1/*
2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
3 *
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.
7 *
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
12 *   more details.
13 */
14
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>
28#include <linux/io.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>
42
43/* <linux/smp.h> doesn't provide this definition. */
44#ifndef CONFIG_SMP
45#define setup_max_cpus 1
46#endif
47
48static inline int ABS(int x) { return x >= 0 ? x : -x; }
49
50/* Chip information */
51char chip_model[64] __write_once;
52
53#ifdef CONFIG_VT
54struct screen_info screen_info;
55#endif
56
57struct pglist_data node_data[MAX_NUMNODES] __read_mostly;
58EXPORT_SYMBOL(node_data);
59
60/* Information on the NUMA nodes that we compute early */
61unsigned long node_start_pfn[MAX_NUMNODES];
62unsigned long node_end_pfn[MAX_NUMNODES];
63unsigned long __initdata node_memmap_pfn[MAX_NUMNODES];
64unsigned long __initdata node_percpu_pfn[MAX_NUMNODES];
65unsigned long __initdata node_free_pfn[MAX_NUMNODES];
66
67static unsigned long __initdata node_percpu[MAX_NUMNODES];
68
69/*
70 * per-CPU stack and boot info.
71 */
72DEFINE_PER_CPU(unsigned long, boot_sp) =
73	(unsigned long)init_stack + THREAD_SIZE;
74
75#ifdef CONFIG_SMP
76DEFINE_PER_CPU(unsigned long, boot_pc) = (unsigned long)start_kernel;
77#else
78/*
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.
81 */
82unsigned long __initdata boot_pc = (unsigned long)start_kernel;
83#endif
84
85#ifdef CONFIG_HIGHMEM
86/* Page frame index of end of lowmem on each controller. */
87unsigned long node_lowmem_end_pfn[MAX_NUMNODES];
88
89/* Number of pages that can be mapped into lowmem. */
90static unsigned long __initdata mappable_physpages;
91#endif
92
93/* Data on which physical memory controller corresponds to which NUMA node */
94int node_controller[MAX_NUMNODES] = { [0 ... MAX_NUMNODES-1] = -1 };
95
96#ifdef CONFIG_HIGHMEM
97/* Map information from VAs to PAs */
98unsigned long pbase_map[1 << (32 - HPAGE_SHIFT)]
99  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
100EXPORT_SYMBOL(pbase_map);
101
102/* Map information from PAs to VAs */
103void *vbase_map[NR_PA_HIGHBIT_VALUES]
104  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
105EXPORT_SYMBOL(vbase_map);
106#endif
107
108/* Node number as a function of the high PA bits */
109int highbits_to_node[NR_PA_HIGHBIT_VALUES] __write_once;
110EXPORT_SYMBOL(highbits_to_node);
111
112static unsigned int __initdata maxmem_pfn = -1U;
113static unsigned int __initdata maxnodemem_pfn[MAX_NUMNODES] = {
114	[0 ... MAX_NUMNODES-1] = -1U
115};
116static nodemask_t __initdata isolnodes;
117
118#if defined(CONFIG_PCI) && !defined(__tilegx__)
119enum { DEFAULT_PCI_RESERVE_MB = 64 };
120static unsigned int __initdata pci_reserve_mb = DEFAULT_PCI_RESERVE_MB;
121unsigned long __initdata pci_reserve_start_pfn = -1U;
122unsigned long __initdata pci_reserve_end_pfn = -1U;
123#endif
124
125static int __init setup_maxmem(char *str)
126{
127	unsigned long long maxmem;
128	if (str == NULL || (maxmem = memparse(str, NULL)) == 0)
129		return -EINVAL;
130
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));
134	return 0;
135}
136early_param("maxmem", setup_maxmem);
137
138static int __init setup_maxnodemem(char *str)
139{
140	char *endp;
141	unsigned long long maxnodemem;
142	long node;
143
144	node = str ? simple_strtoul(str, &endp, 0) : INT_MAX;
145	if (node >= MAX_NUMNODES || *endp != ':')
146		return -EINVAL;
147
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));
153	return 0;
154}
155early_param("maxnodemem", setup_maxnodemem);
156
157struct memmap_entry {
158	u64 addr;	/* start of memory segment */
159	u64 size;	/* size of memory segment */
160};
161static struct memmap_entry memmap_map[64];
162static int memmap_nr;
163
164static void add_memmap_region(u64 addr, u64 size)
165{
166	if (memmap_nr >= ARRAY_SIZE(memmap_map)) {
167		pr_err("Ooops! Too many entries in the memory map!\n");
168		return;
169	}
170	memmap_map[memmap_nr].addr = addr;
171	memmap_map[memmap_nr].size = size;
172	memmap_nr++;
173}
174
175static int __init setup_memmap(char *p)
176{
177	char *oldp;
178	u64 start_at, mem_size;
179
180	if (!p)
181		return -EINVAL;
182
183	if (!strncmp(p, "exactmap", 8)) {
184		pr_err("\"memmap=exactmap\" not valid on tile\n");
185		return 0;
186	}
187
188	oldp = p;
189	mem_size = memparse(p, &p);
190	if (p == oldp)
191		return -EINVAL;
192
193	if (*p == '@') {
194		pr_err("\"memmap=nn@ss\" (force RAM) invalid on tile\n");
195	} else if (*p == '#') {
196		pr_err("\"memmap=nn#ss\" (force ACPI data) invalid on tile\n");
197	} else if (*p == '$') {
198		start_at = memparse(p+1, &p);
199		add_memmap_region(start_at, mem_size);
200	} else {
201		if (mem_size == 0)
202			return -EINVAL;
203		maxmem_pfn = (mem_size >> HPAGE_SHIFT) <<
204			(HPAGE_SHIFT - PAGE_SHIFT);
205	}
206	return *p == '\0' ? 0 : -EINVAL;
207}
208early_param("memmap", setup_memmap);
209
210static int __init setup_mem(char *str)
211{
212	return setup_maxmem(str);
213}
214early_param("mem", setup_mem);  /* compatibility with x86 */
215
216static int __init setup_isolnodes(char *str)
217{
218	char buf[MAX_NUMNODES * 5];
219	if (str == NULL || nodelist_parse(str, isolnodes) != 0)
220		return -EINVAL;
221
222	nodelist_scnprintf(buf, sizeof(buf), isolnodes);
223	pr_info("Set isolnodes value to '%s'\n", buf);
224	return 0;
225}
226early_param("isolnodes", setup_isolnodes);
227
228#if defined(CONFIG_PCI) && !defined(__tilegx__)
229static int __init setup_pci_reserve(char* str)
230{
231	if (str == NULL || kstrtouint(str, 0, &pci_reserve_mb) != 0 ||
232	    pci_reserve_mb > 3 * 1024)
233		return -EINVAL;
234
235	pr_info("Reserving %dMB for PCIE root complex mappings\n",
236		pci_reserve_mb);
237	return 0;
238}
239early_param("pci_reserve", setup_pci_reserve);
240#endif
241
242#ifndef __tilegx__
243/*
244 * vmalloc=size forces the vmalloc area to be exactly 'size' bytes.
245 * This can be used to increase (or decrease) the vmalloc area.
246 */
247static int __init parse_vmalloc(char *arg)
248{
249	if (!arg)
250		return -EINVAL;
251
252	VMALLOC_RESERVE = (memparse(arg, &arg) + PGDIR_SIZE - 1) & PGDIR_MASK;
253
254	/* See validate_va() for more on this test. */
255	if ((long)_VMALLOC_START >= 0)
256		early_panic("\"vmalloc=%#lx\" value too large: maximum %#lx\n",
257			    VMALLOC_RESERVE, _VMALLOC_END - 0x80000000UL);
258
259	return 0;
260}
261early_param("vmalloc", parse_vmalloc);
262#endif
263
264#ifdef CONFIG_HIGHMEM
265/*
266 * Determine for each controller where its lowmem is mapped and how much of
267 * it is mapped there.  On controller zero, the first few megabytes are
268 * already mapped in as code at MEM_SV_START, so in principle we could
269 * start our data mappings higher up, but for now we don't bother, to avoid
270 * additional confusion.
271 *
272 * One question is whether, on systems with more than 768 Mb and
273 * controllers of different sizes, to map in a proportionate amount of
274 * each one, or to try to map the same amount from each controller.
275 * (E.g. if we have three controllers with 256MB, 1GB, and 256MB
276 * respectively, do we map 256MB from each, or do we map 128 MB, 512
277 * MB, and 128 MB respectively?)  For now we use a proportionate
278 * solution like the latter.
279 *
280 * The VA/PA mapping demands that we align our decisions at 16 MB
281 * boundaries so that we can rapidly convert VA to PA.
282 */
283static void *__init setup_pa_va_mapping(void)
284{
285	unsigned long curr_pages = 0;
286	unsigned long vaddr = PAGE_OFFSET;
287	nodemask_t highonlynodes = isolnodes;
288	int i, j;
289
290	memset(pbase_map, -1, sizeof(pbase_map));
291	memset(vbase_map, -1, sizeof(vbase_map));
292
293	/* Node zero cannot be isolated for LOWMEM purposes. */
294	node_clear(0, highonlynodes);
295
296	/* Count up the number of pages on non-highonlynodes controllers. */
297	mappable_physpages = 0;
298	for_each_online_node(i) {
299		if (!node_isset(i, highonlynodes))
300			mappable_physpages +=
301				node_end_pfn[i] - node_start_pfn[i];
302	}
303
304	for_each_online_node(i) {
305		unsigned long start = node_start_pfn[i];
306		unsigned long end = node_end_pfn[i];
307		unsigned long size = end - start;
308		unsigned long vaddr_end;
309
310		if (node_isset(i, highonlynodes)) {
311			/* Mark this controller as having no lowmem. */
312			node_lowmem_end_pfn[i] = start;
313			continue;
314		}
315
316		curr_pages += size;
317		if (mappable_physpages > MAXMEM_PFN) {
318			vaddr_end = PAGE_OFFSET +
319				(((u64)curr_pages * MAXMEM_PFN /
320				  mappable_physpages)
321				 << PAGE_SHIFT);
322		} else {
323			vaddr_end = PAGE_OFFSET + (curr_pages << PAGE_SHIFT);
324		}
325		for (j = 0; vaddr < vaddr_end; vaddr += HPAGE_SIZE, ++j) {
326			unsigned long this_pfn =
327				start + (j << HUGETLB_PAGE_ORDER);
328			pbase_map[vaddr >> HPAGE_SHIFT] = this_pfn;
329			if (vbase_map[__pfn_to_highbits(this_pfn)] ==
330			    (void *)-1)
331				vbase_map[__pfn_to_highbits(this_pfn)] =
332					(void *)(vaddr & HPAGE_MASK);
333		}
334		node_lowmem_end_pfn[i] = start + (j << HUGETLB_PAGE_ORDER);
335		BUG_ON(node_lowmem_end_pfn[i] > end);
336	}
337
338	/* Return highest address of any mapped memory. */
339	return (void *)vaddr;
340}
341#endif /* CONFIG_HIGHMEM */
342
343/*
344 * Register our most important memory mappings with the debug stub.
345 *
346 * This is up to 4 mappings for lowmem, one mapping per memory
347 * controller, plus one for our text segment.
348 */
349static void store_permanent_mappings(void)
350{
351	int i;
352
353	for_each_online_node(i) {
354		HV_PhysAddr pa = ((HV_PhysAddr)node_start_pfn[i]) << PAGE_SHIFT;
355#ifdef CONFIG_HIGHMEM
356		HV_PhysAddr high_mapped_pa = node_lowmem_end_pfn[i];
357#else
358		HV_PhysAddr high_mapped_pa = node_end_pfn[i];
359#endif
360
361		unsigned long pages = high_mapped_pa - node_start_pfn[i];
362		HV_VirtAddr addr = (HV_VirtAddr) __va(pa);
363		hv_store_mapping(addr, pages << PAGE_SHIFT, pa);
364	}
365
366	hv_store_mapping((HV_VirtAddr)_text,
367			 (uint32_t)(_einittext - _text), 0);
368}
369
370/*
371 * Use hv_inquire_physical() to populate node_{start,end}_pfn[]
372 * and node_online_map, doing suitable sanity-checking.
373 * Also set min_low_pfn, max_low_pfn, and max_pfn.
374 */
375static void __init setup_memory(void)
376{
377	int i, j;
378	int highbits_seen[NR_PA_HIGHBIT_VALUES] = { 0 };
379#ifdef CONFIG_HIGHMEM
380	long highmem_pages;
381#endif
382#ifndef __tilegx__
383	int cap;
384#endif
385#if defined(CONFIG_HIGHMEM) || defined(__tilegx__)
386	long lowmem_pages;
387#endif
388	unsigned long physpages = 0;
389
390	/* We are using a char to hold the cpu_2_node[] mapping */
391	BUILD_BUG_ON(MAX_NUMNODES > 127);
392
393	/* Discover the ranges of memory available to us */
394	for (i = 0; ; ++i) {
395		unsigned long start, size, end, highbits;
396		HV_PhysAddrRange range = hv_inquire_physical(i);
397		if (range.size == 0)
398			break;
399#ifdef CONFIG_FLATMEM
400		if (i > 0) {
401			pr_err("Can't use discontiguous PAs: %#llx..%#llx\n",
402			       range.size, range.start + range.size);
403			continue;
404		}
405#endif
406#ifndef __tilegx__
407		if ((unsigned long)range.start) {
408			pr_err("Range not at 4GB multiple: %#llx..%#llx\n",
409			       range.start, range.start + range.size);
410			continue;
411		}
412#endif
413		if ((range.start & (HPAGE_SIZE-1)) != 0 ||
414		    (range.size & (HPAGE_SIZE-1)) != 0) {
415			unsigned long long start_pa = range.start;
416			unsigned long long orig_size = range.size;
417			range.start = (start_pa + HPAGE_SIZE - 1) & HPAGE_MASK;
418			range.size -= (range.start - start_pa);
419			range.size &= HPAGE_MASK;
420			pr_err("Range not hugepage-aligned: %#llx..%#llx:"
421			       " now %#llx-%#llx\n",
422			       start_pa, start_pa + orig_size,
423			       range.start, range.start + range.size);
424		}
425		highbits = __pa_to_highbits(range.start);
426		if (highbits >= NR_PA_HIGHBIT_VALUES) {
427			pr_err("PA high bits too high: %#llx..%#llx\n",
428			       range.start, range.start + range.size);
429			continue;
430		}
431		if (highbits_seen[highbits]) {
432			pr_err("Range overlaps in high bits: %#llx..%#llx\n",
433			       range.start, range.start + range.size);
434			continue;
435		}
436		highbits_seen[highbits] = 1;
437		if (PFN_DOWN(range.size) > maxnodemem_pfn[i]) {
438			int max_size = maxnodemem_pfn[i];
439			if (max_size > 0) {
440				pr_err("Maxnodemem reduced node %d to"
441				       " %d pages\n", i, max_size);
442				range.size = PFN_PHYS(max_size);
443			} else {
444				pr_err("Maxnodemem disabled node %d\n", i);
445				continue;
446			}
447		}
448		if (physpages + PFN_DOWN(range.size) > maxmem_pfn) {
449			int max_size = maxmem_pfn - physpages;
450			if (max_size > 0) {
451				pr_err("Maxmem reduced node %d to %d pages\n",
452				       i, max_size);
453				range.size = PFN_PHYS(max_size);
454			} else {
455				pr_err("Maxmem disabled node %d\n", i);
456				continue;
457			}
458		}
459		if (i >= MAX_NUMNODES) {
460			pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
461			       i, range.size, range.size + range.start);
462			continue;
463		}
464
465		start = range.start >> PAGE_SHIFT;
466		size = range.size >> PAGE_SHIFT;
467		end = start + size;
468
469#ifndef __tilegx__
470		if (((HV_PhysAddr)end << PAGE_SHIFT) !=
471		    (range.start + range.size)) {
472			pr_err("PAs too high to represent: %#llx..%#llx\n",
473			       range.start, range.start + range.size);
474			continue;
475		}
476#endif
477#if defined(CONFIG_PCI) && !defined(__tilegx__)
478		/*
479		 * Blocks that overlap the pci reserved region must
480		 * have enough space to hold the maximum percpu data
481		 * region at the top of the range.  If there isn't
482		 * enough space above the reserved region, just
483		 * truncate the node.
484		 */
485		if (start <= pci_reserve_start_pfn &&
486		    end > pci_reserve_start_pfn) {
487			unsigned int per_cpu_size =
488				__per_cpu_end - __per_cpu_start;
489			unsigned int percpu_pages =
490				NR_CPUS * (PFN_UP(per_cpu_size) >> PAGE_SHIFT);
491			if (end < pci_reserve_end_pfn + percpu_pages) {
492				end = pci_reserve_start_pfn;
493				pr_err("PCI mapping region reduced node %d to"
494				       " %ld pages\n", i, end - start);
495			}
496		}
497#endif
498
499		for (j = __pfn_to_highbits(start);
500		     j <= __pfn_to_highbits(end - 1); j++)
501			highbits_to_node[j] = i;
502
503		node_start_pfn[i] = start;
504		node_end_pfn[i] = end;
505		node_controller[i] = range.controller;
506		physpages += size;
507		max_pfn = end;
508
509		/* Mark node as online */
510		node_set(i, node_online_map);
511		node_set(i, node_possible_map);
512	}
513
514#ifndef __tilegx__
515	/*
516	 * For 4KB pages, mem_map "struct page" data is 1% of the size
517	 * of the physical memory, so can be quite big (640 MB for
518	 * four 16G zones).  These structures must be mapped in
519	 * lowmem, and since we currently cap out at about 768 MB,
520	 * it's impractical to try to use this much address space.
521	 * For now, arbitrarily cap the amount of physical memory
522	 * we're willing to use at 8 million pages (32GB of 4KB pages).
523	 */
524	cap = 8 * 1024 * 1024;  /* 8 million pages */
525	if (physpages > cap) {
526		int num_nodes = num_online_nodes();
527		int cap_each = cap / num_nodes;
528		unsigned long dropped_pages = 0;
529		for (i = 0; i < num_nodes; ++i) {
530			int size = node_end_pfn[i] - node_start_pfn[i];
531			if (size > cap_each) {
532				dropped_pages += (size - cap_each);
533				node_end_pfn[i] = node_start_pfn[i] + cap_each;
534			}
535		}
536		physpages -= dropped_pages;
537		pr_warning("Only using %ldMB memory;"
538		       " ignoring %ldMB.\n",
539		       physpages >> (20 - PAGE_SHIFT),
540		       dropped_pages >> (20 - PAGE_SHIFT));
541		pr_warning("Consider using a larger page size.\n");
542	}
543#endif
544
545	/* Heap starts just above the last loaded address. */
546	min_low_pfn = PFN_UP((unsigned long)_end - PAGE_OFFSET);
547
548#ifdef CONFIG_HIGHMEM
549	/* Find where we map lowmem from each controller. */
550	high_memory = setup_pa_va_mapping();
551
552	/* Set max_low_pfn based on what node 0 can directly address. */
553	max_low_pfn = node_lowmem_end_pfn[0];
554
555	lowmem_pages = (mappable_physpages > MAXMEM_PFN) ?
556		MAXMEM_PFN : mappable_physpages;
557	highmem_pages = (long) (physpages - lowmem_pages);
558
559	pr_notice("%ldMB HIGHMEM available.\n",
560	       pages_to_mb(highmem_pages > 0 ? highmem_pages : 0));
561	pr_notice("%ldMB LOWMEM available.\n",
562			pages_to_mb(lowmem_pages));
563#else
564	/* Set max_low_pfn based on what node 0 can directly address. */
565	max_low_pfn = node_end_pfn[0];
566
567#ifndef __tilegx__
568	if (node_end_pfn[0] > MAXMEM_PFN) {
569		pr_warning("Only using %ldMB LOWMEM.\n",
570		       MAXMEM>>20);
571		pr_warning("Use a HIGHMEM enabled kernel.\n");
572		max_low_pfn = MAXMEM_PFN;
573		max_pfn = MAXMEM_PFN;
574		node_end_pfn[0] = MAXMEM_PFN;
575	} else {
576		pr_notice("%ldMB memory available.\n",
577		       pages_to_mb(node_end_pfn[0]));
578	}
579	for (i = 1; i < MAX_NUMNODES; ++i) {
580		node_start_pfn[i] = 0;
581		node_end_pfn[i] = 0;
582	}
583	high_memory = __va(node_end_pfn[0]);
584#else
585	lowmem_pages = 0;
586	for (i = 0; i < MAX_NUMNODES; ++i) {
587		int pages = node_end_pfn[i] - node_start_pfn[i];
588		lowmem_pages += pages;
589		if (pages)
590			high_memory = pfn_to_kaddr(node_end_pfn[i]);
591	}
592	pr_notice("%ldMB memory available.\n",
593	       pages_to_mb(lowmem_pages));
594#endif
595#endif
596}
597
598/*
599 * On 32-bit machines, we only put bootmem on the low controller,
600 * since PAs > 4GB can't be used in bootmem.  In principle one could
601 * imagine, e.g., multiple 1 GB controllers all of which could support
602 * bootmem, but in practice using controllers this small isn't a
603 * particularly interesting scenario, so we just keep it simple and
604 * use only the first controller for bootmem on 32-bit machines.
605 */
606static inline int node_has_bootmem(int nid)
607{
608#ifdef CONFIG_64BIT
609	return 1;
610#else
611	return nid == 0;
612#endif
613}
614
615static inline unsigned long alloc_bootmem_pfn(int nid,
616					      unsigned long size,
617					      unsigned long goal)
618{
619	void *kva = __alloc_bootmem_node(NODE_DATA(nid), size,
620					 PAGE_SIZE, goal);
621	unsigned long pfn = kaddr_to_pfn(kva);
622	BUG_ON(goal && PFN_PHYS(pfn) != goal);
623	return pfn;
624}
625
626static void __init setup_bootmem_allocator_node(int i)
627{
628	unsigned long start, end, mapsize, mapstart;
629
630	if (node_has_bootmem(i)) {
631		NODE_DATA(i)->bdata = &bootmem_node_data[i];
632	} else {
633		/* Share controller zero's bdata for now. */
634		NODE_DATA(i)->bdata = &bootmem_node_data[0];
635		return;
636	}
637
638	/* Skip up to after the bss in node 0. */
639	start = (i == 0) ? min_low_pfn : node_start_pfn[i];
640
641	/* Only lowmem, if we're a HIGHMEM build. */
642#ifdef CONFIG_HIGHMEM
643	end = node_lowmem_end_pfn[i];
644#else
645	end = node_end_pfn[i];
646#endif
647
648	/* No memory here. */
649	if (end == start)
650		return;
651
652	/* Figure out where the bootmem bitmap is located. */
653	mapsize = bootmem_bootmap_pages(end - start);
654	if (i == 0) {
655		/* Use some space right before the heap on node 0. */
656		mapstart = start;
657		start += mapsize;
658	} else {
659		/* Allocate bitmap on node 0 to avoid page table issues. */
660		mapstart = alloc_bootmem_pfn(0, PFN_PHYS(mapsize), 0);
661	}
662
663	/* Initialize a node. */
664	init_bootmem_node(NODE_DATA(i), mapstart, start, end);
665
666	/* Free all the space back into the allocator. */
667	free_bootmem(PFN_PHYS(start), PFN_PHYS(end - start));
668
669#if defined(CONFIG_PCI) && !defined(__tilegx__)
670	/*
671	 * Throw away any memory aliased by the PCI region.
672	 */
673	if (pci_reserve_start_pfn < end && pci_reserve_end_pfn > start) {
674		start = max(pci_reserve_start_pfn, start);
675		end = min(pci_reserve_end_pfn, end);
676		reserve_bootmem(PFN_PHYS(start), PFN_PHYS(end - start),
677				BOOTMEM_EXCLUSIVE);
678	}
679#endif
680}
681
682static void __init setup_bootmem_allocator(void)
683{
684	int i;
685	for (i = 0; i < MAX_NUMNODES; ++i)
686		setup_bootmem_allocator_node(i);
687
688	/* Reserve any memory excluded by "memmap" arguments. */
689	for (i = 0; i < memmap_nr; ++i) {
690		struct memmap_entry *m = &memmap_map[i];
691		reserve_bootmem(m->addr, m->size, BOOTMEM_DEFAULT);
692	}
693
694#ifdef CONFIG_BLK_DEV_INITRD
695	if (initrd_start) {
696		/* Make sure the initrd memory region is not modified. */
697		if (reserve_bootmem(initrd_start, initrd_end - initrd_start,
698				    BOOTMEM_EXCLUSIVE)) {
699			pr_crit("The initrd memory region has been polluted. Disabling it.\n");
700			initrd_start = 0;
701			initrd_end = 0;
702		} else {
703			/*
704			 * Translate initrd_start & initrd_end from PA to VA for
705			 * future access.
706			 */
707			initrd_start += PAGE_OFFSET;
708			initrd_end += PAGE_OFFSET;
709		}
710	}
711#endif
712
713#ifdef CONFIG_KEXEC
714	if (crashk_res.start != crashk_res.end)
715		reserve_bootmem(crashk_res.start, resource_size(&crashk_res),
716				BOOTMEM_DEFAULT);
717#endif
718}
719
720void *__init alloc_remap(int nid, unsigned long size)
721{
722	int pages = node_end_pfn[nid] - node_start_pfn[nid];
723	void *map = pfn_to_kaddr(node_memmap_pfn[nid]);
724	BUG_ON(size != pages * sizeof(struct page));
725	memset(map, 0, size);
726	return map;
727}
728
729static int __init percpu_size(void)
730{
731	int size = __per_cpu_end - __per_cpu_start;
732	size += PERCPU_MODULE_RESERVE;
733	size += PERCPU_DYNAMIC_EARLY_SIZE;
734	if (size < PCPU_MIN_UNIT_SIZE)
735		size = PCPU_MIN_UNIT_SIZE;
736	size = roundup(size, PAGE_SIZE);
737
738	/* In several places we assume the per-cpu data fits on a huge page. */
739	BUG_ON(kdata_huge && size > HPAGE_SIZE);
740	return size;
741}
742
743static void __init zone_sizes_init(void)
744{
745	unsigned long zones_size[MAX_NR_ZONES] = { 0 };
746	int size = percpu_size();
747	int num_cpus = smp_height * smp_width;
748	const unsigned long dma_end = (1UL << (32 - PAGE_SHIFT));
749
750	int i;
751
752	for (i = 0; i < num_cpus; ++i)
753		node_percpu[cpu_to_node(i)] += size;
754
755	for_each_online_node(i) {
756		unsigned long start = node_start_pfn[i];
757		unsigned long end = node_end_pfn[i];
758#ifdef CONFIG_HIGHMEM
759		unsigned long lowmem_end = node_lowmem_end_pfn[i];
760#else
761		unsigned long lowmem_end = end;
762#endif
763		int memmap_size = (end - start) * sizeof(struct page);
764		node_free_pfn[i] = start;
765
766		/*
767		 * Set aside pages for per-cpu data and the mem_map array.
768		 *
769		 * Since the per-cpu data requires special homecaching,
770		 * if we are in kdata_huge mode, we put it at the end of
771		 * the lowmem region.  If we're not in kdata_huge mode,
772		 * we take the per-cpu pages from the bottom of the
773		 * controller, since that avoids fragmenting a huge page
774		 * that users might want.  We always take the memmap
775		 * from the bottom of the controller, since with
776		 * kdata_huge that lets it be under a huge TLB entry.
777		 *
778		 * If the user has requested isolnodes for a controller,
779		 * though, there'll be no lowmem, so we just alloc_bootmem
780		 * the memmap.  There will be no percpu memory either.
781		 */
782		if (i != 0 && cpu_isset(i, isolnodes)) {
783			node_memmap_pfn[i] =
784				alloc_bootmem_pfn(0, memmap_size, 0);
785			BUG_ON(node_percpu[i] != 0);
786		} else if (node_has_bootmem(start)) {
787			unsigned long goal = 0;
788			node_memmap_pfn[i] =
789				alloc_bootmem_pfn(i, memmap_size, 0);
790			if (kdata_huge)
791				goal = PFN_PHYS(lowmem_end) - node_percpu[i];
792			if (node_percpu[i])
793				node_percpu_pfn[i] =
794					alloc_bootmem_pfn(i, node_percpu[i],
795							  goal);
796		} else {
797			/* In non-bootmem zones, just reserve some pages. */
798			node_memmap_pfn[i] = node_free_pfn[i];
799			node_free_pfn[i] += PFN_UP(memmap_size);
800			if (!kdata_huge) {
801				node_percpu_pfn[i] = node_free_pfn[i];
802				node_free_pfn[i] += PFN_UP(node_percpu[i]);
803			} else {
804				node_percpu_pfn[i] =
805					lowmem_end - PFN_UP(node_percpu[i]);
806			}
807		}
808
809#ifdef CONFIG_HIGHMEM
810		if (start > lowmem_end) {
811			zones_size[ZONE_NORMAL] = 0;
812			zones_size[ZONE_HIGHMEM] = end - start;
813		} else {
814			zones_size[ZONE_NORMAL] = lowmem_end - start;
815			zones_size[ZONE_HIGHMEM] = end - lowmem_end;
816		}
817#else
818		zones_size[ZONE_NORMAL] = end - start;
819#endif
820
821		if (start < dma_end) {
822			zones_size[ZONE_DMA] = min(zones_size[ZONE_NORMAL],
823						   dma_end - start);
824			zones_size[ZONE_NORMAL] -= zones_size[ZONE_DMA];
825		} else {
826			zones_size[ZONE_DMA] = 0;
827		}
828
829		/* Take zone metadata from controller 0 if we're isolnode. */
830		if (node_isset(i, isolnodes))
831			NODE_DATA(i)->bdata = &bootmem_node_data[0];
832
833		free_area_init_node(i, zones_size, start, NULL);
834		printk(KERN_DEBUG "  Normal zone: %ld per-cpu pages\n",
835		       PFN_UP(node_percpu[i]));
836
837		/* Track the type of memory on each node */
838		if (zones_size[ZONE_NORMAL] || zones_size[ZONE_DMA])
839			node_set_state(i, N_NORMAL_MEMORY);
840#ifdef CONFIG_HIGHMEM
841		if (end != start)
842			node_set_state(i, N_HIGH_MEMORY);
843#endif
844
845		node_set_online(i);
846	}
847}
848
849#ifdef CONFIG_NUMA
850
851/* which logical CPUs are on which nodes */
852struct cpumask node_2_cpu_mask[MAX_NUMNODES] __write_once;
853EXPORT_SYMBOL(node_2_cpu_mask);
854
855/* which node each logical CPU is on */
856char cpu_2_node[NR_CPUS] __write_once __attribute__((aligned(L2_CACHE_BYTES)));
857EXPORT_SYMBOL(cpu_2_node);
858
859/* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
860static int __init cpu_to_bound_node(int cpu, struct cpumask* unbound_cpus)
861{
862	if (!cpu_possible(cpu) || cpumask_test_cpu(cpu, unbound_cpus))
863		return -1;
864	else
865		return cpu_to_node(cpu);
866}
867
868/* Return number of immediately-adjacent tiles sharing the same NUMA node. */
869static int __init node_neighbors(int node, int cpu,
870				 struct cpumask *unbound_cpus)
871{
872	int neighbors = 0;
873	int w = smp_width;
874	int h = smp_height;
875	int x = cpu % w;
876	int y = cpu / w;
877	if (x > 0 && cpu_to_bound_node(cpu-1, unbound_cpus) == node)
878		++neighbors;
879	if (x < w-1 && cpu_to_bound_node(cpu+1, unbound_cpus) == node)
880		++neighbors;
881	if (y > 0 && cpu_to_bound_node(cpu-w, unbound_cpus) == node)
882		++neighbors;
883	if (y < h-1 && cpu_to_bound_node(cpu+w, unbound_cpus) == node)
884		++neighbors;
885	return neighbors;
886}
887
888static void __init setup_numa_mapping(void)
889{
890	int distance[MAX_NUMNODES][NR_CPUS];
891	HV_Coord coord;
892	int cpu, node, cpus, i, x, y;
893	int num_nodes = num_online_nodes();
894	struct cpumask unbound_cpus;
895	nodemask_t default_nodes;
896
897	cpumask_clear(&unbound_cpus);
898
899	/* Get set of nodes we will use for defaults */
900	nodes_andnot(default_nodes, node_online_map, isolnodes);
901	if (nodes_empty(default_nodes)) {
902		BUG_ON(!node_isset(0, node_online_map));
903		pr_err("Forcing NUMA node zero available as a default node\n");
904		node_set(0, default_nodes);
905	}
906
907	/* Populate the distance[] array */
908	memset(distance, -1, sizeof(distance));
909	cpu = 0;
910	for (coord.y = 0; coord.y < smp_height; ++coord.y) {
911		for (coord.x = 0; coord.x < smp_width;
912		     ++coord.x, ++cpu) {
913			BUG_ON(cpu >= nr_cpu_ids);
914			if (!cpu_possible(cpu)) {
915				cpu_2_node[cpu] = -1;
916				continue;
917			}
918			for_each_node_mask(node, default_nodes) {
919				HV_MemoryControllerInfo info =
920					hv_inquire_memory_controller(
921						coord, node_controller[node]);
922				distance[node][cpu] =
923					ABS(info.coord.x) + ABS(info.coord.y);
924			}
925			cpumask_set_cpu(cpu, &unbound_cpus);
926		}
927	}
928	cpus = cpu;
929
930	/*
931	 * Round-robin through the NUMA nodes until all the cpus are
932	 * assigned.  We could be more clever here (e.g. create four
933	 * sorted linked lists on the same set of cpu nodes, and pull
934	 * off them in round-robin sequence, removing from all four
935	 * lists each time) but given the relatively small numbers
936	 * involved, O(n^2) seem OK for a one-time cost.
937	 */
938	node = first_node(default_nodes);
939	while (!cpumask_empty(&unbound_cpus)) {
940		int best_cpu = -1;
941		int best_distance = INT_MAX;
942		for (cpu = 0; cpu < cpus; ++cpu) {
943			if (cpumask_test_cpu(cpu, &unbound_cpus)) {
944				/*
945				 * Compute metric, which is how much
946				 * closer the cpu is to this memory
947				 * controller than the others, shifted
948				 * up, and then the number of
949				 * neighbors already in the node as an
950				 * epsilon adjustment to try to keep
951				 * the nodes compact.
952				 */
953				int d = distance[node][cpu] * num_nodes;
954				for_each_node_mask(i, default_nodes) {
955					if (i != node)
956						d -= distance[i][cpu];
957				}
958				d *= 8;  /* allow space for epsilon */
959				d -= node_neighbors(node, cpu, &unbound_cpus);
960				if (d < best_distance) {
961					best_cpu = cpu;
962					best_distance = d;
963				}
964			}
965		}
966		BUG_ON(best_cpu < 0);
967		cpumask_set_cpu(best_cpu, &node_2_cpu_mask[node]);
968		cpu_2_node[best_cpu] = node;
969		cpumask_clear_cpu(best_cpu, &unbound_cpus);
970		node = next_node(node, default_nodes);
971		if (node == MAX_NUMNODES)
972			node = first_node(default_nodes);
973	}
974
975	/* Print out node assignments and set defaults for disabled cpus */
976	cpu = 0;
977	for (y = 0; y < smp_height; ++y) {
978		printk(KERN_DEBUG "NUMA cpu-to-node row %d:", y);
979		for (x = 0; x < smp_width; ++x, ++cpu) {
980			if (cpu_to_node(cpu) < 0) {
981				pr_cont(" -");
982				cpu_2_node[cpu] = first_node(default_nodes);
983			} else {
984				pr_cont(" %d", cpu_to_node(cpu));
985			}
986		}
987		pr_cont("\n");
988	}
989}
990
991static struct cpu cpu_devices[NR_CPUS];
992
993static int __init topology_init(void)
994{
995	int i;
996
997	for_each_online_node(i)
998		register_one_node(i);
999
1000	for (i = 0; i < smp_height * smp_width; ++i)
1001		register_cpu(&cpu_devices[i], i);
1002
1003	return 0;
1004}
1005
1006subsys_initcall(topology_init);
1007
1008#else /* !CONFIG_NUMA */
1009
1010#define setup_numa_mapping() do { } while (0)
1011
1012#endif /* CONFIG_NUMA */
1013
1014/*
1015 * Initialize hugepage support on this cpu.  We do this on all cores
1016 * early in boot: before argument parsing for the boot cpu, and after
1017 * argument parsing but before the init functions run on the secondaries.
1018 * So the values we set up here in the hypervisor may be overridden on
1019 * the boot cpu as arguments are parsed.
1020 */
1021static void init_super_pages(void)
1022{
1023#ifdef CONFIG_HUGETLB_SUPER_PAGES
1024	int i;
1025	for (i = 0; i < HUGE_SHIFT_ENTRIES; ++i)
1026		hv_set_pte_super_shift(i, huge_shift[i]);
1027#endif
1028}
1029
1030/**
1031 * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
1032 * @boot: Is this the boot cpu?
1033 *
1034 * Called from setup_arch() on the boot cpu, or online_secondary().
1035 */
1036void setup_cpu(int boot)
1037{
1038	/* The boot cpu sets up its permanent mappings much earlier. */
1039	if (!boot)
1040		store_permanent_mappings();
1041
1042	/* Allow asynchronous TLB interrupts. */
1043#if CHIP_HAS_TILE_DMA()
1044	arch_local_irq_unmask(INT_DMATLB_MISS);
1045	arch_local_irq_unmask(INT_DMATLB_ACCESS);
1046#endif
1047#ifdef __tilegx__
1048	arch_local_irq_unmask(INT_SINGLE_STEP_K);
1049#endif
1050
1051	/*
1052	 * Allow user access to many generic SPRs, like the cycle
1053	 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
1054	 */
1055	__insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0, 1);
1056
1057#if CHIP_HAS_SN()
1058	/* Static network is not restricted. */
1059	__insn_mtspr(SPR_MPL_SN_ACCESS_SET_0, 1);
1060#endif
1061
1062	/*
1063	 * Set the MPL for interrupt control 0 & 1 to the corresponding
1064	 * values.  This includes access to the SYSTEM_SAVE and EX_CONTEXT
1065	 * SPRs, as well as the interrupt mask.
1066	 */
1067	__insn_mtspr(SPR_MPL_INTCTRL_0_SET_0, 1);
1068	__insn_mtspr(SPR_MPL_INTCTRL_1_SET_1, 1);
1069
1070	/* Initialize IRQ support for this cpu. */
1071	setup_irq_regs();
1072
1073#ifdef CONFIG_HARDWALL
1074	/* Reset the network state on this cpu. */
1075	reset_network_state();
1076#endif
1077
1078	init_super_pages();
1079}
1080
1081#ifdef CONFIG_BLK_DEV_INITRD
1082
1083static int __initdata set_initramfs_file;
1084static char __initdata initramfs_file[128] = "initramfs";
1085
1086static int __init setup_initramfs_file(char *str)
1087{
1088	if (str == NULL)
1089		return -EINVAL;
1090	strncpy(initramfs_file, str, sizeof(initramfs_file) - 1);
1091	set_initramfs_file = 1;
1092
1093	return 0;
1094}
1095early_param("initramfs_file", setup_initramfs_file);
1096
1097/*
1098 * We look for a file called "initramfs" in the hvfs.  If there is one, we
1099 * allocate some memory for it and it will be unpacked to the initramfs.
1100 * If it's compressed, the initd code will uncompress it first.
1101 */
1102static void __init load_hv_initrd(void)
1103{
1104	HV_FS_StatInfo stat;
1105	int fd, rc;
1106	void *initrd;
1107
1108	/* If initrd has already been set, skip initramfs file in hvfs. */
1109	if (initrd_start)
1110		return;
1111
1112	fd = hv_fs_findfile((HV_VirtAddr) initramfs_file);
1113	if (fd == HV_ENOENT) {
1114		if (set_initramfs_file) {
1115			pr_warning("No such hvfs initramfs file '%s'\n",
1116				   initramfs_file);
1117			return;
1118		} else {
1119			/* Try old backwards-compatible name. */
1120			fd = hv_fs_findfile((HV_VirtAddr)"initramfs.cpio.gz");
1121			if (fd == HV_ENOENT)
1122				return;
1123		}
1124	}
1125	BUG_ON(fd < 0);
1126	stat = hv_fs_fstat(fd);
1127	BUG_ON(stat.size < 0);
1128	if (stat.flags & HV_FS_ISDIR) {
1129		pr_warning("Ignoring hvfs file '%s': it's a directory.\n",
1130			   initramfs_file);
1131		return;
1132	}
1133	initrd = alloc_bootmem_pages(stat.size);
1134	rc = hv_fs_pread(fd, (HV_VirtAddr) initrd, stat.size, 0);
1135	if (rc != stat.size) {
1136		pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
1137		       stat.size, initramfs_file, rc);
1138		free_initrd_mem((unsigned long) initrd, stat.size);
1139		return;
1140	}
1141	initrd_start = (unsigned long) initrd;
1142	initrd_end = initrd_start + stat.size;
1143}
1144
1145void __init free_initrd_mem(unsigned long begin, unsigned long end)
1146{
1147	free_bootmem(__pa(begin), end - begin);
1148}
1149
1150static int __init setup_initrd(char *str)
1151{
1152	char *endp;
1153	unsigned long initrd_size;
1154
1155	initrd_size = str ? simple_strtoul(str, &endp, 0) : 0;
1156	if (initrd_size == 0 || *endp != '@')
1157		return -EINVAL;
1158
1159	initrd_start = simple_strtoul(endp+1, &endp, 0);
1160	if (initrd_start == 0)
1161		return -EINVAL;
1162
1163	initrd_end = initrd_start + initrd_size;
1164
1165	return 0;
1166}
1167early_param("initrd", setup_initrd);
1168
1169#else
1170static inline void load_hv_initrd(void) {}
1171#endif /* CONFIG_BLK_DEV_INITRD */
1172
1173static void __init validate_hv(void)
1174{
1175	/*
1176	 * It may already be too late, but let's check our built-in
1177	 * configuration against what the hypervisor is providing.
1178	 */
1179	unsigned long glue_size = hv_sysconf(HV_SYSCONF_GLUE_SIZE);
1180	int hv_page_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL);
1181	int hv_hpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE);
1182	HV_ASIDRange asid_range;
1183
1184#ifndef CONFIG_SMP
1185	HV_Topology topology = hv_inquire_topology();
1186	BUG_ON(topology.coord.x != 0 || topology.coord.y != 0);
1187	if (topology.width != 1 || topology.height != 1) {
1188		pr_warning("Warning: booting UP kernel on %dx%d grid;"
1189			   " will ignore all but first tile.\n",
1190			   topology.width, topology.height);
1191	}
1192#endif
1193
1194	if (PAGE_OFFSET + HV_GLUE_START_CPA + glue_size > (unsigned long)_text)
1195		early_panic("Hypervisor glue size %ld is too big!\n",
1196			    glue_size);
1197	if (hv_page_size != PAGE_SIZE)
1198		early_panic("Hypervisor page size %#x != our %#lx\n",
1199			    hv_page_size, PAGE_SIZE);
1200	if (hv_hpage_size != HPAGE_SIZE)
1201		early_panic("Hypervisor huge page size %#x != our %#lx\n",
1202			    hv_hpage_size, HPAGE_SIZE);
1203
1204#ifdef CONFIG_SMP
1205	/*
1206	 * Some hypervisor APIs take a pointer to a bitmap array
1207	 * whose size is at least the number of cpus on the chip.
1208	 * We use a struct cpumask for this, so it must be big enough.
1209	 */
1210	if ((smp_height * smp_width) > nr_cpu_ids)
1211		early_panic("Hypervisor %d x %d grid too big for Linux"
1212			    " NR_CPUS %d\n", smp_height, smp_width,
1213			    nr_cpu_ids);
1214#endif
1215
1216	/*
1217	 * Check that we're using allowed ASIDs, and initialize the
1218	 * various asid variables to their appropriate initial states.
1219	 */
1220	asid_range = hv_inquire_asid(0);
1221	min_asid = asid_range.start;
1222	__this_cpu_write(current_asid, min_asid);
1223	max_asid = asid_range.start + asid_range.size - 1;
1224
1225	if (hv_confstr(HV_CONFSTR_CHIP_MODEL, (HV_VirtAddr)chip_model,
1226		       sizeof(chip_model)) < 0) {
1227		pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
1228		strlcpy(chip_model, "unknown", sizeof(chip_model));
1229	}
1230}
1231
1232static void __init validate_va(void)
1233{
1234#ifndef __tilegx__   /* FIXME: GX: probably some validation relevant here */
1235	/*
1236	 * Similarly, make sure we're only using allowed VAs.
1237	 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_START,
1238	 * and 0 .. KERNEL_HIGH_VADDR.
1239	 * In addition, make sure we CAN'T use the end of memory, since
1240	 * we use the last chunk of each pgd for the pgd_list.
1241	 */
1242	int i, user_kernel_ok = 0;
1243	unsigned long max_va = 0;
1244	unsigned long list_va =
1245		((PGD_LIST_OFFSET / sizeof(pgd_t)) << PGDIR_SHIFT);
1246
1247	for (i = 0; ; ++i) {
1248		HV_VirtAddrRange range = hv_inquire_virtual(i);
1249		if (range.size == 0)
1250			break;
1251		if (range.start <= MEM_USER_INTRPT &&
1252		    range.start + range.size >= MEM_HV_START)
1253			user_kernel_ok = 1;
1254		if (range.start == 0)
1255			max_va = range.size;
1256		BUG_ON(range.start + range.size > list_va);
1257	}
1258	if (!user_kernel_ok)
1259		early_panic("Hypervisor not configured for user/kernel VAs\n");
1260	if (max_va == 0)
1261		early_panic("Hypervisor not configured for low VAs\n");
1262	if (max_va < KERNEL_HIGH_VADDR)
1263		early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
1264			    max_va, KERNEL_HIGH_VADDR);
1265
1266	/* Kernel PCs must have their high bit set; see intvec.S. */
1267	if ((long)VMALLOC_START >= 0)
1268		early_panic(
1269			"Linux VMALLOC region below the 2GB line (%#lx)!\n"
1270			"Reconfigure the kernel with smaller VMALLOC_RESERVE.\n",
1271			VMALLOC_START);
1272#endif
1273}
1274
1275/*
1276 * cpu_lotar_map lists all the cpus that are valid for the supervisor
1277 * to cache data on at a page level, i.e. what cpus can be placed in
1278 * the LOTAR field of a PTE.  It is equivalent to the set of possible
1279 * cpus plus any other cpus that are willing to share their cache.
1280 * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
1281 */
1282struct cpumask __write_once cpu_lotar_map;
1283EXPORT_SYMBOL(cpu_lotar_map);
1284
1285/*
1286 * hash_for_home_map lists all the tiles that hash-for-home data
1287 * will be cached on.  Note that this may includes tiles that are not
1288 * valid for this supervisor to use otherwise (e.g. if a hypervisor
1289 * device is being shared between multiple supervisors).
1290 * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
1291 */
1292struct cpumask hash_for_home_map;
1293EXPORT_SYMBOL(hash_for_home_map);
1294
1295/*
1296 * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
1297 * flush on our behalf.  It is set to cpu_possible_mask OR'ed with
1298 * hash_for_home_map, and it is what should be passed to
1299 * hv_flush_remote() to flush all caches.  Note that if there are
1300 * dedicated hypervisor driver tiles that have authorized use of their
1301 * cache, those tiles will only appear in cpu_lotar_map, NOT in
1302 * cpu_cacheable_map, as they are a special case.
1303 */
1304struct cpumask __write_once cpu_cacheable_map;
1305EXPORT_SYMBOL(cpu_cacheable_map);
1306
1307static __initdata struct cpumask disabled_map;
1308
1309static int __init disabled_cpus(char *str)
1310{
1311	int boot_cpu = smp_processor_id();
1312
1313	if (str == NULL || cpulist_parse_crop(str, &disabled_map) != 0)
1314		return -EINVAL;
1315	if (cpumask_test_cpu(boot_cpu, &disabled_map)) {
1316		pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu);
1317		cpumask_clear_cpu(boot_cpu, &disabled_map);
1318	}
1319	return 0;
1320}
1321
1322early_param("disabled_cpus", disabled_cpus);
1323
1324void __init print_disabled_cpus(void)
1325{
1326	if (!cpumask_empty(&disabled_map)) {
1327		char buf[100];
1328		cpulist_scnprintf(buf, sizeof(buf), &disabled_map);
1329		pr_info("CPUs not available for Linux: %s\n", buf);
1330	}
1331}
1332
1333static void __init setup_cpu_maps(void)
1334{
1335	struct cpumask hv_disabled_map, cpu_possible_init;
1336	int boot_cpu = smp_processor_id();
1337	int cpus, i, rc;
1338
1339	/* Learn which cpus are allowed by the hypervisor. */
1340	rc = hv_inquire_tiles(HV_INQ_TILES_AVAIL,
1341			      (HV_VirtAddr) cpumask_bits(&cpu_possible_init),
1342			      sizeof(cpu_cacheable_map));
1343	if (rc < 0)
1344		early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc);
1345	if (!cpumask_test_cpu(boot_cpu, &cpu_possible_init))
1346		early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu);
1347
1348	/* Compute the cpus disabled by the hvconfig file. */
1349	cpumask_complement(&hv_disabled_map, &cpu_possible_init);
1350
1351	/* Include them with the cpus disabled by "disabled_cpus". */
1352	cpumask_or(&disabled_map, &disabled_map, &hv_disabled_map);
1353
1354	/*
1355	 * Disable every cpu after "setup_max_cpus".  But don't mark
1356	 * as disabled the cpus that are outside of our initial rectangle,
1357	 * since that turns out to be confusing.
1358	 */
1359	cpus = 1;                          /* this cpu */
1360	cpumask_set_cpu(boot_cpu, &disabled_map);   /* ignore this cpu */
1361	for (i = 0; cpus < setup_max_cpus; ++i)
1362		if (!cpumask_test_cpu(i, &disabled_map))
1363			++cpus;
1364	for (; i < smp_height * smp_width; ++i)
1365		cpumask_set_cpu(i, &disabled_map);
1366	cpumask_clear_cpu(boot_cpu, &disabled_map); /* reset this cpu */
1367	for (i = smp_height * smp_width; i < NR_CPUS; ++i)
1368		cpumask_clear_cpu(i, &disabled_map);
1369
1370	/*
1371	 * Setup cpu_possible map as every cpu allocated to us, minus
1372	 * the results of any "disabled_cpus" settings.
1373	 */
1374	cpumask_andnot(&cpu_possible_init, &cpu_possible_init, &disabled_map);
1375	init_cpu_possible(&cpu_possible_init);
1376
1377	/* Learn which cpus are valid for LOTAR caching. */
1378	rc = hv_inquire_tiles(HV_INQ_TILES_LOTAR,
1379			      (HV_VirtAddr) cpumask_bits(&cpu_lotar_map),
1380			      sizeof(cpu_lotar_map));
1381	if (rc < 0) {
1382		pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
1383		cpu_lotar_map = *cpu_possible_mask;
1384	}
1385
1386	/* Retrieve set of CPUs used for hash-for-home caching */
1387	rc = hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE,
1388			      (HV_VirtAddr) hash_for_home_map.bits,
1389			      sizeof(hash_for_home_map));
1390	if (rc < 0)
1391		early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc);
1392	cpumask_or(&cpu_cacheable_map, cpu_possible_mask, &hash_for_home_map);
1393}
1394
1395
1396static int __init dataplane(char *str)
1397{
1398	pr_warning("WARNING: dataplane support disabled in this kernel\n");
1399	return 0;
1400}
1401
1402early_param("dataplane", dataplane);
1403
1404#ifdef CONFIG_CMDLINE_BOOL
1405static char __initdata builtin_cmdline[COMMAND_LINE_SIZE] = CONFIG_CMDLINE;
1406#endif
1407
1408void __init setup_arch(char **cmdline_p)
1409{
1410	int len;
1411
1412#if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
1413	len = hv_get_command_line((HV_VirtAddr) boot_command_line,
1414				  COMMAND_LINE_SIZE);
1415	if (boot_command_line[0])
1416		pr_warning("WARNING: ignoring dynamic command line \"%s\"\n",
1417			   boot_command_line);
1418	strlcpy(boot_command_line, builtin_cmdline, COMMAND_LINE_SIZE);
1419#else
1420	char *hv_cmdline;
1421#if defined(CONFIG_CMDLINE_BOOL)
1422	if (builtin_cmdline[0]) {
1423		int builtin_len = strlcpy(boot_command_line, builtin_cmdline,
1424					  COMMAND_LINE_SIZE);
1425		if (builtin_len < COMMAND_LINE_SIZE-1)
1426			boot_command_line[builtin_len++] = ' ';
1427		hv_cmdline = &boot_command_line[builtin_len];
1428		len = COMMAND_LINE_SIZE - builtin_len;
1429	} else
1430#endif
1431	{
1432		hv_cmdline = boot_command_line;
1433		len = COMMAND_LINE_SIZE;
1434	}
1435	len = hv_get_command_line((HV_VirtAddr) hv_cmdline, len);
1436	if (len < 0 || len > COMMAND_LINE_SIZE)
1437		early_panic("hv_get_command_line failed: %d\n", len);
1438#endif
1439
1440	*cmdline_p = boot_command_line;
1441
1442	/* Set disabled_map and setup_max_cpus very early */
1443	parse_early_param();
1444
1445	/* Make sure the kernel is compatible with the hypervisor. */
1446	validate_hv();
1447	validate_va();
1448
1449	setup_cpu_maps();
1450
1451
1452#if defined(CONFIG_PCI) && !defined(__tilegx__)
1453	/*
1454	 * Initialize the PCI structures.  This is done before memory
1455	 * setup so that we know whether or not a pci_reserve region
1456	 * is necessary.
1457	 */
1458	if (tile_pci_init() == 0)
1459		pci_reserve_mb = 0;
1460
1461	/* PCI systems reserve a region just below 4GB for mapping iomem. */
1462	pci_reserve_end_pfn  = (1 << (32 - PAGE_SHIFT));
1463	pci_reserve_start_pfn = pci_reserve_end_pfn -
1464		(pci_reserve_mb << (20 - PAGE_SHIFT));
1465#endif
1466
1467	init_mm.start_code = (unsigned long) _text;
1468	init_mm.end_code = (unsigned long) _etext;
1469	init_mm.end_data = (unsigned long) _edata;
1470	init_mm.brk = (unsigned long) _end;
1471
1472	setup_memory();
1473	store_permanent_mappings();
1474	setup_bootmem_allocator();
1475
1476	/*
1477	 * NOTE: before this point _nobody_ is allowed to allocate
1478	 * any memory using the bootmem allocator.
1479	 */
1480
1481#ifdef CONFIG_SWIOTLB
1482	swiotlb_init(0);
1483#endif
1484
1485	paging_init();
1486	setup_numa_mapping();
1487	zone_sizes_init();
1488	set_page_homes();
1489	setup_cpu(1);
1490	setup_clock();
1491	load_hv_initrd();
1492}
1493
1494
1495/*
1496 * Set up per-cpu memory.
1497 */
1498
1499unsigned long __per_cpu_offset[NR_CPUS] __write_once;
1500EXPORT_SYMBOL(__per_cpu_offset);
1501
1502static size_t __initdata pfn_offset[MAX_NUMNODES] = { 0 };
1503static unsigned long __initdata percpu_pfn[NR_CPUS] = { 0 };
1504
1505/*
1506 * As the percpu code allocates pages, we return the pages from the
1507 * end of the node for the specified cpu.
1508 */
1509static void *__init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
1510{
1511	int nid = cpu_to_node(cpu);
1512	unsigned long pfn = node_percpu_pfn[nid] + pfn_offset[nid];
1513
1514	BUG_ON(size % PAGE_SIZE != 0);
1515	pfn_offset[nid] += size / PAGE_SIZE;
1516	BUG_ON(node_percpu[nid] < size);
1517	node_percpu[nid] -= size;
1518	if (percpu_pfn[cpu] == 0)
1519		percpu_pfn[cpu] = pfn;
1520	return pfn_to_kaddr(pfn);
1521}
1522
1523/*
1524 * Pages reserved for percpu memory are not freeable, and in any case we are
1525 * on a short path to panic() in setup_per_cpu_area() at this point anyway.
1526 */
1527static void __init pcpu_fc_free(void *ptr, size_t size)
1528{
1529}
1530
1531/*
1532 * Set up vmalloc page tables using bootmem for the percpu code.
1533 */
1534static void __init pcpu_fc_populate_pte(unsigned long addr)
1535{
1536	pgd_t *pgd;
1537	pud_t *pud;
1538	pmd_t *pmd;
1539	pte_t *pte;
1540
1541	BUG_ON(pgd_addr_invalid(addr));
1542	if (addr < VMALLOC_START || addr >= VMALLOC_END)
1543		panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx;"
1544		      " try increasing CONFIG_VMALLOC_RESERVE\n",
1545		      addr, VMALLOC_START, VMALLOC_END);
1546
1547	pgd = swapper_pg_dir + pgd_index(addr);
1548	pud = pud_offset(pgd, addr);
1549	BUG_ON(!pud_present(*pud));
1550	pmd = pmd_offset(pud, addr);
1551	if (pmd_present(*pmd)) {
1552		BUG_ON(pmd_huge_page(*pmd));
1553	} else {
1554		pte = __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE,
1555				      HV_PAGE_TABLE_ALIGN, 0);
1556		pmd_populate_kernel(&init_mm, pmd, pte);
1557	}
1558}
1559
1560void __init setup_per_cpu_areas(void)
1561{
1562	struct page *pg;
1563	unsigned long delta, pfn, lowmem_va;
1564	unsigned long size = percpu_size();
1565	char *ptr;
1566	int rc, cpu, i;
1567
1568	rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_fc_alloc,
1569				   pcpu_fc_free, pcpu_fc_populate_pte);
1570	if (rc < 0)
1571		panic("Cannot initialize percpu area (err=%d)", rc);
1572
1573	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1574	for_each_possible_cpu(cpu) {
1575		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1576
1577		/* finv the copy out of cache so we can change homecache */
1578		ptr = pcpu_base_addr + pcpu_unit_offsets[cpu];
1579		__finv_buffer(ptr, size);
1580		pfn = percpu_pfn[cpu];
1581
1582		/* Rewrite the page tables to cache on that cpu */
1583		pg = pfn_to_page(pfn);
1584		for (i = 0; i < size; i += PAGE_SIZE, ++pfn, ++pg) {
1585
1586			/* Update the vmalloc mapping and page home. */
1587			unsigned long addr = (unsigned long)ptr + i;
1588			pte_t *ptep = virt_to_kpte(addr);
1589			pte_t pte = *ptep;
1590			BUG_ON(pfn != pte_pfn(pte));
1591			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
1592			pte = set_remote_cache_cpu(pte, cpu);
1593			set_pte_at(&init_mm, addr, ptep, pte);
1594
1595			/* Update the lowmem mapping for consistency. */
1596			lowmem_va = (unsigned long)pfn_to_kaddr(pfn);
1597			ptep = virt_to_kpte(lowmem_va);
1598			if (pte_huge(*ptep)) {
1599				printk(KERN_DEBUG "early shatter of huge page"
1600				       " at %#lx\n", lowmem_va);
1601				shatter_pmd((pmd_t *)ptep);
1602				ptep = virt_to_kpte(lowmem_va);
1603				BUG_ON(pte_huge(*ptep));
1604			}
1605			BUG_ON(pfn != pte_pfn(*ptep));
1606			set_pte_at(&init_mm, lowmem_va, ptep, pte);
1607		}
1608	}
1609
1610	/* Set our thread pointer appropriately. */
1611	set_my_cpu_offset(__per_cpu_offset[smp_processor_id()]);
1612
1613	/* Make sure the finv's have completed. */
1614	mb_incoherent();
1615
1616	/* Flush the TLB so we reference it properly from here on out. */
1617	local_flush_tlb_all();
1618}
1619
1620static struct resource data_resource = {
1621	.name	= "Kernel data",
1622	.start	= 0,
1623	.end	= 0,
1624	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
1625};
1626
1627static struct resource code_resource = {
1628	.name	= "Kernel code",
1629	.start	= 0,
1630	.end	= 0,
1631	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
1632};
1633
1634/*
1635 * On Pro, we reserve all resources above 4GB so that PCI won't try to put
1636 * mappings above 4GB.
1637 */
1638#if defined(CONFIG_PCI) && !defined(__tilegx__)
1639static struct resource* __init
1640insert_non_bus_resource(void)
1641{
1642	struct resource *res =
1643		kzalloc(sizeof(struct resource), GFP_ATOMIC);
1644	if (!res)
1645		return NULL;
1646	res->name = "Non-Bus Physical Address Space";
1647	res->start = (1ULL << 32);
1648	res->end = -1LL;
1649	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1650	if (insert_resource(&iomem_resource, res)) {
1651		kfree(res);
1652		return NULL;
1653	}
1654	return res;
1655}
1656#endif
1657
1658static struct resource* __init
1659insert_ram_resource(u64 start_pfn, u64 end_pfn, bool reserved)
1660{
1661	struct resource *res =
1662		kzalloc(sizeof(struct resource), GFP_ATOMIC);
1663	if (!res)
1664		return NULL;
1665	res->name = reserved ? "Reserved" : "System RAM";
1666	res->start = start_pfn << PAGE_SHIFT;
1667	res->end = (end_pfn << PAGE_SHIFT) - 1;
1668	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1669	if (insert_resource(&iomem_resource, res)) {
1670		kfree(res);
1671		return NULL;
1672	}
1673	return res;
1674}
1675
1676/*
1677 * Request address space for all standard resources
1678 *
1679 * If the system includes PCI root complex drivers, we need to create
1680 * a window just below 4GB where PCI BARs can be mapped.
1681 */
1682static int __init request_standard_resources(void)
1683{
1684	int i;
1685	enum { CODE_DELTA = MEM_SV_START - PAGE_OFFSET };
1686
1687#if defined(CONFIG_PCI) && !defined(__tilegx__)
1688	insert_non_bus_resource();
1689#endif
1690
1691	for_each_online_node(i) {
1692		u64 start_pfn = node_start_pfn[i];
1693		u64 end_pfn = node_end_pfn[i];
1694
1695#if defined(CONFIG_PCI) && !defined(__tilegx__)
1696		if (start_pfn <= pci_reserve_start_pfn &&
1697		    end_pfn > pci_reserve_start_pfn) {
1698			if (end_pfn > pci_reserve_end_pfn)
1699				insert_ram_resource(pci_reserve_end_pfn,
1700						    end_pfn, 0);
1701			end_pfn = pci_reserve_start_pfn;
1702		}
1703#endif
1704		insert_ram_resource(start_pfn, end_pfn, 0);
1705	}
1706
1707	code_resource.start = __pa(_text - CODE_DELTA);
1708	code_resource.end = __pa(_etext - CODE_DELTA)-1;
1709	data_resource.start = __pa(_sdata);
1710	data_resource.end = __pa(_end)-1;
1711
1712	insert_resource(&iomem_resource, &code_resource);
1713	insert_resource(&iomem_resource, &data_resource);
1714
1715	/* Mark any "memmap" regions busy for the resource manager. */
1716	for (i = 0; i < memmap_nr; ++i) {
1717		struct memmap_entry *m = &memmap_map[i];
1718		insert_ram_resource(PFN_DOWN(m->addr),
1719				    PFN_UP(m->addr + m->size - 1), 1);
1720	}
1721
1722#ifdef CONFIG_KEXEC
1723	insert_resource(&iomem_resource, &crashk_res);
1724#endif
1725
1726	return 0;
1727}
1728
1729subsys_initcall(request_standard_resources);
1730