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/swiotlb.h>
27 #include <linux/initrd.h>
29 #include <linux/highmem.h>
30 #include <linux/smp.h>
31 #include <linux/timex.h>
32 #include <linux/hugetlb.h>
33 #include <linux/start_kernel.h>
34 #include <linux/screen_info.h>
35 #include <asm/setup.h>
36 #include <asm/sections.h>
37 #include <asm/cacheflush.h>
38 #include <asm/pgalloc.h>
39 #include <asm/mmu_context.h>
40 #include <hv/hypervisor.h>
41 #include <arch/interrupts.h>
43 /* <linux/smp.h> doesn't provide this definition. */
45 #define setup_max_cpus 1
48 static inline int ABS(int x
) { return x
>= 0 ? x
: -x
; }
50 /* Chip information */
51 char chip_model
[64] __write_once
;
54 struct screen_info screen_info
;
57 struct pglist_data node_data
[MAX_NUMNODES
] __read_mostly
;
58 EXPORT_SYMBOL(node_data
);
60 /* Information on the NUMA nodes that we compute early */
61 unsigned long __cpuinitdata node_start_pfn
[MAX_NUMNODES
];
62 unsigned long __cpuinitdata node_end_pfn
[MAX_NUMNODES
];
63 unsigned long __initdata node_memmap_pfn
[MAX_NUMNODES
];
64 unsigned long __initdata node_percpu_pfn
[MAX_NUMNODES
];
65 unsigned long __initdata node_free_pfn
[MAX_NUMNODES
];
67 static unsigned long __initdata node_percpu
[MAX_NUMNODES
];
70 * per-CPU stack and boot info.
72 DEFINE_PER_CPU(unsigned long, boot_sp
) =
73 (unsigned long)init_stack
+ THREAD_SIZE
;
76 DEFINE_PER_CPU(unsigned long, boot_pc
) = (unsigned long)start_kernel
;
79 * The variable must be __initdata since it references __init code.
80 * With CONFIG_SMP it is per-cpu data, which is exempt from validation.
82 unsigned long __initdata boot_pc
= (unsigned long)start_kernel
;
86 /* Page frame index of end of lowmem on each controller. */
87 unsigned long __cpuinitdata node_lowmem_end_pfn
[MAX_NUMNODES
];
89 /* Number of pages that can be mapped into lowmem. */
90 static unsigned long __initdata mappable_physpages
;
93 /* Data on which physical memory controller corresponds to which NUMA node */
94 int node_controller
[MAX_NUMNODES
] = { [0 ... MAX_NUMNODES
-1] = -1 };
97 /* Map information from VAs to PAs */
98 unsigned long pbase_map
[1 << (32 - HPAGE_SHIFT
)]
99 __write_once
__attribute__((aligned(L2_CACHE_BYTES
)));
100 EXPORT_SYMBOL(pbase_map
);
102 /* Map information from PAs to VAs */
103 void *vbase_map
[NR_PA_HIGHBIT_VALUES
]
104 __write_once
__attribute__((aligned(L2_CACHE_BYTES
)));
105 EXPORT_SYMBOL(vbase_map
);
108 /* Node number as a function of the high PA bits */
109 int highbits_to_node
[NR_PA_HIGHBIT_VALUES
] __write_once
;
110 EXPORT_SYMBOL(highbits_to_node
);
112 static unsigned int __initdata maxmem_pfn
= -1U;
113 static unsigned int __initdata maxnodemem_pfn
[MAX_NUMNODES
] = {
114 [0 ... MAX_NUMNODES
-1] = -1U
116 static nodemask_t __initdata isolnodes
;
118 #if defined(CONFIG_PCI) && !defined(__tilegx__)
119 enum { DEFAULT_PCI_RESERVE_MB
= 64 };
120 static unsigned int __initdata pci_reserve_mb
= DEFAULT_PCI_RESERVE_MB
;
121 unsigned long __initdata pci_reserve_start_pfn
= -1U;
122 unsigned long __initdata pci_reserve_end_pfn
= -1U;
125 static int __init
setup_maxmem(char *str
)
127 unsigned long long maxmem
;
128 if (str
== NULL
|| (maxmem
= memparse(str
, NULL
)) == 0)
131 maxmem_pfn
= (maxmem
>> HPAGE_SHIFT
) << (HPAGE_SHIFT
- PAGE_SHIFT
);
132 pr_info("Forcing RAM used to no more than %dMB\n",
133 maxmem_pfn
>> (20 - PAGE_SHIFT
));
136 early_param("maxmem", setup_maxmem
);
138 static int __init
setup_maxnodemem(char *str
)
141 unsigned long long maxnodemem
;
144 node
= str
? simple_strtoul(str
, &endp
, 0) : INT_MAX
;
145 if (node
>= MAX_NUMNODES
|| *endp
!= ':')
148 maxnodemem
= memparse(endp
+1, NULL
);
149 maxnodemem_pfn
[node
] = (maxnodemem
>> HPAGE_SHIFT
) <<
150 (HPAGE_SHIFT
- PAGE_SHIFT
);
151 pr_info("Forcing RAM used on node %ld to no more than %dMB\n",
152 node
, maxnodemem_pfn
[node
] >> (20 - PAGE_SHIFT
));
155 early_param("maxnodemem", setup_maxnodemem
);
157 static int __init
setup_isolnodes(char *str
)
159 char buf
[MAX_NUMNODES
* 5];
160 if (str
== NULL
|| nodelist_parse(str
, isolnodes
) != 0)
163 nodelist_scnprintf(buf
, sizeof(buf
), isolnodes
);
164 pr_info("Set isolnodes value to '%s'\n", buf
);
167 early_param("isolnodes", setup_isolnodes
);
169 #if defined(CONFIG_PCI) && !defined(__tilegx__)
170 static int __init
setup_pci_reserve(char* str
)
174 if (str
== NULL
|| strict_strtoul(str
, 0, &mb
) != 0 ||
179 pr_info("Reserving %dMB for PCIE root complex mappings\n",
183 early_param("pci_reserve", setup_pci_reserve
);
188 * vmalloc=size forces the vmalloc area to be exactly 'size' bytes.
189 * This can be used to increase (or decrease) the vmalloc area.
191 static int __init
parse_vmalloc(char *arg
)
196 VMALLOC_RESERVE
= (memparse(arg
, &arg
) + PGDIR_SIZE
- 1) & PGDIR_MASK
;
198 /* See validate_va() for more on this test. */
199 if ((long)_VMALLOC_START
>= 0)
200 early_panic("\"vmalloc=%#lx\" value too large: maximum %#lx\n",
201 VMALLOC_RESERVE
, _VMALLOC_END
- 0x80000000UL
);
205 early_param("vmalloc", parse_vmalloc
);
208 #ifdef CONFIG_HIGHMEM
210 * Determine for each controller where its lowmem is mapped and how much of
211 * it is mapped there. On controller zero, the first few megabytes are
212 * already mapped in as code at MEM_SV_INTRPT, so in principle we could
213 * start our data mappings higher up, but for now we don't bother, to avoid
214 * additional confusion.
216 * One question is whether, on systems with more than 768 Mb and
217 * controllers of different sizes, to map in a proportionate amount of
218 * each one, or to try to map the same amount from each controller.
219 * (E.g. if we have three controllers with 256MB, 1GB, and 256MB
220 * respectively, do we map 256MB from each, or do we map 128 MB, 512
221 * MB, and 128 MB respectively?) For now we use a proportionate
222 * solution like the latter.
224 * The VA/PA mapping demands that we align our decisions at 16 MB
225 * boundaries so that we can rapidly convert VA to PA.
227 static void *__init
setup_pa_va_mapping(void)
229 unsigned long curr_pages
= 0;
230 unsigned long vaddr
= PAGE_OFFSET
;
231 nodemask_t highonlynodes
= isolnodes
;
234 memset(pbase_map
, -1, sizeof(pbase_map
));
235 memset(vbase_map
, -1, sizeof(vbase_map
));
237 /* Node zero cannot be isolated for LOWMEM purposes. */
238 node_clear(0, highonlynodes
);
240 /* Count up the number of pages on non-highonlynodes controllers. */
241 mappable_physpages
= 0;
242 for_each_online_node(i
) {
243 if (!node_isset(i
, highonlynodes
))
244 mappable_physpages
+=
245 node_end_pfn
[i
] - node_start_pfn
[i
];
248 for_each_online_node(i
) {
249 unsigned long start
= node_start_pfn
[i
];
250 unsigned long end
= node_end_pfn
[i
];
251 unsigned long size
= end
- start
;
252 unsigned long vaddr_end
;
254 if (node_isset(i
, highonlynodes
)) {
255 /* Mark this controller as having no lowmem. */
256 node_lowmem_end_pfn
[i
] = start
;
261 if (mappable_physpages
> MAXMEM_PFN
) {
262 vaddr_end
= PAGE_OFFSET
+
263 (((u64
)curr_pages
* MAXMEM_PFN
/
267 vaddr_end
= PAGE_OFFSET
+ (curr_pages
<< PAGE_SHIFT
);
269 for (j
= 0; vaddr
< vaddr_end
; vaddr
+= HPAGE_SIZE
, ++j
) {
270 unsigned long this_pfn
=
271 start
+ (j
<< HUGETLB_PAGE_ORDER
);
272 pbase_map
[vaddr
>> HPAGE_SHIFT
] = this_pfn
;
273 if (vbase_map
[__pfn_to_highbits(this_pfn
)] ==
275 vbase_map
[__pfn_to_highbits(this_pfn
)] =
276 (void *)(vaddr
& HPAGE_MASK
);
278 node_lowmem_end_pfn
[i
] = start
+ (j
<< HUGETLB_PAGE_ORDER
);
279 BUG_ON(node_lowmem_end_pfn
[i
] > end
);
282 /* Return highest address of any mapped memory. */
283 return (void *)vaddr
;
285 #endif /* CONFIG_HIGHMEM */
288 * Register our most important memory mappings with the debug stub.
290 * This is up to 4 mappings for lowmem, one mapping per memory
291 * controller, plus one for our text segment.
293 static void __cpuinit
store_permanent_mappings(void)
297 for_each_online_node(i
) {
298 HV_PhysAddr pa
= ((HV_PhysAddr
)node_start_pfn
[i
]) << PAGE_SHIFT
;
299 #ifdef CONFIG_HIGHMEM
300 HV_PhysAddr high_mapped_pa
= node_lowmem_end_pfn
[i
];
302 HV_PhysAddr high_mapped_pa
= node_end_pfn
[i
];
305 unsigned long pages
= high_mapped_pa
- node_start_pfn
[i
];
306 HV_VirtAddr addr
= (HV_VirtAddr
) __va(pa
);
307 hv_store_mapping(addr
, pages
<< PAGE_SHIFT
, pa
);
310 hv_store_mapping((HV_VirtAddr
)_text
,
311 (uint32_t)(_einittext
- _text
), 0);
315 * Use hv_inquire_physical() to populate node_{start,end}_pfn[]
316 * and node_online_map, doing suitable sanity-checking.
317 * Also set min_low_pfn, max_low_pfn, and max_pfn.
319 static void __init
setup_memory(void)
322 int highbits_seen
[NR_PA_HIGHBIT_VALUES
] = { 0 };
323 #ifdef CONFIG_HIGHMEM
329 #if defined(CONFIG_HIGHMEM) || defined(__tilegx__)
332 unsigned long physpages
= 0;
334 /* We are using a char to hold the cpu_2_node[] mapping */
335 BUILD_BUG_ON(MAX_NUMNODES
> 127);
337 /* Discover the ranges of memory available to us */
339 unsigned long start
, size
, end
, highbits
;
340 HV_PhysAddrRange range
= hv_inquire_physical(i
);
343 #ifdef CONFIG_FLATMEM
345 pr_err("Can't use discontiguous PAs: %#llx..%#llx\n",
346 range
.size
, range
.start
+ range
.size
);
351 if ((unsigned long)range
.start
) {
352 pr_err("Range not at 4GB multiple: %#llx..%#llx\n",
353 range
.start
, range
.start
+ range
.size
);
357 if ((range
.start
& (HPAGE_SIZE
-1)) != 0 ||
358 (range
.size
& (HPAGE_SIZE
-1)) != 0) {
359 unsigned long long start_pa
= range
.start
;
360 unsigned long long orig_size
= range
.size
;
361 range
.start
= (start_pa
+ HPAGE_SIZE
- 1) & HPAGE_MASK
;
362 range
.size
-= (range
.start
- start_pa
);
363 range
.size
&= HPAGE_MASK
;
364 pr_err("Range not hugepage-aligned: %#llx..%#llx:"
365 " now %#llx-%#llx\n",
366 start_pa
, start_pa
+ orig_size
,
367 range
.start
, range
.start
+ range
.size
);
369 highbits
= __pa_to_highbits(range
.start
);
370 if (highbits
>= NR_PA_HIGHBIT_VALUES
) {
371 pr_err("PA high bits too high: %#llx..%#llx\n",
372 range
.start
, range
.start
+ range
.size
);
375 if (highbits_seen
[highbits
]) {
376 pr_err("Range overlaps in high bits: %#llx..%#llx\n",
377 range
.start
, range
.start
+ range
.size
);
380 highbits_seen
[highbits
] = 1;
381 if (PFN_DOWN(range
.size
) > maxnodemem_pfn
[i
]) {
382 int max_size
= maxnodemem_pfn
[i
];
384 pr_err("Maxnodemem reduced node %d to"
385 " %d pages\n", i
, max_size
);
386 range
.size
= PFN_PHYS(max_size
);
388 pr_err("Maxnodemem disabled node %d\n", i
);
392 if (physpages
+ PFN_DOWN(range
.size
) > maxmem_pfn
) {
393 int max_size
= maxmem_pfn
- physpages
;
395 pr_err("Maxmem reduced node %d to %d pages\n",
397 range
.size
= PFN_PHYS(max_size
);
399 pr_err("Maxmem disabled node %d\n", i
);
403 if (i
>= MAX_NUMNODES
) {
404 pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
405 i
, range
.size
, range
.size
+ range
.start
);
409 start
= range
.start
>> PAGE_SHIFT
;
410 size
= range
.size
>> PAGE_SHIFT
;
414 if (((HV_PhysAddr
)end
<< PAGE_SHIFT
) !=
415 (range
.start
+ range
.size
)) {
416 pr_err("PAs too high to represent: %#llx..%#llx\n",
417 range
.start
, range
.start
+ range
.size
);
421 #if defined(CONFIG_PCI) && !defined(__tilegx__)
423 * Blocks that overlap the pci reserved region must
424 * have enough space to hold the maximum percpu data
425 * region at the top of the range. If there isn't
426 * enough space above the reserved region, just
429 if (start
<= pci_reserve_start_pfn
&&
430 end
> pci_reserve_start_pfn
) {
431 unsigned int per_cpu_size
=
432 __per_cpu_end
- __per_cpu_start
;
433 unsigned int percpu_pages
=
434 NR_CPUS
* (PFN_UP(per_cpu_size
) >> PAGE_SHIFT
);
435 if (end
< pci_reserve_end_pfn
+ percpu_pages
) {
436 end
= pci_reserve_start_pfn
;
437 pr_err("PCI mapping region reduced node %d to"
438 " %ld pages\n", i
, end
- start
);
443 for (j
= __pfn_to_highbits(start
);
444 j
<= __pfn_to_highbits(end
- 1); j
++)
445 highbits_to_node
[j
] = i
;
447 node_start_pfn
[i
] = start
;
448 node_end_pfn
[i
] = end
;
449 node_controller
[i
] = range
.controller
;
453 /* Mark node as online */
454 node_set(i
, node_online_map
);
455 node_set(i
, node_possible_map
);
460 * For 4KB pages, mem_map "struct page" data is 1% of the size
461 * of the physical memory, so can be quite big (640 MB for
462 * four 16G zones). These structures must be mapped in
463 * lowmem, and since we currently cap out at about 768 MB,
464 * it's impractical to try to use this much address space.
465 * For now, arbitrarily cap the amount of physical memory
466 * we're willing to use at 8 million pages (32GB of 4KB pages).
468 cap
= 8 * 1024 * 1024; /* 8 million pages */
469 if (physpages
> cap
) {
470 int num_nodes
= num_online_nodes();
471 int cap_each
= cap
/ num_nodes
;
472 unsigned long dropped_pages
= 0;
473 for (i
= 0; i
< num_nodes
; ++i
) {
474 int size
= node_end_pfn
[i
] - node_start_pfn
[i
];
475 if (size
> cap_each
) {
476 dropped_pages
+= (size
- cap_each
);
477 node_end_pfn
[i
] = node_start_pfn
[i
] + cap_each
;
480 physpages
-= dropped_pages
;
481 pr_warning("Only using %ldMB memory;"
482 " ignoring %ldMB.\n",
483 physpages
>> (20 - PAGE_SHIFT
),
484 dropped_pages
>> (20 - PAGE_SHIFT
));
485 pr_warning("Consider using a larger page size.\n");
489 /* Heap starts just above the last loaded address. */
490 min_low_pfn
= PFN_UP((unsigned long)_end
- PAGE_OFFSET
);
492 #ifdef CONFIG_HIGHMEM
493 /* Find where we map lowmem from each controller. */
494 high_memory
= setup_pa_va_mapping();
496 /* Set max_low_pfn based on what node 0 can directly address. */
497 max_low_pfn
= node_lowmem_end_pfn
[0];
499 lowmem_pages
= (mappable_physpages
> MAXMEM_PFN
) ?
500 MAXMEM_PFN
: mappable_physpages
;
501 highmem_pages
= (long) (physpages
- lowmem_pages
);
503 pr_notice("%ldMB HIGHMEM available.\n",
504 pages_to_mb(highmem_pages
> 0 ? highmem_pages
: 0));
505 pr_notice("%ldMB LOWMEM available.\n",
506 pages_to_mb(lowmem_pages
));
508 /* Set max_low_pfn based on what node 0 can directly address. */
509 max_low_pfn
= node_end_pfn
[0];
512 if (node_end_pfn
[0] > MAXMEM_PFN
) {
513 pr_warning("Only using %ldMB LOWMEM.\n",
515 pr_warning("Use a HIGHMEM enabled kernel.\n");
516 max_low_pfn
= MAXMEM_PFN
;
517 max_pfn
= MAXMEM_PFN
;
518 node_end_pfn
[0] = MAXMEM_PFN
;
520 pr_notice("%ldMB memory available.\n",
521 pages_to_mb(node_end_pfn
[0]));
523 for (i
= 1; i
< MAX_NUMNODES
; ++i
) {
524 node_start_pfn
[i
] = 0;
527 high_memory
= __va(node_end_pfn
[0]);
530 for (i
= 0; i
< MAX_NUMNODES
; ++i
) {
531 int pages
= node_end_pfn
[i
] - node_start_pfn
[i
];
532 lowmem_pages
+= pages
;
534 high_memory
= pfn_to_kaddr(node_end_pfn
[i
]);
536 pr_notice("%ldMB memory available.\n",
537 pages_to_mb(lowmem_pages
));
543 * On 32-bit machines, we only put bootmem on the low controller,
544 * since PAs > 4GB can't be used in bootmem. In principle one could
545 * imagine, e.g., multiple 1 GB controllers all of which could support
546 * bootmem, but in practice using controllers this small isn't a
547 * particularly interesting scenario, so we just keep it simple and
548 * use only the first controller for bootmem on 32-bit machines.
550 static inline int node_has_bootmem(int nid
)
559 static inline unsigned long alloc_bootmem_pfn(int nid
,
563 void *kva
= __alloc_bootmem_node(NODE_DATA(nid
), size
,
565 unsigned long pfn
= kaddr_to_pfn(kva
);
566 BUG_ON(goal
&& PFN_PHYS(pfn
) != goal
);
570 static void __init
setup_bootmem_allocator_node(int i
)
572 unsigned long start
, end
, mapsize
, mapstart
;
574 if (node_has_bootmem(i
)) {
575 NODE_DATA(i
)->bdata
= &bootmem_node_data
[i
];
577 /* Share controller zero's bdata for now. */
578 NODE_DATA(i
)->bdata
= &bootmem_node_data
[0];
582 /* Skip up to after the bss in node 0. */
583 start
= (i
== 0) ? min_low_pfn
: node_start_pfn
[i
];
585 /* Only lowmem, if we're a HIGHMEM build. */
586 #ifdef CONFIG_HIGHMEM
587 end
= node_lowmem_end_pfn
[i
];
589 end
= node_end_pfn
[i
];
592 /* No memory here. */
596 /* Figure out where the bootmem bitmap is located. */
597 mapsize
= bootmem_bootmap_pages(end
- start
);
599 /* Use some space right before the heap on node 0. */
603 /* Allocate bitmap on node 0 to avoid page table issues. */
604 mapstart
= alloc_bootmem_pfn(0, PFN_PHYS(mapsize
), 0);
607 /* Initialize a node. */
608 init_bootmem_node(NODE_DATA(i
), mapstart
, start
, end
);
610 /* Free all the space back into the allocator. */
611 free_bootmem(PFN_PHYS(start
), PFN_PHYS(end
- start
));
613 #if defined(CONFIG_PCI) && !defined(__tilegx__)
615 * Throw away any memory aliased by the PCI region.
617 if (pci_reserve_start_pfn
< end
&& pci_reserve_end_pfn
> start
) {
618 start
= max(pci_reserve_start_pfn
, start
);
619 end
= min(pci_reserve_end_pfn
, end
);
620 reserve_bootmem(PFN_PHYS(start
), PFN_PHYS(end
- start
),
626 static void __init
setup_bootmem_allocator(void)
629 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
630 setup_bootmem_allocator_node(i
);
633 if (crashk_res
.start
!= crashk_res
.end
)
634 reserve_bootmem(crashk_res
.start
, resource_size(&crashk_res
), 0);
638 void *__init
alloc_remap(int nid
, unsigned long size
)
640 int pages
= node_end_pfn
[nid
] - node_start_pfn
[nid
];
641 void *map
= pfn_to_kaddr(node_memmap_pfn
[nid
]);
642 BUG_ON(size
!= pages
* sizeof(struct page
));
643 memset(map
, 0, size
);
647 static int __init
percpu_size(void)
649 int size
= __per_cpu_end
- __per_cpu_start
;
650 size
+= PERCPU_MODULE_RESERVE
;
651 size
+= PERCPU_DYNAMIC_EARLY_SIZE
;
652 if (size
< PCPU_MIN_UNIT_SIZE
)
653 size
= PCPU_MIN_UNIT_SIZE
;
654 size
= roundup(size
, PAGE_SIZE
);
656 /* In several places we assume the per-cpu data fits on a huge page. */
657 BUG_ON(kdata_huge
&& size
> HPAGE_SIZE
);
661 static void __init
zone_sizes_init(void)
663 unsigned long zones_size
[MAX_NR_ZONES
] = { 0 };
664 int size
= percpu_size();
665 int num_cpus
= smp_height
* smp_width
;
666 const unsigned long dma_end
= (1UL << (32 - PAGE_SHIFT
));
670 for (i
= 0; i
< num_cpus
; ++i
)
671 node_percpu
[cpu_to_node(i
)] += size
;
673 for_each_online_node(i
) {
674 unsigned long start
= node_start_pfn
[i
];
675 unsigned long end
= node_end_pfn
[i
];
676 #ifdef CONFIG_HIGHMEM
677 unsigned long lowmem_end
= node_lowmem_end_pfn
[i
];
679 unsigned long lowmem_end
= end
;
681 int memmap_size
= (end
- start
) * sizeof(struct page
);
682 node_free_pfn
[i
] = start
;
685 * Set aside pages for per-cpu data and the mem_map array.
687 * Since the per-cpu data requires special homecaching,
688 * if we are in kdata_huge mode, we put it at the end of
689 * the lowmem region. If we're not in kdata_huge mode,
690 * we take the per-cpu pages from the bottom of the
691 * controller, since that avoids fragmenting a huge page
692 * that users might want. We always take the memmap
693 * from the bottom of the controller, since with
694 * kdata_huge that lets it be under a huge TLB entry.
696 * If the user has requested isolnodes for a controller,
697 * though, there'll be no lowmem, so we just alloc_bootmem
698 * the memmap. There will be no percpu memory either.
700 if (i
!= 0 && cpu_isset(i
, isolnodes
)) {
702 alloc_bootmem_pfn(0, memmap_size
, 0);
703 BUG_ON(node_percpu
[i
] != 0);
704 } else if (node_has_bootmem(start
)) {
705 unsigned long goal
= 0;
707 alloc_bootmem_pfn(i
, memmap_size
, 0);
709 goal
= PFN_PHYS(lowmem_end
) - node_percpu
[i
];
712 alloc_bootmem_pfn(i
, node_percpu
[i
],
715 /* In non-bootmem zones, just reserve some pages. */
716 node_memmap_pfn
[i
] = node_free_pfn
[i
];
717 node_free_pfn
[i
] += PFN_UP(memmap_size
);
719 node_percpu_pfn
[i
] = node_free_pfn
[i
];
720 node_free_pfn
[i
] += PFN_UP(node_percpu
[i
]);
723 lowmem_end
- PFN_UP(node_percpu
[i
]);
727 #ifdef CONFIG_HIGHMEM
728 if (start
> lowmem_end
) {
729 zones_size
[ZONE_NORMAL
] = 0;
730 zones_size
[ZONE_HIGHMEM
] = end
- start
;
732 zones_size
[ZONE_NORMAL
] = lowmem_end
- start
;
733 zones_size
[ZONE_HIGHMEM
] = end
- lowmem_end
;
736 zones_size
[ZONE_NORMAL
] = end
- start
;
739 if (start
< dma_end
) {
740 zones_size
[ZONE_DMA
] = min(zones_size
[ZONE_NORMAL
],
742 zones_size
[ZONE_NORMAL
] -= zones_size
[ZONE_DMA
];
744 zones_size
[ZONE_DMA
] = 0;
747 /* Take zone metadata from controller 0 if we're isolnode. */
748 if (node_isset(i
, isolnodes
))
749 NODE_DATA(i
)->bdata
= &bootmem_node_data
[0];
751 free_area_init_node(i
, zones_size
, start
, NULL
);
752 printk(KERN_DEBUG
" Normal zone: %ld per-cpu pages\n",
753 PFN_UP(node_percpu
[i
]));
755 /* Track the type of memory on each node */
756 if (zones_size
[ZONE_NORMAL
] || zones_size
[ZONE_DMA
])
757 node_set_state(i
, N_NORMAL_MEMORY
);
758 #ifdef CONFIG_HIGHMEM
760 node_set_state(i
, N_HIGH_MEMORY
);
769 /* which logical CPUs are on which nodes */
770 struct cpumask node_2_cpu_mask
[MAX_NUMNODES
] __write_once
;
771 EXPORT_SYMBOL(node_2_cpu_mask
);
773 /* which node each logical CPU is on */
774 char cpu_2_node
[NR_CPUS
] __write_once
__attribute__((aligned(L2_CACHE_BYTES
)));
775 EXPORT_SYMBOL(cpu_2_node
);
777 /* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
778 static int __init
cpu_to_bound_node(int cpu
, struct cpumask
* unbound_cpus
)
780 if (!cpu_possible(cpu
) || cpumask_test_cpu(cpu
, unbound_cpus
))
783 return cpu_to_node(cpu
);
786 /* Return number of immediately-adjacent tiles sharing the same NUMA node. */
787 static int __init
node_neighbors(int node
, int cpu
,
788 struct cpumask
*unbound_cpus
)
795 if (x
> 0 && cpu_to_bound_node(cpu
-1, unbound_cpus
) == node
)
797 if (x
< w
-1 && cpu_to_bound_node(cpu
+1, unbound_cpus
) == node
)
799 if (y
> 0 && cpu_to_bound_node(cpu
-w
, unbound_cpus
) == node
)
801 if (y
< h
-1 && cpu_to_bound_node(cpu
+w
, unbound_cpus
) == node
)
806 static void __init
setup_numa_mapping(void)
808 int distance
[MAX_NUMNODES
][NR_CPUS
];
810 int cpu
, node
, cpus
, i
, x
, y
;
811 int num_nodes
= num_online_nodes();
812 struct cpumask unbound_cpus
;
813 nodemask_t default_nodes
;
815 cpumask_clear(&unbound_cpus
);
817 /* Get set of nodes we will use for defaults */
818 nodes_andnot(default_nodes
, node_online_map
, isolnodes
);
819 if (nodes_empty(default_nodes
)) {
820 BUG_ON(!node_isset(0, node_online_map
));
821 pr_err("Forcing NUMA node zero available as a default node\n");
822 node_set(0, default_nodes
);
825 /* Populate the distance[] array */
826 memset(distance
, -1, sizeof(distance
));
828 for (coord
.y
= 0; coord
.y
< smp_height
; ++coord
.y
) {
829 for (coord
.x
= 0; coord
.x
< smp_width
;
831 BUG_ON(cpu
>= nr_cpu_ids
);
832 if (!cpu_possible(cpu
)) {
833 cpu_2_node
[cpu
] = -1;
836 for_each_node_mask(node
, default_nodes
) {
837 HV_MemoryControllerInfo info
=
838 hv_inquire_memory_controller(
839 coord
, node_controller
[node
]);
840 distance
[node
][cpu
] =
841 ABS(info
.coord
.x
) + ABS(info
.coord
.y
);
843 cpumask_set_cpu(cpu
, &unbound_cpus
);
849 * Round-robin through the NUMA nodes until all the cpus are
850 * assigned. We could be more clever here (e.g. create four
851 * sorted linked lists on the same set of cpu nodes, and pull
852 * off them in round-robin sequence, removing from all four
853 * lists each time) but given the relatively small numbers
854 * involved, O(n^2) seem OK for a one-time cost.
856 node
= first_node(default_nodes
);
857 while (!cpumask_empty(&unbound_cpus
)) {
859 int best_distance
= INT_MAX
;
860 for (cpu
= 0; cpu
< cpus
; ++cpu
) {
861 if (cpumask_test_cpu(cpu
, &unbound_cpus
)) {
863 * Compute metric, which is how much
864 * closer the cpu is to this memory
865 * controller than the others, shifted
866 * up, and then the number of
867 * neighbors already in the node as an
868 * epsilon adjustment to try to keep
871 int d
= distance
[node
][cpu
] * num_nodes
;
872 for_each_node_mask(i
, default_nodes
) {
874 d
-= distance
[i
][cpu
];
876 d
*= 8; /* allow space for epsilon */
877 d
-= node_neighbors(node
, cpu
, &unbound_cpus
);
878 if (d
< best_distance
) {
884 BUG_ON(best_cpu
< 0);
885 cpumask_set_cpu(best_cpu
, &node_2_cpu_mask
[node
]);
886 cpu_2_node
[best_cpu
] = node
;
887 cpumask_clear_cpu(best_cpu
, &unbound_cpus
);
888 node
= next_node(node
, default_nodes
);
889 if (node
== MAX_NUMNODES
)
890 node
= first_node(default_nodes
);
893 /* Print out node assignments and set defaults for disabled cpus */
895 for (y
= 0; y
< smp_height
; ++y
) {
896 printk(KERN_DEBUG
"NUMA cpu-to-node row %d:", y
);
897 for (x
= 0; x
< smp_width
; ++x
, ++cpu
) {
898 if (cpu_to_node(cpu
) < 0) {
900 cpu_2_node
[cpu
] = first_node(default_nodes
);
902 pr_cont(" %d", cpu_to_node(cpu
));
909 static struct cpu cpu_devices
[NR_CPUS
];
911 static int __init
topology_init(void)
915 for_each_online_node(i
)
916 register_one_node(i
);
918 for (i
= 0; i
< smp_height
* smp_width
; ++i
)
919 register_cpu(&cpu_devices
[i
], i
);
924 subsys_initcall(topology_init
);
926 #else /* !CONFIG_NUMA */
928 #define setup_numa_mapping() do { } while (0)
930 #endif /* CONFIG_NUMA */
933 * Initialize hugepage support on this cpu. We do this on all cores
934 * early in boot: before argument parsing for the boot cpu, and after
935 * argument parsing but before the init functions run on the secondaries.
936 * So the values we set up here in the hypervisor may be overridden on
937 * the boot cpu as arguments are parsed.
939 static __cpuinit
void init_super_pages(void)
941 #ifdef CONFIG_HUGETLB_SUPER_PAGES
943 for (i
= 0; i
< HUGE_SHIFT_ENTRIES
; ++i
)
944 hv_set_pte_super_shift(i
, huge_shift
[i
]);
949 * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
950 * @boot: Is this the boot cpu?
952 * Called from setup_arch() on the boot cpu, or online_secondary().
954 void __cpuinit
setup_cpu(int boot
)
956 /* The boot cpu sets up its permanent mappings much earlier. */
958 store_permanent_mappings();
960 /* Allow asynchronous TLB interrupts. */
961 #if CHIP_HAS_TILE_DMA()
962 arch_local_irq_unmask(INT_DMATLB_MISS
);
963 arch_local_irq_unmask(INT_DMATLB_ACCESS
);
965 #if CHIP_HAS_SN_PROC()
966 arch_local_irq_unmask(INT_SNITLB_MISS
);
969 arch_local_irq_unmask(INT_SINGLE_STEP_K
);
973 * Allow user access to many generic SPRs, like the cycle
974 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
976 __insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0
, 1);
979 /* Static network is not restricted. */
980 __insn_mtspr(SPR_MPL_SN_ACCESS_SET_0
, 1);
982 #if CHIP_HAS_SN_PROC()
983 __insn_mtspr(SPR_MPL_SN_NOTIFY_SET_0
, 1);
984 __insn_mtspr(SPR_MPL_SN_CPL_SET_0
, 1);
988 * Set the MPL for interrupt control 0 & 1 to the corresponding
989 * values. This includes access to the SYSTEM_SAVE and EX_CONTEXT
990 * SPRs, as well as the interrupt mask.
992 __insn_mtspr(SPR_MPL_INTCTRL_0_SET_0
, 1);
993 __insn_mtspr(SPR_MPL_INTCTRL_1_SET_1
, 1);
995 /* Initialize IRQ support for this cpu. */
998 #ifdef CONFIG_HARDWALL
999 /* Reset the network state on this cpu. */
1000 reset_network_state();
1006 #ifdef CONFIG_BLK_DEV_INITRD
1008 static int __initdata set_initramfs_file
;
1009 static char __initdata initramfs_file
[128] = "initramfs";
1011 static int __init
setup_initramfs_file(char *str
)
1015 strncpy(initramfs_file
, str
, sizeof(initramfs_file
) - 1);
1016 set_initramfs_file
= 1;
1020 early_param("initramfs_file", setup_initramfs_file
);
1023 * We look for a file called "initramfs" in the hvfs. If there is one, we
1024 * allocate some memory for it and it will be unpacked to the initramfs.
1025 * If it's compressed, the initd code will uncompress it first.
1027 static void __init
load_hv_initrd(void)
1029 HV_FS_StatInfo stat
;
1033 fd
= hv_fs_findfile((HV_VirtAddr
) initramfs_file
);
1034 if (fd
== HV_ENOENT
) {
1035 if (set_initramfs_file
) {
1036 pr_warning("No such hvfs initramfs file '%s'\n",
1040 /* Try old backwards-compatible name. */
1041 fd
= hv_fs_findfile((HV_VirtAddr
)"initramfs.cpio.gz");
1042 if (fd
== HV_ENOENT
)
1047 stat
= hv_fs_fstat(fd
);
1048 BUG_ON(stat
.size
< 0);
1049 if (stat
.flags
& HV_FS_ISDIR
) {
1050 pr_warning("Ignoring hvfs file '%s': it's a directory.\n",
1054 initrd
= alloc_bootmem_pages(stat
.size
);
1055 rc
= hv_fs_pread(fd
, (HV_VirtAddr
) initrd
, stat
.size
, 0);
1056 if (rc
!= stat
.size
) {
1057 pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
1058 stat
.size
, initramfs_file
, rc
);
1059 free_initrd_mem((unsigned long) initrd
, stat
.size
);
1062 initrd_start
= (unsigned long) initrd
;
1063 initrd_end
= initrd_start
+ stat
.size
;
1066 void __init
free_initrd_mem(unsigned long begin
, unsigned long end
)
1068 free_bootmem(__pa(begin
), end
- begin
);
1072 static inline void load_hv_initrd(void) {}
1073 #endif /* CONFIG_BLK_DEV_INITRD */
1075 static void __init
validate_hv(void)
1078 * It may already be too late, but let's check our built-in
1079 * configuration against what the hypervisor is providing.
1081 unsigned long glue_size
= hv_sysconf(HV_SYSCONF_GLUE_SIZE
);
1082 int hv_page_size
= hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL
);
1083 int hv_hpage_size
= hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE
);
1084 HV_ASIDRange asid_range
;
1087 HV_Topology topology
= hv_inquire_topology();
1088 BUG_ON(topology
.coord
.x
!= 0 || topology
.coord
.y
!= 0);
1089 if (topology
.width
!= 1 || topology
.height
!= 1) {
1090 pr_warning("Warning: booting UP kernel on %dx%d grid;"
1091 " will ignore all but first tile.\n",
1092 topology
.width
, topology
.height
);
1096 if (PAGE_OFFSET
+ HV_GLUE_START_CPA
+ glue_size
> (unsigned long)_text
)
1097 early_panic("Hypervisor glue size %ld is too big!\n",
1099 if (hv_page_size
!= PAGE_SIZE
)
1100 early_panic("Hypervisor page size %#x != our %#lx\n",
1101 hv_page_size
, PAGE_SIZE
);
1102 if (hv_hpage_size
!= HPAGE_SIZE
)
1103 early_panic("Hypervisor huge page size %#x != our %#lx\n",
1104 hv_hpage_size
, HPAGE_SIZE
);
1108 * Some hypervisor APIs take a pointer to a bitmap array
1109 * whose size is at least the number of cpus on the chip.
1110 * We use a struct cpumask for this, so it must be big enough.
1112 if ((smp_height
* smp_width
) > nr_cpu_ids
)
1113 early_panic("Hypervisor %d x %d grid too big for Linux"
1114 " NR_CPUS %d\n", smp_height
, smp_width
,
1119 * Check that we're using allowed ASIDs, and initialize the
1120 * various asid variables to their appropriate initial states.
1122 asid_range
= hv_inquire_asid(0);
1123 __get_cpu_var(current_asid
) = min_asid
= asid_range
.start
;
1124 max_asid
= asid_range
.start
+ asid_range
.size
- 1;
1126 if (hv_confstr(HV_CONFSTR_CHIP_MODEL
, (HV_VirtAddr
)chip_model
,
1127 sizeof(chip_model
)) < 0) {
1128 pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
1129 strlcpy(chip_model
, "unknown", sizeof(chip_model
));
1133 static void __init
validate_va(void)
1135 #ifndef __tilegx__ /* FIXME: GX: probably some validation relevant here */
1137 * Similarly, make sure we're only using allowed VAs.
1138 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_INTRPT,
1139 * and 0 .. KERNEL_HIGH_VADDR.
1140 * In addition, make sure we CAN'T use the end of memory, since
1141 * we use the last chunk of each pgd for the pgd_list.
1143 int i
, user_kernel_ok
= 0;
1144 unsigned long max_va
= 0;
1145 unsigned long list_va
=
1146 ((PGD_LIST_OFFSET
/ sizeof(pgd_t
)) << PGDIR_SHIFT
);
1148 for (i
= 0; ; ++i
) {
1149 HV_VirtAddrRange range
= hv_inquire_virtual(i
);
1150 if (range
.size
== 0)
1152 if (range
.start
<= MEM_USER_INTRPT
&&
1153 range
.start
+ range
.size
>= MEM_HV_INTRPT
)
1155 if (range
.start
== 0)
1156 max_va
= range
.size
;
1157 BUG_ON(range
.start
+ range
.size
> list_va
);
1159 if (!user_kernel_ok
)
1160 early_panic("Hypervisor not configured for user/kernel VAs\n");
1162 early_panic("Hypervisor not configured for low VAs\n");
1163 if (max_va
< KERNEL_HIGH_VADDR
)
1164 early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
1165 max_va
, KERNEL_HIGH_VADDR
);
1167 /* Kernel PCs must have their high bit set; see intvec.S. */
1168 if ((long)VMALLOC_START
>= 0)
1170 "Linux VMALLOC region below the 2GB line (%#lx)!\n"
1171 "Reconfigure the kernel with fewer NR_HUGE_VMAPS\n"
1172 "or smaller VMALLOC_RESERVE.\n",
1178 * cpu_lotar_map lists all the cpus that are valid for the supervisor
1179 * to cache data on at a page level, i.e. what cpus can be placed in
1180 * the LOTAR field of a PTE. It is equivalent to the set of possible
1181 * cpus plus any other cpus that are willing to share their cache.
1182 * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
1184 struct cpumask __write_once cpu_lotar_map
;
1185 EXPORT_SYMBOL(cpu_lotar_map
);
1187 #if CHIP_HAS_CBOX_HOME_MAP()
1189 * hash_for_home_map lists all the tiles that hash-for-home data
1190 * will be cached on. Note that this may includes tiles that are not
1191 * valid for this supervisor to use otherwise (e.g. if a hypervisor
1192 * device is being shared between multiple supervisors).
1193 * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
1195 struct cpumask hash_for_home_map
;
1196 EXPORT_SYMBOL(hash_for_home_map
);
1200 * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
1201 * flush on our behalf. It is set to cpu_possible_mask OR'ed with
1202 * hash_for_home_map, and it is what should be passed to
1203 * hv_flush_remote() to flush all caches. Note that if there are
1204 * dedicated hypervisor driver tiles that have authorized use of their
1205 * cache, those tiles will only appear in cpu_lotar_map, NOT in
1206 * cpu_cacheable_map, as they are a special case.
1208 struct cpumask __write_once cpu_cacheable_map
;
1209 EXPORT_SYMBOL(cpu_cacheable_map
);
1211 static __initdata
struct cpumask disabled_map
;
1213 static int __init
disabled_cpus(char *str
)
1215 int boot_cpu
= smp_processor_id();
1217 if (str
== NULL
|| cpulist_parse_crop(str
, &disabled_map
) != 0)
1219 if (cpumask_test_cpu(boot_cpu
, &disabled_map
)) {
1220 pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu
);
1221 cpumask_clear_cpu(boot_cpu
, &disabled_map
);
1226 early_param("disabled_cpus", disabled_cpus
);
1228 void __init
print_disabled_cpus(void)
1230 if (!cpumask_empty(&disabled_map
)) {
1232 cpulist_scnprintf(buf
, sizeof(buf
), &disabled_map
);
1233 pr_info("CPUs not available for Linux: %s\n", buf
);
1237 static void __init
setup_cpu_maps(void)
1239 struct cpumask hv_disabled_map
, cpu_possible_init
;
1240 int boot_cpu
= smp_processor_id();
1243 /* Learn which cpus are allowed by the hypervisor. */
1244 rc
= hv_inquire_tiles(HV_INQ_TILES_AVAIL
,
1245 (HV_VirtAddr
) cpumask_bits(&cpu_possible_init
),
1246 sizeof(cpu_cacheable_map
));
1248 early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc
);
1249 if (!cpumask_test_cpu(boot_cpu
, &cpu_possible_init
))
1250 early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu
);
1252 /* Compute the cpus disabled by the hvconfig file. */
1253 cpumask_complement(&hv_disabled_map
, &cpu_possible_init
);
1255 /* Include them with the cpus disabled by "disabled_cpus". */
1256 cpumask_or(&disabled_map
, &disabled_map
, &hv_disabled_map
);
1259 * Disable every cpu after "setup_max_cpus". But don't mark
1260 * as disabled the cpus that are outside of our initial rectangle,
1261 * since that turns out to be confusing.
1263 cpus
= 1; /* this cpu */
1264 cpumask_set_cpu(boot_cpu
, &disabled_map
); /* ignore this cpu */
1265 for (i
= 0; cpus
< setup_max_cpus
; ++i
)
1266 if (!cpumask_test_cpu(i
, &disabled_map
))
1268 for (; i
< smp_height
* smp_width
; ++i
)
1269 cpumask_set_cpu(i
, &disabled_map
);
1270 cpumask_clear_cpu(boot_cpu
, &disabled_map
); /* reset this cpu */
1271 for (i
= smp_height
* smp_width
; i
< NR_CPUS
; ++i
)
1272 cpumask_clear_cpu(i
, &disabled_map
);
1275 * Setup cpu_possible map as every cpu allocated to us, minus
1276 * the results of any "disabled_cpus" settings.
1278 cpumask_andnot(&cpu_possible_init
, &cpu_possible_init
, &disabled_map
);
1279 init_cpu_possible(&cpu_possible_init
);
1281 /* Learn which cpus are valid for LOTAR caching. */
1282 rc
= hv_inquire_tiles(HV_INQ_TILES_LOTAR
,
1283 (HV_VirtAddr
) cpumask_bits(&cpu_lotar_map
),
1284 sizeof(cpu_lotar_map
));
1286 pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
1287 cpu_lotar_map
= *cpu_possible_mask
;
1290 #if CHIP_HAS_CBOX_HOME_MAP()
1291 /* Retrieve set of CPUs used for hash-for-home caching */
1292 rc
= hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE
,
1293 (HV_VirtAddr
) hash_for_home_map
.bits
,
1294 sizeof(hash_for_home_map
));
1296 early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc
);
1297 cpumask_or(&cpu_cacheable_map
, cpu_possible_mask
, &hash_for_home_map
);
1299 cpu_cacheable_map
= *cpu_possible_mask
;
1304 static int __init
dataplane(char *str
)
1306 pr_warning("WARNING: dataplane support disabled in this kernel\n");
1310 early_param("dataplane", dataplane
);
1312 #ifdef CONFIG_CMDLINE_BOOL
1313 static char __initdata builtin_cmdline
[COMMAND_LINE_SIZE
] = CONFIG_CMDLINE
;
1316 void __init
setup_arch(char **cmdline_p
)
1320 #if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
1321 len
= hv_get_command_line((HV_VirtAddr
) boot_command_line
,
1323 if (boot_command_line
[0])
1324 pr_warning("WARNING: ignoring dynamic command line \"%s\"\n",
1326 strlcpy(boot_command_line
, builtin_cmdline
, COMMAND_LINE_SIZE
);
1329 #if defined(CONFIG_CMDLINE_BOOL)
1330 if (builtin_cmdline
[0]) {
1331 int builtin_len
= strlcpy(boot_command_line
, builtin_cmdline
,
1333 if (builtin_len
< COMMAND_LINE_SIZE
-1)
1334 boot_command_line
[builtin_len
++] = ' ';
1335 hv_cmdline
= &boot_command_line
[builtin_len
];
1336 len
= COMMAND_LINE_SIZE
- builtin_len
;
1340 hv_cmdline
= boot_command_line
;
1341 len
= COMMAND_LINE_SIZE
;
1343 len
= hv_get_command_line((HV_VirtAddr
) hv_cmdline
, len
);
1344 if (len
< 0 || len
> COMMAND_LINE_SIZE
)
1345 early_panic("hv_get_command_line failed: %d\n", len
);
1348 *cmdline_p
= boot_command_line
;
1350 /* Set disabled_map and setup_max_cpus very early */
1351 parse_early_param();
1353 /* Make sure the kernel is compatible with the hypervisor. */
1360 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1362 * Initialize the PCI structures. This is done before memory
1363 * setup so that we know whether or not a pci_reserve region
1366 if (tile_pci_init() == 0)
1369 /* PCI systems reserve a region just below 4GB for mapping iomem. */
1370 pci_reserve_end_pfn
= (1 << (32 - PAGE_SHIFT
));
1371 pci_reserve_start_pfn
= pci_reserve_end_pfn
-
1372 (pci_reserve_mb
<< (20 - PAGE_SHIFT
));
1375 init_mm
.start_code
= (unsigned long) _text
;
1376 init_mm
.end_code
= (unsigned long) _etext
;
1377 init_mm
.end_data
= (unsigned long) _edata
;
1378 init_mm
.brk
= (unsigned long) _end
;
1381 store_permanent_mappings();
1382 setup_bootmem_allocator();
1385 * NOTE: before this point _nobody_ is allowed to allocate
1386 * any memory using the bootmem allocator.
1389 #ifdef CONFIG_SWIOTLB
1394 setup_numa_mapping();
1404 * Set up per-cpu memory.
1407 unsigned long __per_cpu_offset
[NR_CPUS
] __write_once
;
1408 EXPORT_SYMBOL(__per_cpu_offset
);
1410 static size_t __initdata pfn_offset
[MAX_NUMNODES
] = { 0 };
1411 static unsigned long __initdata percpu_pfn
[NR_CPUS
] = { 0 };
1414 * As the percpu code allocates pages, we return the pages from the
1415 * end of the node for the specified cpu.
1417 static void *__init
pcpu_fc_alloc(unsigned int cpu
, size_t size
, size_t align
)
1419 int nid
= cpu_to_node(cpu
);
1420 unsigned long pfn
= node_percpu_pfn
[nid
] + pfn_offset
[nid
];
1422 BUG_ON(size
% PAGE_SIZE
!= 0);
1423 pfn_offset
[nid
] += size
/ PAGE_SIZE
;
1424 BUG_ON(node_percpu
[nid
] < size
);
1425 node_percpu
[nid
] -= size
;
1426 if (percpu_pfn
[cpu
] == 0)
1427 percpu_pfn
[cpu
] = pfn
;
1428 return pfn_to_kaddr(pfn
);
1432 * Pages reserved for percpu memory are not freeable, and in any case we are
1433 * on a short path to panic() in setup_per_cpu_area() at this point anyway.
1435 static void __init
pcpu_fc_free(void *ptr
, size_t size
)
1440 * Set up vmalloc page tables using bootmem for the percpu code.
1442 static void __init
pcpu_fc_populate_pte(unsigned long addr
)
1449 BUG_ON(pgd_addr_invalid(addr
));
1450 if (addr
< VMALLOC_START
|| addr
>= VMALLOC_END
)
1451 panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx;"
1452 " try increasing CONFIG_VMALLOC_RESERVE\n",
1453 addr
, VMALLOC_START
, VMALLOC_END
);
1455 pgd
= swapper_pg_dir
+ pgd_index(addr
);
1456 pud
= pud_offset(pgd
, addr
);
1457 BUG_ON(!pud_present(*pud
));
1458 pmd
= pmd_offset(pud
, addr
);
1459 if (pmd_present(*pmd
)) {
1460 BUG_ON(pmd_huge_page(*pmd
));
1462 pte
= __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE
,
1463 HV_PAGE_TABLE_ALIGN
, 0);
1464 pmd_populate_kernel(&init_mm
, pmd
, pte
);
1468 void __init
setup_per_cpu_areas(void)
1471 unsigned long delta
, pfn
, lowmem_va
;
1472 unsigned long size
= percpu_size();
1476 rc
= pcpu_page_first_chunk(PERCPU_MODULE_RESERVE
, pcpu_fc_alloc
,
1477 pcpu_fc_free
, pcpu_fc_populate_pte
);
1479 panic("Cannot initialize percpu area (err=%d)", rc
);
1481 delta
= (unsigned long)pcpu_base_addr
- (unsigned long)__per_cpu_start
;
1482 for_each_possible_cpu(cpu
) {
1483 __per_cpu_offset
[cpu
] = delta
+ pcpu_unit_offsets
[cpu
];
1485 /* finv the copy out of cache so we can change homecache */
1486 ptr
= pcpu_base_addr
+ pcpu_unit_offsets
[cpu
];
1487 __finv_buffer(ptr
, size
);
1488 pfn
= percpu_pfn
[cpu
];
1490 /* Rewrite the page tables to cache on that cpu */
1491 pg
= pfn_to_page(pfn
);
1492 for (i
= 0; i
< size
; i
+= PAGE_SIZE
, ++pfn
, ++pg
) {
1494 /* Update the vmalloc mapping and page home. */
1495 unsigned long addr
= (unsigned long)ptr
+ i
;
1496 pte_t
*ptep
= virt_to_pte(NULL
, addr
);
1498 BUG_ON(pfn
!= pte_pfn(pte
));
1499 pte
= hv_pte_set_mode(pte
, HV_PTE_MODE_CACHE_TILE_L3
);
1500 pte
= set_remote_cache_cpu(pte
, cpu
);
1501 set_pte_at(&init_mm
, addr
, ptep
, pte
);
1503 /* Update the lowmem mapping for consistency. */
1504 lowmem_va
= (unsigned long)pfn_to_kaddr(pfn
);
1505 ptep
= virt_to_pte(NULL
, lowmem_va
);
1506 if (pte_huge(*ptep
)) {
1507 printk(KERN_DEBUG
"early shatter of huge page"
1508 " at %#lx\n", lowmem_va
);
1509 shatter_pmd((pmd_t
*)ptep
);
1510 ptep
= virt_to_pte(NULL
, lowmem_va
);
1511 BUG_ON(pte_huge(*ptep
));
1513 BUG_ON(pfn
!= pte_pfn(*ptep
));
1514 set_pte_at(&init_mm
, lowmem_va
, ptep
, pte
);
1518 /* Set our thread pointer appropriately. */
1519 set_my_cpu_offset(__per_cpu_offset
[smp_processor_id()]);
1521 /* Make sure the finv's have completed. */
1524 /* Flush the TLB so we reference it properly from here on out. */
1525 local_flush_tlb_all();
1528 static struct resource data_resource
= {
1529 .name
= "Kernel data",
1532 .flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
1535 static struct resource code_resource
= {
1536 .name
= "Kernel code",
1539 .flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
1543 * On Pro, we reserve all resources above 4GB so that PCI won't try to put
1544 * mappings above 4GB.
1546 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1547 static struct resource
* __init
1548 insert_non_bus_resource(void)
1550 struct resource
*res
=
1551 kzalloc(sizeof(struct resource
), GFP_ATOMIC
);
1552 res
->name
= "Non-Bus Physical Address Space";
1553 res
->start
= (1ULL << 32);
1555 res
->flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
;
1556 if (insert_resource(&iomem_resource
, res
)) {
1564 static struct resource
* __init
1565 insert_ram_resource(u64 start_pfn
, u64 end_pfn
)
1567 struct resource
*res
=
1568 kzalloc(sizeof(struct resource
), GFP_ATOMIC
);
1569 res
->name
= "System RAM";
1570 res
->start
= start_pfn
<< PAGE_SHIFT
;
1571 res
->end
= (end_pfn
<< PAGE_SHIFT
) - 1;
1572 res
->flags
= IORESOURCE_BUSY
| IORESOURCE_MEM
;
1573 if (insert_resource(&iomem_resource
, res
)) {
1581 * Request address space for all standard resources
1583 * If the system includes PCI root complex drivers, we need to create
1584 * a window just below 4GB where PCI BARs can be mapped.
1586 static int __init
request_standard_resources(void)
1589 enum { CODE_DELTA
= MEM_SV_INTRPT
- PAGE_OFFSET
};
1591 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1592 insert_non_bus_resource();
1595 for_each_online_node(i
) {
1596 u64 start_pfn
= node_start_pfn
[i
];
1597 u64 end_pfn
= node_end_pfn
[i
];
1599 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1600 if (start_pfn
<= pci_reserve_start_pfn
&&
1601 end_pfn
> pci_reserve_start_pfn
) {
1602 if (end_pfn
> pci_reserve_end_pfn
)
1603 insert_ram_resource(pci_reserve_end_pfn
,
1605 end_pfn
= pci_reserve_start_pfn
;
1608 insert_ram_resource(start_pfn
, end_pfn
);
1611 code_resource
.start
= __pa(_text
- CODE_DELTA
);
1612 code_resource
.end
= __pa(_etext
- CODE_DELTA
)-1;
1613 data_resource
.start
= __pa(_sdata
);
1614 data_resource
.end
= __pa(_end
)-1;
1616 insert_resource(&iomem_resource
, &code_resource
);
1617 insert_resource(&iomem_resource
, &data_resource
);
1620 insert_resource(&iomem_resource
, &crashk_res
);
1626 subsys_initcall(request_standard_resources
);