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 <linux/tick.h>
36#include <asm/setup.h>
37#include <asm/sections.h>
38#include <asm/cacheflush.h>
39#include <asm/pgalloc.h>
40#include <asm/mmu_context.h>
41#include <hv/hypervisor.h>
42#include <arch/interrupts.h>
43
44/* <linux/smp.h> doesn't provide this definition. */
45#ifndef CONFIG_SMP
46#define setup_max_cpus 1
47#endif
48
49static inline int ABS(int x) { return x >= 0 ? x : -x; }
50
51/* Chip information */
52char chip_model[64] __write_once;
53
54#ifdef CONFIG_VT
55struct screen_info screen_info;
56#endif
57
58struct pglist_data node_data[MAX_NUMNODES] __read_mostly;
59EXPORT_SYMBOL(node_data);
60
61/* Information on the NUMA nodes that we compute early */
62unsigned long node_start_pfn[MAX_NUMNODES];
63unsigned long node_end_pfn[MAX_NUMNODES];
64unsigned long __initdata node_memmap_pfn[MAX_NUMNODES];
65unsigned long __initdata node_percpu_pfn[MAX_NUMNODES];
66unsigned long __initdata node_free_pfn[MAX_NUMNODES];
67
68static unsigned long __initdata node_percpu[MAX_NUMNODES];
69
70/*
71 * per-CPU stack and boot info.
72 */
73DEFINE_PER_CPU(unsigned long, boot_sp) =
74	(unsigned long)init_stack + THREAD_SIZE;
75
76#ifdef CONFIG_SMP
77DEFINE_PER_CPU(unsigned long, boot_pc) = (unsigned long)start_kernel;
78#else
79/*
80 * The variable must be __initdata since it references __init code.
81 * With CONFIG_SMP it is per-cpu data, which is exempt from validation.
82 */
83unsigned long __initdata boot_pc = (unsigned long)start_kernel;
84#endif
85
86#ifdef CONFIG_HIGHMEM
87/* Page frame index of end of lowmem on each controller. */
88unsigned long node_lowmem_end_pfn[MAX_NUMNODES];
89
90/* Number of pages that can be mapped into lowmem. */
91static unsigned long __initdata mappable_physpages;
92#endif
93
94/* Data on which physical memory controller corresponds to which NUMA node */
95int node_controller[MAX_NUMNODES] = { [0 ... MAX_NUMNODES-1] = -1 };
96
97#ifdef CONFIG_HIGHMEM
98/* Map information from VAs to PAs */
99unsigned long pbase_map[1 << (32 - HPAGE_SHIFT)]
100  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
101EXPORT_SYMBOL(pbase_map);
102
103/* Map information from PAs to VAs */
104void *vbase_map[NR_PA_HIGHBIT_VALUES]
105  __write_once __attribute__((aligned(L2_CACHE_BYTES)));
106EXPORT_SYMBOL(vbase_map);
107#endif
108
109/* Node number as a function of the high PA bits */
110int highbits_to_node[NR_PA_HIGHBIT_VALUES] __write_once;
111EXPORT_SYMBOL(highbits_to_node);
112
113static unsigned int __initdata maxmem_pfn = -1U;
114static unsigned int __initdata maxnodemem_pfn[MAX_NUMNODES] = {
115	[0 ... MAX_NUMNODES-1] = -1U
116};
117static nodemask_t __initdata isolnodes;
118
119#if defined(CONFIG_PCI) && !defined(__tilegx__)
120enum { DEFAULT_PCI_RESERVE_MB = 64 };
121static unsigned int __initdata pci_reserve_mb = DEFAULT_PCI_RESERVE_MB;
122unsigned long __initdata pci_reserve_start_pfn = -1U;
123unsigned long __initdata pci_reserve_end_pfn = -1U;
124#endif
125
126static int __init setup_maxmem(char *str)
127{
128	unsigned long long maxmem;
129	if (str == NULL || (maxmem = memparse(str, NULL)) == 0)
130		return -EINVAL;
131
132	maxmem_pfn = (maxmem >> HPAGE_SHIFT) << (HPAGE_SHIFT - PAGE_SHIFT);
133	pr_info("Forcing RAM used to no more than %dMB\n",
134		maxmem_pfn >> (20 - PAGE_SHIFT));
135	return 0;
136}
137early_param("maxmem", setup_maxmem);
138
139static int __init setup_maxnodemem(char *str)
140{
141	char *endp;
142	unsigned long long maxnodemem;
143	long node;
144
145	node = str ? simple_strtoul(str, &endp, 0) : INT_MAX;
146	if (node >= MAX_NUMNODES || *endp != ':')
147		return -EINVAL;
148
149	maxnodemem = memparse(endp+1, NULL);
150	maxnodemem_pfn[node] = (maxnodemem >> HPAGE_SHIFT) <<
151		(HPAGE_SHIFT - PAGE_SHIFT);
152	pr_info("Forcing RAM used on node %ld to no more than %dMB\n",
153		node, maxnodemem_pfn[node] >> (20 - PAGE_SHIFT));
154	return 0;
155}
156early_param("maxnodemem", setup_maxnodemem);
157
158struct memmap_entry {
159	u64 addr;	/* start of memory segment */
160	u64 size;	/* size of memory segment */
161};
162static struct memmap_entry memmap_map[64];
163static int memmap_nr;
164
165static void add_memmap_region(u64 addr, u64 size)
166{
167	if (memmap_nr >= ARRAY_SIZE(memmap_map)) {
168		pr_err("Ooops! Too many entries in the memory map!\n");
169		return;
170	}
171	memmap_map[memmap_nr].addr = addr;
172	memmap_map[memmap_nr].size = size;
173	memmap_nr++;
174}
175
176static int __init setup_memmap(char *p)
177{
178	char *oldp;
179	u64 start_at, mem_size;
180
181	if (!p)
182		return -EINVAL;
183
184	if (!strncmp(p, "exactmap", 8)) {
185		pr_err("\"memmap=exactmap\" not valid on tile\n");
186		return 0;
187	}
188
189	oldp = p;
190	mem_size = memparse(p, &p);
191	if (p == oldp)
192		return -EINVAL;
193
194	if (*p == '@') {
195		pr_err("\"memmap=nn@ss\" (force RAM) invalid on tile\n");
196	} else if (*p == '#') {
197		pr_err("\"memmap=nn#ss\" (force ACPI data) invalid on tile\n");
198	} else if (*p == '$') {
199		start_at = memparse(p+1, &p);
200		add_memmap_region(start_at, mem_size);
201	} else {
202		if (mem_size == 0)
203			return -EINVAL;
204		maxmem_pfn = (mem_size >> HPAGE_SHIFT) <<
205			(HPAGE_SHIFT - PAGE_SHIFT);
206	}
207	return *p == '\0' ? 0 : -EINVAL;
208}
209early_param("memmap", setup_memmap);
210
211static int __init setup_mem(char *str)
212{
213	return setup_maxmem(str);
214}
215early_param("mem", setup_mem);  /* compatibility with x86 */
216
217static int __init setup_isolnodes(char *str)
218{
219	if (str == NULL || nodelist_parse(str, isolnodes) != 0)
220		return -EINVAL;
221
222	pr_info("Set isolnodes value to '%*pbl'\n",
223		nodemask_pr_args(&isolnodes));
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: now %#llx-%#llx\n",
421			       start_pa, start_pa + orig_size,
422			       range.start, range.start + range.size);
423		}
424		highbits = __pa_to_highbits(range.start);
425		if (highbits >= NR_PA_HIGHBIT_VALUES) {
426			pr_err("PA high bits too high: %#llx..%#llx\n",
427			       range.start, range.start + range.size);
428			continue;
429		}
430		if (highbits_seen[highbits]) {
431			pr_err("Range overlaps in high bits: %#llx..%#llx\n",
432			       range.start, range.start + range.size);
433			continue;
434		}
435		highbits_seen[highbits] = 1;
436		if (PFN_DOWN(range.size) > maxnodemem_pfn[i]) {
437			int max_size = maxnodemem_pfn[i];
438			if (max_size > 0) {
439				pr_err("Maxnodemem reduced node %d to %d pages\n",
440				       i, max_size);
441				range.size = PFN_PHYS(max_size);
442			} else {
443				pr_err("Maxnodemem disabled node %d\n", i);
444				continue;
445			}
446		}
447		if (physpages + PFN_DOWN(range.size) > maxmem_pfn) {
448			int max_size = maxmem_pfn - physpages;
449			if (max_size > 0) {
450				pr_err("Maxmem reduced node %d to %d pages\n",
451				       i, max_size);
452				range.size = PFN_PHYS(max_size);
453			} else {
454				pr_err("Maxmem disabled node %d\n", i);
455				continue;
456			}
457		}
458		if (i >= MAX_NUMNODES) {
459			pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
460			       i, range.size, range.size + range.start);
461			continue;
462		}
463
464		start = range.start >> PAGE_SHIFT;
465		size = range.size >> PAGE_SHIFT;
466		end = start + size;
467
468#ifndef __tilegx__
469		if (((HV_PhysAddr)end << PAGE_SHIFT) !=
470		    (range.start + range.size)) {
471			pr_err("PAs too high to represent: %#llx..%#llx\n",
472			       range.start, range.start + range.size);
473			continue;
474		}
475#endif
476#if defined(CONFIG_PCI) && !defined(__tilegx__)
477		/*
478		 * Blocks that overlap the pci reserved region must
479		 * have enough space to hold the maximum percpu data
480		 * region at the top of the range.  If there isn't
481		 * enough space above the reserved region, just
482		 * truncate the node.
483		 */
484		if (start <= pci_reserve_start_pfn &&
485		    end > pci_reserve_start_pfn) {
486			unsigned int per_cpu_size =
487				__per_cpu_end - __per_cpu_start;
488			unsigned int percpu_pages =
489				NR_CPUS * (PFN_UP(per_cpu_size) >> PAGE_SHIFT);
490			if (end < pci_reserve_end_pfn + percpu_pages) {
491				end = pci_reserve_start_pfn;
492				pr_err("PCI mapping region reduced node %d to %ld pages\n",
493				       i, end - start);
494			}
495		}
496#endif
497
498		for (j = __pfn_to_highbits(start);
499		     j <= __pfn_to_highbits(end - 1); j++)
500			highbits_to_node[j] = i;
501
502		node_start_pfn[i] = start;
503		node_end_pfn[i] = end;
504		node_controller[i] = range.controller;
505		physpages += size;
506		max_pfn = end;
507
508		/* Mark node as online */
509		node_set(i, node_online_map);
510		node_set(i, node_possible_map);
511	}
512
513#ifndef __tilegx__
514	/*
515	 * For 4KB pages, mem_map "struct page" data is 1% of the size
516	 * of the physical memory, so can be quite big (640 MB for
517	 * four 16G zones).  These structures must be mapped in
518	 * lowmem, and since we currently cap out at about 768 MB,
519	 * it's impractical to try to use this much address space.
520	 * For now, arbitrarily cap the amount of physical memory
521	 * we're willing to use at 8 million pages (32GB of 4KB pages).
522	 */
523	cap = 8 * 1024 * 1024;  /* 8 million pages */
524	if (physpages > cap) {
525		int num_nodes = num_online_nodes();
526		int cap_each = cap / num_nodes;
527		unsigned long dropped_pages = 0;
528		for (i = 0; i < num_nodes; ++i) {
529			int size = node_end_pfn[i] - node_start_pfn[i];
530			if (size > cap_each) {
531				dropped_pages += (size - cap_each);
532				node_end_pfn[i] = node_start_pfn[i] + cap_each;
533			}
534		}
535		physpages -= dropped_pages;
536		pr_warn("Only using %ldMB memory - ignoring %ldMB\n",
537			physpages >> (20 - PAGE_SHIFT),
538			dropped_pages >> (20 - PAGE_SHIFT));
539		pr_warn("Consider using a larger page size\n");
540	}
541#endif
542
543	/* Heap starts just above the last loaded address. */
544	min_low_pfn = PFN_UP((unsigned long)_end - PAGE_OFFSET);
545
546#ifdef CONFIG_HIGHMEM
547	/* Find where we map lowmem from each controller. */
548	high_memory = setup_pa_va_mapping();
549
550	/* Set max_low_pfn based on what node 0 can directly address. */
551	max_low_pfn = node_lowmem_end_pfn[0];
552
553	lowmem_pages = (mappable_physpages > MAXMEM_PFN) ?
554		MAXMEM_PFN : mappable_physpages;
555	highmem_pages = (long) (physpages - lowmem_pages);
556
557	pr_notice("%ldMB HIGHMEM available\n",
558		  pages_to_mb(highmem_pages > 0 ? highmem_pages : 0));
559	pr_notice("%ldMB LOWMEM available\n", pages_to_mb(lowmem_pages));
560#else
561	/* Set max_low_pfn based on what node 0 can directly address. */
562	max_low_pfn = node_end_pfn[0];
563
564#ifndef __tilegx__
565	if (node_end_pfn[0] > MAXMEM_PFN) {
566		pr_warn("Only using %ldMB LOWMEM\n", MAXMEM >> 20);
567		pr_warn("Use a HIGHMEM enabled kernel\n");
568		max_low_pfn = MAXMEM_PFN;
569		max_pfn = MAXMEM_PFN;
570		node_end_pfn[0] = MAXMEM_PFN;
571	} else {
572		pr_notice("%ldMB memory available\n",
573			  pages_to_mb(node_end_pfn[0]));
574	}
575	for (i = 1; i < MAX_NUMNODES; ++i) {
576		node_start_pfn[i] = 0;
577		node_end_pfn[i] = 0;
578	}
579	high_memory = __va(node_end_pfn[0]);
580#else
581	lowmem_pages = 0;
582	for (i = 0; i < MAX_NUMNODES; ++i) {
583		int pages = node_end_pfn[i] - node_start_pfn[i];
584		lowmem_pages += pages;
585		if (pages)
586			high_memory = pfn_to_kaddr(node_end_pfn[i]);
587	}
588	pr_notice("%ldMB memory available\n", pages_to_mb(lowmem_pages));
589#endif
590#endif
591}
592
593/*
594 * On 32-bit machines, we only put bootmem on the low controller,
595 * since PAs > 4GB can't be used in bootmem.  In principle one could
596 * imagine, e.g., multiple 1 GB controllers all of which could support
597 * bootmem, but in practice using controllers this small isn't a
598 * particularly interesting scenario, so we just keep it simple and
599 * use only the first controller for bootmem on 32-bit machines.
600 */
601static inline int node_has_bootmem(int nid)
602{
603#ifdef CONFIG_64BIT
604	return 1;
605#else
606	return nid == 0;
607#endif
608}
609
610static inline unsigned long alloc_bootmem_pfn(int nid,
611					      unsigned long size,
612					      unsigned long goal)
613{
614	void *kva = __alloc_bootmem_node(NODE_DATA(nid), size,
615					 PAGE_SIZE, goal);
616	unsigned long pfn = kaddr_to_pfn(kva);
617	BUG_ON(goal && PFN_PHYS(pfn) != goal);
618	return pfn;
619}
620
621static void __init setup_bootmem_allocator_node(int i)
622{
623	unsigned long start, end, mapsize, mapstart;
624
625	if (node_has_bootmem(i)) {
626		NODE_DATA(i)->bdata = &bootmem_node_data[i];
627	} else {
628		/* Share controller zero's bdata for now. */
629		NODE_DATA(i)->bdata = &bootmem_node_data[0];
630		return;
631	}
632
633	/* Skip up to after the bss in node 0. */
634	start = (i == 0) ? min_low_pfn : node_start_pfn[i];
635
636	/* Only lowmem, if we're a HIGHMEM build. */
637#ifdef CONFIG_HIGHMEM
638	end = node_lowmem_end_pfn[i];
639#else
640	end = node_end_pfn[i];
641#endif
642
643	/* No memory here. */
644	if (end == start)
645		return;
646
647	/* Figure out where the bootmem bitmap is located. */
648	mapsize = bootmem_bootmap_pages(end - start);
649	if (i == 0) {
650		/* Use some space right before the heap on node 0. */
651		mapstart = start;
652		start += mapsize;
653	} else {
654		/* Allocate bitmap on node 0 to avoid page table issues. */
655		mapstart = alloc_bootmem_pfn(0, PFN_PHYS(mapsize), 0);
656	}
657
658	/* Initialize a node. */
659	init_bootmem_node(NODE_DATA(i), mapstart, start, end);
660
661	/* Free all the space back into the allocator. */
662	free_bootmem(PFN_PHYS(start), PFN_PHYS(end - start));
663
664#if defined(CONFIG_PCI) && !defined(__tilegx__)
665	/*
666	 * Throw away any memory aliased by the PCI region.
667	 */
668	if (pci_reserve_start_pfn < end && pci_reserve_end_pfn > start) {
669		start = max(pci_reserve_start_pfn, start);
670		end = min(pci_reserve_end_pfn, end);
671		reserve_bootmem(PFN_PHYS(start), PFN_PHYS(end - start),
672				BOOTMEM_EXCLUSIVE);
673	}
674#endif
675}
676
677static void __init setup_bootmem_allocator(void)
678{
679	int i;
680	for (i = 0; i < MAX_NUMNODES; ++i)
681		setup_bootmem_allocator_node(i);
682
683	/* Reserve any memory excluded by "memmap" arguments. */
684	for (i = 0; i < memmap_nr; ++i) {
685		struct memmap_entry *m = &memmap_map[i];
686		reserve_bootmem(m->addr, m->size, BOOTMEM_DEFAULT);
687	}
688
689#ifdef CONFIG_BLK_DEV_INITRD
690	if (initrd_start) {
691		/* Make sure the initrd memory region is not modified. */
692		if (reserve_bootmem(initrd_start, initrd_end - initrd_start,
693				    BOOTMEM_EXCLUSIVE)) {
694			pr_crit("The initrd memory region has been polluted. Disabling it.\n");
695			initrd_start = 0;
696			initrd_end = 0;
697		} else {
698			/*
699			 * Translate initrd_start & initrd_end from PA to VA for
700			 * future access.
701			 */
702			initrd_start += PAGE_OFFSET;
703			initrd_end += PAGE_OFFSET;
704		}
705	}
706#endif
707
708#ifdef CONFIG_KEXEC
709	if (crashk_res.start != crashk_res.end)
710		reserve_bootmem(crashk_res.start, resource_size(&crashk_res),
711				BOOTMEM_DEFAULT);
712#endif
713}
714
715void *__init alloc_remap(int nid, unsigned long size)
716{
717	int pages = node_end_pfn[nid] - node_start_pfn[nid];
718	void *map = pfn_to_kaddr(node_memmap_pfn[nid]);
719	BUG_ON(size != pages * sizeof(struct page));
720	memset(map, 0, size);
721	return map;
722}
723
724static int __init percpu_size(void)
725{
726	int size = __per_cpu_end - __per_cpu_start;
727	size += PERCPU_MODULE_RESERVE;
728	size += PERCPU_DYNAMIC_EARLY_SIZE;
729	if (size < PCPU_MIN_UNIT_SIZE)
730		size = PCPU_MIN_UNIT_SIZE;
731	size = roundup(size, PAGE_SIZE);
732
733	/* In several places we assume the per-cpu data fits on a huge page. */
734	BUG_ON(kdata_huge && size > HPAGE_SIZE);
735	return size;
736}
737
738static void __init zone_sizes_init(void)
739{
740	unsigned long zones_size[MAX_NR_ZONES] = { 0 };
741	int size = percpu_size();
742	int num_cpus = smp_height * smp_width;
743	const unsigned long dma_end = (1UL << (32 - PAGE_SHIFT));
744
745	int i;
746
747	for (i = 0; i < num_cpus; ++i)
748		node_percpu[cpu_to_node(i)] += size;
749
750	for_each_online_node(i) {
751		unsigned long start = node_start_pfn[i];
752		unsigned long end = node_end_pfn[i];
753#ifdef CONFIG_HIGHMEM
754		unsigned long lowmem_end = node_lowmem_end_pfn[i];
755#else
756		unsigned long lowmem_end = end;
757#endif
758		int memmap_size = (end - start) * sizeof(struct page);
759		node_free_pfn[i] = start;
760
761		/*
762		 * Set aside pages for per-cpu data and the mem_map array.
763		 *
764		 * Since the per-cpu data requires special homecaching,
765		 * if we are in kdata_huge mode, we put it at the end of
766		 * the lowmem region.  If we're not in kdata_huge mode,
767		 * we take the per-cpu pages from the bottom of the
768		 * controller, since that avoids fragmenting a huge page
769		 * that users might want.  We always take the memmap
770		 * from the bottom of the controller, since with
771		 * kdata_huge that lets it be under a huge TLB entry.
772		 *
773		 * If the user has requested isolnodes for a controller,
774		 * though, there'll be no lowmem, so we just alloc_bootmem
775		 * the memmap.  There will be no percpu memory either.
776		 */
777		if (i != 0 && node_isset(i, isolnodes)) {
778			node_memmap_pfn[i] =
779				alloc_bootmem_pfn(0, memmap_size, 0);
780			BUG_ON(node_percpu[i] != 0);
781		} else if (node_has_bootmem(start)) {
782			unsigned long goal = 0;
783			node_memmap_pfn[i] =
784				alloc_bootmem_pfn(i, memmap_size, 0);
785			if (kdata_huge)
786				goal = PFN_PHYS(lowmem_end) - node_percpu[i];
787			if (node_percpu[i])
788				node_percpu_pfn[i] =
789					alloc_bootmem_pfn(i, node_percpu[i],
790							  goal);
791		} else {
792			/* In non-bootmem zones, just reserve some pages. */
793			node_memmap_pfn[i] = node_free_pfn[i];
794			node_free_pfn[i] += PFN_UP(memmap_size);
795			if (!kdata_huge) {
796				node_percpu_pfn[i] = node_free_pfn[i];
797				node_free_pfn[i] += PFN_UP(node_percpu[i]);
798			} else {
799				node_percpu_pfn[i] =
800					lowmem_end - PFN_UP(node_percpu[i]);
801			}
802		}
803
804#ifdef CONFIG_HIGHMEM
805		if (start > lowmem_end) {
806			zones_size[ZONE_NORMAL] = 0;
807			zones_size[ZONE_HIGHMEM] = end - start;
808		} else {
809			zones_size[ZONE_NORMAL] = lowmem_end - start;
810			zones_size[ZONE_HIGHMEM] = end - lowmem_end;
811		}
812#else
813		zones_size[ZONE_NORMAL] = end - start;
814#endif
815
816		if (start < dma_end) {
817			zones_size[ZONE_DMA] = min(zones_size[ZONE_NORMAL],
818						   dma_end - start);
819			zones_size[ZONE_NORMAL] -= zones_size[ZONE_DMA];
820		} else {
821			zones_size[ZONE_DMA] = 0;
822		}
823
824		/* Take zone metadata from controller 0 if we're isolnode. */
825		if (node_isset(i, isolnodes))
826			NODE_DATA(i)->bdata = &bootmem_node_data[0];
827
828		free_area_init_node(i, zones_size, start, NULL);
829		printk(KERN_DEBUG "  Normal zone: %ld per-cpu pages\n",
830		       PFN_UP(node_percpu[i]));
831
832		/* Track the type of memory on each node */
833		if (zones_size[ZONE_NORMAL] || zones_size[ZONE_DMA])
834			node_set_state(i, N_NORMAL_MEMORY);
835#ifdef CONFIG_HIGHMEM
836		if (end != start)
837			node_set_state(i, N_HIGH_MEMORY);
838#endif
839
840		node_set_online(i);
841	}
842}
843
844#ifdef CONFIG_NUMA
845
846/* which logical CPUs are on which nodes */
847struct cpumask node_2_cpu_mask[MAX_NUMNODES] __write_once;
848EXPORT_SYMBOL(node_2_cpu_mask);
849
850/* which node each logical CPU is on */
851char cpu_2_node[NR_CPUS] __write_once __attribute__((aligned(L2_CACHE_BYTES)));
852EXPORT_SYMBOL(cpu_2_node);
853
854/* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
855static int __init cpu_to_bound_node(int cpu, struct cpumask* unbound_cpus)
856{
857	if (!cpu_possible(cpu) || cpumask_test_cpu(cpu, unbound_cpus))
858		return -1;
859	else
860		return cpu_to_node(cpu);
861}
862
863/* Return number of immediately-adjacent tiles sharing the same NUMA node. */
864static int __init node_neighbors(int node, int cpu,
865				 struct cpumask *unbound_cpus)
866{
867	int neighbors = 0;
868	int w = smp_width;
869	int h = smp_height;
870	int x = cpu % w;
871	int y = cpu / w;
872	if (x > 0 && cpu_to_bound_node(cpu-1, unbound_cpus) == node)
873		++neighbors;
874	if (x < w-1 && cpu_to_bound_node(cpu+1, unbound_cpus) == node)
875		++neighbors;
876	if (y > 0 && cpu_to_bound_node(cpu-w, unbound_cpus) == node)
877		++neighbors;
878	if (y < h-1 && cpu_to_bound_node(cpu+w, unbound_cpus) == node)
879		++neighbors;
880	return neighbors;
881}
882
883static void __init setup_numa_mapping(void)
884{
885	int distance[MAX_NUMNODES][NR_CPUS];
886	HV_Coord coord;
887	int cpu, node, cpus, i, x, y;
888	int num_nodes = num_online_nodes();
889	struct cpumask unbound_cpus;
890	nodemask_t default_nodes;
891
892	cpumask_clear(&unbound_cpus);
893
894	/* Get set of nodes we will use for defaults */
895	nodes_andnot(default_nodes, node_online_map, isolnodes);
896	if (nodes_empty(default_nodes)) {
897		BUG_ON(!node_isset(0, node_online_map));
898		pr_err("Forcing NUMA node zero available as a default node\n");
899		node_set(0, default_nodes);
900	}
901
902	/* Populate the distance[] array */
903	memset(distance, -1, sizeof(distance));
904	cpu = 0;
905	for (coord.y = 0; coord.y < smp_height; ++coord.y) {
906		for (coord.x = 0; coord.x < smp_width;
907		     ++coord.x, ++cpu) {
908			BUG_ON(cpu >= nr_cpu_ids);
909			if (!cpu_possible(cpu)) {
910				cpu_2_node[cpu] = -1;
911				continue;
912			}
913			for_each_node_mask(node, default_nodes) {
914				HV_MemoryControllerInfo info =
915					hv_inquire_memory_controller(
916						coord, node_controller[node]);
917				distance[node][cpu] =
918					ABS(info.coord.x) + ABS(info.coord.y);
919			}
920			cpumask_set_cpu(cpu, &unbound_cpus);
921		}
922	}
923	cpus = cpu;
924
925	/*
926	 * Round-robin through the NUMA nodes until all the cpus are
927	 * assigned.  We could be more clever here (e.g. create four
928	 * sorted linked lists on the same set of cpu nodes, and pull
929	 * off them in round-robin sequence, removing from all four
930	 * lists each time) but given the relatively small numbers
931	 * involved, O(n^2) seem OK for a one-time cost.
932	 */
933	node = first_node(default_nodes);
934	while (!cpumask_empty(&unbound_cpus)) {
935		int best_cpu = -1;
936		int best_distance = INT_MAX;
937		for (cpu = 0; cpu < cpus; ++cpu) {
938			if (cpumask_test_cpu(cpu, &unbound_cpus)) {
939				/*
940				 * Compute metric, which is how much
941				 * closer the cpu is to this memory
942				 * controller than the others, shifted
943				 * up, and then the number of
944				 * neighbors already in the node as an
945				 * epsilon adjustment to try to keep
946				 * the nodes compact.
947				 */
948				int d = distance[node][cpu] * num_nodes;
949				for_each_node_mask(i, default_nodes) {
950					if (i != node)
951						d -= distance[i][cpu];
952				}
953				d *= 8;  /* allow space for epsilon */
954				d -= node_neighbors(node, cpu, &unbound_cpus);
955				if (d < best_distance) {
956					best_cpu = cpu;
957					best_distance = d;
958				}
959			}
960		}
961		BUG_ON(best_cpu < 0);
962		cpumask_set_cpu(best_cpu, &node_2_cpu_mask[node]);
963		cpu_2_node[best_cpu] = node;
964		cpumask_clear_cpu(best_cpu, &unbound_cpus);
965		node = next_node(node, default_nodes);
966		if (node == MAX_NUMNODES)
967			node = first_node(default_nodes);
968	}
969
970	/* Print out node assignments and set defaults for disabled cpus */
971	cpu = 0;
972	for (y = 0; y < smp_height; ++y) {
973		printk(KERN_DEBUG "NUMA cpu-to-node row %d:", y);
974		for (x = 0; x < smp_width; ++x, ++cpu) {
975			if (cpu_to_node(cpu) < 0) {
976				pr_cont(" -");
977				cpu_2_node[cpu] = first_node(default_nodes);
978			} else {
979				pr_cont(" %d", cpu_to_node(cpu));
980			}
981		}
982		pr_cont("\n");
983	}
984}
985
986static struct cpu cpu_devices[NR_CPUS];
987
988static int __init topology_init(void)
989{
990	int i;
991
992	for_each_online_node(i)
993		register_one_node(i);
994
995	for (i = 0; i < smp_height * smp_width; ++i)
996		register_cpu(&cpu_devices[i], i);
997
998	return 0;
999}
1000
1001subsys_initcall(topology_init);
1002
1003#else /* !CONFIG_NUMA */
1004
1005#define setup_numa_mapping() do { } while (0)
1006
1007#endif /* CONFIG_NUMA */
1008
1009/*
1010 * Initialize hugepage support on this cpu.  We do this on all cores
1011 * early in boot: before argument parsing for the boot cpu, and after
1012 * argument parsing but before the init functions run on the secondaries.
1013 * So the values we set up here in the hypervisor may be overridden on
1014 * the boot cpu as arguments are parsed.
1015 */
1016static void init_super_pages(void)
1017{
1018#ifdef CONFIG_HUGETLB_SUPER_PAGES
1019	int i;
1020	for (i = 0; i < HUGE_SHIFT_ENTRIES; ++i)
1021		hv_set_pte_super_shift(i, huge_shift[i]);
1022#endif
1023}
1024
1025/**
1026 * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
1027 * @boot: Is this the boot cpu?
1028 *
1029 * Called from setup_arch() on the boot cpu, or online_secondary().
1030 */
1031void setup_cpu(int boot)
1032{
1033	/* The boot cpu sets up its permanent mappings much earlier. */
1034	if (!boot)
1035		store_permanent_mappings();
1036
1037	/* Allow asynchronous TLB interrupts. */
1038#if CHIP_HAS_TILE_DMA()
1039	arch_local_irq_unmask(INT_DMATLB_MISS);
1040	arch_local_irq_unmask(INT_DMATLB_ACCESS);
1041#endif
1042#ifdef __tilegx__
1043	arch_local_irq_unmask(INT_SINGLE_STEP_K);
1044#endif
1045
1046	/*
1047	 * Allow user access to many generic SPRs, like the cycle
1048	 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
1049	 */
1050	__insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0, 1);
1051
1052#if CHIP_HAS_SN()
1053	/* Static network is not restricted. */
1054	__insn_mtspr(SPR_MPL_SN_ACCESS_SET_0, 1);
1055#endif
1056
1057	/*
1058	 * Set the MPL for interrupt control 0 & 1 to the corresponding
1059	 * values.  This includes access to the SYSTEM_SAVE and EX_CONTEXT
1060	 * SPRs, as well as the interrupt mask.
1061	 */
1062	__insn_mtspr(SPR_MPL_INTCTRL_0_SET_0, 1);
1063	__insn_mtspr(SPR_MPL_INTCTRL_1_SET_1, 1);
1064
1065	/* Initialize IRQ support for this cpu. */
1066	setup_irq_regs();
1067
1068#ifdef CONFIG_HARDWALL
1069	/* Reset the network state on this cpu. */
1070	reset_network_state();
1071#endif
1072
1073	init_super_pages();
1074}
1075
1076#ifdef CONFIG_BLK_DEV_INITRD
1077
1078static int __initdata set_initramfs_file;
1079static char __initdata initramfs_file[128] = "initramfs";
1080
1081static int __init setup_initramfs_file(char *str)
1082{
1083	if (str == NULL)
1084		return -EINVAL;
1085	strncpy(initramfs_file, str, sizeof(initramfs_file) - 1);
1086	set_initramfs_file = 1;
1087
1088	return 0;
1089}
1090early_param("initramfs_file", setup_initramfs_file);
1091
1092/*
1093 * We look for a file called "initramfs" in the hvfs.  If there is one, we
1094 * allocate some memory for it and it will be unpacked to the initramfs.
1095 * If it's compressed, the initd code will uncompress it first.
1096 */
1097static void __init load_hv_initrd(void)
1098{
1099	HV_FS_StatInfo stat;
1100	int fd, rc;
1101	void *initrd;
1102
1103	/* If initrd has already been set, skip initramfs file in hvfs. */
1104	if (initrd_start)
1105		return;
1106
1107	fd = hv_fs_findfile((HV_VirtAddr) initramfs_file);
1108	if (fd == HV_ENOENT) {
1109		if (set_initramfs_file) {
1110			pr_warn("No such hvfs initramfs file '%s'\n",
1111				initramfs_file);
1112			return;
1113		} else {
1114			/* Try old backwards-compatible name. */
1115			fd = hv_fs_findfile((HV_VirtAddr)"initramfs.cpio.gz");
1116			if (fd == HV_ENOENT)
1117				return;
1118		}
1119	}
1120	BUG_ON(fd < 0);
1121	stat = hv_fs_fstat(fd);
1122	BUG_ON(stat.size < 0);
1123	if (stat.flags & HV_FS_ISDIR) {
1124		pr_warn("Ignoring hvfs file '%s': it's a directory\n",
1125			initramfs_file);
1126		return;
1127	}
1128	initrd = alloc_bootmem_pages(stat.size);
1129	rc = hv_fs_pread(fd, (HV_VirtAddr) initrd, stat.size, 0);
1130	if (rc != stat.size) {
1131		pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
1132		       stat.size, initramfs_file, rc);
1133		free_initrd_mem((unsigned long) initrd, stat.size);
1134		return;
1135	}
1136	initrd_start = (unsigned long) initrd;
1137	initrd_end = initrd_start + stat.size;
1138}
1139
1140void __init free_initrd_mem(unsigned long begin, unsigned long end)
1141{
1142	free_bootmem_late(__pa(begin), end - begin);
1143}
1144
1145static int __init setup_initrd(char *str)
1146{
1147	char *endp;
1148	unsigned long initrd_size;
1149
1150	initrd_size = str ? simple_strtoul(str, &endp, 0) : 0;
1151	if (initrd_size == 0 || *endp != '@')
1152		return -EINVAL;
1153
1154	initrd_start = simple_strtoul(endp+1, &endp, 0);
1155	if (initrd_start == 0)
1156		return -EINVAL;
1157
1158	initrd_end = initrd_start + initrd_size;
1159
1160	return 0;
1161}
1162early_param("initrd", setup_initrd);
1163
1164#else
1165static inline void load_hv_initrd(void) {}
1166#endif /* CONFIG_BLK_DEV_INITRD */
1167
1168static void __init validate_hv(void)
1169{
1170	/*
1171	 * It may already be too late, but let's check our built-in
1172	 * configuration against what the hypervisor is providing.
1173	 */
1174	unsigned long glue_size = hv_sysconf(HV_SYSCONF_GLUE_SIZE);
1175	int hv_page_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL);
1176	int hv_hpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE);
1177	HV_ASIDRange asid_range;
1178
1179#ifndef CONFIG_SMP
1180	HV_Topology topology = hv_inquire_topology();
1181	BUG_ON(topology.coord.x != 0 || topology.coord.y != 0);
1182	if (topology.width != 1 || topology.height != 1) {
1183		pr_warn("Warning: booting UP kernel on %dx%d grid; will ignore all but first tile\n",
1184			topology.width, topology.height);
1185	}
1186#endif
1187
1188	if (PAGE_OFFSET + HV_GLUE_START_CPA + glue_size > (unsigned long)_text)
1189		early_panic("Hypervisor glue size %ld is too big!\n",
1190			    glue_size);
1191	if (hv_page_size != PAGE_SIZE)
1192		early_panic("Hypervisor page size %#x != our %#lx\n",
1193			    hv_page_size, PAGE_SIZE);
1194	if (hv_hpage_size != HPAGE_SIZE)
1195		early_panic("Hypervisor huge page size %#x != our %#lx\n",
1196			    hv_hpage_size, HPAGE_SIZE);
1197
1198#ifdef CONFIG_SMP
1199	/*
1200	 * Some hypervisor APIs take a pointer to a bitmap array
1201	 * whose size is at least the number of cpus on the chip.
1202	 * We use a struct cpumask for this, so it must be big enough.
1203	 */
1204	if ((smp_height * smp_width) > nr_cpu_ids)
1205		early_panic("Hypervisor %d x %d grid too big for Linux NR_CPUS %d\n",
1206			    smp_height, smp_width, nr_cpu_ids);
1207#endif
1208
1209	/*
1210	 * Check that we're using allowed ASIDs, and initialize the
1211	 * various asid variables to their appropriate initial states.
1212	 */
1213	asid_range = hv_inquire_asid(0);
1214	min_asid = asid_range.start;
1215	__this_cpu_write(current_asid, min_asid);
1216	max_asid = asid_range.start + asid_range.size - 1;
1217
1218	if (hv_confstr(HV_CONFSTR_CHIP_MODEL, (HV_VirtAddr)chip_model,
1219		       sizeof(chip_model)) < 0) {
1220		pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
1221		strlcpy(chip_model, "unknown", sizeof(chip_model));
1222	}
1223}
1224
1225static void __init validate_va(void)
1226{
1227#ifndef __tilegx__   /* FIXME: GX: probably some validation relevant here */
1228	/*
1229	 * Similarly, make sure we're only using allowed VAs.
1230	 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_START,
1231	 * and 0 .. KERNEL_HIGH_VADDR.
1232	 * In addition, make sure we CAN'T use the end of memory, since
1233	 * we use the last chunk of each pgd for the pgd_list.
1234	 */
1235	int i, user_kernel_ok = 0;
1236	unsigned long max_va = 0;
1237	unsigned long list_va =
1238		((PGD_LIST_OFFSET / sizeof(pgd_t)) << PGDIR_SHIFT);
1239
1240	for (i = 0; ; ++i) {
1241		HV_VirtAddrRange range = hv_inquire_virtual(i);
1242		if (range.size == 0)
1243			break;
1244		if (range.start <= MEM_USER_INTRPT &&
1245		    range.start + range.size >= MEM_HV_START)
1246			user_kernel_ok = 1;
1247		if (range.start == 0)
1248			max_va = range.size;
1249		BUG_ON(range.start + range.size > list_va);
1250	}
1251	if (!user_kernel_ok)
1252		early_panic("Hypervisor not configured for user/kernel VAs\n");
1253	if (max_va == 0)
1254		early_panic("Hypervisor not configured for low VAs\n");
1255	if (max_va < KERNEL_HIGH_VADDR)
1256		early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
1257			    max_va, KERNEL_HIGH_VADDR);
1258
1259	/* Kernel PCs must have their high bit set; see intvec.S. */
1260	if ((long)VMALLOC_START >= 0)
1261		early_panic("Linux VMALLOC region below the 2GB line (%#lx)!\n"
1262			    "Reconfigure the kernel with smaller VMALLOC_RESERVE\n",
1263			    VMALLOC_START);
1264#endif
1265}
1266
1267/*
1268 * cpu_lotar_map lists all the cpus that are valid for the supervisor
1269 * to cache data on at a page level, i.e. what cpus can be placed in
1270 * the LOTAR field of a PTE.  It is equivalent to the set of possible
1271 * cpus plus any other cpus that are willing to share their cache.
1272 * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
1273 */
1274struct cpumask __write_once cpu_lotar_map;
1275EXPORT_SYMBOL(cpu_lotar_map);
1276
1277/*
1278 * hash_for_home_map lists all the tiles that hash-for-home data
1279 * will be cached on.  Note that this may includes tiles that are not
1280 * valid for this supervisor to use otherwise (e.g. if a hypervisor
1281 * device is being shared between multiple supervisors).
1282 * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
1283 */
1284struct cpumask hash_for_home_map;
1285EXPORT_SYMBOL(hash_for_home_map);
1286
1287/*
1288 * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
1289 * flush on our behalf.  It is set to cpu_possible_mask OR'ed with
1290 * hash_for_home_map, and it is what should be passed to
1291 * hv_flush_remote() to flush all caches.  Note that if there are
1292 * dedicated hypervisor driver tiles that have authorized use of their
1293 * cache, those tiles will only appear in cpu_lotar_map, NOT in
1294 * cpu_cacheable_map, as they are a special case.
1295 */
1296struct cpumask __write_once cpu_cacheable_map;
1297EXPORT_SYMBOL(cpu_cacheable_map);
1298
1299static __initdata struct cpumask disabled_map;
1300
1301static int __init disabled_cpus(char *str)
1302{
1303	int boot_cpu = smp_processor_id();
1304
1305	if (str == NULL || cpulist_parse_crop(str, &disabled_map) != 0)
1306		return -EINVAL;
1307	if (cpumask_test_cpu(boot_cpu, &disabled_map)) {
1308		pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu);
1309		cpumask_clear_cpu(boot_cpu, &disabled_map);
1310	}
1311	return 0;
1312}
1313
1314early_param("disabled_cpus", disabled_cpus);
1315
1316void __init print_disabled_cpus(void)
1317{
1318	if (!cpumask_empty(&disabled_map))
1319		pr_info("CPUs not available for Linux: %*pbl\n",
1320			cpumask_pr_args(&disabled_map));
1321}
1322
1323static void __init setup_cpu_maps(void)
1324{
1325	struct cpumask hv_disabled_map, cpu_possible_init;
1326	int boot_cpu = smp_processor_id();
1327	int cpus, i, rc;
1328
1329	/* Learn which cpus are allowed by the hypervisor. */
1330	rc = hv_inquire_tiles(HV_INQ_TILES_AVAIL,
1331			      (HV_VirtAddr) cpumask_bits(&cpu_possible_init),
1332			      sizeof(cpu_cacheable_map));
1333	if (rc < 0)
1334		early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc);
1335	if (!cpumask_test_cpu(boot_cpu, &cpu_possible_init))
1336		early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu);
1337
1338	/* Compute the cpus disabled by the hvconfig file. */
1339	cpumask_complement(&hv_disabled_map, &cpu_possible_init);
1340
1341	/* Include them with the cpus disabled by "disabled_cpus". */
1342	cpumask_or(&disabled_map, &disabled_map, &hv_disabled_map);
1343
1344	/*
1345	 * Disable every cpu after "setup_max_cpus".  But don't mark
1346	 * as disabled the cpus that are outside of our initial rectangle,
1347	 * since that turns out to be confusing.
1348	 */
1349	cpus = 1;                          /* this cpu */
1350	cpumask_set_cpu(boot_cpu, &disabled_map);   /* ignore this cpu */
1351	for (i = 0; cpus < setup_max_cpus; ++i)
1352		if (!cpumask_test_cpu(i, &disabled_map))
1353			++cpus;
1354	for (; i < smp_height * smp_width; ++i)
1355		cpumask_set_cpu(i, &disabled_map);
1356	cpumask_clear_cpu(boot_cpu, &disabled_map); /* reset this cpu */
1357	for (i = smp_height * smp_width; i < NR_CPUS; ++i)
1358		cpumask_clear_cpu(i, &disabled_map);
1359
1360	/*
1361	 * Setup cpu_possible map as every cpu allocated to us, minus
1362	 * the results of any "disabled_cpus" settings.
1363	 */
1364	cpumask_andnot(&cpu_possible_init, &cpu_possible_init, &disabled_map);
1365	init_cpu_possible(&cpu_possible_init);
1366
1367	/* Learn which cpus are valid for LOTAR caching. */
1368	rc = hv_inquire_tiles(HV_INQ_TILES_LOTAR,
1369			      (HV_VirtAddr) cpumask_bits(&cpu_lotar_map),
1370			      sizeof(cpu_lotar_map));
1371	if (rc < 0) {
1372		pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
1373		cpu_lotar_map = *cpu_possible_mask;
1374	}
1375
1376	/* Retrieve set of CPUs used for hash-for-home caching */
1377	rc = hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE,
1378			      (HV_VirtAddr) hash_for_home_map.bits,
1379			      sizeof(hash_for_home_map));
1380	if (rc < 0)
1381		early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc);
1382	cpumask_or(&cpu_cacheable_map, cpu_possible_mask, &hash_for_home_map);
1383}
1384
1385
1386static int __init dataplane(char *str)
1387{
1388	pr_warn("WARNING: dataplane support disabled in this kernel\n");
1389	return 0;
1390}
1391
1392early_param("dataplane", dataplane);
1393
1394#ifdef CONFIG_NO_HZ_FULL
1395/* Warn if hypervisor shared cpus are marked as nohz_full. */
1396static int __init check_nohz_full_cpus(void)
1397{
1398	struct cpumask shared;
1399	int cpu;
1400
1401	if (hv_inquire_tiles(HV_INQ_TILES_SHARED,
1402			     (HV_VirtAddr) shared.bits, sizeof(shared)) < 0) {
1403		pr_warn("WARNING: No support for inquiring hv shared tiles\n");
1404		return 0;
1405	}
1406	for_each_cpu(cpu, &shared) {
1407		if (tick_nohz_full_cpu(cpu))
1408			pr_warn("WARNING: nohz_full cpu %d receives hypervisor interrupts!\n",
1409			       cpu);
1410	}
1411	return 0;
1412}
1413arch_initcall(check_nohz_full_cpus);
1414#endif
1415
1416#ifdef CONFIG_CMDLINE_BOOL
1417static char __initdata builtin_cmdline[COMMAND_LINE_SIZE] = CONFIG_CMDLINE;
1418#endif
1419
1420void __init setup_arch(char **cmdline_p)
1421{
1422	int len;
1423
1424#if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
1425	len = hv_get_command_line((HV_VirtAddr) boot_command_line,
1426				  COMMAND_LINE_SIZE);
1427	if (boot_command_line[0])
1428		pr_warn("WARNING: ignoring dynamic command line \"%s\"\n",
1429			boot_command_line);
1430	strlcpy(boot_command_line, builtin_cmdline, COMMAND_LINE_SIZE);
1431#else
1432	char *hv_cmdline;
1433#if defined(CONFIG_CMDLINE_BOOL)
1434	if (builtin_cmdline[0]) {
1435		int builtin_len = strlcpy(boot_command_line, builtin_cmdline,
1436					  COMMAND_LINE_SIZE);
1437		if (builtin_len < COMMAND_LINE_SIZE-1)
1438			boot_command_line[builtin_len++] = ' ';
1439		hv_cmdline = &boot_command_line[builtin_len];
1440		len = COMMAND_LINE_SIZE - builtin_len;
1441	} else
1442#endif
1443	{
1444		hv_cmdline = boot_command_line;
1445		len = COMMAND_LINE_SIZE;
1446	}
1447	len = hv_get_command_line((HV_VirtAddr) hv_cmdline, len);
1448	if (len < 0 || len > COMMAND_LINE_SIZE)
1449		early_panic("hv_get_command_line failed: %d\n", len);
1450#endif
1451
1452	*cmdline_p = boot_command_line;
1453
1454	/* Set disabled_map and setup_max_cpus very early */
1455	parse_early_param();
1456
1457	/* Make sure the kernel is compatible with the hypervisor. */
1458	validate_hv();
1459	validate_va();
1460
1461	setup_cpu_maps();
1462
1463
1464#if defined(CONFIG_PCI) && !defined(__tilegx__)
1465	/*
1466	 * Initialize the PCI structures.  This is done before memory
1467	 * setup so that we know whether or not a pci_reserve region
1468	 * is necessary.
1469	 */
1470	if (tile_pci_init() == 0)
1471		pci_reserve_mb = 0;
1472
1473	/* PCI systems reserve a region just below 4GB for mapping iomem. */
1474	pci_reserve_end_pfn  = (1 << (32 - PAGE_SHIFT));
1475	pci_reserve_start_pfn = pci_reserve_end_pfn -
1476		(pci_reserve_mb << (20 - PAGE_SHIFT));
1477#endif
1478
1479	init_mm.start_code = (unsigned long) _text;
1480	init_mm.end_code = (unsigned long) _etext;
1481	init_mm.end_data = (unsigned long) _edata;
1482	init_mm.brk = (unsigned long) _end;
1483
1484	setup_memory();
1485	store_permanent_mappings();
1486	setup_bootmem_allocator();
1487
1488	/*
1489	 * NOTE: before this point _nobody_ is allowed to allocate
1490	 * any memory using the bootmem allocator.
1491	 */
1492
1493#ifdef CONFIG_SWIOTLB
1494	swiotlb_init(0);
1495#endif
1496
1497	paging_init();
1498	setup_numa_mapping();
1499	zone_sizes_init();
1500	set_page_homes();
1501	setup_cpu(1);
1502	setup_clock();
1503	load_hv_initrd();
1504}
1505
1506
1507/*
1508 * Set up per-cpu memory.
1509 */
1510
1511unsigned long __per_cpu_offset[NR_CPUS] __write_once;
1512EXPORT_SYMBOL(__per_cpu_offset);
1513
1514static size_t __initdata pfn_offset[MAX_NUMNODES] = { 0 };
1515static unsigned long __initdata percpu_pfn[NR_CPUS] = { 0 };
1516
1517/*
1518 * As the percpu code allocates pages, we return the pages from the
1519 * end of the node for the specified cpu.
1520 */
1521static void *__init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
1522{
1523	int nid = cpu_to_node(cpu);
1524	unsigned long pfn = node_percpu_pfn[nid] + pfn_offset[nid];
1525
1526	BUG_ON(size % PAGE_SIZE != 0);
1527	pfn_offset[nid] += size / PAGE_SIZE;
1528	BUG_ON(node_percpu[nid] < size);
1529	node_percpu[nid] -= size;
1530	if (percpu_pfn[cpu] == 0)
1531		percpu_pfn[cpu] = pfn;
1532	return pfn_to_kaddr(pfn);
1533}
1534
1535/*
1536 * Pages reserved for percpu memory are not freeable, and in any case we are
1537 * on a short path to panic() in setup_per_cpu_area() at this point anyway.
1538 */
1539static void __init pcpu_fc_free(void *ptr, size_t size)
1540{
1541}
1542
1543/*
1544 * Set up vmalloc page tables using bootmem for the percpu code.
1545 */
1546static void __init pcpu_fc_populate_pte(unsigned long addr)
1547{
1548	pgd_t *pgd;
1549	pud_t *pud;
1550	pmd_t *pmd;
1551	pte_t *pte;
1552
1553	BUG_ON(pgd_addr_invalid(addr));
1554	if (addr < VMALLOC_START || addr >= VMALLOC_END)
1555		panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx; try increasing CONFIG_VMALLOC_RESERVE\n",
1556		      addr, VMALLOC_START, VMALLOC_END);
1557
1558	pgd = swapper_pg_dir + pgd_index(addr);
1559	pud = pud_offset(pgd, addr);
1560	BUG_ON(!pud_present(*pud));
1561	pmd = pmd_offset(pud, addr);
1562	if (pmd_present(*pmd)) {
1563		BUG_ON(pmd_huge_page(*pmd));
1564	} else {
1565		pte = __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE,
1566				      HV_PAGE_TABLE_ALIGN, 0);
1567		pmd_populate_kernel(&init_mm, pmd, pte);
1568	}
1569}
1570
1571void __init setup_per_cpu_areas(void)
1572{
1573	struct page *pg;
1574	unsigned long delta, pfn, lowmem_va;
1575	unsigned long size = percpu_size();
1576	char *ptr;
1577	int rc, cpu, i;
1578
1579	rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_fc_alloc,
1580				   pcpu_fc_free, pcpu_fc_populate_pte);
1581	if (rc < 0)
1582		panic("Cannot initialize percpu area (err=%d)", rc);
1583
1584	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1585	for_each_possible_cpu(cpu) {
1586		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1587
1588		/* finv the copy out of cache so we can change homecache */
1589		ptr = pcpu_base_addr + pcpu_unit_offsets[cpu];
1590		__finv_buffer(ptr, size);
1591		pfn = percpu_pfn[cpu];
1592
1593		/* Rewrite the page tables to cache on that cpu */
1594		pg = pfn_to_page(pfn);
1595		for (i = 0; i < size; i += PAGE_SIZE, ++pfn, ++pg) {
1596
1597			/* Update the vmalloc mapping and page home. */
1598			unsigned long addr = (unsigned long)ptr + i;
1599			pte_t *ptep = virt_to_kpte(addr);
1600			pte_t pte = *ptep;
1601			BUG_ON(pfn != pte_pfn(pte));
1602			pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
1603			pte = set_remote_cache_cpu(pte, cpu);
1604			set_pte_at(&init_mm, addr, ptep, pte);
1605
1606			/* Update the lowmem mapping for consistency. */
1607			lowmem_va = (unsigned long)pfn_to_kaddr(pfn);
1608			ptep = virt_to_kpte(lowmem_va);
1609			if (pte_huge(*ptep)) {
1610				printk(KERN_DEBUG "early shatter of huge page at %#lx\n",
1611				       lowmem_va);
1612				shatter_pmd((pmd_t *)ptep);
1613				ptep = virt_to_kpte(lowmem_va);
1614				BUG_ON(pte_huge(*ptep));
1615			}
1616			BUG_ON(pfn != pte_pfn(*ptep));
1617			set_pte_at(&init_mm, lowmem_va, ptep, pte);
1618		}
1619	}
1620
1621	/* Set our thread pointer appropriately. */
1622	set_my_cpu_offset(__per_cpu_offset[smp_processor_id()]);
1623
1624	/* Make sure the finv's have completed. */
1625	mb_incoherent();
1626
1627	/* Flush the TLB so we reference it properly from here on out. */
1628	local_flush_tlb_all();
1629}
1630
1631static struct resource data_resource = {
1632	.name	= "Kernel data",
1633	.start	= 0,
1634	.end	= 0,
1635	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
1636};
1637
1638static struct resource code_resource = {
1639	.name	= "Kernel code",
1640	.start	= 0,
1641	.end	= 0,
1642	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
1643};
1644
1645/*
1646 * On Pro, we reserve all resources above 4GB so that PCI won't try to put
1647 * mappings above 4GB.
1648 */
1649#if defined(CONFIG_PCI) && !defined(__tilegx__)
1650static struct resource* __init
1651insert_non_bus_resource(void)
1652{
1653	struct resource *res =
1654		kzalloc(sizeof(struct resource), GFP_ATOMIC);
1655	if (!res)
1656		return NULL;
1657	res->name = "Non-Bus Physical Address Space";
1658	res->start = (1ULL << 32);
1659	res->end = -1LL;
1660	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1661	if (insert_resource(&iomem_resource, res)) {
1662		kfree(res);
1663		return NULL;
1664	}
1665	return res;
1666}
1667#endif
1668
1669static struct resource* __init
1670insert_ram_resource(u64 start_pfn, u64 end_pfn, bool reserved)
1671{
1672	struct resource *res =
1673		kzalloc(sizeof(struct resource), GFP_ATOMIC);
1674	if (!res)
1675		return NULL;
1676	res->name = reserved ? "Reserved" : "System RAM";
1677	res->start = start_pfn << PAGE_SHIFT;
1678	res->end = (end_pfn << PAGE_SHIFT) - 1;
1679	res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1680	if (insert_resource(&iomem_resource, res)) {
1681		kfree(res);
1682		return NULL;
1683	}
1684	return res;
1685}
1686
1687/*
1688 * Request address space for all standard resources
1689 *
1690 * If the system includes PCI root complex drivers, we need to create
1691 * a window just below 4GB where PCI BARs can be mapped.
1692 */
1693static int __init request_standard_resources(void)
1694{
1695	int i;
1696	enum { CODE_DELTA = MEM_SV_START - PAGE_OFFSET };
1697
1698#if defined(CONFIG_PCI) && !defined(__tilegx__)
1699	insert_non_bus_resource();
1700#endif
1701
1702	for_each_online_node(i) {
1703		u64 start_pfn = node_start_pfn[i];
1704		u64 end_pfn = node_end_pfn[i];
1705
1706#if defined(CONFIG_PCI) && !defined(__tilegx__)
1707		if (start_pfn <= pci_reserve_start_pfn &&
1708		    end_pfn > pci_reserve_start_pfn) {
1709			if (end_pfn > pci_reserve_end_pfn)
1710				insert_ram_resource(pci_reserve_end_pfn,
1711						    end_pfn, 0);
1712			end_pfn = pci_reserve_start_pfn;
1713		}
1714#endif
1715		insert_ram_resource(start_pfn, end_pfn, 0);
1716	}
1717
1718	code_resource.start = __pa(_text - CODE_DELTA);
1719	code_resource.end = __pa(_etext - CODE_DELTA)-1;
1720	data_resource.start = __pa(_sdata);
1721	data_resource.end = __pa(_end)-1;
1722
1723	insert_resource(&iomem_resource, &code_resource);
1724	insert_resource(&iomem_resource, &data_resource);
1725
1726	/* Mark any "memmap" regions busy for the resource manager. */
1727	for (i = 0; i < memmap_nr; ++i) {
1728		struct memmap_entry *m = &memmap_map[i];
1729		insert_ram_resource(PFN_DOWN(m->addr),
1730				    PFN_UP(m->addr + m->size - 1), 1);
1731	}
1732
1733#ifdef CONFIG_KEXEC
1734	insert_resource(&iomem_resource, &crashk_res);
1735#endif
1736
1737	return 0;
1738}
1739
1740subsys_initcall(request_standard_resources);
1741