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)_stext,
311 (uint32_t)(_einittext - _stext), 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__)
333 /* We are using a char to hold the cpu_2_node[] mapping */
334 BUILD_BUG_ON(MAX_NUMNODES > 127);
336 /* Discover the ranges of memory available to us */
338 unsigned long start, size, end, highbits;
339 HV_PhysAddrRange range = hv_inquire_physical(i);
342 #ifdef CONFIG_FLATMEM
344 pr_err("Can't use discontiguous PAs: %#llx..%#llx\n",
345 range.size, range.start + range.size);
350 if ((unsigned long)range.start) {
351 pr_err("Range not at 4GB multiple: %#llx..%#llx\n",
352 range.start, range.start + range.size);
356 if ((range.start & (HPAGE_SIZE-1)) != 0 ||
357 (range.size & (HPAGE_SIZE-1)) != 0) {
358 unsigned long long start_pa = range.start;
359 unsigned long long orig_size = range.size;
360 range.start = (start_pa + HPAGE_SIZE - 1) & HPAGE_MASK;
361 range.size -= (range.start - start_pa);
362 range.size &= HPAGE_MASK;
363 pr_err("Range not hugepage-aligned: %#llx..%#llx:"
364 " now %#llx-%#llx\n",
365 start_pa, start_pa + orig_size,
366 range.start, range.start + range.size);
368 highbits = __pa_to_highbits(range.start);
369 if (highbits >= NR_PA_HIGHBIT_VALUES) {
370 pr_err("PA high bits too high: %#llx..%#llx\n",
371 range.start, range.start + range.size);
374 if (highbits_seen[highbits]) {
375 pr_err("Range overlaps in high bits: %#llx..%#llx\n",
376 range.start, range.start + range.size);
379 highbits_seen[highbits] = 1;
380 if (PFN_DOWN(range.size) > maxnodemem_pfn[i]) {
381 int max_size = maxnodemem_pfn[i];
383 pr_err("Maxnodemem reduced node %d to"
384 " %d pages\n", i, max_size);
385 range.size = PFN_PHYS(max_size);
387 pr_err("Maxnodemem disabled node %d\n", i);
391 if (num_physpages + PFN_DOWN(range.size) > maxmem_pfn) {
392 int max_size = maxmem_pfn - num_physpages;
394 pr_err("Maxmem reduced node %d to %d pages\n",
396 range.size = PFN_PHYS(max_size);
398 pr_err("Maxmem disabled node %d\n", i);
402 if (i >= MAX_NUMNODES) {
403 pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
404 i, range.size, range.size + range.start);
408 start = range.start >> PAGE_SHIFT;
409 size = range.size >> PAGE_SHIFT;
413 if (((HV_PhysAddr)end << PAGE_SHIFT) !=
414 (range.start + range.size)) {
415 pr_err("PAs too high to represent: %#llx..%#llx\n",
416 range.start, range.start + range.size);
420 #if defined(CONFIG_PCI) && !defined(__tilegx__)
422 * Blocks that overlap the pci reserved region must
423 * have enough space to hold the maximum percpu data
424 * region at the top of the range. If there isn't
425 * enough space above the reserved region, just
428 if (start <= pci_reserve_start_pfn &&
429 end > pci_reserve_start_pfn) {
430 unsigned int per_cpu_size =
431 __per_cpu_end - __per_cpu_start;
432 unsigned int percpu_pages =
433 NR_CPUS * (PFN_UP(per_cpu_size) >> PAGE_SHIFT);
434 if (end < pci_reserve_end_pfn + percpu_pages) {
435 end = pci_reserve_start_pfn;
436 pr_err("PCI mapping region reduced node %d to"
437 " %ld pages\n", i, end - start);
442 for (j = __pfn_to_highbits(start);
443 j <= __pfn_to_highbits(end - 1); j++)
444 highbits_to_node[j] = i;
446 node_start_pfn[i] = start;
447 node_end_pfn[i] = end;
448 node_controller[i] = range.controller;
449 num_physpages += size;
452 /* Mark node as online */
453 node_set(i, node_online_map);
454 node_set(i, node_possible_map);
459 * For 4KB pages, mem_map "struct page" data is 1% of the size
460 * of the physical memory, so can be quite big (640 MB for
461 * four 16G zones). These structures must be mapped in
462 * lowmem, and since we currently cap out at about 768 MB,
463 * it's impractical to try to use this much address space.
464 * For now, arbitrarily cap the amount of physical memory
465 * we're willing to use at 8 million pages (32GB of 4KB pages).
467 cap = 8 * 1024 * 1024; /* 8 million pages */
468 if (num_physpages > cap) {
469 int num_nodes = num_online_nodes();
470 int cap_each = cap / num_nodes;
471 unsigned long dropped_pages = 0;
472 for (i = 0; i < num_nodes; ++i) {
473 int size = node_end_pfn[i] - node_start_pfn[i];
474 if (size > cap_each) {
475 dropped_pages += (size - cap_each);
476 node_end_pfn[i] = node_start_pfn[i] + cap_each;
479 num_physpages -= dropped_pages;
480 pr_warning("Only using %ldMB memory;"
481 " ignoring %ldMB.\n",
482 num_physpages >> (20 - PAGE_SHIFT),
483 dropped_pages >> (20 - PAGE_SHIFT));
484 pr_warning("Consider using a larger page size.\n");
488 /* Heap starts just above the last loaded address. */
489 min_low_pfn = PFN_UP((unsigned long)_end - PAGE_OFFSET);
491 #ifdef CONFIG_HIGHMEM
492 /* Find where we map lowmem from each controller. */
493 high_memory = setup_pa_va_mapping();
495 /* Set max_low_pfn based on what node 0 can directly address. */
496 max_low_pfn = node_lowmem_end_pfn[0];
498 lowmem_pages = (mappable_physpages > MAXMEM_PFN) ?
499 MAXMEM_PFN : mappable_physpages;
500 highmem_pages = (long) (num_physpages - lowmem_pages);
502 pr_notice("%ldMB HIGHMEM available.\n",
503 pages_to_mb(highmem_pages > 0 ? highmem_pages : 0));
504 pr_notice("%ldMB LOWMEM available.\n",
505 pages_to_mb(lowmem_pages));
507 /* Set max_low_pfn based on what node 0 can directly address. */
508 max_low_pfn = node_end_pfn[0];
511 if (node_end_pfn[0] > MAXMEM_PFN) {
512 pr_warning("Only using %ldMB LOWMEM.\n",
514 pr_warning("Use a HIGHMEM enabled kernel.\n");
515 max_low_pfn = MAXMEM_PFN;
516 max_pfn = MAXMEM_PFN;
517 num_physpages = 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 reserve_bootmem(PFN_PHYS(pci_reserve_start_pfn),
619 PFN_PHYS(pci_reserve_end_pfn -
620 pci_reserve_start_pfn),
625 static void __init setup_bootmem_allocator(void)
628 for (i = 0; i < MAX_NUMNODES; ++i)
629 setup_bootmem_allocator_node(i);
632 if (crashk_res.start != crashk_res.end)
633 reserve_bootmem(crashk_res.start, resource_size(&crashk_res), 0);
637 void *__init alloc_remap(int nid, unsigned long size)
639 int pages = node_end_pfn[nid] - node_start_pfn[nid];
640 void *map = pfn_to_kaddr(node_memmap_pfn[nid]);
641 BUG_ON(size != pages * sizeof(struct page));
642 memset(map, 0, size);
646 static int __init percpu_size(void)
648 int size = __per_cpu_end - __per_cpu_start;
649 size += PERCPU_MODULE_RESERVE;
650 size += PERCPU_DYNAMIC_EARLY_SIZE;
651 if (size < PCPU_MIN_UNIT_SIZE)
652 size = PCPU_MIN_UNIT_SIZE;
653 size = roundup(size, PAGE_SIZE);
655 /* In several places we assume the per-cpu data fits on a huge page. */
656 BUG_ON(kdata_huge && size > HPAGE_SIZE);
660 static void __init zone_sizes_init(void)
662 unsigned long zones_size[MAX_NR_ZONES] = { 0 };
663 int size = percpu_size();
664 int num_cpus = smp_height * smp_width;
665 const unsigned long dma_end = (1UL << (32 - PAGE_SHIFT));
669 for (i = 0; i < num_cpus; ++i)
670 node_percpu[cpu_to_node(i)] += size;
672 for_each_online_node(i) {
673 unsigned long start = node_start_pfn[i];
674 unsigned long end = node_end_pfn[i];
675 #ifdef CONFIG_HIGHMEM
676 unsigned long lowmem_end = node_lowmem_end_pfn[i];
678 unsigned long lowmem_end = end;
680 int memmap_size = (end - start) * sizeof(struct page);
681 node_free_pfn[i] = start;
684 * Set aside pages for per-cpu data and the mem_map array.
686 * Since the per-cpu data requires special homecaching,
687 * if we are in kdata_huge mode, we put it at the end of
688 * the lowmem region. If we're not in kdata_huge mode,
689 * we take the per-cpu pages from the bottom of the
690 * controller, since that avoids fragmenting a huge page
691 * that users might want. We always take the memmap
692 * from the bottom of the controller, since with
693 * kdata_huge that lets it be under a huge TLB entry.
695 * If the user has requested isolnodes for a controller,
696 * though, there'll be no lowmem, so we just alloc_bootmem
697 * the memmap. There will be no percpu memory either.
699 if (i != 0 && cpu_isset(i, isolnodes)) {
701 alloc_bootmem_pfn(0, memmap_size, 0);
702 BUG_ON(node_percpu[i] != 0);
703 } else if (node_has_bootmem(start)) {
704 unsigned long goal = 0;
706 alloc_bootmem_pfn(i, memmap_size, 0);
708 goal = PFN_PHYS(lowmem_end) - node_percpu[i];
711 alloc_bootmem_pfn(i, node_percpu[i],
714 /* In non-bootmem zones, just reserve some pages. */
715 node_memmap_pfn[i] = node_free_pfn[i];
716 node_free_pfn[i] += PFN_UP(memmap_size);
718 node_percpu_pfn[i] = node_free_pfn[i];
719 node_free_pfn[i] += PFN_UP(node_percpu[i]);
722 lowmem_end - PFN_UP(node_percpu[i]);
726 #ifdef CONFIG_HIGHMEM
727 if (start > lowmem_end) {
728 zones_size[ZONE_NORMAL] = 0;
729 zones_size[ZONE_HIGHMEM] = end - start;
731 zones_size[ZONE_NORMAL] = lowmem_end - start;
732 zones_size[ZONE_HIGHMEM] = end - lowmem_end;
735 zones_size[ZONE_NORMAL] = end - start;
738 if (start < dma_end) {
739 zones_size[ZONE_DMA] = min(zones_size[ZONE_NORMAL],
741 zones_size[ZONE_NORMAL] -= zones_size[ZONE_DMA];
743 zones_size[ZONE_DMA] = 0;
746 /* Take zone metadata from controller 0 if we're isolnode. */
747 if (node_isset(i, isolnodes))
748 NODE_DATA(i)->bdata = &bootmem_node_data[0];
750 free_area_init_node(i, zones_size, start, NULL);
751 printk(KERN_DEBUG " Normal zone: %ld per-cpu pages\n",
752 PFN_UP(node_percpu[i]));
754 /* Track the type of memory on each node */
755 if (zones_size[ZONE_NORMAL] || zones_size[ZONE_DMA])
756 node_set_state(i, N_NORMAL_MEMORY);
757 #ifdef CONFIG_HIGHMEM
759 node_set_state(i, N_HIGH_MEMORY);
768 /* which logical CPUs are on which nodes */
769 struct cpumask node_2_cpu_mask[MAX_NUMNODES] __write_once;
770 EXPORT_SYMBOL(node_2_cpu_mask);
772 /* which node each logical CPU is on */
773 char cpu_2_node[NR_CPUS] __write_once __attribute__((aligned(L2_CACHE_BYTES)));
774 EXPORT_SYMBOL(cpu_2_node);
776 /* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
777 static int __init cpu_to_bound_node(int cpu, struct cpumask* unbound_cpus)
779 if (!cpu_possible(cpu) || cpumask_test_cpu(cpu, unbound_cpus))
782 return cpu_to_node(cpu);
785 /* Return number of immediately-adjacent tiles sharing the same NUMA node. */
786 static int __init node_neighbors(int node, int cpu,
787 struct cpumask *unbound_cpus)
794 if (x > 0 && cpu_to_bound_node(cpu-1, unbound_cpus) == node)
796 if (x < w-1 && cpu_to_bound_node(cpu+1, unbound_cpus) == node)
798 if (y > 0 && cpu_to_bound_node(cpu-w, unbound_cpus) == node)
800 if (y < h-1 && cpu_to_bound_node(cpu+w, unbound_cpus) == node)
805 static void __init setup_numa_mapping(void)
807 int distance[MAX_NUMNODES][NR_CPUS];
809 int cpu, node, cpus, i, x, y;
810 int num_nodes = num_online_nodes();
811 struct cpumask unbound_cpus;
812 nodemask_t default_nodes;
814 cpumask_clear(&unbound_cpus);
816 /* Get set of nodes we will use for defaults */
817 nodes_andnot(default_nodes, node_online_map, isolnodes);
818 if (nodes_empty(default_nodes)) {
819 BUG_ON(!node_isset(0, node_online_map));
820 pr_err("Forcing NUMA node zero available as a default node\n");
821 node_set(0, default_nodes);
824 /* Populate the distance[] array */
825 memset(distance, -1, sizeof(distance));
827 for (coord.y = 0; coord.y < smp_height; ++coord.y) {
828 for (coord.x = 0; coord.x < smp_width;
830 BUG_ON(cpu >= nr_cpu_ids);
831 if (!cpu_possible(cpu)) {
832 cpu_2_node[cpu] = -1;
835 for_each_node_mask(node, default_nodes) {
836 HV_MemoryControllerInfo info =
837 hv_inquire_memory_controller(
838 coord, node_controller[node]);
839 distance[node][cpu] =
840 ABS(info.coord.x) + ABS(info.coord.y);
842 cpumask_set_cpu(cpu, &unbound_cpus);
848 * Round-robin through the NUMA nodes until all the cpus are
849 * assigned. We could be more clever here (e.g. create four
850 * sorted linked lists on the same set of cpu nodes, and pull
851 * off them in round-robin sequence, removing from all four
852 * lists each time) but given the relatively small numbers
853 * involved, O(n^2) seem OK for a one-time cost.
855 node = first_node(default_nodes);
856 while (!cpumask_empty(&unbound_cpus)) {
858 int best_distance = INT_MAX;
859 for (cpu = 0; cpu < cpus; ++cpu) {
860 if (cpumask_test_cpu(cpu, &unbound_cpus)) {
862 * Compute metric, which is how much
863 * closer the cpu is to this memory
864 * controller than the others, shifted
865 * up, and then the number of
866 * neighbors already in the node as an
867 * epsilon adjustment to try to keep
870 int d = distance[node][cpu] * num_nodes;
871 for_each_node_mask(i, default_nodes) {
873 d -= distance[i][cpu];
875 d *= 8; /* allow space for epsilon */
876 d -= node_neighbors(node, cpu, &unbound_cpus);
877 if (d < best_distance) {
883 BUG_ON(best_cpu < 0);
884 cpumask_set_cpu(best_cpu, &node_2_cpu_mask[node]);
885 cpu_2_node[best_cpu] = node;
886 cpumask_clear_cpu(best_cpu, &unbound_cpus);
887 node = next_node(node, default_nodes);
888 if (node == MAX_NUMNODES)
889 node = first_node(default_nodes);
892 /* Print out node assignments and set defaults for disabled cpus */
894 for (y = 0; y < smp_height; ++y) {
895 printk(KERN_DEBUG "NUMA cpu-to-node row %d:", y);
896 for (x = 0; x < smp_width; ++x, ++cpu) {
897 if (cpu_to_node(cpu) < 0) {
899 cpu_2_node[cpu] = first_node(default_nodes);
901 pr_cont(" %d", cpu_to_node(cpu));
908 static struct cpu cpu_devices[NR_CPUS];
910 static int __init topology_init(void)
914 for_each_online_node(i)
915 register_one_node(i);
917 for (i = 0; i < smp_height * smp_width; ++i)
918 register_cpu(&cpu_devices[i], i);
923 subsys_initcall(topology_init);
925 #else /* !CONFIG_NUMA */
927 #define setup_numa_mapping() do { } while (0)
929 #endif /* CONFIG_NUMA */
932 * Initialize hugepage support on this cpu. We do this on all cores
933 * early in boot: before argument parsing for the boot cpu, and after
934 * argument parsing but before the init functions run on the secondaries.
935 * So the values we set up here in the hypervisor may be overridden on
936 * the boot cpu as arguments are parsed.
938 static __cpuinit void init_super_pages(void)
940 #ifdef CONFIG_HUGETLB_SUPER_PAGES
942 for (i = 0; i < HUGE_SHIFT_ENTRIES; ++i)
943 hv_set_pte_super_shift(i, huge_shift[i]);
948 * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
949 * @boot: Is this the boot cpu?
951 * Called from setup_arch() on the boot cpu, or online_secondary().
953 void __cpuinit setup_cpu(int boot)
955 /* The boot cpu sets up its permanent mappings much earlier. */
957 store_permanent_mappings();
959 /* Allow asynchronous TLB interrupts. */
960 #if CHIP_HAS_TILE_DMA()
961 arch_local_irq_unmask(INT_DMATLB_MISS);
962 arch_local_irq_unmask(INT_DMATLB_ACCESS);
964 #if CHIP_HAS_SN_PROC()
965 arch_local_irq_unmask(INT_SNITLB_MISS);
968 arch_local_irq_unmask(INT_SINGLE_STEP_K);
972 * Allow user access to many generic SPRs, like the cycle
973 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
975 __insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0, 1);
978 /* Static network is not restricted. */
979 __insn_mtspr(SPR_MPL_SN_ACCESS_SET_0, 1);
981 #if CHIP_HAS_SN_PROC()
982 __insn_mtspr(SPR_MPL_SN_NOTIFY_SET_0, 1);
983 __insn_mtspr(SPR_MPL_SN_CPL_SET_0, 1);
987 * Set the MPL for interrupt control 0 & 1 to the corresponding
988 * values. This includes access to the SYSTEM_SAVE and EX_CONTEXT
989 * SPRs, as well as the interrupt mask.
991 __insn_mtspr(SPR_MPL_INTCTRL_0_SET_0, 1);
992 __insn_mtspr(SPR_MPL_INTCTRL_1_SET_1, 1);
994 /* Initialize IRQ support for this cpu. */
997 #ifdef CONFIG_HARDWALL
998 /* Reset the network state on this cpu. */
999 reset_network_state();
1005 #ifdef CONFIG_BLK_DEV_INITRD
1007 static int __initdata set_initramfs_file;
1008 static char __initdata initramfs_file[128] = "initramfs";
1010 static int __init setup_initramfs_file(char *str)
1014 strncpy(initramfs_file, str, sizeof(initramfs_file) - 1);
1015 set_initramfs_file = 1;
1019 early_param("initramfs_file", setup_initramfs_file);
1022 * We look for a file called "initramfs" in the hvfs. If there is one, we
1023 * allocate some memory for it and it will be unpacked to the initramfs.
1024 * If it's compressed, the initd code will uncompress it first.
1026 static void __init load_hv_initrd(void)
1028 HV_FS_StatInfo stat;
1032 fd = hv_fs_findfile((HV_VirtAddr) initramfs_file);
1033 if (fd == HV_ENOENT) {
1034 if (set_initramfs_file) {
1035 pr_warning("No such hvfs initramfs file '%s'\n",
1039 /* Try old backwards-compatible name. */
1040 fd = hv_fs_findfile((HV_VirtAddr)"initramfs.cpio.gz");
1041 if (fd == HV_ENOENT)
1046 stat = hv_fs_fstat(fd);
1047 BUG_ON(stat.size < 0);
1048 if (stat.flags & HV_FS_ISDIR) {
1049 pr_warning("Ignoring hvfs file '%s': it's a directory.\n",
1053 initrd = alloc_bootmem_pages(stat.size);
1054 rc = hv_fs_pread(fd, (HV_VirtAddr) initrd, stat.size, 0);
1055 if (rc != stat.size) {
1056 pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
1057 stat.size, initramfs_file, rc);
1058 free_initrd_mem((unsigned long) initrd, stat.size);
1061 initrd_start = (unsigned long) initrd;
1062 initrd_end = initrd_start + stat.size;
1065 void __init free_initrd_mem(unsigned long begin, unsigned long end)
1067 free_bootmem(__pa(begin), end - begin);
1071 static inline void load_hv_initrd(void) {}
1072 #endif /* CONFIG_BLK_DEV_INITRD */
1074 static void __init validate_hv(void)
1077 * It may already be too late, but let's check our built-in
1078 * configuration against what the hypervisor is providing.
1080 unsigned long glue_size = hv_sysconf(HV_SYSCONF_GLUE_SIZE);
1081 int hv_page_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL);
1082 int hv_hpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE);
1083 HV_ASIDRange asid_range;
1086 HV_Topology topology = hv_inquire_topology();
1087 BUG_ON(topology.coord.x != 0 || topology.coord.y != 0);
1088 if (topology.width != 1 || topology.height != 1) {
1089 pr_warning("Warning: booting UP kernel on %dx%d grid;"
1090 " will ignore all but first tile.\n",
1091 topology.width, topology.height);
1095 if (PAGE_OFFSET + HV_GLUE_START_CPA + glue_size > (unsigned long)_text)
1096 early_panic("Hypervisor glue size %ld is too big!\n",
1098 if (hv_page_size != PAGE_SIZE)
1099 early_panic("Hypervisor page size %#x != our %#lx\n",
1100 hv_page_size, PAGE_SIZE);
1101 if (hv_hpage_size != HPAGE_SIZE)
1102 early_panic("Hypervisor huge page size %#x != our %#lx\n",
1103 hv_hpage_size, HPAGE_SIZE);
1107 * Some hypervisor APIs take a pointer to a bitmap array
1108 * whose size is at least the number of cpus on the chip.
1109 * We use a struct cpumask for this, so it must be big enough.
1111 if ((smp_height * smp_width) > nr_cpu_ids)
1112 early_panic("Hypervisor %d x %d grid too big for Linux"
1113 " NR_CPUS %d\n", smp_height, smp_width,
1118 * Check that we're using allowed ASIDs, and initialize the
1119 * various asid variables to their appropriate initial states.
1121 asid_range = hv_inquire_asid(0);
1122 __get_cpu_var(current_asid) = min_asid = asid_range.start;
1123 max_asid = asid_range.start + asid_range.size - 1;
1125 if (hv_confstr(HV_CONFSTR_CHIP_MODEL, (HV_VirtAddr)chip_model,
1126 sizeof(chip_model)) < 0) {
1127 pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
1128 strlcpy(chip_model, "unknown", sizeof(chip_model));
1132 static void __init validate_va(void)
1134 #ifndef __tilegx__ /* FIXME: GX: probably some validation relevant here */
1136 * Similarly, make sure we're only using allowed VAs.
1137 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_INTRPT,
1138 * and 0 .. KERNEL_HIGH_VADDR.
1139 * In addition, make sure we CAN'T use the end of memory, since
1140 * we use the last chunk of each pgd for the pgd_list.
1142 int i, user_kernel_ok = 0;
1143 unsigned long max_va = 0;
1144 unsigned long list_va =
1145 ((PGD_LIST_OFFSET / sizeof(pgd_t)) << PGDIR_SHIFT);
1147 for (i = 0; ; ++i) {
1148 HV_VirtAddrRange range = hv_inquire_virtual(i);
1149 if (range.size == 0)
1151 if (range.start <= MEM_USER_INTRPT &&
1152 range.start + range.size >= MEM_HV_INTRPT)
1154 if (range.start == 0)
1155 max_va = range.size;
1156 BUG_ON(range.start + range.size > list_va);
1158 if (!user_kernel_ok)
1159 early_panic("Hypervisor not configured for user/kernel VAs\n");
1161 early_panic("Hypervisor not configured for low VAs\n");
1162 if (max_va < KERNEL_HIGH_VADDR)
1163 early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
1164 max_va, KERNEL_HIGH_VADDR);
1166 /* Kernel PCs must have their high bit set; see intvec.S. */
1167 if ((long)VMALLOC_START >= 0)
1169 "Linux VMALLOC region below the 2GB line (%#lx)!\n"
1170 "Reconfigure the kernel with fewer NR_HUGE_VMAPS\n"
1171 "or smaller VMALLOC_RESERVE.\n",
1177 * cpu_lotar_map lists all the cpus that are valid for the supervisor
1178 * to cache data on at a page level, i.e. what cpus can be placed in
1179 * the LOTAR field of a PTE. It is equivalent to the set of possible
1180 * cpus plus any other cpus that are willing to share their cache.
1181 * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
1183 struct cpumask __write_once cpu_lotar_map;
1184 EXPORT_SYMBOL(cpu_lotar_map);
1186 #if CHIP_HAS_CBOX_HOME_MAP()
1188 * hash_for_home_map lists all the tiles that hash-for-home data
1189 * will be cached on. Note that this may includes tiles that are not
1190 * valid for this supervisor to use otherwise (e.g. if a hypervisor
1191 * device is being shared between multiple supervisors).
1192 * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
1194 struct cpumask hash_for_home_map;
1195 EXPORT_SYMBOL(hash_for_home_map);
1199 * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
1200 * flush on our behalf. It is set to cpu_possible_mask OR'ed with
1201 * hash_for_home_map, and it is what should be passed to
1202 * hv_flush_remote() to flush all caches. Note that if there are
1203 * dedicated hypervisor driver tiles that have authorized use of their
1204 * cache, those tiles will only appear in cpu_lotar_map, NOT in
1205 * cpu_cacheable_map, as they are a special case.
1207 struct cpumask __write_once cpu_cacheable_map;
1208 EXPORT_SYMBOL(cpu_cacheable_map);
1210 static __initdata struct cpumask disabled_map;
1212 static int __init disabled_cpus(char *str)
1214 int boot_cpu = smp_processor_id();
1216 if (str == NULL || cpulist_parse_crop(str, &disabled_map) != 0)
1218 if (cpumask_test_cpu(boot_cpu, &disabled_map)) {
1219 pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu);
1220 cpumask_clear_cpu(boot_cpu, &disabled_map);
1225 early_param("disabled_cpus", disabled_cpus);
1227 void __init print_disabled_cpus(void)
1229 if (!cpumask_empty(&disabled_map)) {
1231 cpulist_scnprintf(buf, sizeof(buf), &disabled_map);
1232 pr_info("CPUs not available for Linux: %s\n", buf);
1236 static void __init setup_cpu_maps(void)
1238 struct cpumask hv_disabled_map, cpu_possible_init;
1239 int boot_cpu = smp_processor_id();
1242 /* Learn which cpus are allowed by the hypervisor. */
1243 rc = hv_inquire_tiles(HV_INQ_TILES_AVAIL,
1244 (HV_VirtAddr) cpumask_bits(&cpu_possible_init),
1245 sizeof(cpu_cacheable_map));
1247 early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc);
1248 if (!cpumask_test_cpu(boot_cpu, &cpu_possible_init))
1249 early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu);
1251 /* Compute the cpus disabled by the hvconfig file. */
1252 cpumask_complement(&hv_disabled_map, &cpu_possible_init);
1254 /* Include them with the cpus disabled by "disabled_cpus". */
1255 cpumask_or(&disabled_map, &disabled_map, &hv_disabled_map);
1258 * Disable every cpu after "setup_max_cpus". But don't mark
1259 * as disabled the cpus that are outside of our initial rectangle,
1260 * since that turns out to be confusing.
1262 cpus = 1; /* this cpu */
1263 cpumask_set_cpu(boot_cpu, &disabled_map); /* ignore this cpu */
1264 for (i = 0; cpus < setup_max_cpus; ++i)
1265 if (!cpumask_test_cpu(i, &disabled_map))
1267 for (; i < smp_height * smp_width; ++i)
1268 cpumask_set_cpu(i, &disabled_map);
1269 cpumask_clear_cpu(boot_cpu, &disabled_map); /* reset this cpu */
1270 for (i = smp_height * smp_width; i < NR_CPUS; ++i)
1271 cpumask_clear_cpu(i, &disabled_map);
1274 * Setup cpu_possible map as every cpu allocated to us, minus
1275 * the results of any "disabled_cpus" settings.
1277 cpumask_andnot(&cpu_possible_init, &cpu_possible_init, &disabled_map);
1278 init_cpu_possible(&cpu_possible_init);
1280 /* Learn which cpus are valid for LOTAR caching. */
1281 rc = hv_inquire_tiles(HV_INQ_TILES_LOTAR,
1282 (HV_VirtAddr) cpumask_bits(&cpu_lotar_map),
1283 sizeof(cpu_lotar_map));
1285 pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
1286 cpu_lotar_map = *cpu_possible_mask;
1289 #if CHIP_HAS_CBOX_HOME_MAP()
1290 /* Retrieve set of CPUs used for hash-for-home caching */
1291 rc = hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE,
1292 (HV_VirtAddr) hash_for_home_map.bits,
1293 sizeof(hash_for_home_map));
1295 early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc);
1296 cpumask_or(&cpu_cacheable_map, cpu_possible_mask, &hash_for_home_map);
1298 cpu_cacheable_map = *cpu_possible_mask;
1303 static int __init dataplane(char *str)
1305 pr_warning("WARNING: dataplane support disabled in this kernel\n");
1309 early_param("dataplane", dataplane);
1311 #ifdef CONFIG_CMDLINE_BOOL
1312 static char __initdata builtin_cmdline[COMMAND_LINE_SIZE] = CONFIG_CMDLINE;
1315 void __init setup_arch(char **cmdline_p)
1319 #if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
1320 len = hv_get_command_line((HV_VirtAddr) boot_command_line,
1322 if (boot_command_line[0])
1323 pr_warning("WARNING: ignoring dynamic command line \"%s\"\n",
1325 strlcpy(boot_command_line, builtin_cmdline, COMMAND_LINE_SIZE);
1328 #if defined(CONFIG_CMDLINE_BOOL)
1329 if (builtin_cmdline[0]) {
1330 int builtin_len = strlcpy(boot_command_line, builtin_cmdline,
1332 if (builtin_len < COMMAND_LINE_SIZE-1)
1333 boot_command_line[builtin_len++] = ' ';
1334 hv_cmdline = &boot_command_line[builtin_len];
1335 len = COMMAND_LINE_SIZE - builtin_len;
1339 hv_cmdline = boot_command_line;
1340 len = COMMAND_LINE_SIZE;
1342 len = hv_get_command_line((HV_VirtAddr) hv_cmdline, len);
1343 if (len < 0 || len > COMMAND_LINE_SIZE)
1344 early_panic("hv_get_command_line failed: %d\n", len);
1347 *cmdline_p = boot_command_line;
1349 /* Set disabled_map and setup_max_cpus very early */
1350 parse_early_param();
1352 /* Make sure the kernel is compatible with the hypervisor. */
1359 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1361 * Initialize the PCI structures. This is done before memory
1362 * setup so that we know whether or not a pci_reserve region
1365 if (tile_pci_init() == 0)
1368 /* PCI systems reserve a region just below 4GB for mapping iomem. */
1369 pci_reserve_end_pfn = (1 << (32 - PAGE_SHIFT));
1370 pci_reserve_start_pfn = pci_reserve_end_pfn -
1371 (pci_reserve_mb << (20 - PAGE_SHIFT));
1374 init_mm.start_code = (unsigned long) _text;
1375 init_mm.end_code = (unsigned long) _etext;
1376 init_mm.end_data = (unsigned long) _edata;
1377 init_mm.brk = (unsigned long) _end;
1380 store_permanent_mappings();
1381 setup_bootmem_allocator();
1384 * NOTE: before this point _nobody_ is allowed to allocate
1385 * any memory using the bootmem allocator.
1388 #ifdef CONFIG_SWIOTLB
1393 setup_numa_mapping();
1403 * Set up per-cpu memory.
1406 unsigned long __per_cpu_offset[NR_CPUS] __write_once;
1407 EXPORT_SYMBOL(__per_cpu_offset);
1409 static size_t __initdata pfn_offset[MAX_NUMNODES] = { 0 };
1410 static unsigned long __initdata percpu_pfn[NR_CPUS] = { 0 };
1413 * As the percpu code allocates pages, we return the pages from the
1414 * end of the node for the specified cpu.
1416 static void *__init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
1418 int nid = cpu_to_node(cpu);
1419 unsigned long pfn = node_percpu_pfn[nid] + pfn_offset[nid];
1421 BUG_ON(size % PAGE_SIZE != 0);
1422 pfn_offset[nid] += size / PAGE_SIZE;
1423 BUG_ON(node_percpu[nid] < size);
1424 node_percpu[nid] -= size;
1425 if (percpu_pfn[cpu] == 0)
1426 percpu_pfn[cpu] = pfn;
1427 return pfn_to_kaddr(pfn);
1431 * Pages reserved for percpu memory are not freeable, and in any case we are
1432 * on a short path to panic() in setup_per_cpu_area() at this point anyway.
1434 static void __init pcpu_fc_free(void *ptr, size_t size)
1439 * Set up vmalloc page tables using bootmem for the percpu code.
1441 static void __init pcpu_fc_populate_pte(unsigned long addr)
1448 BUG_ON(pgd_addr_invalid(addr));
1449 if (addr < VMALLOC_START || addr >= VMALLOC_END)
1450 panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx;"
1451 " try increasing CONFIG_VMALLOC_RESERVE\n",
1452 addr, VMALLOC_START, VMALLOC_END);
1454 pgd = swapper_pg_dir + pgd_index(addr);
1455 pud = pud_offset(pgd, addr);
1456 BUG_ON(!pud_present(*pud));
1457 pmd = pmd_offset(pud, addr);
1458 if (pmd_present(*pmd)) {
1459 BUG_ON(pmd_huge_page(*pmd));
1461 pte = __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE,
1462 HV_PAGE_TABLE_ALIGN, 0);
1463 pmd_populate_kernel(&init_mm, pmd, pte);
1467 void __init setup_per_cpu_areas(void)
1470 unsigned long delta, pfn, lowmem_va;
1471 unsigned long size = percpu_size();
1475 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_fc_alloc,
1476 pcpu_fc_free, pcpu_fc_populate_pte);
1478 panic("Cannot initialize percpu area (err=%d)", rc);
1480 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1481 for_each_possible_cpu(cpu) {
1482 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1484 /* finv the copy out of cache so we can change homecache */
1485 ptr = pcpu_base_addr + pcpu_unit_offsets[cpu];
1486 __finv_buffer(ptr, size);
1487 pfn = percpu_pfn[cpu];
1489 /* Rewrite the page tables to cache on that cpu */
1490 pg = pfn_to_page(pfn);
1491 for (i = 0; i < size; i += PAGE_SIZE, ++pfn, ++pg) {
1493 /* Update the vmalloc mapping and page home. */
1494 unsigned long addr = (unsigned long)ptr + i;
1495 pte_t *ptep = virt_to_pte(NULL, addr);
1497 BUG_ON(pfn != pte_pfn(pte));
1498 pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
1499 pte = set_remote_cache_cpu(pte, cpu);
1500 set_pte_at(&init_mm, addr, ptep, pte);
1502 /* Update the lowmem mapping for consistency. */
1503 lowmem_va = (unsigned long)pfn_to_kaddr(pfn);
1504 ptep = virt_to_pte(NULL, lowmem_va);
1505 if (pte_huge(*ptep)) {
1506 printk(KERN_DEBUG "early shatter of huge page"
1507 " at %#lx\n", lowmem_va);
1508 shatter_pmd((pmd_t *)ptep);
1509 ptep = virt_to_pte(NULL, lowmem_va);
1510 BUG_ON(pte_huge(*ptep));
1512 BUG_ON(pfn != pte_pfn(*ptep));
1513 set_pte_at(&init_mm, lowmem_va, ptep, pte);
1517 /* Set our thread pointer appropriately. */
1518 set_my_cpu_offset(__per_cpu_offset[smp_processor_id()]);
1520 /* Make sure the finv's have completed. */
1523 /* Flush the TLB so we reference it properly from here on out. */
1524 local_flush_tlb_all();
1527 static struct resource data_resource = {
1528 .name = "Kernel data",
1531 .flags = IORESOURCE_BUSY | IORESOURCE_MEM
1534 static struct resource code_resource = {
1535 .name = "Kernel code",
1538 .flags = IORESOURCE_BUSY | IORESOURCE_MEM
1542 * On Pro, we reserve all resources above 4GB so that PCI won't try to put
1543 * mappings above 4GB.
1545 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1546 static struct resource* __init
1547 insert_non_bus_resource(void)
1549 struct resource *res =
1550 kzalloc(sizeof(struct resource), GFP_ATOMIC);
1551 res->name = "Non-Bus Physical Address Space";
1552 res->start = (1ULL << 32);
1554 res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1555 if (insert_resource(&iomem_resource, res)) {
1563 static struct resource* __init
1564 insert_ram_resource(u64 start_pfn, u64 end_pfn)
1566 struct resource *res =
1567 kzalloc(sizeof(struct resource), GFP_ATOMIC);
1568 res->name = "System RAM";
1569 res->start = start_pfn << PAGE_SHIFT;
1570 res->end = (end_pfn << PAGE_SHIFT) - 1;
1571 res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1572 if (insert_resource(&iomem_resource, res)) {
1580 * Request address space for all standard resources
1582 * If the system includes PCI root complex drivers, we need to create
1583 * a window just below 4GB where PCI BARs can be mapped.
1585 static int __init request_standard_resources(void)
1588 enum { CODE_DELTA = MEM_SV_INTRPT - PAGE_OFFSET };
1590 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1591 insert_non_bus_resource();
1594 for_each_online_node(i) {
1595 u64 start_pfn = node_start_pfn[i];
1596 u64 end_pfn = node_end_pfn[i];
1598 #if defined(CONFIG_PCI) && !defined(__tilegx__)
1599 if (start_pfn <= pci_reserve_start_pfn &&
1600 end_pfn > pci_reserve_start_pfn) {
1601 if (end_pfn > pci_reserve_end_pfn)
1602 insert_ram_resource(pci_reserve_end_pfn,
1604 end_pfn = pci_reserve_start_pfn;
1607 insert_ram_resource(start_pfn, end_pfn);
1610 code_resource.start = __pa(_text - CODE_DELTA);
1611 code_resource.end = __pa(_etext - CODE_DELTA)-1;
1612 data_resource.start = __pa(_sdata);
1613 data_resource.end = __pa(_end)-1;
1615 insert_resource(&iomem_resource, &code_resource);
1616 insert_resource(&iomem_resource, &data_resource);
1619 insert_resource(&iomem_resource, &crashk_res);
1625 subsys_initcall(request_standard_resources);