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/initrd.h>
27 #include <linux/io.h>
28 #include <linux/highmem.h>
29 #include <linux/smp.h>
30 #include <linux/timex.h>
31 #include <asm/setup.h>
32 #include <asm/sections.h>
33 #include <asm/cacheflush.h>
34 #include <asm/pgalloc.h>
35 #include <asm/mmu_context.h>
36 #include <hv/hypervisor.h>
37 #include <arch/interrupts.h>
38
39 /* <linux/smp.h> doesn't provide this definition. */
40 #ifndef CONFIG_SMP
41 #define setup_max_cpus 1
42 #endif
43
ABS(int x)44 static inline int ABS(int x) { return x >= 0 ? x : -x; }
45
46 /* Chip information */
47 char chip_model[64] __write_once;
48
49 struct pglist_data node_data[MAX_NUMNODES] __read_mostly;
50 EXPORT_SYMBOL(node_data);
51
52 /* We only create bootmem data on node 0. */
53 static bootmem_data_t __initdata node0_bdata;
54
55 /* Information on the NUMA nodes that we compute early */
56 unsigned long __cpuinitdata node_start_pfn[MAX_NUMNODES];
57 unsigned long __cpuinitdata node_end_pfn[MAX_NUMNODES];
58 unsigned long __initdata node_memmap_pfn[MAX_NUMNODES];
59 unsigned long __initdata node_percpu_pfn[MAX_NUMNODES];
60 unsigned long __initdata node_free_pfn[MAX_NUMNODES];
61
62 static unsigned long __initdata node_percpu[MAX_NUMNODES];
63
64 #ifdef CONFIG_HIGHMEM
65 /* Page frame index of end of lowmem on each controller. */
66 unsigned long __cpuinitdata node_lowmem_end_pfn[MAX_NUMNODES];
67
68 /* Number of pages that can be mapped into lowmem. */
69 static unsigned long __initdata mappable_physpages;
70 #endif
71
72 /* Data on which physical memory controller corresponds to which NUMA node */
73 int node_controller[MAX_NUMNODES] = { [0 ... MAX_NUMNODES-1] = -1 };
74
75 #ifdef CONFIG_HIGHMEM
76 /* Map information from VAs to PAs */
77 unsigned long pbase_map[1 << (32 - HPAGE_SHIFT)]
78 __write_once __attribute__((aligned(L2_CACHE_BYTES)));
79 EXPORT_SYMBOL(pbase_map);
80
81 /* Map information from PAs to VAs */
82 void *vbase_map[NR_PA_HIGHBIT_VALUES]
83 __write_once __attribute__((aligned(L2_CACHE_BYTES)));
84 EXPORT_SYMBOL(vbase_map);
85 #endif
86
87 /* Node number as a function of the high PA bits */
88 int highbits_to_node[NR_PA_HIGHBIT_VALUES] __write_once;
89 EXPORT_SYMBOL(highbits_to_node);
90
91 static unsigned int __initdata maxmem_pfn = -1U;
92 static unsigned int __initdata maxnodemem_pfn[MAX_NUMNODES] = {
93 [0 ... MAX_NUMNODES-1] = -1U
94 };
95 static nodemask_t __initdata isolnodes;
96
97 #ifdef CONFIG_PCI
98 enum { DEFAULT_PCI_RESERVE_MB = 64 };
99 static unsigned int __initdata pci_reserve_mb = DEFAULT_PCI_RESERVE_MB;
100 unsigned long __initdata pci_reserve_start_pfn = -1U;
101 unsigned long __initdata pci_reserve_end_pfn = -1U;
102 #endif
103
setup_maxmem(char * str)104 static int __init setup_maxmem(char *str)
105 {
106 long maxmem_mb;
107 if (str == NULL || strict_strtol(str, 0, &maxmem_mb) != 0 ||
108 maxmem_mb == 0)
109 return -EINVAL;
110
111 maxmem_pfn = (maxmem_mb >> (HPAGE_SHIFT - 20)) <<
112 (HPAGE_SHIFT - PAGE_SHIFT);
113 pr_info("Forcing RAM used to no more than %dMB\n",
114 maxmem_pfn >> (20 - PAGE_SHIFT));
115 return 0;
116 }
117 early_param("maxmem", setup_maxmem);
118
setup_maxnodemem(char * str)119 static int __init setup_maxnodemem(char *str)
120 {
121 char *endp;
122 long maxnodemem_mb, node;
123
124 node = str ? simple_strtoul(str, &endp, 0) : INT_MAX;
125 if (node >= MAX_NUMNODES || *endp != ':' ||
126 strict_strtol(endp+1, 0, &maxnodemem_mb) != 0)
127 return -EINVAL;
128
129 maxnodemem_pfn[node] = (maxnodemem_mb >> (HPAGE_SHIFT - 20)) <<
130 (HPAGE_SHIFT - PAGE_SHIFT);
131 pr_info("Forcing RAM used on node %ld to no more than %dMB\n",
132 node, maxnodemem_pfn[node] >> (20 - PAGE_SHIFT));
133 return 0;
134 }
135 early_param("maxnodemem", setup_maxnodemem);
136
setup_isolnodes(char * str)137 static int __init setup_isolnodes(char *str)
138 {
139 char buf[MAX_NUMNODES * 5];
140 if (str == NULL || nodelist_parse(str, isolnodes) != 0)
141 return -EINVAL;
142
143 nodelist_scnprintf(buf, sizeof(buf), isolnodes);
144 pr_info("Set isolnodes value to '%s'\n", buf);
145 return 0;
146 }
147 early_param("isolnodes", setup_isolnodes);
148
149 #ifdef CONFIG_PCI
setup_pci_reserve(char * str)150 static int __init setup_pci_reserve(char* str)
151 {
152 unsigned long mb;
153
154 if (str == NULL || strict_strtoul(str, 0, &mb) != 0 ||
155 mb > 3 * 1024)
156 return -EINVAL;
157
158 pci_reserve_mb = mb;
159 pr_info("Reserving %dMB for PCIE root complex mappings\n",
160 pci_reserve_mb);
161 return 0;
162 }
163 early_param("pci_reserve", setup_pci_reserve);
164 #endif
165
166 #ifndef __tilegx__
167 /*
168 * vmalloc=size forces the vmalloc area to be exactly 'size' bytes.
169 * This can be used to increase (or decrease) the vmalloc area.
170 */
parse_vmalloc(char * arg)171 static int __init parse_vmalloc(char *arg)
172 {
173 if (!arg)
174 return -EINVAL;
175
176 VMALLOC_RESERVE = (memparse(arg, &arg) + PGDIR_SIZE - 1) & PGDIR_MASK;
177
178 /* See validate_va() for more on this test. */
179 if ((long)_VMALLOC_START >= 0)
180 early_panic("\"vmalloc=%#lx\" value too large: maximum %#lx\n",
181 VMALLOC_RESERVE, _VMALLOC_END - 0x80000000UL);
182
183 return 0;
184 }
185 early_param("vmalloc", parse_vmalloc);
186 #endif
187
188 #ifdef CONFIG_HIGHMEM
189 /*
190 * Determine for each controller where its lowmem is mapped and how much of
191 * it is mapped there. On controller zero, the first few megabytes are
192 * already mapped in as code at MEM_SV_INTRPT, so in principle we could
193 * start our data mappings higher up, but for now we don't bother, to avoid
194 * additional confusion.
195 *
196 * One question is whether, on systems with more than 768 Mb and
197 * controllers of different sizes, to map in a proportionate amount of
198 * each one, or to try to map the same amount from each controller.
199 * (E.g. if we have three controllers with 256MB, 1GB, and 256MB
200 * respectively, do we map 256MB from each, or do we map 128 MB, 512
201 * MB, and 128 MB respectively?) For now we use a proportionate
202 * solution like the latter.
203 *
204 * The VA/PA mapping demands that we align our decisions at 16 MB
205 * boundaries so that we can rapidly convert VA to PA.
206 */
setup_pa_va_mapping(void)207 static void *__init setup_pa_va_mapping(void)
208 {
209 unsigned long curr_pages = 0;
210 unsigned long vaddr = PAGE_OFFSET;
211 nodemask_t highonlynodes = isolnodes;
212 int i, j;
213
214 memset(pbase_map, -1, sizeof(pbase_map));
215 memset(vbase_map, -1, sizeof(vbase_map));
216
217 /* Node zero cannot be isolated for LOWMEM purposes. */
218 node_clear(0, highonlynodes);
219
220 /* Count up the number of pages on non-highonlynodes controllers. */
221 mappable_physpages = 0;
222 for_each_online_node(i) {
223 if (!node_isset(i, highonlynodes))
224 mappable_physpages +=
225 node_end_pfn[i] - node_start_pfn[i];
226 }
227
228 for_each_online_node(i) {
229 unsigned long start = node_start_pfn[i];
230 unsigned long end = node_end_pfn[i];
231 unsigned long size = end - start;
232 unsigned long vaddr_end;
233
234 if (node_isset(i, highonlynodes)) {
235 /* Mark this controller as having no lowmem. */
236 node_lowmem_end_pfn[i] = start;
237 continue;
238 }
239
240 curr_pages += size;
241 if (mappable_physpages > MAXMEM_PFN) {
242 vaddr_end = PAGE_OFFSET +
243 (((u64)curr_pages * MAXMEM_PFN /
244 mappable_physpages)
245 << PAGE_SHIFT);
246 } else {
247 vaddr_end = PAGE_OFFSET + (curr_pages << PAGE_SHIFT);
248 }
249 for (j = 0; vaddr < vaddr_end; vaddr += HPAGE_SIZE, ++j) {
250 unsigned long this_pfn =
251 start + (j << HUGETLB_PAGE_ORDER);
252 pbase_map[vaddr >> HPAGE_SHIFT] = this_pfn;
253 if (vbase_map[__pfn_to_highbits(this_pfn)] ==
254 (void *)-1)
255 vbase_map[__pfn_to_highbits(this_pfn)] =
256 (void *)(vaddr & HPAGE_MASK);
257 }
258 node_lowmem_end_pfn[i] = start + (j << HUGETLB_PAGE_ORDER);
259 BUG_ON(node_lowmem_end_pfn[i] > end);
260 }
261
262 /* Return highest address of any mapped memory. */
263 return (void *)vaddr;
264 }
265 #endif /* CONFIG_HIGHMEM */
266
267 /*
268 * Register our most important memory mappings with the debug stub.
269 *
270 * This is up to 4 mappings for lowmem, one mapping per memory
271 * controller, plus one for our text segment.
272 */
store_permanent_mappings(void)273 static void __cpuinit store_permanent_mappings(void)
274 {
275 int i;
276
277 for_each_online_node(i) {
278 HV_PhysAddr pa = ((HV_PhysAddr)node_start_pfn[i]) << PAGE_SHIFT;
279 #ifdef CONFIG_HIGHMEM
280 HV_PhysAddr high_mapped_pa = node_lowmem_end_pfn[i];
281 #else
282 HV_PhysAddr high_mapped_pa = node_end_pfn[i];
283 #endif
284
285 unsigned long pages = high_mapped_pa - node_start_pfn[i];
286 HV_VirtAddr addr = (HV_VirtAddr) __va(pa);
287 hv_store_mapping(addr, pages << PAGE_SHIFT, pa);
288 }
289
290 hv_store_mapping((HV_VirtAddr)_stext,
291 (uint32_t)(_einittext - _stext), 0);
292 }
293
294 /*
295 * Use hv_inquire_physical() to populate node_{start,end}_pfn[]
296 * and node_online_map, doing suitable sanity-checking.
297 * Also set min_low_pfn, max_low_pfn, and max_pfn.
298 */
setup_memory(void)299 static void __init setup_memory(void)
300 {
301 int i, j;
302 int highbits_seen[NR_PA_HIGHBIT_VALUES] = { 0 };
303 #ifdef CONFIG_HIGHMEM
304 long highmem_pages;
305 #endif
306 #ifndef __tilegx__
307 int cap;
308 #endif
309 #if defined(CONFIG_HIGHMEM) || defined(__tilegx__)
310 long lowmem_pages;
311 #endif
312
313 /* We are using a char to hold the cpu_2_node[] mapping */
314 BUILD_BUG_ON(MAX_NUMNODES > 127);
315
316 /* Discover the ranges of memory available to us */
317 for (i = 0; ; ++i) {
318 unsigned long start, size, end, highbits;
319 HV_PhysAddrRange range = hv_inquire_physical(i);
320 if (range.size == 0)
321 break;
322 #ifdef CONFIG_FLATMEM
323 if (i > 0) {
324 pr_err("Can't use discontiguous PAs: %#llx..%#llx\n",
325 range.size, range.start + range.size);
326 continue;
327 }
328 #endif
329 #ifndef __tilegx__
330 if ((unsigned long)range.start) {
331 pr_err("Range not at 4GB multiple: %#llx..%#llx\n",
332 range.start, range.start + range.size);
333 continue;
334 }
335 #endif
336 if ((range.start & (HPAGE_SIZE-1)) != 0 ||
337 (range.size & (HPAGE_SIZE-1)) != 0) {
338 unsigned long long start_pa = range.start;
339 unsigned long long orig_size = range.size;
340 range.start = (start_pa + HPAGE_SIZE - 1) & HPAGE_MASK;
341 range.size -= (range.start - start_pa);
342 range.size &= HPAGE_MASK;
343 pr_err("Range not hugepage-aligned: %#llx..%#llx:"
344 " now %#llx-%#llx\n",
345 start_pa, start_pa + orig_size,
346 range.start, range.start + range.size);
347 }
348 highbits = __pa_to_highbits(range.start);
349 if (highbits >= NR_PA_HIGHBIT_VALUES) {
350 pr_err("PA high bits too high: %#llx..%#llx\n",
351 range.start, range.start + range.size);
352 continue;
353 }
354 if (highbits_seen[highbits]) {
355 pr_err("Range overlaps in high bits: %#llx..%#llx\n",
356 range.start, range.start + range.size);
357 continue;
358 }
359 highbits_seen[highbits] = 1;
360 if (PFN_DOWN(range.size) > maxnodemem_pfn[i]) {
361 int max_size = maxnodemem_pfn[i];
362 if (max_size > 0) {
363 pr_err("Maxnodemem reduced node %d to"
364 " %d pages\n", i, max_size);
365 range.size = PFN_PHYS(max_size);
366 } else {
367 pr_err("Maxnodemem disabled node %d\n", i);
368 continue;
369 }
370 }
371 if (num_physpages + PFN_DOWN(range.size) > maxmem_pfn) {
372 int max_size = maxmem_pfn - num_physpages;
373 if (max_size > 0) {
374 pr_err("Maxmem reduced node %d to %d pages\n",
375 i, max_size);
376 range.size = PFN_PHYS(max_size);
377 } else {
378 pr_err("Maxmem disabled node %d\n", i);
379 continue;
380 }
381 }
382 if (i >= MAX_NUMNODES) {
383 pr_err("Too many PA nodes (#%d): %#llx...%#llx\n",
384 i, range.size, range.size + range.start);
385 continue;
386 }
387
388 start = range.start >> PAGE_SHIFT;
389 size = range.size >> PAGE_SHIFT;
390 end = start + size;
391
392 #ifndef __tilegx__
393 if (((HV_PhysAddr)end << PAGE_SHIFT) !=
394 (range.start + range.size)) {
395 pr_err("PAs too high to represent: %#llx..%#llx\n",
396 range.start, range.start + range.size);
397 continue;
398 }
399 #endif
400 #ifdef CONFIG_PCI
401 /*
402 * Blocks that overlap the pci reserved region must
403 * have enough space to hold the maximum percpu data
404 * region at the top of the range. If there isn't
405 * enough space above the reserved region, just
406 * truncate the node.
407 */
408 if (start <= pci_reserve_start_pfn &&
409 end > pci_reserve_start_pfn) {
410 unsigned int per_cpu_size =
411 __per_cpu_end - __per_cpu_start;
412 unsigned int percpu_pages =
413 NR_CPUS * (PFN_UP(per_cpu_size) >> PAGE_SHIFT);
414 if (end < pci_reserve_end_pfn + percpu_pages) {
415 end = pci_reserve_start_pfn;
416 pr_err("PCI mapping region reduced node %d to"
417 " %ld pages\n", i, end - start);
418 }
419 }
420 #endif
421
422 for (j = __pfn_to_highbits(start);
423 j <= __pfn_to_highbits(end - 1); j++)
424 highbits_to_node[j] = i;
425
426 node_start_pfn[i] = start;
427 node_end_pfn[i] = end;
428 node_controller[i] = range.controller;
429 num_physpages += size;
430 max_pfn = end;
431
432 /* Mark node as online */
433 node_set(i, node_online_map);
434 node_set(i, node_possible_map);
435 }
436
437 #ifndef __tilegx__
438 /*
439 * For 4KB pages, mem_map "struct page" data is 1% of the size
440 * of the physical memory, so can be quite big (640 MB for
441 * four 16G zones). These structures must be mapped in
442 * lowmem, and since we currently cap out at about 768 MB,
443 * it's impractical to try to use this much address space.
444 * For now, arbitrarily cap the amount of physical memory
445 * we're willing to use at 8 million pages (32GB of 4KB pages).
446 */
447 cap = 8 * 1024 * 1024; /* 8 million pages */
448 if (num_physpages > cap) {
449 int num_nodes = num_online_nodes();
450 int cap_each = cap / num_nodes;
451 unsigned long dropped_pages = 0;
452 for (i = 0; i < num_nodes; ++i) {
453 int size = node_end_pfn[i] - node_start_pfn[i];
454 if (size > cap_each) {
455 dropped_pages += (size - cap_each);
456 node_end_pfn[i] = node_start_pfn[i] + cap_each;
457 }
458 }
459 num_physpages -= dropped_pages;
460 pr_warning("Only using %ldMB memory;"
461 " ignoring %ldMB.\n",
462 num_physpages >> (20 - PAGE_SHIFT),
463 dropped_pages >> (20 - PAGE_SHIFT));
464 pr_warning("Consider using a larger page size.\n");
465 }
466 #endif
467
468 /* Heap starts just above the last loaded address. */
469 min_low_pfn = PFN_UP((unsigned long)_end - PAGE_OFFSET);
470
471 #ifdef CONFIG_HIGHMEM
472 /* Find where we map lowmem from each controller. */
473 high_memory = setup_pa_va_mapping();
474
475 /* Set max_low_pfn based on what node 0 can directly address. */
476 max_low_pfn = node_lowmem_end_pfn[0];
477
478 lowmem_pages = (mappable_physpages > MAXMEM_PFN) ?
479 MAXMEM_PFN : mappable_physpages;
480 highmem_pages = (long) (num_physpages - lowmem_pages);
481
482 pr_notice("%ldMB HIGHMEM available.\n",
483 pages_to_mb(highmem_pages > 0 ? highmem_pages : 0));
484 pr_notice("%ldMB LOWMEM available.\n",
485 pages_to_mb(lowmem_pages));
486 #else
487 /* Set max_low_pfn based on what node 0 can directly address. */
488 max_low_pfn = node_end_pfn[0];
489
490 #ifndef __tilegx__
491 if (node_end_pfn[0] > MAXMEM_PFN) {
492 pr_warning("Only using %ldMB LOWMEM.\n",
493 MAXMEM>>20);
494 pr_warning("Use a HIGHMEM enabled kernel.\n");
495 max_low_pfn = MAXMEM_PFN;
496 max_pfn = MAXMEM_PFN;
497 num_physpages = MAXMEM_PFN;
498 node_end_pfn[0] = MAXMEM_PFN;
499 } else {
500 pr_notice("%ldMB memory available.\n",
501 pages_to_mb(node_end_pfn[0]));
502 }
503 for (i = 1; i < MAX_NUMNODES; ++i) {
504 node_start_pfn[i] = 0;
505 node_end_pfn[i] = 0;
506 }
507 high_memory = __va(node_end_pfn[0]);
508 #else
509 lowmem_pages = 0;
510 for (i = 0; i < MAX_NUMNODES; ++i) {
511 int pages = node_end_pfn[i] - node_start_pfn[i];
512 lowmem_pages += pages;
513 if (pages)
514 high_memory = pfn_to_kaddr(node_end_pfn[i]);
515 }
516 pr_notice("%ldMB memory available.\n",
517 pages_to_mb(lowmem_pages));
518 #endif
519 #endif
520 }
521
setup_bootmem_allocator(void)522 static void __init setup_bootmem_allocator(void)
523 {
524 unsigned long bootmap_size, first_alloc_pfn, last_alloc_pfn;
525
526 /* Provide a node 0 bdata. */
527 NODE_DATA(0)->bdata = &node0_bdata;
528
529 #ifdef CONFIG_PCI
530 /* Don't let boot memory alias the PCI region. */
531 last_alloc_pfn = min(max_low_pfn, pci_reserve_start_pfn);
532 #else
533 last_alloc_pfn = max_low_pfn;
534 #endif
535
536 /*
537 * Initialize the boot-time allocator (with low memory only):
538 * The first argument says where to put the bitmap, and the
539 * second says where the end of allocatable memory is.
540 */
541 bootmap_size = init_bootmem(min_low_pfn, last_alloc_pfn);
542
543 /*
544 * Let the bootmem allocator use all the space we've given it
545 * except for its own bitmap.
546 */
547 first_alloc_pfn = min_low_pfn + PFN_UP(bootmap_size);
548 if (first_alloc_pfn >= last_alloc_pfn)
549 early_panic("Not enough memory on controller 0 for bootmem\n");
550
551 free_bootmem(PFN_PHYS(first_alloc_pfn),
552 PFN_PHYS(last_alloc_pfn - first_alloc_pfn));
553
554 #ifdef CONFIG_KEXEC
555 if (crashk_res.start != crashk_res.end)
556 reserve_bootmem(crashk_res.start,
557 crashk_res.end - crashk_res.start + 1, 0);
558 #endif
559 }
560
alloc_remap(int nid,unsigned long size)561 void *__init alloc_remap(int nid, unsigned long size)
562 {
563 int pages = node_end_pfn[nid] - node_start_pfn[nid];
564 void *map = pfn_to_kaddr(node_memmap_pfn[nid]);
565 BUG_ON(size != pages * sizeof(struct page));
566 memset(map, 0, size);
567 return map;
568 }
569
percpu_size(void)570 static int __init percpu_size(void)
571 {
572 int size = __per_cpu_end - __per_cpu_start;
573 size += PERCPU_MODULE_RESERVE;
574 size += PERCPU_DYNAMIC_EARLY_SIZE;
575 if (size < PCPU_MIN_UNIT_SIZE)
576 size = PCPU_MIN_UNIT_SIZE;
577 size = roundup(size, PAGE_SIZE);
578
579 /* In several places we assume the per-cpu data fits on a huge page. */
580 BUG_ON(kdata_huge && size > HPAGE_SIZE);
581 return size;
582 }
583
alloc_bootmem_pfn(int size,unsigned long goal)584 static inline unsigned long alloc_bootmem_pfn(int size, unsigned long goal)
585 {
586 void *kva = __alloc_bootmem(size, PAGE_SIZE, goal);
587 unsigned long pfn = kaddr_to_pfn(kva);
588 BUG_ON(goal && PFN_PHYS(pfn) != goal);
589 return pfn;
590 }
591
zone_sizes_init(void)592 static void __init zone_sizes_init(void)
593 {
594 unsigned long zones_size[MAX_NR_ZONES] = { 0 };
595 int size = percpu_size();
596 int num_cpus = smp_height * smp_width;
597 int i;
598
599 for (i = 0; i < num_cpus; ++i)
600 node_percpu[cpu_to_node(i)] += size;
601
602 for_each_online_node(i) {
603 unsigned long start = node_start_pfn[i];
604 unsigned long end = node_end_pfn[i];
605 #ifdef CONFIG_HIGHMEM
606 unsigned long lowmem_end = node_lowmem_end_pfn[i];
607 #else
608 unsigned long lowmem_end = end;
609 #endif
610 int memmap_size = (end - start) * sizeof(struct page);
611 node_free_pfn[i] = start;
612
613 /*
614 * Set aside pages for per-cpu data and the mem_map array.
615 *
616 * Since the per-cpu data requires special homecaching,
617 * if we are in kdata_huge mode, we put it at the end of
618 * the lowmem region. If we're not in kdata_huge mode,
619 * we take the per-cpu pages from the bottom of the
620 * controller, since that avoids fragmenting a huge page
621 * that users might want. We always take the memmap
622 * from the bottom of the controller, since with
623 * kdata_huge that lets it be under a huge TLB entry.
624 *
625 * If the user has requested isolnodes for a controller,
626 * though, there'll be no lowmem, so we just alloc_bootmem
627 * the memmap. There will be no percpu memory either.
628 */
629 if (__pfn_to_highbits(start) == 0) {
630 /* In low PAs, allocate via bootmem. */
631 unsigned long goal = 0;
632 node_memmap_pfn[i] =
633 alloc_bootmem_pfn(memmap_size, goal);
634 if (kdata_huge)
635 goal = PFN_PHYS(lowmem_end) - node_percpu[i];
636 if (node_percpu[i])
637 node_percpu_pfn[i] =
638 alloc_bootmem_pfn(node_percpu[i], goal);
639 } else if (cpu_isset(i, isolnodes)) {
640 node_memmap_pfn[i] = alloc_bootmem_pfn(memmap_size, 0);
641 BUG_ON(node_percpu[i] != 0);
642 } else {
643 /* In high PAs, just reserve some pages. */
644 node_memmap_pfn[i] = node_free_pfn[i];
645 node_free_pfn[i] += PFN_UP(memmap_size);
646 if (!kdata_huge) {
647 node_percpu_pfn[i] = node_free_pfn[i];
648 node_free_pfn[i] += PFN_UP(node_percpu[i]);
649 } else {
650 node_percpu_pfn[i] =
651 lowmem_end - PFN_UP(node_percpu[i]);
652 }
653 }
654
655 #ifdef CONFIG_HIGHMEM
656 if (start > lowmem_end) {
657 zones_size[ZONE_NORMAL] = 0;
658 zones_size[ZONE_HIGHMEM] = end - start;
659 } else {
660 zones_size[ZONE_NORMAL] = lowmem_end - start;
661 zones_size[ZONE_HIGHMEM] = end - lowmem_end;
662 }
663 #else
664 zones_size[ZONE_NORMAL] = end - start;
665 #endif
666
667 /*
668 * Everyone shares node 0's bootmem allocator, but
669 * we use alloc_remap(), above, to put the actual
670 * struct page array on the individual controllers,
671 * which is most of the data that we actually care about.
672 * We can't place bootmem allocators on the other
673 * controllers since the bootmem allocator can only
674 * operate on 32-bit physical addresses.
675 */
676 NODE_DATA(i)->bdata = NODE_DATA(0)->bdata;
677
678 free_area_init_node(i, zones_size, start, NULL);
679 printk(KERN_DEBUG " Normal zone: %ld per-cpu pages\n",
680 PFN_UP(node_percpu[i]));
681
682 /* Track the type of memory on each node */
683 if (zones_size[ZONE_NORMAL])
684 node_set_state(i, N_NORMAL_MEMORY);
685 #ifdef CONFIG_HIGHMEM
686 if (end != start)
687 node_set_state(i, N_HIGH_MEMORY);
688 #endif
689
690 node_set_online(i);
691 }
692 }
693
694 #ifdef CONFIG_NUMA
695
696 /* which logical CPUs are on which nodes */
697 struct cpumask node_2_cpu_mask[MAX_NUMNODES] __write_once;
698 EXPORT_SYMBOL(node_2_cpu_mask);
699
700 /* which node each logical CPU is on */
701 char cpu_2_node[NR_CPUS] __write_once __attribute__((aligned(L2_CACHE_BYTES)));
702 EXPORT_SYMBOL(cpu_2_node);
703
704 /* Return cpu_to_node() except for cpus not yet assigned, which return -1 */
cpu_to_bound_node(int cpu,struct cpumask * unbound_cpus)705 static int __init cpu_to_bound_node(int cpu, struct cpumask* unbound_cpus)
706 {
707 if (!cpu_possible(cpu) || cpumask_test_cpu(cpu, unbound_cpus))
708 return -1;
709 else
710 return cpu_to_node(cpu);
711 }
712
713 /* Return number of immediately-adjacent tiles sharing the same NUMA node. */
node_neighbors(int node,int cpu,struct cpumask * unbound_cpus)714 static int __init node_neighbors(int node, int cpu,
715 struct cpumask *unbound_cpus)
716 {
717 int neighbors = 0;
718 int w = smp_width;
719 int h = smp_height;
720 int x = cpu % w;
721 int y = cpu / w;
722 if (x > 0 && cpu_to_bound_node(cpu-1, unbound_cpus) == node)
723 ++neighbors;
724 if (x < w-1 && cpu_to_bound_node(cpu+1, unbound_cpus) == node)
725 ++neighbors;
726 if (y > 0 && cpu_to_bound_node(cpu-w, unbound_cpus) == node)
727 ++neighbors;
728 if (y < h-1 && cpu_to_bound_node(cpu+w, unbound_cpus) == node)
729 ++neighbors;
730 return neighbors;
731 }
732
setup_numa_mapping(void)733 static void __init setup_numa_mapping(void)
734 {
735 int distance[MAX_NUMNODES][NR_CPUS];
736 HV_Coord coord;
737 int cpu, node, cpus, i, x, y;
738 int num_nodes = num_online_nodes();
739 struct cpumask unbound_cpus;
740 nodemask_t default_nodes;
741
742 cpumask_clear(&unbound_cpus);
743
744 /* Get set of nodes we will use for defaults */
745 nodes_andnot(default_nodes, node_online_map, isolnodes);
746 if (nodes_empty(default_nodes)) {
747 BUG_ON(!node_isset(0, node_online_map));
748 pr_err("Forcing NUMA node zero available as a default node\n");
749 node_set(0, default_nodes);
750 }
751
752 /* Populate the distance[] array */
753 memset(distance, -1, sizeof(distance));
754 cpu = 0;
755 for (coord.y = 0; coord.y < smp_height; ++coord.y) {
756 for (coord.x = 0; coord.x < smp_width;
757 ++coord.x, ++cpu) {
758 BUG_ON(cpu >= nr_cpu_ids);
759 if (!cpu_possible(cpu)) {
760 cpu_2_node[cpu] = -1;
761 continue;
762 }
763 for_each_node_mask(node, default_nodes) {
764 HV_MemoryControllerInfo info =
765 hv_inquire_memory_controller(
766 coord, node_controller[node]);
767 distance[node][cpu] =
768 ABS(info.coord.x) + ABS(info.coord.y);
769 }
770 cpumask_set_cpu(cpu, &unbound_cpus);
771 }
772 }
773 cpus = cpu;
774
775 /*
776 * Round-robin through the NUMA nodes until all the cpus are
777 * assigned. We could be more clever here (e.g. create four
778 * sorted linked lists on the same set of cpu nodes, and pull
779 * off them in round-robin sequence, removing from all four
780 * lists each time) but given the relatively small numbers
781 * involved, O(n^2) seem OK for a one-time cost.
782 */
783 node = first_node(default_nodes);
784 while (!cpumask_empty(&unbound_cpus)) {
785 int best_cpu = -1;
786 int best_distance = INT_MAX;
787 for (cpu = 0; cpu < cpus; ++cpu) {
788 if (cpumask_test_cpu(cpu, &unbound_cpus)) {
789 /*
790 * Compute metric, which is how much
791 * closer the cpu is to this memory
792 * controller than the others, shifted
793 * up, and then the number of
794 * neighbors already in the node as an
795 * epsilon adjustment to try to keep
796 * the nodes compact.
797 */
798 int d = distance[node][cpu] * num_nodes;
799 for_each_node_mask(i, default_nodes) {
800 if (i != node)
801 d -= distance[i][cpu];
802 }
803 d *= 8; /* allow space for epsilon */
804 d -= node_neighbors(node, cpu, &unbound_cpus);
805 if (d < best_distance) {
806 best_cpu = cpu;
807 best_distance = d;
808 }
809 }
810 }
811 BUG_ON(best_cpu < 0);
812 cpumask_set_cpu(best_cpu, &node_2_cpu_mask[node]);
813 cpu_2_node[best_cpu] = node;
814 cpumask_clear_cpu(best_cpu, &unbound_cpus);
815 node = next_node(node, default_nodes);
816 if (node == MAX_NUMNODES)
817 node = first_node(default_nodes);
818 }
819
820 /* Print out node assignments and set defaults for disabled cpus */
821 cpu = 0;
822 for (y = 0; y < smp_height; ++y) {
823 printk(KERN_DEBUG "NUMA cpu-to-node row %d:", y);
824 for (x = 0; x < smp_width; ++x, ++cpu) {
825 if (cpu_to_node(cpu) < 0) {
826 pr_cont(" -");
827 cpu_2_node[cpu] = first_node(default_nodes);
828 } else {
829 pr_cont(" %d", cpu_to_node(cpu));
830 }
831 }
832 pr_cont("\n");
833 }
834 }
835
836 static struct cpu cpu_devices[NR_CPUS];
837
topology_init(void)838 static int __init topology_init(void)
839 {
840 int i;
841
842 for_each_online_node(i)
843 register_one_node(i);
844
845 for (i = 0; i < smp_height * smp_width; ++i)
846 register_cpu(&cpu_devices[i], i);
847
848 return 0;
849 }
850
851 subsys_initcall(topology_init);
852
853 #else /* !CONFIG_NUMA */
854
855 #define setup_numa_mapping() do { } while (0)
856
857 #endif /* CONFIG_NUMA */
858
859 /**
860 * setup_cpu() - Do all necessary per-cpu, tile-specific initialization.
861 * @boot: Is this the boot cpu?
862 *
863 * Called from setup_arch() on the boot cpu, or online_secondary().
864 */
setup_cpu(int boot)865 void __cpuinit setup_cpu(int boot)
866 {
867 /* The boot cpu sets up its permanent mappings much earlier. */
868 if (!boot)
869 store_permanent_mappings();
870
871 /* Allow asynchronous TLB interrupts. */
872 #if CHIP_HAS_TILE_DMA()
873 arch_local_irq_unmask(INT_DMATLB_MISS);
874 arch_local_irq_unmask(INT_DMATLB_ACCESS);
875 #endif
876 #if CHIP_HAS_SN_PROC()
877 arch_local_irq_unmask(INT_SNITLB_MISS);
878 #endif
879 #ifdef __tilegx__
880 arch_local_irq_unmask(INT_SINGLE_STEP_K);
881 #endif
882
883 /*
884 * Allow user access to many generic SPRs, like the cycle
885 * counter, PASS/FAIL/DONE, INTERRUPT_CRITICAL_SECTION, etc.
886 */
887 __insn_mtspr(SPR_MPL_WORLD_ACCESS_SET_0, 1);
888
889 #if CHIP_HAS_SN()
890 /* Static network is not restricted. */
891 __insn_mtspr(SPR_MPL_SN_ACCESS_SET_0, 1);
892 #endif
893 #if CHIP_HAS_SN_PROC()
894 __insn_mtspr(SPR_MPL_SN_NOTIFY_SET_0, 1);
895 __insn_mtspr(SPR_MPL_SN_CPL_SET_0, 1);
896 #endif
897
898 /*
899 * Set the MPL for interrupt control 0 & 1 to the corresponding
900 * values. This includes access to the SYSTEM_SAVE and EX_CONTEXT
901 * SPRs, as well as the interrupt mask.
902 */
903 __insn_mtspr(SPR_MPL_INTCTRL_0_SET_0, 1);
904 __insn_mtspr(SPR_MPL_INTCTRL_1_SET_1, 1);
905
906 /* Initialize IRQ support for this cpu. */
907 setup_irq_regs();
908
909 #ifdef CONFIG_HARDWALL
910 /* Reset the network state on this cpu. */
911 reset_network_state();
912 #endif
913 }
914
915 static int __initdata set_initramfs_file;
916 static char __initdata initramfs_file[128] = "initramfs.cpio.gz";
917
setup_initramfs_file(char * str)918 static int __init setup_initramfs_file(char *str)
919 {
920 if (str == NULL)
921 return -EINVAL;
922 strncpy(initramfs_file, str, sizeof(initramfs_file) - 1);
923 set_initramfs_file = 1;
924
925 return 0;
926 }
927 early_param("initramfs_file", setup_initramfs_file);
928
929 /*
930 * We look for an additional "initramfs.cpio.gz" file in the hvfs.
931 * If there is one, we allocate some memory for it and it will be
932 * unpacked to the initramfs after any built-in initramfs_data.
933 */
load_hv_initrd(void)934 static void __init load_hv_initrd(void)
935 {
936 HV_FS_StatInfo stat;
937 int fd, rc;
938 void *initrd;
939
940 fd = hv_fs_findfile((HV_VirtAddr) initramfs_file);
941 if (fd == HV_ENOENT) {
942 if (set_initramfs_file)
943 pr_warning("No such hvfs initramfs file '%s'\n",
944 initramfs_file);
945 return;
946 }
947 BUG_ON(fd < 0);
948 stat = hv_fs_fstat(fd);
949 BUG_ON(stat.size < 0);
950 if (stat.flags & HV_FS_ISDIR) {
951 pr_warning("Ignoring hvfs file '%s': it's a directory.\n",
952 initramfs_file);
953 return;
954 }
955 initrd = alloc_bootmem_pages(stat.size);
956 rc = hv_fs_pread(fd, (HV_VirtAddr) initrd, stat.size, 0);
957 if (rc != stat.size) {
958 pr_err("Error reading %d bytes from hvfs file '%s': %d\n",
959 stat.size, initramfs_file, rc);
960 free_initrd_mem((unsigned long) initrd, stat.size);
961 return;
962 }
963 initrd_start = (unsigned long) initrd;
964 initrd_end = initrd_start + stat.size;
965 }
966
free_initrd_mem(unsigned long begin,unsigned long end)967 void __init free_initrd_mem(unsigned long begin, unsigned long end)
968 {
969 free_bootmem(__pa(begin), end - begin);
970 }
971
validate_hv(void)972 static void __init validate_hv(void)
973 {
974 /*
975 * It may already be too late, but let's check our built-in
976 * configuration against what the hypervisor is providing.
977 */
978 unsigned long glue_size = hv_sysconf(HV_SYSCONF_GLUE_SIZE);
979 int hv_page_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_SMALL);
980 int hv_hpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_LARGE);
981 HV_ASIDRange asid_range;
982
983 #ifndef CONFIG_SMP
984 HV_Topology topology = hv_inquire_topology();
985 BUG_ON(topology.coord.x != 0 || topology.coord.y != 0);
986 if (topology.width != 1 || topology.height != 1) {
987 pr_warning("Warning: booting UP kernel on %dx%d grid;"
988 " will ignore all but first tile.\n",
989 topology.width, topology.height);
990 }
991 #endif
992
993 if (PAGE_OFFSET + HV_GLUE_START_CPA + glue_size > (unsigned long)_text)
994 early_panic("Hypervisor glue size %ld is too big!\n",
995 glue_size);
996 if (hv_page_size != PAGE_SIZE)
997 early_panic("Hypervisor page size %#x != our %#lx\n",
998 hv_page_size, PAGE_SIZE);
999 if (hv_hpage_size != HPAGE_SIZE)
1000 early_panic("Hypervisor huge page size %#x != our %#lx\n",
1001 hv_hpage_size, HPAGE_SIZE);
1002
1003 #ifdef CONFIG_SMP
1004 /*
1005 * Some hypervisor APIs take a pointer to a bitmap array
1006 * whose size is at least the number of cpus on the chip.
1007 * We use a struct cpumask for this, so it must be big enough.
1008 */
1009 if ((smp_height * smp_width) > nr_cpu_ids)
1010 early_panic("Hypervisor %d x %d grid too big for Linux"
1011 " NR_CPUS %d\n", smp_height, smp_width,
1012 nr_cpu_ids);
1013 #endif
1014
1015 /*
1016 * Check that we're using allowed ASIDs, and initialize the
1017 * various asid variables to their appropriate initial states.
1018 */
1019 asid_range = hv_inquire_asid(0);
1020 __get_cpu_var(current_asid) = min_asid = asid_range.start;
1021 max_asid = asid_range.start + asid_range.size - 1;
1022
1023 if (hv_confstr(HV_CONFSTR_CHIP_MODEL, (HV_VirtAddr)chip_model,
1024 sizeof(chip_model)) < 0) {
1025 pr_err("Warning: HV_CONFSTR_CHIP_MODEL not available\n");
1026 strlcpy(chip_model, "unknown", sizeof(chip_model));
1027 }
1028 }
1029
validate_va(void)1030 static void __init validate_va(void)
1031 {
1032 #ifndef __tilegx__ /* FIXME: GX: probably some validation relevant here */
1033 /*
1034 * Similarly, make sure we're only using allowed VAs.
1035 * We assume we can contiguously use MEM_USER_INTRPT .. MEM_HV_INTRPT,
1036 * and 0 .. KERNEL_HIGH_VADDR.
1037 * In addition, make sure we CAN'T use the end of memory, since
1038 * we use the last chunk of each pgd for the pgd_list.
1039 */
1040 int i, user_kernel_ok = 0;
1041 unsigned long max_va = 0;
1042 unsigned long list_va =
1043 ((PGD_LIST_OFFSET / sizeof(pgd_t)) << PGDIR_SHIFT);
1044
1045 for (i = 0; ; ++i) {
1046 HV_VirtAddrRange range = hv_inquire_virtual(i);
1047 if (range.size == 0)
1048 break;
1049 if (range.start <= MEM_USER_INTRPT &&
1050 range.start + range.size >= MEM_HV_INTRPT)
1051 user_kernel_ok = 1;
1052 if (range.start == 0)
1053 max_va = range.size;
1054 BUG_ON(range.start + range.size > list_va);
1055 }
1056 if (!user_kernel_ok)
1057 early_panic("Hypervisor not configured for user/kernel VAs\n");
1058 if (max_va == 0)
1059 early_panic("Hypervisor not configured for low VAs\n");
1060 if (max_va < KERNEL_HIGH_VADDR)
1061 early_panic("Hypervisor max VA %#lx smaller than %#lx\n",
1062 max_va, KERNEL_HIGH_VADDR);
1063
1064 /* Kernel PCs must have their high bit set; see intvec.S. */
1065 if ((long)VMALLOC_START >= 0)
1066 early_panic(
1067 "Linux VMALLOC region below the 2GB line (%#lx)!\n"
1068 "Reconfigure the kernel with fewer NR_HUGE_VMAPS\n"
1069 "or smaller VMALLOC_RESERVE.\n",
1070 VMALLOC_START);
1071 #endif
1072 }
1073
1074 /*
1075 * cpu_lotar_map lists all the cpus that are valid for the supervisor
1076 * to cache data on at a page level, i.e. what cpus can be placed in
1077 * the LOTAR field of a PTE. It is equivalent to the set of possible
1078 * cpus plus any other cpus that are willing to share their cache.
1079 * It is set by hv_inquire_tiles(HV_INQ_TILES_LOTAR).
1080 */
1081 struct cpumask __write_once cpu_lotar_map;
1082 EXPORT_SYMBOL(cpu_lotar_map);
1083
1084 #if CHIP_HAS_CBOX_HOME_MAP()
1085 /*
1086 * hash_for_home_map lists all the tiles that hash-for-home data
1087 * will be cached on. Note that this may includes tiles that are not
1088 * valid for this supervisor to use otherwise (e.g. if a hypervisor
1089 * device is being shared between multiple supervisors).
1090 * It is set by hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE).
1091 */
1092 struct cpumask hash_for_home_map;
1093 EXPORT_SYMBOL(hash_for_home_map);
1094 #endif
1095
1096 /*
1097 * cpu_cacheable_map lists all the cpus whose caches the hypervisor can
1098 * flush on our behalf. It is set to cpu_possible_map OR'ed with
1099 * hash_for_home_map, and it is what should be passed to
1100 * hv_flush_remote() to flush all caches. Note that if there are
1101 * dedicated hypervisor driver tiles that have authorized use of their
1102 * cache, those tiles will only appear in cpu_lotar_map, NOT in
1103 * cpu_cacheable_map, as they are a special case.
1104 */
1105 struct cpumask __write_once cpu_cacheable_map;
1106 EXPORT_SYMBOL(cpu_cacheable_map);
1107
1108 static __initdata struct cpumask disabled_map;
1109
disabled_cpus(char * str)1110 static int __init disabled_cpus(char *str)
1111 {
1112 int boot_cpu = smp_processor_id();
1113
1114 if (str == NULL || cpulist_parse_crop(str, &disabled_map) != 0)
1115 return -EINVAL;
1116 if (cpumask_test_cpu(boot_cpu, &disabled_map)) {
1117 pr_err("disabled_cpus: can't disable boot cpu %d\n", boot_cpu);
1118 cpumask_clear_cpu(boot_cpu, &disabled_map);
1119 }
1120 return 0;
1121 }
1122
1123 early_param("disabled_cpus", disabled_cpus);
1124
print_disabled_cpus(void)1125 void __init print_disabled_cpus(void)
1126 {
1127 if (!cpumask_empty(&disabled_map)) {
1128 char buf[100];
1129 cpulist_scnprintf(buf, sizeof(buf), &disabled_map);
1130 pr_info("CPUs not available for Linux: %s\n", buf);
1131 }
1132 }
1133
setup_cpu_maps(void)1134 static void __init setup_cpu_maps(void)
1135 {
1136 struct cpumask hv_disabled_map, cpu_possible_init;
1137 int boot_cpu = smp_processor_id();
1138 int cpus, i, rc;
1139
1140 /* Learn which cpus are allowed by the hypervisor. */
1141 rc = hv_inquire_tiles(HV_INQ_TILES_AVAIL,
1142 (HV_VirtAddr) cpumask_bits(&cpu_possible_init),
1143 sizeof(cpu_cacheable_map));
1144 if (rc < 0)
1145 early_panic("hv_inquire_tiles(AVAIL) failed: rc %d\n", rc);
1146 if (!cpumask_test_cpu(boot_cpu, &cpu_possible_init))
1147 early_panic("Boot CPU %d disabled by hypervisor!\n", boot_cpu);
1148
1149 /* Compute the cpus disabled by the hvconfig file. */
1150 cpumask_complement(&hv_disabled_map, &cpu_possible_init);
1151
1152 /* Include them with the cpus disabled by "disabled_cpus". */
1153 cpumask_or(&disabled_map, &disabled_map, &hv_disabled_map);
1154
1155 /*
1156 * Disable every cpu after "setup_max_cpus". But don't mark
1157 * as disabled the cpus that are outside of our initial rectangle,
1158 * since that turns out to be confusing.
1159 */
1160 cpus = 1; /* this cpu */
1161 cpumask_set_cpu(boot_cpu, &disabled_map); /* ignore this cpu */
1162 for (i = 0; cpus < setup_max_cpus; ++i)
1163 if (!cpumask_test_cpu(i, &disabled_map))
1164 ++cpus;
1165 for (; i < smp_height * smp_width; ++i)
1166 cpumask_set_cpu(i, &disabled_map);
1167 cpumask_clear_cpu(boot_cpu, &disabled_map); /* reset this cpu */
1168 for (i = smp_height * smp_width; i < NR_CPUS; ++i)
1169 cpumask_clear_cpu(i, &disabled_map);
1170
1171 /*
1172 * Setup cpu_possible map as every cpu allocated to us, minus
1173 * the results of any "disabled_cpus" settings.
1174 */
1175 cpumask_andnot(&cpu_possible_init, &cpu_possible_init, &disabled_map);
1176 init_cpu_possible(&cpu_possible_init);
1177
1178 /* Learn which cpus are valid for LOTAR caching. */
1179 rc = hv_inquire_tiles(HV_INQ_TILES_LOTAR,
1180 (HV_VirtAddr) cpumask_bits(&cpu_lotar_map),
1181 sizeof(cpu_lotar_map));
1182 if (rc < 0) {
1183 pr_err("warning: no HV_INQ_TILES_LOTAR; using AVAIL\n");
1184 cpu_lotar_map = cpu_possible_map;
1185 }
1186
1187 #if CHIP_HAS_CBOX_HOME_MAP()
1188 /* Retrieve set of CPUs used for hash-for-home caching */
1189 rc = hv_inquire_tiles(HV_INQ_TILES_HFH_CACHE,
1190 (HV_VirtAddr) hash_for_home_map.bits,
1191 sizeof(hash_for_home_map));
1192 if (rc < 0)
1193 early_panic("hv_inquire_tiles(HFH_CACHE) failed: rc %d\n", rc);
1194 cpumask_or(&cpu_cacheable_map, &cpu_possible_map, &hash_for_home_map);
1195 #else
1196 cpu_cacheable_map = cpu_possible_map;
1197 #endif
1198 }
1199
1200
dataplane(char * str)1201 static int __init dataplane(char *str)
1202 {
1203 pr_warning("WARNING: dataplane support disabled in this kernel\n");
1204 return 0;
1205 }
1206
1207 early_param("dataplane", dataplane);
1208
1209 #ifdef CONFIG_CMDLINE_BOOL
1210 static char __initdata builtin_cmdline[COMMAND_LINE_SIZE] = CONFIG_CMDLINE;
1211 #endif
1212
setup_arch(char ** cmdline_p)1213 void __init setup_arch(char **cmdline_p)
1214 {
1215 int len;
1216
1217 #if defined(CONFIG_CMDLINE_BOOL) && defined(CONFIG_CMDLINE_OVERRIDE)
1218 len = hv_get_command_line((HV_VirtAddr) boot_command_line,
1219 COMMAND_LINE_SIZE);
1220 if (boot_command_line[0])
1221 pr_warning("WARNING: ignoring dynamic command line \"%s\"\n",
1222 boot_command_line);
1223 strlcpy(boot_command_line, builtin_cmdline, COMMAND_LINE_SIZE);
1224 #else
1225 char *hv_cmdline;
1226 #if defined(CONFIG_CMDLINE_BOOL)
1227 if (builtin_cmdline[0]) {
1228 int builtin_len = strlcpy(boot_command_line, builtin_cmdline,
1229 COMMAND_LINE_SIZE);
1230 if (builtin_len < COMMAND_LINE_SIZE-1)
1231 boot_command_line[builtin_len++] = ' ';
1232 hv_cmdline = &boot_command_line[builtin_len];
1233 len = COMMAND_LINE_SIZE - builtin_len;
1234 } else
1235 #endif
1236 {
1237 hv_cmdline = boot_command_line;
1238 len = COMMAND_LINE_SIZE;
1239 }
1240 len = hv_get_command_line((HV_VirtAddr) hv_cmdline, len);
1241 if (len < 0 || len > COMMAND_LINE_SIZE)
1242 early_panic("hv_get_command_line failed: %d\n", len);
1243 #endif
1244
1245 *cmdline_p = boot_command_line;
1246
1247 /* Set disabled_map and setup_max_cpus very early */
1248 parse_early_param();
1249
1250 /* Make sure the kernel is compatible with the hypervisor. */
1251 validate_hv();
1252 validate_va();
1253
1254 setup_cpu_maps();
1255
1256
1257 #ifdef CONFIG_PCI
1258 /*
1259 * Initialize the PCI structures. This is done before memory
1260 * setup so that we know whether or not a pci_reserve region
1261 * is necessary.
1262 */
1263 if (tile_pci_init() == 0)
1264 pci_reserve_mb = 0;
1265
1266 /* PCI systems reserve a region just below 4GB for mapping iomem. */
1267 pci_reserve_end_pfn = (1 << (32 - PAGE_SHIFT));
1268 pci_reserve_start_pfn = pci_reserve_end_pfn -
1269 (pci_reserve_mb << (20 - PAGE_SHIFT));
1270 #endif
1271
1272 init_mm.start_code = (unsigned long) _text;
1273 init_mm.end_code = (unsigned long) _etext;
1274 init_mm.end_data = (unsigned long) _edata;
1275 init_mm.brk = (unsigned long) _end;
1276
1277 setup_memory();
1278 store_permanent_mappings();
1279 setup_bootmem_allocator();
1280
1281 /*
1282 * NOTE: before this point _nobody_ is allowed to allocate
1283 * any memory using the bootmem allocator.
1284 */
1285
1286 paging_init();
1287 setup_numa_mapping();
1288 zone_sizes_init();
1289 set_page_homes();
1290 setup_cpu(1);
1291 setup_clock();
1292 load_hv_initrd();
1293 }
1294
1295
1296 /*
1297 * Set up per-cpu memory.
1298 */
1299
1300 unsigned long __per_cpu_offset[NR_CPUS] __write_once;
1301 EXPORT_SYMBOL(__per_cpu_offset);
1302
1303 static size_t __initdata pfn_offset[MAX_NUMNODES] = { 0 };
1304 static unsigned long __initdata percpu_pfn[NR_CPUS] = { 0 };
1305
1306 /*
1307 * As the percpu code allocates pages, we return the pages from the
1308 * end of the node for the specified cpu.
1309 */
pcpu_fc_alloc(unsigned int cpu,size_t size,size_t align)1310 static void *__init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align)
1311 {
1312 int nid = cpu_to_node(cpu);
1313 unsigned long pfn = node_percpu_pfn[nid] + pfn_offset[nid];
1314
1315 BUG_ON(size % PAGE_SIZE != 0);
1316 pfn_offset[nid] += size / PAGE_SIZE;
1317 BUG_ON(node_percpu[nid] < size);
1318 node_percpu[nid] -= size;
1319 if (percpu_pfn[cpu] == 0)
1320 percpu_pfn[cpu] = pfn;
1321 return pfn_to_kaddr(pfn);
1322 }
1323
1324 /*
1325 * Pages reserved for percpu memory are not freeable, and in any case we are
1326 * on a short path to panic() in setup_per_cpu_area() at this point anyway.
1327 */
pcpu_fc_free(void * ptr,size_t size)1328 static void __init pcpu_fc_free(void *ptr, size_t size)
1329 {
1330 }
1331
1332 /*
1333 * Set up vmalloc page tables using bootmem for the percpu code.
1334 */
pcpu_fc_populate_pte(unsigned long addr)1335 static void __init pcpu_fc_populate_pte(unsigned long addr)
1336 {
1337 pgd_t *pgd;
1338 pud_t *pud;
1339 pmd_t *pmd;
1340 pte_t *pte;
1341
1342 BUG_ON(pgd_addr_invalid(addr));
1343 if (addr < VMALLOC_START || addr >= VMALLOC_END)
1344 panic("PCPU addr %#lx outside vmalloc range %#lx..%#lx;"
1345 " try increasing CONFIG_VMALLOC_RESERVE\n",
1346 addr, VMALLOC_START, VMALLOC_END);
1347
1348 pgd = swapper_pg_dir + pgd_index(addr);
1349 pud = pud_offset(pgd, addr);
1350 BUG_ON(!pud_present(*pud));
1351 pmd = pmd_offset(pud, addr);
1352 if (pmd_present(*pmd)) {
1353 BUG_ON(pmd_huge_page(*pmd));
1354 } else {
1355 pte = __alloc_bootmem(L2_KERNEL_PGTABLE_SIZE,
1356 HV_PAGE_TABLE_ALIGN, 0);
1357 pmd_populate_kernel(&init_mm, pmd, pte);
1358 }
1359 }
1360
setup_per_cpu_areas(void)1361 void __init setup_per_cpu_areas(void)
1362 {
1363 struct page *pg;
1364 unsigned long delta, pfn, lowmem_va;
1365 unsigned long size = percpu_size();
1366 char *ptr;
1367 int rc, cpu, i;
1368
1369 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_fc_alloc,
1370 pcpu_fc_free, pcpu_fc_populate_pte);
1371 if (rc < 0)
1372 panic("Cannot initialize percpu area (err=%d)", rc);
1373
1374 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1375 for_each_possible_cpu(cpu) {
1376 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1377
1378 /* finv the copy out of cache so we can change homecache */
1379 ptr = pcpu_base_addr + pcpu_unit_offsets[cpu];
1380 __finv_buffer(ptr, size);
1381 pfn = percpu_pfn[cpu];
1382
1383 /* Rewrite the page tables to cache on that cpu */
1384 pg = pfn_to_page(pfn);
1385 for (i = 0; i < size; i += PAGE_SIZE, ++pfn, ++pg) {
1386
1387 /* Update the vmalloc mapping and page home. */
1388 pte_t *ptep =
1389 virt_to_pte(NULL, (unsigned long)ptr + i);
1390 pte_t pte = *ptep;
1391 BUG_ON(pfn != pte_pfn(pte));
1392 pte = hv_pte_set_mode(pte, HV_PTE_MODE_CACHE_TILE_L3);
1393 pte = set_remote_cache_cpu(pte, cpu);
1394 set_pte(ptep, pte);
1395
1396 /* Update the lowmem mapping for consistency. */
1397 lowmem_va = (unsigned long)pfn_to_kaddr(pfn);
1398 ptep = virt_to_pte(NULL, lowmem_va);
1399 if (pte_huge(*ptep)) {
1400 printk(KERN_DEBUG "early shatter of huge page"
1401 " at %#lx\n", lowmem_va);
1402 shatter_pmd((pmd_t *)ptep);
1403 ptep = virt_to_pte(NULL, lowmem_va);
1404 BUG_ON(pte_huge(*ptep));
1405 }
1406 BUG_ON(pfn != pte_pfn(*ptep));
1407 set_pte(ptep, pte);
1408 }
1409 }
1410
1411 /* Set our thread pointer appropriately. */
1412 set_my_cpu_offset(__per_cpu_offset[smp_processor_id()]);
1413
1414 /* Make sure the finv's have completed. */
1415 mb_incoherent();
1416
1417 /* Flush the TLB so we reference it properly from here on out. */
1418 local_flush_tlb_all();
1419 }
1420
1421 static struct resource data_resource = {
1422 .name = "Kernel data",
1423 .start = 0,
1424 .end = 0,
1425 .flags = IORESOURCE_BUSY | IORESOURCE_MEM
1426 };
1427
1428 static struct resource code_resource = {
1429 .name = "Kernel code",
1430 .start = 0,
1431 .end = 0,
1432 .flags = IORESOURCE_BUSY | IORESOURCE_MEM
1433 };
1434
1435 /*
1436 * We reserve all resources above 4GB so that PCI won't try to put
1437 * mappings above 4GB; the standard allows that for some devices but
1438 * the probing code trunates values to 32 bits.
1439 */
1440 #ifdef CONFIG_PCI
1441 static struct resource* __init
insert_non_bus_resource(void)1442 insert_non_bus_resource(void)
1443 {
1444 struct resource *res =
1445 kzalloc(sizeof(struct resource), GFP_ATOMIC);
1446 res->name = "Non-Bus Physical Address Space";
1447 res->start = (1ULL << 32);
1448 res->end = -1LL;
1449 res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1450 if (insert_resource(&iomem_resource, res)) {
1451 kfree(res);
1452 return NULL;
1453 }
1454 return res;
1455 }
1456 #endif
1457
1458 static struct resource* __init
insert_ram_resource(u64 start_pfn,u64 end_pfn)1459 insert_ram_resource(u64 start_pfn, u64 end_pfn)
1460 {
1461 struct resource *res =
1462 kzalloc(sizeof(struct resource), GFP_ATOMIC);
1463 res->name = "System RAM";
1464 res->start = start_pfn << PAGE_SHIFT;
1465 res->end = (end_pfn << PAGE_SHIFT) - 1;
1466 res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
1467 if (insert_resource(&iomem_resource, res)) {
1468 kfree(res);
1469 return NULL;
1470 }
1471 return res;
1472 }
1473
1474 /*
1475 * Request address space for all standard resources
1476 *
1477 * If the system includes PCI root complex drivers, we need to create
1478 * a window just below 4GB where PCI BARs can be mapped.
1479 */
request_standard_resources(void)1480 static int __init request_standard_resources(void)
1481 {
1482 int i;
1483 enum { CODE_DELTA = MEM_SV_INTRPT - PAGE_OFFSET };
1484
1485 iomem_resource.end = -1LL;
1486 #ifdef CONFIG_PCI
1487 insert_non_bus_resource();
1488 #endif
1489
1490 for_each_online_node(i) {
1491 u64 start_pfn = node_start_pfn[i];
1492 u64 end_pfn = node_end_pfn[i];
1493
1494 #ifdef CONFIG_PCI
1495 if (start_pfn <= pci_reserve_start_pfn &&
1496 end_pfn > pci_reserve_start_pfn) {
1497 if (end_pfn > pci_reserve_end_pfn)
1498 insert_ram_resource(pci_reserve_end_pfn,
1499 end_pfn);
1500 end_pfn = pci_reserve_start_pfn;
1501 }
1502 #endif
1503 insert_ram_resource(start_pfn, end_pfn);
1504 }
1505
1506 code_resource.start = __pa(_text - CODE_DELTA);
1507 code_resource.end = __pa(_etext - CODE_DELTA)-1;
1508 data_resource.start = __pa(_sdata);
1509 data_resource.end = __pa(_end)-1;
1510
1511 insert_resource(&iomem_resource, &code_resource);
1512 insert_resource(&iomem_resource, &data_resource);
1513
1514 #ifdef CONFIG_KEXEC
1515 insert_resource(&iomem_resource, &crashk_res);
1516 #endif
1517
1518 return 0;
1519 }
1520
1521 subsys_initcall(request_standard_resources);
1522