2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation, version 2.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11 * NON INFRINGEMENT. See the GNU General Public License for
15 #include <linux/sched.h>
16 #include <linux/kernel.h>
17 #include <linux/mmzone.h>
18 #include <linux/bootmem.h>
19 #include <linux/module.h>
20 #include <linux/node.h>
21 #include <linux/cpu.h>
22 #include <linux/ioport.h>
23 #include <linux/irq.h>
24 #include <linux/kexec.h>
25 #include <linux/pci.h>
26 #include <linux/initrd.h>
28 #include <linux/highmem.h>
29 #include <linux/smp.h>
30 #include <linux/timex.h>
31 #include <asm/setup.h>
32 #include <asm/sections.h>
33 #include <asm/cacheflush.h>
34 #include <asm/pgalloc.h>
35 #include <asm/mmu_context.h>
36 #include <hv/hypervisor.h>
37 #include <arch/interrupts.h>
39 /* <linux/smp.h> doesn't provide this definition. */
41 #define setup_max_cpus 1
44 static inline int ABS(int x
) { return x
>= 0 ? x
: -x
; }
46 /* Chip information */
47 char chip_model
[64] __write_once
;
49 struct pglist_data node_data
[MAX_NUMNODES
] __read_mostly
;
50 EXPORT_SYMBOL(node_data
);
52 /* We only create bootmem data on node 0. */
53 static bootmem_data_t __initdata node0_bdata
;
55 /* Information on the NUMA nodes that we compute early */
56 unsigned long __cpuinitdata node_start_pfn
[MAX_NUMNODES
];
57 unsigned long __cpuinitdata node_end_pfn
[MAX_NUMNODES
];
58 unsigned long __initdata node_memmap_pfn
[MAX_NUMNODES
];
59 unsigned long __initdata node_percpu_pfn
[MAX_NUMNODES
];
60 unsigned long __initdata node_free_pfn
[MAX_NUMNODES
];
62 static unsigned long __initdata node_percpu
[MAX_NUMNODES
];
65 * per-CPU stack and boot info.
67 DEFINE_PER_CPU(unsigned long, boot_sp
) =
68 (unsigned long)init_stack
+ THREAD_SIZE
;
71 DEFINE_PER_CPU(unsigned long, boot_pc
) = (unsigned long)start_kernel
;
74 * The variable must be __initdata since it references __init code.
75 * With CONFIG_SMP it is per-cpu data, which is exempt from validation.
77 unsigned long __initdata boot_pc
= (unsigned long)start_kernel
;
81 /* Page frame index of end of lowmem on each controller. */
82 unsigned long __cpuinitdata node_lowmem_end_pfn
[MAX_NUMNODES
];
84 /* Number of pages that can be mapped into lowmem. */
85 static unsigned long __initdata mappable_physpages
;
88 /* Data on which physical memory controller corresponds to which NUMA node */
89 int node_controller
[MAX_NUMNODES
] = { [0 ... MAX_NUMNODES
-1] = -1 };
92 /* Map information from VAs to PAs */
93 unsigned long pbase_map
[1 << (32 - HPAGE_SHIFT
)]
94 __write_once
__attribute__((aligned(L2_CACHE_BYTES
)));
95 EXPORT_SYMBOL(pbase_map
);
97 /* Map information from PAs to VAs */
98 void *vbase_map
[NR_PA_HIGHBIT_VALUES
]
99 __write_once
__attribute__((aligned(L2_CACHE_BYTES
)));
100 EXPORT_SYMBOL(vbase_map
);
103 /* Node number as a function of the high PA bits */
104 int highbits_to_node
[NR_PA_HIGHBIT_VALUES
] __write_once
;
105 EXPORT_SYMBOL(highbits_to_node
);
107 static unsigned int __initdata maxmem_pfn
= -1U;
108 static unsigned int __initdata maxnodemem_pfn
[MAX_NUMNODES
] = {
109 [0 ... MAX_NUMNODES
-1] = -1U
111 static nodemask_t __initdata isolnodes
;
114 enum { DEFAULT_PCI_RESERVE_MB
= 64 };
115 static unsigned int __initdata pci_reserve_mb
= DEFAULT_PCI_RESERVE_MB
;
116 unsigned long __initdata pci_reserve_start_pfn
= -1U;
117 unsigned long __initdata pci_reserve_end_pfn
= -1U;
120 static int __init
setup_maxmem(char *str
)
122 unsigned long long maxmem
;
123 if (str
== NULL
|| (maxmem
= memparse(str
, NULL
)) == 0)
126 maxmem_pfn
= (maxmem
>> HPAGE_SHIFT
) << (HPAGE_SHIFT
- PAGE_SHIFT
);
127 pr_info("Forcing RAM used to no more than %dMB\n",
128 maxmem_pfn
>> (20 - PAGE_SHIFT
));
131 early_param("maxmem", setup_maxmem
);
133 static int __init
setup_maxnodemem(char *str
)
136 unsigned long long maxnodemem
;
139 node
= str
? simple_strtoul(str
, &endp
, 0) : INT_MAX
;
140 if (node
>= MAX_NUMNODES
|| *endp
!= ':')
143 maxnodemem
= memparse(endp
+1, NULL
);
144 maxnodemem_pfn
[node
] = (maxnodemem
>> HPAGE_SHIFT
) <<
145 (HPAGE_SHIFT
- PAGE_SHIFT
);
146 pr_info("Forcing RAM used on node %ld to no more than %dMB\n",
147 node
, maxnodemem_pfn
[node
] >> (20 - PAGE_SHIFT
));
150 early_param("maxnodemem", setup_maxnodemem
);
152 static int __init
setup_isolnodes(char *str
)
154 char buf
[MAX_NUMNODES
* 5];
155 if (str
== NULL
|| nodelist_parse(str
, isolnodes
) != 0)
158 nodelist_scnprintf(buf
, sizeof(buf
), isolnodes
);
159 pr_info("Set isolnodes value to '%s'\n", buf
);
162 early_param("isolnodes", setup_isolnodes
);
165 static int __init
setup_pci_reserve(char* str
)
169 if (str
== NULL
|| strict_strtoul(str
, 0, &mb
) != 0 ||
174 pr_info("Reserving %dMB for PCIE root complex mappings\n",
178 early_param("pci_reserve", setup_pci_reserve
);
183 * vmalloc=size forces the vmalloc area to be exactly 'size' bytes.
184 * This can be used to increase (or decrease) the vmalloc area.
186 static int __init
parse_vmalloc(char *arg
)
191 VMALLOC_RESERVE
= (memparse(arg
, &arg
) + PGDIR_SIZE
- 1) & PGDIR_MASK
;
193 /* See validate_va() for more on this test. */
194 if ((long)_VMALLOC_START
>= 0)
195 early_panic("\"vmalloc=%#lx\" value too large: maximum %#lx\n",
196 VMALLOC_RESERVE
, _VMALLOC_END
- 0x80000000UL
);
200 early_param("vmalloc", parse_vmalloc
);
203 #ifdef CONFIG_HIGHMEM
205 * Determine for each controller where its lowmem is mapped and how much of
206 * it is mapped there. On controller zero, the first few megabytes are
207 * already mapped in as code at MEM_SV_INTRPT, so in principle we could
208 * start our data mappings higher up, but for now we don't bother, to avoid
209 * additional confusion.
211 * One question is whether, on systems with more than 768 Mb and
212 * controllers of different sizes, to map in a proportionate amount of
213 * each one, or to try to map the same amount from each controller.
214 * (E.g. if we have three controllers with 256MB, 1GB, and 256MB
215 * respectively, do we map 256MB from each, or do we map 128 MB, 512
216 * MB, and 128 MB respectively?) For now we use a proportionate
217 * solution like the latter.
219 * The VA/PA mapping demands that we align our decisions at 16 MB
220 * boundaries so that we can rapidly convert VA to PA.
222 static void *__init
setup_pa_va_mapping(void)
224 unsigned long curr_pages
= 0;
225 unsigned long vaddr
= PAGE_OFFSET
;
226 nodemask_t highonlynodes
= isolnodes
;
229 memset(pbase_map
, -1, sizeof(pbase_map
));
230 memset(vbase_map
, -1, sizeof(vbase_map
));
232 /* Node zero cannot be isolated for LOWMEM purposes. */
233 node_clear(0, highonlynodes
);
235 /* Count up the number of pages on non-highonlynodes controllers. */
236 mappable_physpages
= 0;
237 for_each_online_node(i
) {
238 if (!node_isset(i
, highonlynodes
))
239 mappable_physpages
+=
240 node_end_pfn
[i
] - node_start_pfn
[i
];
243 for_each_online_node(i
) {
244 unsigned long start
= node_start_pfn
[i
];
245 unsigned long end
= node_end_pfn
[i
];
246 unsigned long size
= end
- start
;
247 unsigned long vaddr_end
;
249 if (node_isset(i
, highonlynodes
)) {
250 /* Mark this controller as having no lowmem. */
251 node_lowmem_end_pfn
[i
] = start
;
256 if (mappable_physpages
> MAXMEM_PFN
) {
257 vaddr_end
= PAGE_OFFSET
+
258 (((u64
)curr_pages
* MAXMEM_PFN
/
262 vaddr_end
= PAGE_OFFSET
+ (curr_pages
<< PAGE_SHIFT
);
264 for (j
= 0; vaddr
< vaddr_end
; vaddr
+= HPAGE_SIZE
, ++j
) {
265 unsigned long this_pfn
=
266 start
+ (j
<< HUGETLB_PAGE_ORDER
);
267 pbase_map
[vaddr
>> HPAGE_SHIFT
] = this_pfn
;
268 if (vbase_map
[__pfn_to_highbits(this_pfn
)] ==
270 vbase_map
[__pfn_to_highbits(this_pfn
)] =
271 (void *)(vaddr
& HPAGE_MASK
);
273 node_lowmem_end_pfn
[i
] = start
+ (j
<< HUGETLB_PAGE_ORDER
);
274 BUG_ON(node_lowmem_end_pfn
[i
] > end
);
277 /* Return highest address of any mapped memory. */
278 return (void *)vaddr
;
280 #endif /* CONFIG_HIGHMEM */
283 * Register our most important memory mappings with the debug stub.
285 * This is up to 4 mappings for lowmem, one mapping per memory
286 * controller, plus one for our text segment.
288 static void __cpuinit
store_permanent_mappings(void)
292 for_each_online_node(i
) {
293 HV_PhysAddr pa
= ((HV_PhysAddr
)node_start_pfn
[i
]) << PAGE_SHIFT
;
294 #ifdef CONFIG_HIGHMEM
295 HV_PhysAddr high_mapped_pa
= node_lowmem_end_pfn
[i
];
297 HV_PhysAddr high_mapped_pa
= node_end_pfn
[i
];
300 unsigned long pages
= high_mapped_pa
- node_start_pfn
[i
];
301 HV_VirtAddr addr
= (HV_VirtAddr
) __va(pa
);
302 hv_store_mapping(addr
, pages
<< PAGE_SHIFT
, pa
);
305 hv_store_mapping((HV_VirtAddr
)_stext
,
306 (uint32_t)(_einittext
- _stext
), 0);
310 * Use hv_inquire_physical() to populate node_{start,end}_pfn[]
311 * and node_online_map, doing suitable sanity-checking.
312 * Also set min_low_pfn, max_low_pfn, and max_pfn.
314 static void __init
setup_memory(void)
317 int highbits_seen
[NR_PA_HIGHBIT_VALUES
] = { 0 };
318 #ifdef CONFIG_HIGHMEM
324 #if defined(CONFIG_HIGHMEM) || defined(__tilegx__)
328 /* We are using a char to hold the cpu_2_node[] mapping */
329 BUILD_BUG_ON(MAX_NUMNODES
> 127);
331 /* Discover the ranges of memory available to us */
333 unsigned long start
, size
, end
, highbits
;
334 HV_PhysAddrRange range
= hv_inquire_physical(i
);
337 #ifdef CONFIG_FLATMEM
339 pr_err("Can't use discontiguous PAs: %#llx..%#llx\n",
340 range
.size
, range
.start
+ range
.size
);
345 if ((unsigned long)range
.start
) {
346 pr_err("Range not at 4GB multiple: %#llx..%#llx\n",
347 range
.start
, range
.start
+ range
.size
);
351 if ((range
.start
& (HPAGE_SIZE
-1)) != 0 ||
352 (range
.size
& (HPAGE_SIZE
-1)) != 0) {
353 unsigned long long start_pa
= range
.start
;
354 unsigned long long orig_size
= range
.size
;
355 range
.start
= (start_pa
+ HPAGE_SIZE
- 1) & HPAGE_MASK
;
356 range
.size
-= (range
.start
- start_pa
);
357 range
.size
&= HPAGE_MASK
;
358 pr_err("Range not hugepage-aligned: %#llx..%#llx:"
359 " now %#llx-%#llx\n",
360 start_pa
, start_pa
+ orig_size
,
361 range
.start
, range
.start
+ range
.size
);
363 highbits
= __pa_to_highbits(range
.start
);
364 if (highbits
>= NR_PA_HIGHBIT_VALUES
) {
365 pr_err("PA high bits too high: %#llx..%#llx\n",
366 range
.start
, range
.start
+ range
.size
);
369 if (highbits_seen
[highbits
]) {
370 pr_err("Range overlaps in high bits: %#llx..%#llx\n",
371 range
.start
, range
.start
+ range
.size
);
374 highbits_seen
[highbits
] = 1;
375 if (PFN_DOWN(range
.size
) > maxnodemem_pfn
[i
]) {
376 int max_size
= maxnodemem_pfn
[i
];
378 pr_err("Maxnodemem reduced node %d to"
379 " %d pages\n", i
, max_size
);
380 range
.size
= PFN_PHYS(max_size
);
382 pr_err("Maxnodemem disabled node %d\n", i
);
386 if (num_physpages
+ PFN_DOWN(range
.size
) > maxmem_pfn
) {
387 int max_size
= maxmem_pfn
- num_physpages
;
389 pr_err("Maxmem reduced node %d to %d pages\n",
391 range
.size
= PFN_PHYS(max_size
);
393 pr_err("Maxmem disabled node %d\n", i
);
397 if (i
>= MAX_NUMNODES
) {
398 pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
399 i
, range
.size
, range
.size
+ range
.start
);
403 start
= range
.start
>> PAGE_SHIFT
;
404 size
= range
.size
>> PAGE_SHIFT
;
408 if (((HV_PhysAddr
)end
<< PAGE_SHIFT
) !=
409 (range
.start
+ range
.size
)) {
410 pr_err("PAs too high to represent: %#llx..%#llx\n",
411 range
.start
, range
.start
+ range
.size
);
417 * Blocks that overlap the pci reserved region must
418 * have enough space to hold the maximum percpu data
419 * region at the top of the range. If there isn't
420 * enough space above the reserved region, just
423 if (start
<= pci_reserve_start_pfn
&&
424 end
> pci_reserve_start_pfn
) {
425 unsigned int per_cpu_size
=
426 __per_cpu_end
- __per_cpu_start
;
427 unsigned int percpu_pages
=
428 NR_CPUS
* (PFN_UP(per_cpu_size
) >> PAGE_SHIFT
);
429 if (end
< pci_reserve_end_pfn
+ percpu_pages
) {
430 end
= pci_reserve_start_pfn
;
431 pr_err("PCI mapping region reduced node %d to"
432 " %ld pages\n", i
, end
- start
);
437 for (j
= __pfn_to_highbits(start
);
438 j
<= __pfn_to_highbits(end
- 1); j
++)
439 highbits_to_node
[j
] = i
;
441 node_start_pfn
[i
] = start
;
442 node_end_pfn
[i
] = end
;
443 node_controller
[i
] = range
.controller
;
444 num_physpages
+= size
;
447 /* Mark node as online */
448 node_set(i
, node_online_map
);
449 node_set(i
, node_possible_map
);
454 * For 4KB pages, mem_map "struct page" data is 1% of the size
455 * of the physical memory, so can be quite big (640 MB for
456 * four 16G zones). These structures must be mapped in
457 * lowmem, and since we currently cap out at about 768 MB,
458 * it's impractical to try to use this much address space.
459 * For now, arbitrarily cap the amount of physical memory
460 * we're willing to use at 8 million pages (32GB of 4KB pages).
462 cap
= 8 * 1024 * 1024; /* 8 million pages */
463 if (num_physpages
> cap
) {
464 int num_nodes
= num_online_nodes();
465 int cap_each
= cap
/ num_nodes
;
466 unsigned long dropped_pages
= 0;
467 for (i
= 0; i
< num_nodes
; ++i
) {
468 int size
= node_end_pfn
[i
] - node_start_pfn
[i
];
469 if (size
> cap_each
) {
470 dropped_pages
+= (size
- cap_each
);
471 node_end_pfn
[i
] = node_start_pfn
[i
] + cap_each
;
474 num_physpages
-= dropped_pages
;
475 pr_warning("Only using %ldMB memory;"
476 " ignoring %ldMB.\n",
477 num_physpages
>> (20 - PAGE_SHIFT
),
478 dropped_pages
>> (20 - PAGE_SHIFT
));
479 pr_warning("Consider using a larger page size.\n");
483 /* Heap starts just above the last loaded address. */
484 min_low_pfn
= PFN_UP((unsigned long)_end
- PAGE_OFFSET
);
486 #ifdef CONFIG_HIGHMEM
487 /* Find where we map lowmem from each controller. */
488 high_memory
= setup_pa_va_mapping();
490 /* Set max_low_pfn based on what node 0 can directly address. */
491 max_low_pfn
= node_lowmem_end_pfn
[0];
493 lowmem_pages
= (mappable_physpages
> MAXMEM_PFN
) ?
494 MAXMEM_PFN
: mappable_physpages
;
495 highmem_pages
= (long) (num_physpages
- lowmem_pages
);
497 pr_notice("%ldMB HIGHMEM available.\n",
498 pages_to_mb(highmem_pages
> 0 ? highmem_pages
: 0));
499 pr_notice("%ldMB LOWMEM available.\n",
500 pages_to_mb(lowmem_pages
));
502 /* Set max_low_pfn based on what node 0 can directly address. */
503 max_low_pfn
= node_end_pfn
[0];
506 if (node_end_pfn
[0] > MAXMEM_PFN
) {
507 pr_warning("Only using %ldMB LOWMEM.\n",
509 pr_warning("Use a HIGHMEM enabled kernel.\n");
510 max_low_pfn
= MAXMEM_PFN
;
511 max_pfn
= MAXMEM_PFN
;
512 num_physpages
= MAXMEM_PFN
;
513 node_end_pfn
[0] = MAXMEM_PFN
;
515 pr_notice("%ldMB memory available.\n",
516 pages_to_mb(node_end_pfn
[0]));
518 for (i
= 1; i
< MAX_NUMNODES
; ++i
) {
519 node_start_pfn
[i
] = 0;
522 high_memory
= __va(node_end_pfn
[0]);
525 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
526 int pages
= node_end_pfn
[i
] - node_start_pfn
[i
];
527 lowmem_pages
+= pages
;
529 high_memory
= pfn_to_kaddr(node_end_pfn
[i
]);
531 pr_notice("%ldMB memory available.\n",
532 pages_to_mb(lowmem_pages
));
537 static void __init
setup_bootmem_allocator(void)
539 unsigned long bootmap_size
, first_alloc_pfn
, last_alloc_pfn
;
541 /* Provide a node 0 bdata. */
542 NODE_DATA(0)->bdata
= &node0_bdata
;
545 /* Don't let boot memory alias the PCI region. */
546 last_alloc_pfn
= min(max_low_pfn
, pci_reserve_start_pfn
);
548 last_alloc_pfn
= max_low_pfn
;
552 * Initialize the boot-time allocator (with low memory only):
553 * The first argument says where to put the bitmap, and the
554 * second says where the end of allocatable memory is.
556 bootmap_size
= init_bootmem(min_low_pfn
, last_alloc_pfn
);
559 * Let the bootmem allocator use all the space we've given it
560 * except for its own bitmap.
562 first_alloc_pfn
= min_low_pfn
+ PFN_UP(bootmap_size
);
563 if (first_alloc_pfn
>= last_alloc_pfn
)
564 early_panic("Not enough memory on controller 0 for bootmem\n");
566 free_bootmem(PFN_PHYS(first_alloc_pfn
),
567 PFN_PHYS(last_alloc_pfn
- first_alloc_pfn
));
570 if (crashk_res
.start
!= crashk_res
.end
)
571 reserve_bootmem(crashk_res
.start
, resource_size(&crashk_res
), 0);
575 void *__init
alloc_remap(int nid
, unsigned long size
)
577 int pages
= node_end_pfn
[nid
] - node_start_pfn
[nid
];
578 void *map
= pfn_to_kaddr(node_memmap_pfn
[nid
]);
579 BUG_ON(size
!= pages
* sizeof(struct page
));
580 memset(map
, 0, size
);
584 static int __init
percpu_size(void)
586 int size
= __per_cpu_end
- __per_cpu_start
;
587 size
+= PERCPU_MODULE_RESERVE
;
588 size
+= PERCPU_DYNAMIC_EARLY_SIZE
;
589 if (size
< PCPU_MIN_UNIT_SIZE
)
590 size
= PCPU_MIN_UNIT_SIZE
;
591 size
= roundup(size
, PAGE_SIZE
);
593 /* In several places we assume the per-cpu data fits on a huge page. */
594 BUG_ON(kdata_huge
&& size
> HPAGE_SIZE
);
598 static inline unsigned long alloc_bootmem_pfn(int size
, unsigned long goal
)
600 void *kva
= __alloc_bootmem(size
, PAGE_SIZE
, goal
);
601 unsigned long pfn
= kaddr_to_pfn(kva
);
602 BUG_ON(goal
&& PFN_PHYS(pfn
) != goal
);
606 static void __init
zone_sizes_init(void)
608 unsigned long zones_size
[MAX_NR_ZONES
] = { 0 };
609 int size
= percpu_size();
610 int num_cpus
= smp_height
* smp_width
;
613 for (i
= 0; i
< num_cpus
; ++i
)
614 node_percpu
[cpu_to_node(i
)] += size
;
616 for_each_online_node(i
) {
617 unsigned long start
= node_start_pfn
[i
];
618 unsigned long end
= node_end_pfn
[i
];
619 #ifdef CONFIG_HIGHMEM
620 unsigned long lowmem_end
= node_lowmem_end_pfn
[i
];
622 unsigned long lowmem_end
= end
;
624 int memmap_size
= (end
- start
) * sizeof(struct page
);
625 node_free_pfn
[i
] = start
;
628 * Set aside pages for per-cpu data and the mem_map array.
630 * Since the per-cpu data requires special homecaching,
631 * if we are in kdata_huge mode, we put it at the end of
632 * the lowmem region. If we're not in kdata_huge mode,
633 * we take the per-cpu pages from the bottom of the
634 * controller, since that avoids fragmenting a huge page
635 * that users might want. We always take the memmap
636 * from the bottom of the controller, since with
637 * kdata_huge that lets it be under a huge TLB entry.
639 * If the user has requested isolnodes for a controller,
640 * though, there'll be no lowmem, so we just alloc_bootmem
641 * the memmap. There will be no percpu memory either.
643 if (__pfn_to_highbits(start
) == 0) {
644 /* In low PAs, allocate via bootmem. */
645 unsigned long goal
= 0;
647 alloc_bootmem_pfn(memmap_size
, goal
);
649 goal
= PFN_PHYS(lowmem_end
) - node_percpu
[i
];
652 alloc_bootmem_pfn(node_percpu
[i
], goal
);
653 } else if (cpu_isset(i
, isolnodes
)) {
654 node_memmap_pfn
[i
] = alloc_bootmem_pfn(memmap_size
, 0);
655 BUG_ON(node_percpu
[i
] != 0);
657 /* In high PAs, just reserve some pages. */
658 node_memmap_pfn
[i
] = node_free_pfn
[i
];
659 node_free_pfn
[i
] += PFN_UP(memmap_size
);
661 node_percpu_pfn
[i
] = node_free_pfn
[i
];
662 node_free_pfn
[i
] += PFN_UP(node_percpu
[i
]);
665 lowmem_end
- PFN_UP(node_percpu
[i
]);
669 #ifdef CONFIG_HIGHMEM
670 if (start
> lowmem_end
) {
671 zones_size
[ZONE_NORMAL
] = 0;
672 zones_size
[ZONE_HIGHMEM
] = end
- start
;
674 zones_size
[ZONE_NORMAL
] = lowmem_end
- start
;
675 zones_size
[ZONE_HIGHMEM
] = end
- lowmem_end
;
678 zones_size
[ZONE_NORMAL
] = end
- start
;
682 * Everyone shares node 0's bootmem allocator, but
683 * we use alloc_remap(), above, to put the actual
684 * struct page array on the individual controllers,
685 * which is most of the data that we actually care about.
686 * We can't place bootmem allocators on the other
687 * controllers since the bootmem allocator can only
688 * operate on 32-bit physical addresses.
690 NODE_DATA(i
)->bdata
= NODE_DATA(0)->bdata
;
692 free_area_init_node(i
, zones_size
, start
, NULL
);
693 printk(KERN_DEBUG
" Normal zone: %ld per-cpu pages\n",
694 PFN_UP(node_percpu
[i
]));
696 /* Track the type of memory on each node */
697 if (zones_size
[ZONE_NORMAL
])
698 node_set_state(i
, N_NORMAL_MEMORY
);
699 #ifdef CONFIG_HIGHMEM
701 node_set_state(i
, N_HIGH_MEMORY
);
710 /* which logical CPUs are on which nodes */
711 struct cpumask node_2_cpu_mask
[MAX_NUMNODES
] __write_once
;
712 EXPORT_SYMBOL(node_2_cpu_mask
);
714 /* which node each logical CPU is on */
715 char cpu_2_node
[NR_CPUS
] __write_once
__attribute__((aligned(L2_CACHE_BYTES
)));
716 EXPORT_SYMBOL(cpu_2_node
);
718 /* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
719 static int __init
cpu_to_bound_node(int cpu
, struct cpumask
* unbound_cpus
)
721 if (!cpu_possible(cpu
) || cpumask_test_cpu(cpu
, unbound_cpus
))
724 return cpu_to_node(cpu
);
727 /* Return number of immediately-adjacent tiles sharing the same NUMA node. */
728 static int __init
node_neighbors(int node
, int cpu
,
729 struct cpumask
*unbound_cpus
)
736 if (x
> 0 && cpu_to_bound_node(cpu
-1, unbound_cpus
) == node
)
738 if (x
< w
-1 && cpu_to_bound_node(cpu
+1, unbound_cpus
) == node
)
740 if (y
> 0 && cpu_to_bound_node(cpu
-w
, unbound_cpus
) == node
)
742 if (y
< h
-1 && cpu_to_bound_node(cpu
+w
, unbound_cpus
) == node
)
747 static void __init
setup_numa_mapping(void)
749 int distance
[MAX_NUMNODES
][NR_CPUS
];
751 int cpu
, node
, cpus
, i
, x
, y
;
752 int num_nodes
= num_online_nodes();
753 struct cpumask unbound_cpus
;
754 nodemask_t default_nodes
;
756 cpumask_clear(&unbound_cpus
);
758 /* Get set of nodes we will use for defaults */
759 nodes_andnot(default_nodes
, node_online_map
, isolnodes
);
760 if (nodes_empty(default_nodes
)) {
761 BUG_ON(!node_isset(0, node_online_map
));
762 pr_err("Forcing NUMA node zero available as a default node\n");
763 node_set(0, default_nodes
);
766 /* Populate the distance[] array */
767 memset(distance
, -1, sizeof(distance
));
769 for (coord
.y
= 0; coord
.y
< smp_height
; ++coord
.y
) {
770 for (coord
.x
= 0; coord
.x
< smp_width
;
772 BUG_ON(cpu
>= nr_cpu_ids
);
773 if (!cpu_possible(cpu
)) {
774 cpu_2_node
[cpu
] = -1;
777 for_each_node_mask(node
, default_nodes
) {
778 HV_MemoryControllerInfo info
=
779 hv_inquire_memory_controller(
780 coord
, node_controller
[node
]);
781 distance
[node
][cpu
] =
782 ABS(info
.coord
.x
) + ABS(info
.coord
.y
);
784 cpumask_set_cpu(cpu
, &unbound_cpus
);
790 * Round-robin through the NUMA nodes until all the cpus are
791 * assigned. We could be more clever here (e.g. create four
792 * sorted linked lists on the same set of cpu nodes, and pull
793 * off them in round-robin sequence, removing from all four
794 * lists each time) but given the relatively small numbers
795 * involved, O(n^2) seem OK for a one-time cost.
797 node
= first_node(default_nodes
);
798 while (!cpumask_empty(&unbound_cpus
)) {
800 int best_distance
= INT_MAX
;
801 for (cpu
= 0; cpu
< cpus
; ++cpu
) {
802 if (cpumask_test_cpu(cpu
, &unbound_cpus
)) {
804 * Compute metric, which is how much
805 * closer the cpu is to this memory
806 * controller than the others, shifted
807 * up, and then the number of
808 * neighbors already in the node as an
809 * epsilon adjustment to try to keep
812 int d
= distance
[node
][cpu
] * num_nodes
;
813 for_each_node_mask(i
, default_nodes
) {
815 d
-= distance
[i
][cpu
];
817 d
*= 8; /* allow space for epsilon */
818 d
-= node_neighbors(node
, cpu
, &unbound_cpus
);
819 if (d
< best_distance
) {
825 BUG_ON(best_cpu
< 0);
826 cpumask_set_cpu(best_cpu
, &node_2_cpu_mask
[node
]);
827 cpu_2_node
[best_cpu
] = node
;
828 cpumask_clear_cpu(best_cpu
, &unbound_cpus
);
829 node
= next_node(node
, default_nodes
);
830 if (node
== MAX_NUMNODES
)
831 node
= first_node(default_nodes
);
834 /* Print out node assignments and set defaults for disabled cpus */
836 for (y
= 0; y
< smp_height
; ++y
) {
837 printk(KERN_DEBUG
"NUMA cpu-to-node row %d:", y
);
838 for (x
= 0; x
< smp_width
; ++x
, ++cpu
) {
839 if (cpu_to_node(cpu
) < 0) {
841 cpu_2_node
[cpu
] = first_node(default_nodes
);
843 pr_cont(" %d", cpu_to_node(cpu
));
850 static struct cpu cpu_devices
[NR_CPUS
];
852 static int __init
topology_init(void)
856 for_each_online_node(i
)
857 register_one_node(i
);
859 for (i
= 0; i
< smp_height
* smp_width
; ++i
)
860 register_cpu(&cpu_devices
[i
], i
);
865 subsys_initcall(topology_init
);
867 #else /* !CONFIG_NUMA */
869 #define setup_numa_mapping() do { } while (0)
871 #endif /* CONFIG_NUMA */
874 * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
875 * @boot: Is this the boot cpu?
877 * Called from setup_arch() on the boot cpu, or online_secondary().
879 void __cpuinit
setup_cpu(int boot
)
881 /* The boot cpu sets up its permanent mappings much earlier. */
883 store_permanent_mappings();
885 /* Allow asynchronous TLB interrupts. */
886 #if CHIP_HAS_TILE_DMA()
887 arch_local_irq_unmask(INT_DMATLB_MISS
);
888 arch_local_irq_unmask(INT_DMATLB_ACCESS
);
890 #if CHIP_HAS_SN_PROC()
891 arch_local_irq_unmask(INT_SNITLB_MISS
);
894 arch_local_irq_unmask(INT_SINGLE_STEP_K
);
898 * Allow user access to many generic SPRs, like the cycle
899 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
901 __insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0
, 1);
904 /* Static network is not restricted. */
905 __insn_mtspr(SPR_MPL_SN_ACCESS_SET_0
, 1);
907 #if CHIP_HAS_SN_PROC()
908 __insn_mtspr(SPR_MPL_SN_NOTIFY_SET_0
, 1);
909 __insn_mtspr(SPR_MPL_SN_CPL_SET_0
, 1);
913 * Set the MPL for interrupt control 0 & 1 to the corresponding
914 * values. This includes access to the SYSTEM_SAVE and EX_CONTEXT
915 * SPRs, as well as the interrupt mask.
917 __insn_mtspr(SPR_MPL_INTCTRL_0_SET_0
, 1);
918 __insn_mtspr(SPR_MPL_INTCTRL_1_SET_1
, 1);
920 /* Initialize IRQ support for this cpu. */
923 #ifdef CONFIG_HARDWALL
924 /* Reset the network state on this cpu. */
925 reset_network_state();
929 #ifdef CONFIG_BLK_DEV_INITRD
932 * Note that the kernel can potentially support other compression
933 * techniques than gz, though we don't do so by default. If we ever
934 * decide to do so we can either look for other filename extensions,
935 * or just allow a file with this name to be compressed with an
936 * arbitrary compressor (somewhat counterintuitively).
938 static int __initdata set_initramfs_file
;
939 static char __initdata initramfs_file
[128] = "initramfs.cpio.gz";
941 static int __init
setup_initramfs_file(char *str
)
945 strncpy(initramfs_file
, str
, sizeof(initramfs_file
) - 1);
946 set_initramfs_file
= 1;
950 early_param("initramfs_file", setup_initramfs_file
);
953 * We look for an "initramfs.cpio.gz" file in the hvfs.
954 * If there is one, we allocate some memory for it and it will be
955 * unpacked to the initramfs.
957 static void __init
load_hv_initrd(void)
963 fd
= hv_fs_findfile((HV_VirtAddr
) initramfs_file
);
964 if (fd
== HV_ENOENT
) {
965 if (set_initramfs_file
)
966 pr_warning("No such hvfs initramfs file '%s'\n",
971 stat
= hv_fs_fstat(fd
);
972 BUG_ON(stat
.size
< 0);
973 if (stat
.flags
& HV_FS_ISDIR
) {
974 pr_warning("Ignoring hvfs file '%s': it's a directory.\n",
978 initrd
= alloc_bootmem_pages(stat
.size
);
979 rc
= hv_fs_pread(fd
, (HV_VirtAddr
) initrd
, stat
.size
, 0);
980 if (rc
!= stat
.size
) {
981 pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
982 stat
.size
, initramfs_file
, rc
);
983 free_initrd_mem((unsigned long) initrd
, stat
.size
);
986 initrd_start
= (unsigned long) initrd
;
987 initrd_end
= initrd_start
+ stat
.size
;
990 void __init
free_initrd_mem(unsigned long begin
, unsigned long end
)
992 free_bootmem(__pa(begin
), end
- begin
);
996 static inline void load_hv_initrd(void) {}
997 #endif /* CONFIG_BLK_DEV_INITRD */
999 static void __init
validate_hv(void)
1002 * It may already be too late, but let's check our built-in
1003 * configuration against what the hypervisor is providing.
1005 unsigned long glue_size
= hv_sysconf(HV_SYSCONF_GLUE_SIZE
);
1006 int hv_page_size
= hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL
);
1007 int hv_hpage_size
= hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE
);
1008 HV_ASIDRange asid_range
;
1011 HV_Topology topology
= hv_inquire_topology();
1012 BUG_ON(topology
.coord
.x
!= 0 || topology
.coord
.y
!= 0);
1013 if (topology
.width
!= 1 || topology
.height
!= 1) {
1014 pr_warning("Warning: booting UP kernel on %dx%d grid;"
1015 " will ignore all but first tile.\n",
1016 topology
.width
, topology
.height
);
1020 if (PAGE_OFFSET
+ HV_GLUE_START_CPA
+ glue_size
> (unsigned long)_text
)
1021 early_panic("Hypervisor glue size %ld is too big!\n",
1023 if (hv_page_size
!= PAGE_SIZE
)
1024 early_panic("Hypervisor page size %#x != our %#lx\n",
1025 hv_page_size
, PAGE_SIZE
);
1026 if (hv_hpage_size
!= HPAGE_SIZE
)
1027 early_panic("Hypervisor huge page size %#x != our %#lx\n",
1028 hv_hpage_size
, HPAGE_SIZE
);
1032 * Some hypervisor APIs take a pointer to a bitmap array
1033 * whose size is at least the number of cpus on the chip.
1034 * We use a struct cpumask for this, so it must be big enough.
1036 if ((smp_height
* smp_width
) > nr_cpu_ids
)
1037 early_panic("Hypervisor %d x %d grid too big for Linux"
1038 " NR_CPUS %d\n", smp_height
, smp_width
,
1043 * Check that we're using allowed ASIDs, and initialize the
1044 * various asid variables to their appropriate initial states.
1046 asid_range
= hv_inquire_asid(0);
1047 __get_cpu_var(current_asid
) = min_asid
= asid_range
.start
;
1048 max_asid
= asid_range
.start
+ asid_range
.size
- 1;
1050 if (hv_confstr(HV_CONFSTR_CHIP_MODEL
, (HV_VirtAddr
)chip_model
,
1051 sizeof(chip_model
)) < 0) {
1052 pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
1053 strlcpy(chip_model
, "unknown", sizeof(chip_model
));
1057 static void __init
validate_va(void)
1059 #ifndef __tilegx__ /* FIXME: GX: probably some validation relevant here */
1061 * Similarly, make sure we're only using allowed VAs.
1062 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_INTRPT,
1063 * and 0 .. KERNEL_HIGH_VADDR.
1064 * In addition, make sure we CAN'T use the end of memory, since
1065 * we use the last chunk of each pgd for the pgd_list.
1067 int i
, user_kernel_ok
= 0;
1068 unsigned long max_va
= 0;
1069 unsigned long list_va
=
1070 ((PGD_LIST_OFFSET
/ sizeof(pgd_t
)) << PGDIR_SHIFT
);
1072 for (i
= 0; ; ++i
) {
1073 HV_VirtAddrRange range
= hv_inquire_virtual(i
);
1074 if (range
.size
== 0)
1076 if (range
.start
<= MEM_USER_INTRPT
&&
1077 range
.start
+ range
.size
>= MEM_HV_INTRPT
)
1079 if (range
.start
== 0)
1080 max_va
= range
.size
;
1081 BUG_ON(range
.start
+ range
.size
> list_va
);
1083 if (!user_kernel_ok
)
1084 early_panic("Hypervisor not configured for user/kernel VAs\n");
1086 early_panic("Hypervisor not configured for low VAs\n");
1087 if (max_va
< KERNEL_HIGH_VADDR
)
1088 early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
1089 max_va
, KERNEL_HIGH_VADDR
);
1091 /* Kernel PCs must have their high bit set; see intvec.S. */
1092 if ((long)VMALLOC_START
>= 0)
1094 "Linux VMALLOC region below the 2GB line (%#lx)!\n"
1095 "Reconfigure the kernel with fewer NR_HUGE_VMAPS\n"
1096 "or smaller VMALLOC_RESERVE.\n",
1102 * cpu_lotar_map lists all the cpus that are valid for the supervisor
1103 * to cache data on at a page level, i.e. what cpus can be placed in
1104 * the LOTAR field of a PTE. It is equivalent to the set of possible
1105 * cpus plus any other cpus that are willing to share their cache.
1106 * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
1108 struct cpumask __write_once cpu_lotar_map
;
1109 EXPORT_SYMBOL(cpu_lotar_map
);
1111 #if CHIP_HAS_CBOX_HOME_MAP()
1113 * hash_for_home_map lists all the tiles that hash-for-home data
1114 * will be cached on. Note that this may includes tiles that are not
1115 * valid for this supervisor to use otherwise (e.g. if a hypervisor
1116 * device is being shared between multiple supervisors).
1117 * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
1119 struct cpumask hash_for_home_map
;
1120 EXPORT_SYMBOL(hash_for_home_map
);
1124 * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
1125 * flush on our behalf. It is set to cpu_possible_mask OR'ed with
1126 * hash_for_home_map, and it is what should be passed to
1127 * hv_flush_remote() to flush all caches. Note that if there are
1128 * dedicated hypervisor driver tiles that have authorized use of their
1129 * cache, those tiles will only appear in cpu_lotar_map, NOT in
1130 * cpu_cacheable_map, as they are a special case.
1132 struct cpumask __write_once cpu_cacheable_map
;
1133 EXPORT_SYMBOL(cpu_cacheable_map
);
1135 static __initdata
struct cpumask disabled_map
;
1137 static int __init
disabled_cpus(char *str
)
1139 int boot_cpu
= smp_processor_id();
1141 if (str
== NULL
|| cpulist_parse_crop(str
, &disabled_map
) != 0)
1143 if (cpumask_test_cpu(boot_cpu
, &disabled_map
)) {
1144 pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu
);
1145 cpumask_clear_cpu(boot_cpu
, &disabled_map
);
1150 early_param("disabled_cpus", disabled_cpus
);
1152 void __init
print_disabled_cpus(void)
1154 if (!cpumask_empty(&disabled_map
)) {
1156 cpulist_scnprintf(buf
, sizeof(buf
), &disabled_map
);
1157 pr_info("CPUs not available for Linux: %s\n", buf
);
1161 static void __init
setup_cpu_maps(void)
1163 struct cpumask hv_disabled_map
, cpu_possible_init
;
1164 int boot_cpu
= smp_processor_id();
1167 /* Learn which cpus are allowed by the hypervisor. */
1168 rc
= hv_inquire_tiles(HV_INQ_TILES_AVAIL
,
1169 (HV_VirtAddr
) cpumask_bits(&cpu_possible_init
),
1170 sizeof(cpu_cacheable_map
));
1172 early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc
);
1173 if (!cpumask_test_cpu(boot_cpu
, &cpu_possible_init
))
1174 early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu
);
1176 /* Compute the cpus disabled by the hvconfig file. */
1177 cpumask_complement(&hv_disabled_map
, &cpu_possible_init
);
1179 /* Include them with the cpus disabled by "disabled_cpus". */
1180 cpumask_or(&disabled_map
, &disabled_map
, &hv_disabled_map
);
1183 * Disable every cpu after "setup_max_cpus". But don't mark
1184 * as disabled the cpus that are outside of our initial rectangle,
1185 * since that turns out to be confusing.
1187 cpus
= 1; /* this cpu */
1188 cpumask_set_cpu(boot_cpu
, &disabled_map
); /* ignore this cpu */
1189 for (i
= 0; cpus
< setup_max_cpus
; ++i
)
1190 if (!cpumask_test_cpu(i
, &disabled_map
))
1192 for (; i
< smp_height
* smp_width
; ++i
)
1193 cpumask_set_cpu(i
, &disabled_map
);
1194 cpumask_clear_cpu(boot_cpu
, &disabled_map
); /* reset this cpu */
1195 for (i
= smp_height
* smp_width
; i
< NR_CPUS
; ++i
)
1196 cpumask_clear_cpu(i
, &disabled_map
);
1199 * Setup cpu_possible map as every cpu allocated to us, minus
1200 * the results of any "disabled_cpus" settings.
1202 cpumask_andnot(&cpu_possible_init
, &cpu_possible_init
, &disabled_map
);
1203 init_cpu_possible(&cpu_possible_init
);
1205 /* Learn which cpus are valid for LOTAR caching. */
1206 rc
= hv_inquire_tiles(HV_INQ_TILES_LOTAR
,
1207 (HV_VirtAddr
) cpumask_bits(&cpu_lotar_map
),
1208 sizeof(cpu_lotar_map
));
1210 pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
1211 cpu_lotar_map
= *cpu_possible_mask
;
1214 #if CHIP_HAS_CBOX_HOME_MAP()
1215 /* Retrieve set of CPUs used for hash-for-home caching */
1216 rc
= hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE
,
1217 (HV_VirtAddr
) hash_for_home_map
.bits
,
1218 sizeof(hash_for_home_map
));
1220 early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc
);
1221 cpumask_or(&cpu_cacheable_map
, cpu_possible_mask
, &hash_for_home_map
);
1223 cpu_cacheable_map
= *cpu_possible_mask
;
1228 static int __init
dataplane(char *str
)
1230 pr_warning("WARNING: dataplane support disabled in this kernel\n");
1234 early_param("dataplane", dataplane
);
1236 #ifdef CONFIG_CMDLINE_BOOL
1237 static char __initdata builtin_cmdline
[COMMAND_LINE_SIZE
] = CONFIG_CMDLINE
;
1240 void __init
setup_arch(char **cmdline_p
)
1244 #if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
1245 len
= hv_get_command_line((HV_VirtAddr
) boot_command_line
,
1247 if (boot_command_line
[0])
1248 pr_warning("WARNING: ignoring dynamic command line \"%s\"\n",
1250 strlcpy(boot_command_line
, builtin_cmdline
, COMMAND_LINE_SIZE
);
1253 #if defined(CONFIG_CMDLINE_BOOL)
1254 if (builtin_cmdline
[0]) {
1255 int builtin_len
= strlcpy(boot_command_line
, builtin_cmdline
,
1257 if (builtin_len
< COMMAND_LINE_SIZE
-1)
1258 boot_command_line
[builtin_len
++] = ' ';
1259 hv_cmdline
= &boot_command_line
[builtin_len
];
1260 len
= COMMAND_LINE_SIZE
- builtin_len
;
1264 hv_cmdline
= boot_command_line
;
1265 len
= COMMAND_LINE_SIZE
;
1267 len
= hv_get_command_line((HV_VirtAddr
) hv_cmdline
, len
);
1268 if (len
< 0 || len
> COMMAND_LINE_SIZE
)
1269 early_panic("hv_get_command_line failed: %d\n", len
);
1272 *cmdline_p
= boot_command_line
;
1274 /* Set disabled_map and setup_max_cpus very early */
1275 parse_early_param();
1277 /* Make sure the kernel is compatible with the hypervisor. */
1286 * Initialize the PCI structures. This is done before memory
1287 * setup so that we know whether or not a pci_reserve region
1290 if (tile_pci_init() == 0)
1293 /* PCI systems reserve a region just below 4GB for mapping iomem. */
1294 pci_reserve_end_pfn
= (1 << (32 - PAGE_SHIFT
));
1295 pci_reserve_start_pfn
= pci_reserve_end_pfn
-
1296 (pci_reserve_mb
<< (20 - PAGE_SHIFT
));
1299 init_mm
.start_code
= (unsigned long) _text
;
1300 init_mm
.end_code
= (unsigned long) _etext
;
1301 init_mm
.end_data
= (unsigned long) _edata
;
1302 init_mm
.brk
= (unsigned long) _end
;
1305 store_permanent_mappings();
1306 setup_bootmem_allocator();
1309 * NOTE: before this point _nobody_ is allowed to allocate
1310 * any memory using the bootmem allocator.
1314 setup_numa_mapping();
1324 * Set up per-cpu memory.
1327 unsigned long __per_cpu_offset
[NR_CPUS
] __write_once
;
1328 EXPORT_SYMBOL(__per_cpu_offset
);
1330 static size_t __initdata pfn_offset
[MAX_NUMNODES
] = { 0 };
1331 static unsigned long __initdata percpu_pfn
[NR_CPUS
] = { 0 };
1334 * As the percpu code allocates pages, we return the pages from the
1335 * end of the node for the specified cpu.
1337 static void *__init
pcpu_fc_alloc(unsigned int cpu
, size_t size
, size_t align
)
1339 int nid
= cpu_to_node(cpu
);
1340 unsigned long pfn
= node_percpu_pfn
[nid
] + pfn_offset
[nid
];
1342 BUG_ON(size
% PAGE_SIZE
!= 0);
1343 pfn_offset
[nid
] += size
/ PAGE_SIZE
;
1344 BUG_ON(node_percpu
[nid
] < size
);
1345 node_percpu
[nid
] -= size
;
1346 if (percpu_pfn
[cpu
] == 0)
1347 percpu_pfn
[cpu
] = pfn
;
1348 return pfn_to_kaddr(pfn
);
1352 * Pages reserved for percpu memory are not freeable, and in any case we are
1353 * on a short path to panic() in setup_per_cpu_area() at this point anyway.
1355 static void __init
pcpu_fc_free(void *ptr
, size_t size
)
1360 * Set up vmalloc page tables using bootmem for the percpu code.
1362 static void __init
pcpu_fc_populate_pte(unsigned long addr
)
1369 BUG_ON(pgd_addr_invalid(addr
));
1370 if (addr
< VMALLOC_START
|| addr
>= VMALLOC_END
)
1371 panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx;"
1372 " try increasing CONFIG_VMALLOC_RESERVE\n",
1373 addr
, VMALLOC_START
, VMALLOC_END
);
1375 pgd
= swapper_pg_dir
+ pgd_index(addr
);
1376 pud
= pud_offset(pgd
, addr
);
1377 BUG_ON(!pud_present(*pud
));
1378 pmd
= pmd_offset(pud
, addr
);
1379 if (pmd_present(*pmd
)) {
1380 BUG_ON(pmd_huge_page(*pmd
));
1382 pte
= __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE
,
1383 HV_PAGE_TABLE_ALIGN
, 0);
1384 pmd_populate_kernel(&init_mm
, pmd
, pte
);
1388 void __init
setup_per_cpu_areas(void)
1391 unsigned long delta
, pfn
, lowmem_va
;
1392 unsigned long size
= percpu_size();
1396 rc
= pcpu_page_first_chunk(PERCPU_MODULE_RESERVE
, pcpu_fc_alloc
,
1397 pcpu_fc_free
, pcpu_fc_populate_pte
);
1399 panic("Cannot initialize percpu area (err=%d)", rc
);
1401 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
1402 for_each_possible_cpu(cpu
) {
1403 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
1405 /* finv the copy out of cache so we can change homecache */
1406 ptr
= pcpu_base_addr
+ pcpu_unit_offsets
[cpu
];
1407 __finv_buffer(ptr
, size
);
1408 pfn
= percpu_pfn
[cpu
];
1410 /* Rewrite the page tables to cache on that cpu */
1411 pg
= pfn_to_page(pfn
);
1412 for (i
= 0; i
< size
; i
+= PAGE_SIZE
, ++pfn
, ++pg
) {
1414 /* Update the vmalloc mapping and page home. */
1416 virt_to_pte(NULL
, (unsigned long)ptr
+ i
);
1418 BUG_ON(pfn
!= pte_pfn(pte
));
1419 pte
= hv_pte_set_mode(pte
, HV_PTE_MODE_CACHE_TILE_L3
);
1420 pte
= set_remote_cache_cpu(pte
, cpu
);
1423 /* Update the lowmem mapping for consistency. */
1424 lowmem_va
= (unsigned long)pfn_to_kaddr(pfn
);
1425 ptep
= virt_to_pte(NULL
, lowmem_va
);
1426 if (pte_huge(*ptep
)) {
1427 printk(KERN_DEBUG
"early shatter of huge page"
1428 " at %#lx\n", lowmem_va
);
1429 shatter_pmd((pmd_t
*)ptep
);
1430 ptep
= virt_to_pte(NULL
, lowmem_va
);
1431 BUG_ON(pte_huge(*ptep
));
1433 BUG_ON(pfn
!= pte_pfn(*ptep
));
1438 /* Set our thread pointer appropriately. */
1439 set_my_cpu_offset(__per_cpu_offset
[smp_processor_id()]);
1441 /* Make sure the finv's have completed. */
1444 /* Flush the TLB so we reference it properly from here on out. */
1445 local_flush_tlb_all();
1448 static struct resource data_resource
= {
1449 .name
= "Kernel data",
1452 .flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
1455 static struct resource code_resource
= {
1456 .name
= "Kernel code",
1459 .flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
1463 * We reserve all resources above 4GB so that PCI won't try to put
1464 * mappings above 4GB; the standard allows that for some devices but
1465 * the probing code trunates values to 32 bits.
1468 static struct resource
* __init
1469 insert_non_bus_resource(void)
1471 struct resource
*res
=
1472 kzalloc(sizeof(struct resource
), GFP_ATOMIC
);
1473 res
->name
= "Non-Bus Physical Address Space";
1474 res
->start
= (1ULL << 32);
1476 res
->flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
;
1477 if (insert_resource(&iomem_resource
, res
)) {
1485 static struct resource
* __init
1486 insert_ram_resource(u64 start_pfn
, u64 end_pfn
)
1488 struct resource
*res
=
1489 kzalloc(sizeof(struct resource
), GFP_ATOMIC
);
1490 res
->name
= "System RAM";
1491 res
->start
= start_pfn
<< PAGE_SHIFT
;
1492 res
->end
= (end_pfn
<< PAGE_SHIFT
) - 1;
1493 res
->flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
;
1494 if (insert_resource(&iomem_resource
, res
)) {
1502 * Request address space for all standard resources
1504 * If the system includes PCI root complex drivers, we need to create
1505 * a window just below 4GB where PCI BARs can be mapped.
1507 static int __init
request_standard_resources(void)
1510 enum { CODE_DELTA
= MEM_SV_INTRPT
- PAGE_OFFSET
};
1512 iomem_resource
.end
= -1LL;
1514 insert_non_bus_resource();
1517 for_each_online_node(i
) {
1518 u64 start_pfn
= node_start_pfn
[i
];
1519 u64 end_pfn
= node_end_pfn
[i
];
1522 if (start_pfn
<= pci_reserve_start_pfn
&&
1523 end_pfn
> pci_reserve_start_pfn
) {
1524 if (end_pfn
> pci_reserve_end_pfn
)
1525 insert_ram_resource(pci_reserve_end_pfn
,
1527 end_pfn
= pci_reserve_start_pfn
;
1530 insert_ram_resource(start_pfn
, end_pfn
);
1533 code_resource
.start
= __pa(_text
- CODE_DELTA
);
1534 code_resource
.end
= __pa(_etext
- CODE_DELTA
)-1;
1535 data_resource
.start
= __pa(_sdata
);
1536 data_resource
.end
= __pa(_end
)-1;
1538 insert_resource(&iomem_resource
, &code_resource
);
1539 insert_resource(&iomem_resource
, &data_resource
);
1542 insert_resource(&iomem_resource
, &crashk_res
);
1548 subsys_initcall(request_standard_resources
);