xref: /linux/arch/powerpc/mm/init_64.c (revision 2fe05e1139a555ae91f00a812cb9520e7d3022ab)
1 /*
2  *  PowerPC version
3  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
4  *
5  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
6  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
7  *    Copyright (C) 1996 Paul Mackerras
8  *
9  *  Derived from "arch/i386/mm/init.c"
10  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
11  *
12  *  Dave Engebretsen <engebret@us.ibm.com>
13  *      Rework for PPC64 port.
14  *
15  *  This program is free software; you can redistribute it and/or
16  *  modify it under the terms of the GNU General Public License
17  *  as published by the Free Software Foundation; either version
18  *  2 of the License, or (at your option) any later version.
19  *
20  */
21 
22 #undef DEBUG
23 
24 #include <linux/signal.h>
25 #include <linux/sched.h>
26 #include <linux/kernel.h>
27 #include <linux/errno.h>
28 #include <linux/string.h>
29 #include <linux/types.h>
30 #include <linux/mman.h>
31 #include <linux/mm.h>
32 #include <linux/swap.h>
33 #include <linux/stddef.h>
34 #include <linux/vmalloc.h>
35 #include <linux/init.h>
36 #include <linux/delay.h>
37 #include <linux/highmem.h>
38 #include <linux/idr.h>
39 #include <linux/nodemask.h>
40 #include <linux/module.h>
41 #include <linux/poison.h>
42 #include <linux/memblock.h>
43 #include <linux/hugetlb.h>
44 #include <linux/slab.h>
45 #include <linux/of_fdt.h>
46 #include <linux/libfdt.h>
47 #include <linux/memremap.h>
48 
49 #include <asm/pgalloc.h>
50 #include <asm/page.h>
51 #include <asm/prom.h>
52 #include <asm/rtas.h>
53 #include <asm/io.h>
54 #include <asm/mmu_context.h>
55 #include <asm/pgtable.h>
56 #include <asm/mmu.h>
57 #include <linux/uaccess.h>
58 #include <asm/smp.h>
59 #include <asm/machdep.h>
60 #include <asm/tlb.h>
61 #include <asm/eeh.h>
62 #include <asm/processor.h>
63 #include <asm/mmzone.h>
64 #include <asm/cputable.h>
65 #include <asm/sections.h>
66 #include <asm/iommu.h>
67 #include <asm/vdso.h>
68 
69 #include "mmu_decl.h"
70 
71 #ifdef CONFIG_PPC_STD_MMU_64
72 #if H_PGTABLE_RANGE > USER_VSID_RANGE
73 #warning Limited user VSID range means pagetable space is wasted
74 #endif
75 #endif /* CONFIG_PPC_STD_MMU_64 */
76 
77 phys_addr_t memstart_addr = ~0;
78 EXPORT_SYMBOL_GPL(memstart_addr);
79 phys_addr_t kernstart_addr;
80 EXPORT_SYMBOL_GPL(kernstart_addr);
81 
82 #ifdef CONFIG_SPARSEMEM_VMEMMAP
83 /*
84  * Given an address within the vmemmap, determine the pfn of the page that
85  * represents the start of the section it is within.  Note that we have to
86  * do this by hand as the proffered address may not be correctly aligned.
87  * Subtraction of non-aligned pointers produces undefined results.
88  */
89 static unsigned long __meminit vmemmap_section_start(unsigned long page)
90 {
91 	unsigned long offset = page - ((unsigned long)(vmemmap));
92 
93 	/* Return the pfn of the start of the section. */
94 	return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
95 }
96 
97 /*
98  * Check if this vmemmap page is already initialised.  If any section
99  * which overlaps this vmemmap page is initialised then this page is
100  * initialised already.
101  */
102 static int __meminit vmemmap_populated(unsigned long start, int page_size)
103 {
104 	unsigned long end = start + page_size;
105 	start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
106 
107 	for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
108 		if (pfn_valid(page_to_pfn((struct page *)start)))
109 			return 1;
110 
111 	return 0;
112 }
113 
114 /*
115  * vmemmap virtual address space management does not have a traditonal page
116  * table to track which virtual struct pages are backed by physical mapping.
117  * The virtual to physical mappings are tracked in a simple linked list
118  * format. 'vmemmap_list' maintains the entire vmemmap physical mapping at
119  * all times where as the 'next' list maintains the available
120  * vmemmap_backing structures which have been deleted from the
121  * 'vmemmap_global' list during system runtime (memory hotplug remove
122  * operation). The freed 'vmemmap_backing' structures are reused later when
123  * new requests come in without allocating fresh memory. This pointer also
124  * tracks the allocated 'vmemmap_backing' structures as we allocate one
125  * full page memory at a time when we dont have any.
126  */
127 struct vmemmap_backing *vmemmap_list;
128 static struct vmemmap_backing *next;
129 
130 /*
131  * The same pointer 'next' tracks individual chunks inside the allocated
132  * full page during the boot time and again tracks the freeed nodes during
133  * runtime. It is racy but it does not happen as they are separated by the
134  * boot process. Will create problem if some how we have memory hotplug
135  * operation during boot !!
136  */
137 static int num_left;
138 static int num_freed;
139 
140 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
141 {
142 	struct vmemmap_backing *vmem_back;
143 	/* get from freed entries first */
144 	if (num_freed) {
145 		num_freed--;
146 		vmem_back = next;
147 		next = next->list;
148 
149 		return vmem_back;
150 	}
151 
152 	/* allocate a page when required and hand out chunks */
153 	if (!num_left) {
154 		next = vmemmap_alloc_block(PAGE_SIZE, node);
155 		if (unlikely(!next)) {
156 			WARN_ON(1);
157 			return NULL;
158 		}
159 		num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
160 	}
161 
162 	num_left--;
163 
164 	return next++;
165 }
166 
167 static __meminit void vmemmap_list_populate(unsigned long phys,
168 					    unsigned long start,
169 					    int node)
170 {
171 	struct vmemmap_backing *vmem_back;
172 
173 	vmem_back = vmemmap_list_alloc(node);
174 	if (unlikely(!vmem_back)) {
175 		WARN_ON(1);
176 		return;
177 	}
178 
179 	vmem_back->phys = phys;
180 	vmem_back->virt_addr = start;
181 	vmem_back->list = vmemmap_list;
182 
183 	vmemmap_list = vmem_back;
184 }
185 
186 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
187 {
188 	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
189 
190 	/* Align to the page size of the linear mapping. */
191 	start = _ALIGN_DOWN(start, page_size);
192 
193 	pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
194 
195 	for (; start < end; start += page_size) {
196 		struct vmem_altmap *altmap;
197 		void *p;
198 		int rc;
199 
200 		if (vmemmap_populated(start, page_size))
201 			continue;
202 
203 		/* altmap lookups only work at section boundaries */
204 		altmap = to_vmem_altmap(SECTION_ALIGN_DOWN(start));
205 
206 		p =  __vmemmap_alloc_block_buf(page_size, node, altmap);
207 		if (!p)
208 			return -ENOMEM;
209 
210 		vmemmap_list_populate(__pa(p), start, node);
211 
212 		pr_debug("      * %016lx..%016lx allocated at %p\n",
213 			 start, start + page_size, p);
214 
215 		rc = vmemmap_create_mapping(start, page_size, __pa(p));
216 		if (rc < 0) {
217 			pr_warning(
218 				"vmemmap_populate: Unable to create vmemmap mapping: %d\n",
219 				rc);
220 			return -EFAULT;
221 		}
222 	}
223 
224 	return 0;
225 }
226 
227 #ifdef CONFIG_MEMORY_HOTPLUG
228 static unsigned long vmemmap_list_free(unsigned long start)
229 {
230 	struct vmemmap_backing *vmem_back, *vmem_back_prev;
231 
232 	vmem_back_prev = vmem_back = vmemmap_list;
233 
234 	/* look for it with prev pointer recorded */
235 	for (; vmem_back; vmem_back = vmem_back->list) {
236 		if (vmem_back->virt_addr == start)
237 			break;
238 		vmem_back_prev = vmem_back;
239 	}
240 
241 	if (unlikely(!vmem_back)) {
242 		WARN_ON(1);
243 		return 0;
244 	}
245 
246 	/* remove it from vmemmap_list */
247 	if (vmem_back == vmemmap_list) /* remove head */
248 		vmemmap_list = vmem_back->list;
249 	else
250 		vmem_back_prev->list = vmem_back->list;
251 
252 	/* next point to this freed entry */
253 	vmem_back->list = next;
254 	next = vmem_back;
255 	num_freed++;
256 
257 	return vmem_back->phys;
258 }
259 
260 void __ref vmemmap_free(unsigned long start, unsigned long end)
261 {
262 	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
263 	unsigned long page_order = get_order(page_size);
264 
265 	start = _ALIGN_DOWN(start, page_size);
266 
267 	pr_debug("vmemmap_free %lx...%lx\n", start, end);
268 
269 	for (; start < end; start += page_size) {
270 		unsigned long nr_pages, addr;
271 		struct vmem_altmap *altmap;
272 		struct page *section_base;
273 		struct page *page;
274 
275 		/*
276 		 * the section has already be marked as invalid, so
277 		 * vmemmap_populated() true means some other sections still
278 		 * in this page, so skip it.
279 		 */
280 		if (vmemmap_populated(start, page_size))
281 			continue;
282 
283 		addr = vmemmap_list_free(start);
284 		if (!addr)
285 			continue;
286 
287 		page = pfn_to_page(addr >> PAGE_SHIFT);
288 		section_base = pfn_to_page(vmemmap_section_start(start));
289 		nr_pages = 1 << page_order;
290 
291 		altmap = to_vmem_altmap((unsigned long) section_base);
292 		if (altmap) {
293 			vmem_altmap_free(altmap, nr_pages);
294 		} else if (PageReserved(page)) {
295 			/* allocated from bootmem */
296 			if (page_size < PAGE_SIZE) {
297 				/*
298 				 * this shouldn't happen, but if it is
299 				 * the case, leave the memory there
300 				 */
301 				WARN_ON_ONCE(1);
302 			} else {
303 				while (nr_pages--)
304 					free_reserved_page(page++);
305 			}
306 		} else {
307 			free_pages((unsigned long)(__va(addr)), page_order);
308 		}
309 
310 		vmemmap_remove_mapping(start, page_size);
311 	}
312 }
313 #endif
314 void register_page_bootmem_memmap(unsigned long section_nr,
315 				  struct page *start_page, unsigned long size)
316 {
317 }
318 
319 /*
320  * We do not have access to the sparsemem vmemmap, so we fallback to
321  * walking the list of sparsemem blocks which we already maintain for
322  * the sake of crashdump. In the long run, we might want to maintain
323  * a tree if performance of that linear walk becomes a problem.
324  *
325  * realmode_pfn_to_page functions can fail due to:
326  * 1) As real sparsemem blocks do not lay in RAM continously (they
327  * are in virtual address space which is not available in the real mode),
328  * the requested page struct can be split between blocks so get_page/put_page
329  * may fail.
330  * 2) When huge pages are used, the get_page/put_page API will fail
331  * in real mode as the linked addresses in the page struct are virtual
332  * too.
333  */
334 struct page *realmode_pfn_to_page(unsigned long pfn)
335 {
336 	struct vmemmap_backing *vmem_back;
337 	struct page *page;
338 	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
339 	unsigned long pg_va = (unsigned long) pfn_to_page(pfn);
340 
341 	for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) {
342 		if (pg_va < vmem_back->virt_addr)
343 			continue;
344 
345 		/* After vmemmap_list entry free is possible, need check all */
346 		if ((pg_va + sizeof(struct page)) <=
347 				(vmem_back->virt_addr + page_size)) {
348 			page = (struct page *) (vmem_back->phys + pg_va -
349 				vmem_back->virt_addr);
350 			return page;
351 		}
352 	}
353 
354 	/* Probably that page struct is split between real pages */
355 	return NULL;
356 }
357 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
358 
359 #elif defined(CONFIG_FLATMEM)
360 
361 struct page *realmode_pfn_to_page(unsigned long pfn)
362 {
363 	struct page *page = pfn_to_page(pfn);
364 	return page;
365 }
366 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
367 
368 #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */
369 
370 #ifdef CONFIG_PPC_STD_MMU_64
371 static bool disable_radix;
372 static int __init parse_disable_radix(char *p)
373 {
374 	disable_radix = true;
375 	return 0;
376 }
377 early_param("disable_radix", parse_disable_radix);
378 
379 /*
380  * If we're running under a hypervisor, we need to check the contents of
381  * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do
382  * radix.  If not, we clear the radix feature bit so we fall back to hash.
383  */
384 static void early_check_vec5(void)
385 {
386 	unsigned long root, chosen;
387 	int size;
388 	const u8 *vec5;
389 	u8 mmu_supported;
390 
391 	root = of_get_flat_dt_root();
392 	chosen = of_get_flat_dt_subnode_by_name(root, "chosen");
393 	if (chosen == -FDT_ERR_NOTFOUND) {
394 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
395 		return;
396 	}
397 	vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size);
398 	if (!vec5) {
399 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
400 		return;
401 	}
402 	if (size <= OV5_INDX(OV5_MMU_SUPPORT)) {
403 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
404 		return;
405 	}
406 
407 	/* Check for supported configuration */
408 	mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] &
409 			OV5_FEAT(OV5_MMU_SUPPORT);
410 	if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) {
411 		/* Hypervisor only supports radix - check enabled && GTSE */
412 		if (!early_radix_enabled()) {
413 			pr_warn("WARNING: Ignoring cmdline option disable_radix\n");
414 		}
415 		if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] &
416 						OV5_FEAT(OV5_RADIX_GTSE))) {
417 			pr_warn("WARNING: Hypervisor doesn't support RADIX with GTSE\n");
418 		}
419 		/* Do radix anyway - the hypervisor said we had to */
420 		cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX;
421 	} else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) {
422 		/* Hypervisor only supports hash - disable radix */
423 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
424 	}
425 }
426 
427 void __init mmu_early_init_devtree(void)
428 {
429 	/* Disable radix mode based on kernel command line. */
430 	if (disable_radix)
431 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
432 
433 	/*
434 	 * Check /chosen/ibm,architecture-vec-5 if running as a guest.
435 	 * When running bare-metal, we can use radix if we like
436 	 * even though the ibm,architecture-vec-5 property created by
437 	 * skiboot doesn't have the necessary bits set.
438 	 */
439 	if (!(mfmsr() & MSR_HV))
440 		early_check_vec5();
441 
442 	if (early_radix_enabled())
443 		radix__early_init_devtree();
444 	else
445 		hash__early_init_devtree();
446 }
447 #endif /* CONFIG_PPC_STD_MMU_64 */
448