384, 4+(sizeof(long)/8), 1<<1 },
105 {128, 32768, 5+(sizeof(long)/8), 1<<2 },
106 {256, 65536, 6+(sizeof(long)/8), 1<<1 | 1<<2 },
107 {512, 131072, 7+(sizeof(long)/8), 1<<3 },
108 {1024, 262144, 8+(sizeof(long)/8), 1<<1 | 1<<3},
109 {2048, 524288, 9+(sizeof(long)/8), 1<<2 | 1<<3}
114 * Get the current Aperture size from the x86-64.
115 * Note, that there may be multiple x86-64's, but we just return
116 * the value from the first one we find. The set_size functions
117 * keep the rest coherent anyway. Or at least should do.
119 static int amd64_fetch_size(void)
124 struct aper_size_info_32 *values;
126 dev = k8_northbridges[0];
130 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &temp);
132 values = A_SIZE_32(amd64_aperture_sizes);
134 for (i = 0; i < agp_bridge->driver->num_aperture_sizes; i++) {
135 if (temp == values[i].size_value) {
136 agp_bridge->previous_size =
137 agp_bridge->current_size = (void *) (values + i);
139 agp_bridge->aperture_size_idx = i;
140 return values[i].size;
147 * In a multiprocessor x86-64 system, this function gets
148 * called once for each CPU.
150 static u64 amd64_configure(struct pci_dev *hammer, u64 gatt_table)
156 /* Address to map to */
157 pci_read_config_dword(hammer, AMD64_GARTAPERTUREBASE, &tmp);
158 aperturebase = tmp << 25;
159 aper_base = (aperturebase & PCI_BASE_ADDRESS_MEM_MASK);
161 enable_gart_translation(hammer, gatt_table);
167 static const struct aper_size_info_32 amd_8151_sizes[7] =
169 {2048, 524288, 9, 0x00000000 }, /* 0 0 0 0 0 0 */
170 {1024, 262144, 8, 0x00000400 }, /* 1 0 0 0 0 0 */
171 {512, 131072, 7, 0x00000600 }, /* 1 1 0 0 0 0 */
172 {256, 65536, 6, 0x00000700 }, /* 1 1 1 0 0 0 */
173 {128, 32768, 5, 0x00000720 }, /* 1 1 1 1 0 0 */
174 {64, 16384, 4, 0x00000730 }, /* 1 1 1 1 1 0 */
175 {32, 8192, 3, 0x00000738 } /* 1 1 1 1 1 1 */
178 static int amd_8151_configure(void)
180 unsigned long gatt_bus = virt_to_gart(agp_bridge->gatt_table_real);
183 /* Configure AGP regs in each x86-64 host bridge. */
184 for (i = 0; i < num_k8_northbridges; i++) {
185 agp_bridge->gart_bus_addr =
186 amd64_configure(k8_northbridges[i], gatt_bus);
193 static void amd64_cleanup(void)
197 for (i = 0; i < num_k8_northbridges; i++) {
198 struct pci_dev *dev = k8_northbridges[i];
199 /* disable gart translation */
200 pci_read_config_dword(dev, AMD64_GARTAPERTURECTL, &tmp);
201 tmp &= ~AMD64_GARTEN;
202 pci_write_config_dword(dev, AMD64_GARTAPERTURECTL, tmp);
207 static const struct agp_bridge_driver amd_8151_driver = {
208 .owner = THIS_MODULE,
209 .aperture_sizes = amd_8151_sizes,
210 .size_type = U32_APER_SIZE,
211 .num_aperture_sizes = 7,
212 .configure = amd_8151_configure,
213 .fetch_size = amd64_fetch_size,
214 .cleanup = amd64_cleanup,
215 .tlb_flush = amd64_tlbflush,
216 .mask_memory = agp_generic_mask_memory,
218 .agp_enable = agp_generic_enable,
219 .cache_flush = global_cache_flush,
220 .create_gatt_table = agp_generic_create_gatt_table,
221 .free_gatt_table = agp_generic_free_gatt_table,
222 .insert_memory = amd64_insert_memory,
223 .remove_memory = agp_generic_remove_memory,
224 .alloc_by_type = agp_generic_alloc_by_type,
225 .free_by_type = agp_generic_free_by_type,
226 .agp_alloc_page = agp_generic_alloc_page,
227 .agp_alloc_pages = agp_generic_alloc_pages,
228 .agp_destroy_page = agp_generic_destroy_page,
229 .agp_destroy_pages = agp_generic_destroy_pages,
230 .agp_type_to_mask_type = agp_generic_type_to_mask_type,
233 /* Some basic sanity checks for the aperture. */
234 static int __devinit agp_aperture_valid(u64 aper, u32 size)
236 if (!aperture_valid(aper, size, 32*1024*1024))
239 /* Request the Aperture. This catches cases when someone else
240 already put a mapping in there - happens with some very broken BIOS
242 Maybe better to use pci_assign_resource/pci_enable_device instead
243 trusting the bridges? */
244 if (!aperture_resource &&
245 !(aperture_resource = request_mem_region(aper, size, "aperture"))) {
246 printk(KERN_ERR PFX "Aperture conflicts with PCI mapping.\n");
253 * W*s centric BIOS sometimes only set up the aperture in the AGP
254 * bridge, not the northbridge. On AMD64 this is handled early
255 * in aperture.c, but when IOMMU is not enabled or we run
256 * on a 32bit kernel this needs to be redone.
257 * Unfortunately it is impossible to fix the aperture here because it's too late
258 * to allocate that much memory. But at least error out cleanly instead of
261 static __devinit int fix_northbridge(struct pci_dev *nb, struct pci_dev *agp,
264 u32 aper_low, aper_hi;
267 u32 nb_order, nb_base;
270 pci_read_config_dword(nb, AMD64_GARTAPERTURECTL, &nb_order);
271 nb_order = (nb_order >> 1) & 7;
272 pci_read_config_dword(nb, AMD64_GARTAPERTUREBASE, &nb_base);
273 nb_aper = nb_base << 25;
275 /* Northbridge seems to contain crap. Try the AGP bridge. */
277 pci_read_config_word(agp, cap+0x14, &apsize);
278 if (apsize == 0xffff) {
279 if (agp_aperture_valid(nb_aper, (32*1024*1024)<<nb_order))
285 /* Some BIOS use weird encodings not in the AGPv3 table. */
288 order = 7 - hweight16(apsize);
290 pci_read_config_dword(agp, 0x10, &aper_low);
291 pci_read_config_dword(agp, 0x14, &aper_hi);
292 aper = (aper_low & ~((1<<22)-1)) | ((u64)aper_hi << 32);
295 * On some sick chips APSIZE is 0. This means it wants 4G
296 * so let double check that order, and lets trust the AMD NB settings
298 if (order >=0 && aper + (32ULL<<(20 + order)) > 0x100000000ULL) {
299 dev_info(&agp->dev, "aperture size %u MB is not right, using settings from NB\n",
304 if (nb_order >= order) {
305 if (agp_aperture_valid(nb_aper, (32*1024*1024)<<nb_order))
309 dev_info(&agp->dev, "aperture from AGP @ %Lx size %u MB\n",
311 if (order < 0 || !agp_aperture_valid(aper, (32*1024*1024)<<order))
314 pci_write_config_dword(nb, AMD64_GARTAPERTURECTL, order << 1);
315 pci_write_config_dword(nb, AMD64_GARTAPERTUREBASE, aper >> 25);
320 static __devinit int cache_nbs (struct pci_dev *pdev, u32 cap_ptr)
324 if (cache_k8_northbridges() < 0)
328 for (i = 0; i < num_k8_northbridges; i++) {
329 struct pci_dev *dev = k8_northbridges[i];
330 if (fix_northbridge(dev, pdev, cap_ptr) < 0) {
331 dev_err(&dev->dev, "no usable aperture found\n");
333 /* should port this to i386 */
334 dev_err(&dev->dev, "consider rebooting with iommu=memaper=2 to get a good aperture\n");
342 /* Handle AMD 8151 quirks */
343 static void __devinit amd8151_init(struct pci_dev *pdev, struct agp_bridge_data *bridge)
347 switch (pdev->revision) {
348 case 0x01: revstring="A0"; break;
349 case 0x02: revstring="A1"; break;
350 case 0x11: revstring="B0"; break;
351 case 0x12: revstring="B1"; break;
352 case 0x13: revstring="B2"; break;
353 case 0x14: revstring="B3"; break;
354 default: revstring="??"; break;
357 dev_info(&pdev->dev, "AMD 8151 AGP Bridge rev %s\n", revstring);
360 * Work around errata.
361 * Chips before B2 stepping incorrectly reporting v3.5
363 if (pdev->revision < 0x13) {
364 dev_info(&pdev->dev, "correcting AGP revision (reports 3.5, is really 3.0)\n");
365 bridge->major_version = 3;
366 bridge->minor_version = 0;
371 static const struct aper_size_info_32 uli_sizes[7] =
381 static int __devinit uli_agp_init(struct pci_dev *pdev)
383 u32 httfea,baseaddr,enuscr;
384 struct pci_dev *dev1;
386 unsigned size = amd64_fetch_size();
388 dev_info(&pdev->dev, "setting up ULi AGP\n");
389 dev1 = pci_get_slot (pdev->bus,PCI_DEVFN(0,0));
391 dev_info(&pdev->dev, "can't find ULi secondary device\n");
395 for (i = 0; i < ARRAY_SIZE(uli_sizes); i++)
396 if (uli_sizes[i].size == size)
399 if (i == ARRAY_SIZE(uli_sizes)) {
400 dev_info(&pdev->dev, "no ULi size found for %d\n", size);
404 /* shadow x86-64 registers into ULi registers */
405 pci_read_config_dword (k8_northbridges[0], AMD64_GARTAPERTUREBASE, &httfea);
407 /* if x86-64 aperture base is beyond 4G, exit here */
408 if ((httfea & 0x7fff) >> (32 - 25))
411 httfea = (httfea& 0x7fff) << 25;
413 pci_read_config_dword(pdev, ULI_X86_64_BASE_ADDR, &baseaddr);
414 baseaddr&= ~PCI_BASE_ADDRESS_MEM_MASK;
416 pci_write_config_dword(pdev, ULI_X86_64_BASE_ADDR, baseaddr);
418 enuscr= httfea+ (size * 1024 * 1024) - 1;
419 pci_write_config_dword(dev1, ULI_X86_64_HTT_FEA_REG, httfea);
420 pci_write_config_dword(dev1, ULI_X86_64_ENU_SCR_REG, enuscr);
427 static const struct aper_size_info_32 nforce3_sizes[5] =
429 {512, 131072, 7, 0x00000000 },
430 {256, 65536, 6, 0x00000008 },
431 {128, 32768, 5, 0x0000000C },
432 {64, 16384, 4, 0x0000000E },
433 {32, 8192, 3, 0x0000000F }
436 /* Handle shadow device of the Nvidia NForce3 */
437 /* CHECK-ME original 2.4 version set up some IORRs. Check if that is needed. */
438 static int nforce3_agp_init(struct pci_dev *pdev)
440 u32 tmp, apbase, apbar, aplimit;
441 struct pci_dev *dev1;
443 unsigned size = amd64_fetch_size();
445 dev_info(&pdev->dev, "setting up Nforce3 AGP\n");
447 dev1 = pci_get_slot(pdev->bus, PCI_DEVFN(11, 0));
449 dev_info(&pdev->dev, "can't find Nforce3 secondary device\n");
453 for (i = 0; i < ARRAY_SIZE(nforce3_sizes); i++)
454 if (nforce3_sizes[i].size == size)
457 if (i == ARRAY_SIZE(nforce3_sizes)) {
458 dev_info(&pdev->dev, "no NForce3 size found for %d\n", size);
462 pci_read_config_dword(dev1, NVIDIA_X86_64_1_APSIZE, &tmp);
464 tmp |= nforce3_sizes[i].size_value;
465 pci_write_config_dword(dev1, NVIDIA_X86_64_1_APSIZE, tmp);
467 /* shadow x86-64 registers into NVIDIA registers */
468 pci_read_config_dword (k8_northbridges[0], AMD64_GARTAPERTUREBASE, &apbase);
470 /* if x86-64 aperture base is beyond 4G, exit here */
471 if ( (apbase & 0x7fff) >> (32 - 25) ) {
472 dev_info(&pdev->dev, "aperture base > 4G\n");
476 apbase = (apbase & 0x7fff) << 25;
478 pci_read_config_dword(pdev, NVIDIA_X86_64_0_APBASE, &apbar);
479 apbar &= ~PCI_BASE_ADDRESS_MEM_MASK;
481 pci_write_config_dword(pdev, NVIDIA_X86_64_0_APBASE, apbar);
483 aplimit = apbase + (size * 1024 * 1024) - 1;
484 pci_write_config_dword(dev1, NVIDIA_X86_64_1_APBASE1, apbase);
485 pci_write_config_dword(dev1, NVIDIA_X86_64_1_APLIMIT1, aplimit);
486 pci_write_config_dword(dev1, NVIDIA_X86_64_1_APBASE2, apbase);
487 pci_write_config_dword(dev1, NVIDIA_X86_64_1_APLIMIT2, aplimit);
494 static int __devinit agp_amd64_probe(struct pci_dev *pdev,
495 const struct pci_device_id *ent)
497 struct agp_bridge_data *bridge;
501 cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP);
505 /* Could check for AGPv3 here */
507 bridge = agp_alloc_bridge();
511 if (pdev->vendor == PCI_VENDOR_ID_AMD &&
512 pdev->device == PCI_DEVICE_ID_AMD_8151_0) {
513 amd8151_init(pdev, bridge);
515 dev_info(&pdev->dev, "AGP bridge [%04x/%04x]\n",
516 pdev->vendor, pdev->device);
519 bridge->driver = &amd_8151_driver;
521 bridge->capndx = cap_ptr;
523 /* Fill in the mode register */
524 pci_read_config_dword(pdev, bridge->capndx+PCI_AGP_STATUS, &bridge->mode);
526 if (cache_nbs(pdev, cap_ptr) == -1) {
527 agp_put_bridge(bridge);
531 if (pdev->vendor == PCI_VENDOR_ID_NVIDIA) {
532 int ret = nforce3_agp_init(pdev);
534 agp_put_bridge(bridge);
539 if (pdev->vendor == PCI_VENDOR_ID_AL) {
540 int ret = uli_agp_init(pdev);
542 agp_put_bridge(bridge);
547 pci_set_drvdata(pdev, bridge);
548 err = agp_add_bridge(bridge);
556 static void __devexit agp_amd64_remove(struct pci_dev *pdev)
558 struct agp_bridge_data *bridge = pci_get_drvdata(pdev);
560 release_mem_region(virt_to_gart(bridge->gatt_table_real),
561 amd64_aperture_sizes[bridge->aperture_size_idx].size);
562 agp_remove_bridge(bridge);
563 agp_put_bridge(bridge);
568 static int agp_amd64_suspend(struct pci_dev *pdev, pm_message_t state)
570 pci_save_state(pdev);
571 pci_set_power_state(pdev, pci_choose_state(pdev, state));
576 static int agp_amd64_resume(struct pci_dev *pdev)
578 pci_set_power_state(pdev, PCI_D0);
579 pci_restore_state(pdev);
581 if (pdev->vendor == PCI_VENDOR_ID_NVIDIA)
582 nforce3_agp_init(pdev);
584 return amd_8151_configure();
587 #endif /* CONFIG_PM */
589 static struct pci_device_id agp_amd64_pci_table[] = {
591 .class = (PCI_CLASS_BRIDGE_HOST << 8),
593 .vendor = PCI_VENDOR_ID_AMD,
594 .device = PCI_DEVICE_ID_AMD_8151_0,
595 .subvendor = PCI_ANY_ID,
596 .subdevice = PCI_ANY_ID,
600 .class = (PCI_CLASS_BRIDGE_HOST << 8),
602 .vendor = PCI_VENDOR_ID_AL,
603 .device = PCI_DEVICE_ID_AL_M1689,
604 .subvendor = PCI_ANY_ID,
605 .subdevice = PCI_ANY_ID,
609 .class = (PCI_CLASS_BRIDGE_HOST << 8),
611 .vendor = PCI_VENDOR_ID_VIA,
612 .device = PCI_DEVICE_ID_VIA_K8T800PRO_0,
613 .subvendor = PCI_ANY_ID,
614 .subdevice = PCI_ANY_ID,
618 .class = (PCI_CLASS_BRIDGE_HOST << 8),
620 .vendor = PCI_VENDOR_ID_VIA,
621 .device = PCI_DEVICE_ID_VIA_8385_0,
622 .subvendor = PCI_ANY_ID,
623 .subdevice = PCI_ANY_ID,
625 /* VIA K8M800 / K8N800 */
627 .class = (PCI_CLASS_BRIDGE_HOST << 8),
629 .vendor = PCI_VENDOR_ID_VIA,
630 .device = PCI_DEVICE_ID_VIA_8380_0,
631 .subvendor = PCI_ANY_ID,
632 .subdevice = PCI_ANY_ID,
634 /* VIA K8M890 / K8N890 */
636 .class = (PCI_CLASS_BRIDGE_HOST << 8),
638 .vendor = PCI_VENDOR_ID_VIA,
639 .device = PCI_DEVICE_ID_VIA_VT3336,
640 .subvendor = PCI_ANY_ID,
641 .subdevice = PCI_ANY_ID,
645 .class = (PCI_CLASS_BRIDGE_HOST << 8),
647 .vendor = PCI_VENDOR_ID_VIA,
648 .device = PCI_DEVICE_ID_VIA_3238_0,
649 .subvendor = PCI_ANY_ID,
650 .subdevice = PCI_ANY_ID,
652 /* VIA K8T800/K8M800/K8N800 */
654 .class = (PCI_CLASS_BRIDGE_HOST << 8),
656 .vendor = PCI_VENDOR_ID_VIA,
657 .device = PCI_DEVICE_ID_VIA_838X_1,
658 .subvendor = PCI_ANY_ID,
659 .subdevice = PCI_ANY_ID,
663 .class = (PCI_CLASS_BRIDGE_HOST << 8),
665 .vendor = PCI_VENDOR_ID_NVIDIA,
666 .device = PCI_DEVICE_ID_NVIDIA_NFORCE3,
667 .subvendor = PCI_ANY_ID,
668 .subdevice = PCI_ANY_ID,
671 .class = (PCI_CLASS_BRIDGE_HOST << 8),
673 .vendor = PCI_VENDOR_ID_NVIDIA,
674 .device = PCI_DEVICE_ID_NVIDIA_NFORCE3S,
675 .subvendor = PCI_ANY_ID,
676 .subdevice = PCI_ANY_ID,
680 .class = (PCI_CLASS_BRIDGE_HOST << 8),
682 .vendor = PCI_VENDOR_ID_SI,
683 .device = PCI_DEVICE_ID_SI_755,
684 .subvendor = PCI_ANY_ID,
685 .subdevice = PCI_ANY_ID,
689 .class = (PCI_CLASS_BRIDGE_HOST << 8),
691 .vendor = PCI_VENDOR_ID_SI,
692 .device = PCI_DEVICE_ID_SI_760,
693 .subvendor = PCI_ANY_ID,
694 .subdevice = PCI_ANY_ID,
698 .class = (PCI_CLASS_BRIDGE_HOST << 8),
700 .vendor = PCI_VENDOR_ID_AL,
702 .subvendor = PCI_ANY_ID,
703 .subdevice = PCI_ANY_ID,
709 MODULE_DEVICE_TABLE(pci, agp_amd64_pci_table);
711 static struct pci_driver agp_amd64_pci_driver = {
712 .name = "agpgart-amd64",
713 .id_table = agp_amd64_pci_table,
714 .probe = agp_amd64_probe,
715 .remove = agp_amd64_remove,
717 .suspend = agp_amd64_suspend,
718 .resume = agp_amd64_resume,
723 /* Not static due to IOMMU code calling it early. */
724 int __init agp_amd64_init(void)
730 err = pci_register_driver(&agp_amd64_pci_driver);
734 if (agp_bridges_found == 0) {
736 if (!agp_try_unsupported && !agp_try_unsupported_boot) {
737 printk(KERN_INFO PFX "No supported AGP bridge found.\n");
739 printk(KERN_INFO PFX "You can try agp_try_unsupported=1\n");
741 printk(KERN_INFO PFX "You can boot with agp=try_unsupported\n");
746 /* First check that we have at least one AMD64 NB */
747 if (!pci_dev_present(k8_nb_ids))
750 /* Look for any AGP bridge */
753 for_each_pci_dev(dev) {
754 if (!pci_find_capability(dev, PCI_CAP_ID_AGP))
756 /* Only one bridge supported right now */
757 if (agp_amd64_probe(dev, NULL) == 0) {
766 static void __exit agp_amd64_cleanup(void)
768 if (aperture_resource)
769 release_resource(aperture_resource);
770 pci_unregister_driver(&agp_amd64_pci_driver);
773 /* On AMD64 the PCI driver needs to initialize this driver early
774 for the IOMMU, so it has to be called via a backdoor. */
775 #ifndef CONFIG_GART_IOMMU
776 module_init(agp_amd64_init);
777 module_exit(agp_amd64_cleanup);
780 MODULE_AUTHOR("Dave Jones <davej@redhat.com>, Andi Kleen");
781 module_param(agp_try_unsupported, bool, 0);
782 MODULE_LICENSE("GPL");