2 * mmconfig-shared.c - Low-level direct PCI config space access via
3 * MMCONFIG - common code between i386 and x86-64.
6 * - known chipset handling
7 * - ACPI decoding and validation
9 * Per-architecture code takes care of the mappings and accesses
13 #include <linux/pci.h>
14 #include <linux/init.h>
15 #include <linux/sfi_acpi.h>
16 #include <linux/bitmap.h>
17 #include <linux/dmi.h>
18 #include <linux/slab.h>
19 #include <linux/mutex.h>
20 #include <linux/rculist.h>
22 #include <asm/pci_x86.h>
25 #define PREFIX "PCI: "
27 /* Indicate if the mmcfg resources have been placed into the resource table. */
28 static bool pci_mmcfg_running_state
;
29 static bool pci_mmcfg_arch_init_failed
;
30 static DEFINE_MUTEX(pci_mmcfg_lock
);
32 LIST_HEAD(pci_mmcfg_list
);
34 static __init
void pci_mmconfig_remove(struct pci_mmcfg_region
*cfg
)
37 release_resource(&cfg
->res
);
42 static __init
void free_all_mmcfg(void)
44 struct pci_mmcfg_region
*cfg
, *tmp
;
46 pci_mmcfg_arch_free();
47 list_for_each_entry_safe(cfg
, tmp
, &pci_mmcfg_list
, list
)
48 pci_mmconfig_remove(cfg
);
51 static void list_add_sorted(struct pci_mmcfg_region
*new)
53 struct pci_mmcfg_region
*cfg
;
55 /* keep list sorted by segment and starting bus number */
56 list_for_each_entry_rcu(cfg
, &pci_mmcfg_list
, list
) {
57 if (cfg
->segment
> new->segment
||
58 (cfg
->segment
== new->segment
&&
59 cfg
->start_bus
>= new->start_bus
)) {
60 list_add_tail_rcu(&new->list
, &cfg
->list
);
64 list_add_tail_rcu(&new->list
, &pci_mmcfg_list
);
67 static struct pci_mmcfg_region
*pci_mmconfig_alloc(int segment
, int start
,
70 struct pci_mmcfg_region
*new;
76 new = kzalloc(sizeof(*new), GFP_KERNEL
);
81 new->segment
= segment
;
82 new->start_bus
= start
;
86 res
->start
= addr
+ PCI_MMCFG_BUS_OFFSET(start
);
87 res
->end
= addr
+ PCI_MMCFG_BUS_OFFSET(end
+ 1) - 1;
88 res
->flags
= IORESOURCE_MEM
| IORESOURCE_BUSY
;
89 snprintf(new->name
, PCI_MMCFG_RESOURCE_NAME_LEN
,
90 "PCI MMCONFIG %04x [bus %02x-%02x]", segment
, start
, end
);
91 res
->name
= new->name
;
96 static __init
struct pci_mmcfg_region
*pci_mmconfig_add(int segment
, int start
,
99 struct pci_mmcfg_region
*new;
101 new = pci_mmconfig_alloc(segment
, start
, end
, addr
);
103 mutex_lock(&pci_mmcfg_lock
);
104 list_add_sorted(new);
105 mutex_unlock(&pci_mmcfg_lock
);
108 "MMCONFIG for domain %04x [bus %02x-%02x] at %pR "
110 segment
, start
, end
, &new->res
, (unsigned long)addr
);
116 struct pci_mmcfg_region
*pci_mmconfig_lookup(int segment
, int bus
)
118 struct pci_mmcfg_region
*cfg
;
120 list_for_each_entry_rcu(cfg
, &pci_mmcfg_list
, list
)
121 if (cfg
->segment
== segment
&&
122 cfg
->start_bus
<= bus
&& bus
<= cfg
->end_bus
)
128 static const char __init
*pci_mmcfg_e7520(void)
131 raw_pci_ops
->read(0, 0, PCI_DEVFN(0, 0), 0xce, 2, &win
);
134 if (win
== 0x0000 || win
== 0xf000)
137 if (pci_mmconfig_add(0, 0, 255, win
<< 16) == NULL
)
140 return "Intel Corporation E7520 Memory Controller Hub";
143 static const char __init
*pci_mmcfg_intel_945(void)
145 u32 pciexbar
, mask
= 0, len
= 0;
147 raw_pci_ops
->read(0, 0, PCI_DEVFN(0, 0), 0x48, 4, &pciexbar
);
154 switch ((pciexbar
>> 1) & 3) {
171 /* Errata #2, things break when not aligned on a 256Mb boundary */
172 /* Can only happen in 64M/128M mode */
174 if ((pciexbar
& mask
) & 0x0fffffffU
)
177 /* Don't hit the APIC registers and their friends */
178 if ((pciexbar
& mask
) >= 0xf0000000U
)
181 if (pci_mmconfig_add(0, 0, (len
>> 20) - 1, pciexbar
& mask
) == NULL
)
184 return "Intel Corporation 945G/GZ/P/PL Express Memory Controller Hub";
187 static const char __init
*pci_mmcfg_amd_fam10h(void)
189 u32 low
, high
, address
;
192 unsigned segnbits
= 0, busnbits
, end_bus
;
194 if (!(pci_probe
& PCI_CHECK_ENABLE_AMD_MMCONF
))
197 address
= MSR_FAM10H_MMIO_CONF_BASE
;
198 if (rdmsr_safe(address
, &low
, &high
))
205 /* mmconfig is not enable */
206 if (!(msr
& FAM10H_MMIO_CONF_ENABLE
))
209 base
= msr
& (FAM10H_MMIO_CONF_BASE_MASK
<<FAM10H_MMIO_CONF_BASE_SHIFT
);
211 busnbits
= (msr
>> FAM10H_MMIO_CONF_BUSRANGE_SHIFT
) &
212 FAM10H_MMIO_CONF_BUSRANGE_MASK
;
215 * only handle bus 0 ?
222 segnbits
= busnbits
- 8;
226 end_bus
= (1 << busnbits
) - 1;
227 for (i
= 0; i
< (1 << segnbits
); i
++)
228 if (pci_mmconfig_add(i
, 0, end_bus
,
229 base
+ (1<<28) * i
) == NULL
) {
234 return "AMD Family 10h NB";
237 static bool __initdata mcp55_checked
;
238 static const char __init
*pci_mmcfg_nvidia_mcp55(void)
241 int mcp55_mmconf_found
= 0;
243 static const u32 extcfg_regnum
= 0x90;
244 static const u32 extcfg_regsize
= 4;
245 static const u32 extcfg_enable_mask
= 1<<31;
246 static const u32 extcfg_start_mask
= 0xff<<16;
247 static const int extcfg_start_shift
= 16;
248 static const u32 extcfg_size_mask
= 0x3<<28;
249 static const int extcfg_size_shift
= 28;
250 static const int extcfg_sizebus
[] = {0x100, 0x80, 0x40, 0x20};
251 static const u32 extcfg_base_mask
[] = {0x7ff8, 0x7ffc, 0x7ffe, 0x7fff};
252 static const int extcfg_base_lshift
= 25;
255 * do check if amd fam10h already took over
257 if (!acpi_disabled
|| !list_empty(&pci_mmcfg_list
) || mcp55_checked
)
260 mcp55_checked
= true;
261 for (bus
= 0; bus
< 256; bus
++) {
265 int start
, size_index
, end
;
267 raw_pci_ops
->read(0, bus
, PCI_DEVFN(0, 0), 0, 4, &l
);
269 device
= (l
>> 16) & 0xffff;
271 if (PCI_VENDOR_ID_NVIDIA
!= vendor
|| 0x0369 != device
)
274 raw_pci_ops
->read(0, bus
, PCI_DEVFN(0, 0), extcfg_regnum
,
275 extcfg_regsize
, &extcfg
);
277 if (!(extcfg
& extcfg_enable_mask
))
280 size_index
= (extcfg
& extcfg_size_mask
) >> extcfg_size_shift
;
281 base
= extcfg
& extcfg_base_mask
[size_index
];
282 /* base could > 4G */
283 base
<<= extcfg_base_lshift
;
284 start
= (extcfg
& extcfg_start_mask
) >> extcfg_start_shift
;
285 end
= start
+ extcfg_sizebus
[size_index
] - 1;
286 if (pci_mmconfig_add(0, start
, end
, base
) == NULL
)
288 mcp55_mmconf_found
++;
291 if (!mcp55_mmconf_found
)
294 return "nVidia MCP55";
297 struct pci_mmcfg_hostbridge_probe
{
302 const char *(*probe
)(void);
305 static struct pci_mmcfg_hostbridge_probe pci_mmcfg_probes
[] __initdata
= {
306 { 0, PCI_DEVFN(0, 0), PCI_VENDOR_ID_INTEL
,
307 PCI_DEVICE_ID_INTEL_E7520_MCH
, pci_mmcfg_e7520
},
308 { 0, PCI_DEVFN(0, 0), PCI_VENDOR_ID_INTEL
,
309 PCI_DEVICE_ID_INTEL_82945G_HB
, pci_mmcfg_intel_945
},
310 { 0, PCI_DEVFN(0x18, 0), PCI_VENDOR_ID_AMD
,
311 0x1200, pci_mmcfg_amd_fam10h
},
312 { 0xff, PCI_DEVFN(0, 0), PCI_VENDOR_ID_AMD
,
313 0x1200, pci_mmcfg_amd_fam10h
},
314 { 0, PCI_DEVFN(0, 0), PCI_VENDOR_ID_NVIDIA
,
315 0x0369, pci_mmcfg_nvidia_mcp55
},
318 static void __init
pci_mmcfg_check_end_bus_number(void)
320 struct pci_mmcfg_region
*cfg
, *cfgx
;
323 list_for_each_entry(cfg
, &pci_mmcfg_list
, list
) {
324 if (cfg
->end_bus
< cfg
->start_bus
)
327 /* Don't access the list head ! */
328 if (cfg
->list
.next
== &pci_mmcfg_list
)
331 cfgx
= list_entry(cfg
->list
.next
, typeof(*cfg
), list
);
332 if (cfg
->end_bus
>= cfgx
->start_bus
)
333 cfg
->end_bus
= cfgx
->start_bus
- 1;
337 static int __init
pci_mmcfg_check_hostbridge(void)
350 for (i
= 0; i
< ARRAY_SIZE(pci_mmcfg_probes
); i
++) {
351 bus
= pci_mmcfg_probes
[i
].bus
;
352 devfn
= pci_mmcfg_probes
[i
].devfn
;
353 raw_pci_ops
->read(0, bus
, devfn
, 0, 4, &l
);
355 device
= (l
>> 16) & 0xffff;
358 if (pci_mmcfg_probes
[i
].vendor
== vendor
&&
359 pci_mmcfg_probes
[i
].device
== device
)
360 name
= pci_mmcfg_probes
[i
].probe();
363 pr_info(PREFIX
"%s with MMCONFIG support\n", name
);
366 /* some end_bus_number is crazy, fix it */
367 pci_mmcfg_check_end_bus_number();
369 return !list_empty(&pci_mmcfg_list
);
372 static acpi_status
check_mcfg_resource(struct acpi_resource
*res
, void *data
)
374 struct resource
*mcfg_res
= data
;
375 struct acpi_resource_address64 address
;
378 if (res
->type
== ACPI_RESOURCE_TYPE_FIXED_MEMORY32
) {
379 struct acpi_resource_fixed_memory32
*fixmem32
=
380 &res
->data
.fixed_memory32
;
383 if ((mcfg_res
->start
>= fixmem32
->address
) &&
384 (mcfg_res
->end
< (fixmem32
->address
+
385 fixmem32
->address_length
))) {
387 return AE_CTRL_TERMINATE
;
390 if ((res
->type
!= ACPI_RESOURCE_TYPE_ADDRESS32
) &&
391 (res
->type
!= ACPI_RESOURCE_TYPE_ADDRESS64
))
394 status
= acpi_resource_to_address64(res
, &address
);
395 if (ACPI_FAILURE(status
) ||
396 (address
.address_length
<= 0) ||
397 (address
.resource_type
!= ACPI_MEMORY_RANGE
))
400 if ((mcfg_res
->start
>= address
.minimum
) &&
401 (mcfg_res
->end
< (address
.minimum
+ address
.address_length
))) {
403 return AE_CTRL_TERMINATE
;
408 static acpi_status
find_mboard_resource(acpi_handle handle
, u32 lvl
,
409 void *context
, void **rv
)
411 struct resource
*mcfg_res
= context
;
413 acpi_walk_resources(handle
, METHOD_NAME__CRS
,
414 check_mcfg_resource
, context
);
417 return AE_CTRL_TERMINATE
;
422 static int is_acpi_reserved(u64 start
, u64 end
, unsigned not_used
)
424 struct resource mcfg_res
;
426 mcfg_res
.start
= start
;
427 mcfg_res
.end
= end
- 1;
430 acpi_get_devices("PNP0C01", find_mboard_resource
, &mcfg_res
, NULL
);
433 acpi_get_devices("PNP0C02", find_mboard_resource
, &mcfg_res
,
436 return mcfg_res
.flags
;
439 typedef int (*check_reserved_t
)(u64 start
, u64 end
, unsigned type
);
441 static int __ref
is_mmconf_reserved(check_reserved_t is_reserved
,
442 struct pci_mmcfg_region
*cfg
,
443 struct device
*dev
, int with_e820
)
445 u64 addr
= cfg
->res
.start
;
446 u64 size
= resource_size(&cfg
->res
);
449 char *method
= with_e820
? "E820" : "ACPI motherboard resources";
451 while (!is_reserved(addr
, addr
+ size
, E820_RESERVED
)) {
453 if (size
< (16UL<<20))
457 if (size
< (16UL<<20) && size
!= old_size
)
461 dev_info(dev
, "MMCONFIG at %pR reserved in %s\n",
464 pr_info(PREFIX
"MMCONFIG at %pR reserved in %s\n",
467 if (old_size
!= size
) {
469 cfg
->end_bus
= cfg
->start_bus
+ ((size
>>20) - 1);
470 num_buses
= cfg
->end_bus
- cfg
->start_bus
+ 1;
471 cfg
->res
.end
= cfg
->res
.start
+
472 PCI_MMCFG_BUS_OFFSET(num_buses
) - 1;
473 snprintf(cfg
->name
, PCI_MMCFG_RESOURCE_NAME_LEN
,
474 "PCI MMCONFIG %04x [bus %02x-%02x]",
475 cfg
->segment
, cfg
->start_bus
, cfg
->end_bus
);
480 "at %pR (base %#lx) (size reduced!)\n",
481 &cfg
->res
, (unsigned long) cfg
->address
);
484 "MMCONFIG for %04x [bus%02x-%02x] "
485 "at %pR (base %#lx) (size reduced!)\n",
486 cfg
->segment
, cfg
->start_bus
, cfg
->end_bus
,
487 &cfg
->res
, (unsigned long) cfg
->address
);
493 static int __ref
pci_mmcfg_check_reserved(struct device
*dev
,
494 struct pci_mmcfg_region
*cfg
, int early
)
496 if (!early
&& !acpi_disabled
) {
497 if (is_mmconf_reserved(is_acpi_reserved
, cfg
, dev
, 0))
501 dev_info(dev
, FW_INFO
502 "MMCONFIG at %pR not reserved in "
503 "ACPI motherboard resources\n",
506 pr_info(FW_INFO PREFIX
507 "MMCONFIG at %pR not reserved in "
508 "ACPI motherboard resources\n",
513 * e820_all_mapped() is marked as __init.
514 * All entries from ACPI MCFG table have been checked at boot time.
515 * For MCFG information constructed from hotpluggable host bridge's
516 * _CBA method, just assume it's reserved.
518 if (pci_mmcfg_running_state
)
521 /* Don't try to do this check unless configuration
522 type 1 is available. how about type 2 ?*/
524 return is_mmconf_reserved(e820_all_mapped
, cfg
, dev
, 1);
529 static void __init
pci_mmcfg_reject_broken(int early
)
531 struct pci_mmcfg_region
*cfg
;
533 list_for_each_entry(cfg
, &pci_mmcfg_list
, list
) {
534 if (pci_mmcfg_check_reserved(NULL
, cfg
, early
) == 0) {
535 pr_info(PREFIX
"not using MMCONFIG\n");
542 static int __init
acpi_mcfg_check_entry(struct acpi_table_mcfg
*mcfg
,
543 struct acpi_mcfg_allocation
*cfg
)
547 if (cfg
->address
< 0xFFFFFFFF)
550 if (!strncmp(mcfg
->header
.oem_id
, "SGI", 3))
553 if (mcfg
->header
.revision
>= 1) {
554 if (dmi_get_date(DMI_BIOS_DATE
, &year
, NULL
, NULL
) &&
559 pr_err(PREFIX
"MCFG region for %04x [bus %02x-%02x] at %#llx "
560 "is above 4GB, ignored\n", cfg
->pci_segment
,
561 cfg
->start_bus_number
, cfg
->end_bus_number
, cfg
->address
);
565 static int __init
pci_parse_mcfg(struct acpi_table_header
*header
)
567 struct acpi_table_mcfg
*mcfg
;
568 struct acpi_mcfg_allocation
*cfg_table
, *cfg
;
575 mcfg
= (struct acpi_table_mcfg
*)header
;
577 /* how many config structures do we have */
580 i
= header
->length
- sizeof(struct acpi_table_mcfg
);
581 while (i
>= sizeof(struct acpi_mcfg_allocation
)) {
583 i
-= sizeof(struct acpi_mcfg_allocation
);
586 pr_err(PREFIX
"MMCONFIG has no entries\n");
590 cfg_table
= (struct acpi_mcfg_allocation
*) &mcfg
[1];
591 for (i
= 0; i
< entries
; i
++) {
593 if (acpi_mcfg_check_entry(mcfg
, cfg
)) {
598 if (pci_mmconfig_add(cfg
->pci_segment
, cfg
->start_bus_number
,
599 cfg
->end_bus_number
, cfg
->address
) == NULL
) {
600 pr_warn(PREFIX
"no memory for MCFG entries\n");
609 static void __init
__pci_mmcfg_init(int early
)
611 pci_mmcfg_reject_broken(early
);
612 if (list_empty(&pci_mmcfg_list
))
615 if (pcibios_last_bus
< 0) {
616 const struct pci_mmcfg_region
*cfg
;
618 list_for_each_entry(cfg
, &pci_mmcfg_list
, list
) {
621 pcibios_last_bus
= cfg
->end_bus
;
625 if (pci_mmcfg_arch_init())
626 pci_probe
= (pci_probe
& ~PCI_PROBE_MASK
) | PCI_PROBE_MMCONF
;
629 pci_mmcfg_arch_init_failed
= true;
633 static int __initdata known_bridge
;
635 void __init
pci_mmcfg_early_init(void)
637 if (pci_probe
& PCI_PROBE_MMCONF
) {
638 if (pci_mmcfg_check_hostbridge())
641 acpi_sfi_table_parse(ACPI_SIG_MCFG
, pci_parse_mcfg
);
646 void __init
pci_mmcfg_late_init(void)
648 /* MMCONFIG disabled */
649 if ((pci_probe
& PCI_PROBE_MMCONF
) == 0)
655 /* MMCONFIG hasn't been enabled yet, try again */
656 if (pci_probe
& PCI_PROBE_MASK
& ~PCI_PROBE_MMCONF
) {
657 acpi_sfi_table_parse(ACPI_SIG_MCFG
, pci_parse_mcfg
);
662 static int __init
pci_mmcfg_late_insert_resources(void)
664 struct pci_mmcfg_region
*cfg
;
666 pci_mmcfg_running_state
= true;
668 /* If we are not using MMCONFIG, don't insert the resources. */
669 if ((pci_probe
& PCI_PROBE_MMCONF
) == 0)
673 * Attempt to insert the mmcfg resources but not with the busy flag
674 * marked so it won't cause request errors when __request_region is
677 list_for_each_entry(cfg
, &pci_mmcfg_list
, list
)
678 if (!cfg
->res
.parent
)
679 insert_resource(&iomem_resource
, &cfg
->res
);
685 * Perform MMCONFIG resource insertion after PCI initialization to allow for
686 * misprogrammed MCFG tables that state larger sizes but actually conflict
687 * with other system resources.
689 late_initcall(pci_mmcfg_late_insert_resources
);
691 /* Add MMCFG information for host bridges */
692 int pci_mmconfig_insert(struct device
*dev
, u16 seg
, u8 start
, u8 end
,
696 struct resource
*tmp
= NULL
;
697 struct pci_mmcfg_region
*cfg
;
699 if (!(pci_probe
& PCI_PROBE_MMCONF
) || pci_mmcfg_arch_init_failed
)
705 mutex_lock(&pci_mmcfg_lock
);
706 cfg
= pci_mmconfig_lookup(seg
, start
);
708 if (cfg
->end_bus
< end
)
709 dev_info(dev
, FW_INFO
711 "domain %04x [bus %02x-%02x] "
712 "only partially covers this bridge\n",
713 cfg
->segment
, cfg
->start_bus
, cfg
->end_bus
);
714 mutex_unlock(&pci_mmcfg_lock
);
719 mutex_unlock(&pci_mmcfg_lock
);
724 cfg
= pci_mmconfig_alloc(seg
, start
, end
, addr
);
726 dev_warn(dev
, "fail to add MMCONFIG (out of memory)\n");
728 } else if (!pci_mmcfg_check_reserved(dev
, cfg
, 0)) {
729 dev_warn(dev
, FW_BUG
"MMCONFIG %pR isn't reserved\n",
732 /* Insert resource if it's not in boot stage */
733 if (pci_mmcfg_running_state
)
734 tmp
= insert_resource_conflict(&iomem_resource
,
739 "MMCONFIG %pR conflicts with "
741 &cfg
->res
, tmp
->name
, tmp
);
742 } else if (pci_mmcfg_arch_map(cfg
)) {
743 dev_warn(dev
, "fail to map MMCONFIG %pR.\n",
746 list_add_sorted(cfg
);
747 dev_info(dev
, "MMCONFIG at %pR (base %#lx)\n",
748 &cfg
->res
, (unsigned long)addr
);
756 release_resource(&cfg
->res
);
760 mutex_unlock(&pci_mmcfg_lock
);
765 /* Delete MMCFG information for host bridges */
766 int pci_mmconfig_delete(u16 seg
, u8 start
, u8 end
)
768 struct pci_mmcfg_region
*cfg
;
770 mutex_lock(&pci_mmcfg_lock
);
771 list_for_each_entry_rcu(cfg
, &pci_mmcfg_list
, list
)
772 if (cfg
->segment
== seg
&& cfg
->start_bus
== start
&&
773 cfg
->end_bus
== end
) {
774 list_del_rcu(&cfg
->list
);
776 pci_mmcfg_arch_unmap(cfg
);
778 release_resource(&cfg
->res
);
779 mutex_unlock(&pci_mmcfg_lock
);
783 mutex_unlock(&pci_mmcfg_lock
);