xref: /openbmc/linux/arch/powerpc/mm/init_64.c (revision 04b3c795)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  *  PowerPC version
4  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
5  *
6  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
7  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
8  *    Copyright (C) 1996 Paul Mackerras
9  *
10  *  Derived from "arch/i386/mm/init.c"
11  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
12  *
13  *  Dave Engebretsen <engebret@us.ibm.com>
14  *      Rework for PPC64 port.
15  */
16 
17 #undef DEBUG
18 
19 #include <linux/signal.h>
20 #include <linux/sched.h>
21 #include <linux/kernel.h>
22 #include <linux/errno.h>
23 #include <linux/string.h>
24 #include <linux/types.h>
25 #include <linux/mman.h>
26 #include <linux/mm.h>
27 #include <linux/swap.h>
28 #include <linux/stddef.h>
29 #include <linux/vmalloc.h>
30 #include <linux/init.h>
31 #include <linux/delay.h>
32 #include <linux/highmem.h>
33 #include <linux/idr.h>
34 #include <linux/nodemask.h>
35 #include <linux/module.h>
36 #include <linux/poison.h>
37 #include <linux/memblock.h>
38 #include <linux/hugetlb.h>
39 #include <linux/slab.h>
40 #include <linux/of_fdt.h>
41 #include <linux/libfdt.h>
42 #include <linux/memremap.h>
43 
44 #include <asm/pgalloc.h>
45 #include <asm/page.h>
46 #include <asm/prom.h>
47 #include <asm/rtas.h>
48 #include <asm/io.h>
49 #include <asm/mmu_context.h>
50 #include <asm/mmu.h>
51 #include <linux/uaccess.h>
52 #include <asm/smp.h>
53 #include <asm/machdep.h>
54 #include <asm/tlb.h>
55 #include <asm/eeh.h>
56 #include <asm/processor.h>
57 #include <asm/mmzone.h>
58 #include <asm/cputable.h>
59 #include <asm/sections.h>
60 #include <asm/iommu.h>
61 #include <asm/vdso.h>
62 
63 #include <mm/mmu_decl.h>
64 
65 #ifdef CONFIG_SPARSEMEM_VMEMMAP
66 /*
67  * Given an address within the vmemmap, determine the page that
68  * represents the start of the subsection it is within.  Note that we have to
69  * do this by hand as the proffered address may not be correctly aligned.
70  * Subtraction of non-aligned pointers produces undefined results.
71  */
72 static struct page * __meminit vmemmap_subsection_start(unsigned long vmemmap_addr)
73 {
74 	unsigned long start_pfn;
75 	unsigned long offset = vmemmap_addr - ((unsigned long)(vmemmap));
76 
77 	/* Return the pfn of the start of the section. */
78 	start_pfn = (offset / sizeof(struct page)) & PAGE_SUBSECTION_MASK;
79 	return pfn_to_page(start_pfn);
80 }
81 
82 /*
83  * Since memory is added in sub-section chunks, before creating a new vmemmap
84  * mapping, the kernel should check whether there is an existing memmap mapping
85  * covering the new subsection added. This is needed because kernel can map
86  * vmemmap area using 16MB pages which will cover a memory range of 16G. Such
87  * a range covers multiple subsections (2M)
88  *
89  * If any subsection in the 16G range mapped by vmemmap is valid we consider the
90  * vmemmap populated (There is a page table entry already present). We can't do
91  * a page table lookup here because with the hash translation we don't keep
92  * vmemmap details in linux page table.
93  */
94 static int __meminit vmemmap_populated(unsigned long vmemmap_addr, int vmemmap_map_size)
95 {
96 	struct page *start;
97 	unsigned long vmemmap_end = vmemmap_addr + vmemmap_map_size;
98 	start = vmemmap_subsection_start(vmemmap_addr);
99 
100 	for (; (unsigned long)start < vmemmap_end; start += PAGES_PER_SUBSECTION)
101 		/*
102 		 * pfn valid check here is intended to really check
103 		 * whether we have any subsection already initialized
104 		 * in this range.
105 		 */
106 		if (pfn_valid(page_to_pfn(start)))
107 			return 1;
108 
109 	return 0;
110 }
111 
112 /*
113  * vmemmap virtual address space management does not have a traditonal page
114  * table to track which virtual struct pages are backed by physical mapping.
115  * The virtual to physical mappings are tracked in a simple linked list
116  * format. 'vmemmap_list' maintains the entire vmemmap physical mapping at
117  * all times where as the 'next' list maintains the available
118  * vmemmap_backing structures which have been deleted from the
119  * 'vmemmap_global' list during system runtime (memory hotplug remove
120  * operation). The freed 'vmemmap_backing' structures are reused later when
121  * new requests come in without allocating fresh memory. This pointer also
122  * tracks the allocated 'vmemmap_backing' structures as we allocate one
123  * full page memory at a time when we dont have any.
124  */
125 struct vmemmap_backing *vmemmap_list;
126 static struct vmemmap_backing *next;
127 
128 /*
129  * The same pointer 'next' tracks individual chunks inside the allocated
130  * full page during the boot time and again tracks the freeed nodes during
131  * runtime. It is racy but it does not happen as they are separated by the
132  * boot process. Will create problem if some how we have memory hotplug
133  * operation during boot !!
134  */
135 static int num_left;
136 static int num_freed;
137 
138 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
139 {
140 	struct vmemmap_backing *vmem_back;
141 	/* get from freed entries first */
142 	if (num_freed) {
143 		num_freed--;
144 		vmem_back = next;
145 		next = next->list;
146 
147 		return vmem_back;
148 	}
149 
150 	/* allocate a page when required and hand out chunks */
151 	if (!num_left) {
152 		next = vmemmap_alloc_block(PAGE_SIZE, node);
153 		if (unlikely(!next)) {
154 			WARN_ON(1);
155 			return NULL;
156 		}
157 		num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
158 	}
159 
160 	num_left--;
161 
162 	return next++;
163 }
164 
165 static __meminit void vmemmap_list_populate(unsigned long phys,
166 					    unsigned long start,
167 					    int node)
168 {
169 	struct vmemmap_backing *vmem_back;
170 
171 	vmem_back = vmemmap_list_alloc(node);
172 	if (unlikely(!vmem_back)) {
173 		WARN_ON(1);
174 		return;
175 	}
176 
177 	vmem_back->phys = phys;
178 	vmem_back->virt_addr = start;
179 	vmem_back->list = vmemmap_list;
180 
181 	vmemmap_list = vmem_back;
182 }
183 
184 static bool altmap_cross_boundary(struct vmem_altmap *altmap, unsigned long start,
185 				unsigned long page_size)
186 {
187 	unsigned long nr_pfn = page_size / sizeof(struct page);
188 	unsigned long start_pfn = page_to_pfn((struct page *)start);
189 
190 	if ((start_pfn + nr_pfn) > altmap->end_pfn)
191 		return true;
192 
193 	if (start_pfn < altmap->base_pfn)
194 		return true;
195 
196 	return false;
197 }
198 
199 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
200 		struct vmem_altmap *altmap)
201 {
202 	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
203 
204 	/* Align to the page size of the linear mapping. */
205 	start = ALIGN_DOWN(start, page_size);
206 
207 	pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
208 
209 	for (; start < end; start += page_size) {
210 		void *p = NULL;
211 		int rc;
212 
213 		/*
214 		 * This vmemmap range is backing different subsections. If any
215 		 * of that subsection is marked valid, that means we already
216 		 * have initialized a page table covering this range and hence
217 		 * the vmemmap range is populated.
218 		 */
219 		if (vmemmap_populated(start, page_size))
220 			continue;
221 
222 		/*
223 		 * Allocate from the altmap first if we have one. This may
224 		 * fail due to alignment issues when using 16MB hugepages, so
225 		 * fall back to system memory if the altmap allocation fail.
226 		 */
227 		if (altmap && !altmap_cross_boundary(altmap, start, page_size)) {
228 			p = altmap_alloc_block_buf(page_size, altmap);
229 			if (!p)
230 				pr_debug("altmap block allocation failed, falling back to system memory");
231 		}
232 		if (!p)
233 			p = vmemmap_alloc_block_buf(page_size, node);
234 		if (!p)
235 			return -ENOMEM;
236 
237 		vmemmap_list_populate(__pa(p), start, node);
238 
239 		pr_debug("      * %016lx..%016lx allocated at %p\n",
240 			 start, start + page_size, p);
241 
242 		rc = vmemmap_create_mapping(start, page_size, __pa(p));
243 		if (rc < 0) {
244 			pr_warn("%s: Unable to create vmemmap mapping: %d\n",
245 				__func__, rc);
246 			return -EFAULT;
247 		}
248 	}
249 
250 	return 0;
251 }
252 
253 #ifdef CONFIG_MEMORY_HOTPLUG
254 static unsigned long vmemmap_list_free(unsigned long start)
255 {
256 	struct vmemmap_backing *vmem_back, *vmem_back_prev;
257 
258 	vmem_back_prev = vmem_back = vmemmap_list;
259 
260 	/* look for it with prev pointer recorded */
261 	for (; vmem_back; vmem_back = vmem_back->list) {
262 		if (vmem_back->virt_addr == start)
263 			break;
264 		vmem_back_prev = vmem_back;
265 	}
266 
267 	if (unlikely(!vmem_back)) {
268 		WARN_ON(1);
269 		return 0;
270 	}
271 
272 	/* remove it from vmemmap_list */
273 	if (vmem_back == vmemmap_list) /* remove head */
274 		vmemmap_list = vmem_back->list;
275 	else
276 		vmem_back_prev->list = vmem_back->list;
277 
278 	/* next point to this freed entry */
279 	vmem_back->list = next;
280 	next = vmem_back;
281 	num_freed++;
282 
283 	return vmem_back->phys;
284 }
285 
286 void __ref vmemmap_free(unsigned long start, unsigned long end,
287 		struct vmem_altmap *altmap)
288 {
289 	unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
290 	unsigned long page_order = get_order(page_size);
291 	unsigned long alt_start = ~0, alt_end = ~0;
292 	unsigned long base_pfn;
293 
294 	start = ALIGN_DOWN(start, page_size);
295 	if (altmap) {
296 		alt_start = altmap->base_pfn;
297 		alt_end = altmap->base_pfn + altmap->reserve +
298 			  altmap->free + altmap->alloc + altmap->align;
299 	}
300 
301 	pr_debug("vmemmap_free %lx...%lx\n", start, end);
302 
303 	for (; start < end; start += page_size) {
304 		unsigned long nr_pages, addr;
305 		struct page *page;
306 
307 		/*
308 		 * We have already marked the subsection we are trying to remove
309 		 * invalid. So if we want to remove the vmemmap range, we
310 		 * need to make sure there is no subsection marked valid
311 		 * in this range.
312 		 */
313 		if (vmemmap_populated(start, page_size))
314 			continue;
315 
316 		addr = vmemmap_list_free(start);
317 		if (!addr)
318 			continue;
319 
320 		page = pfn_to_page(addr >> PAGE_SHIFT);
321 		nr_pages = 1 << page_order;
322 		base_pfn = PHYS_PFN(addr);
323 
324 		if (base_pfn >= alt_start && base_pfn < alt_end) {
325 			vmem_altmap_free(altmap, nr_pages);
326 		} else if (PageReserved(page)) {
327 			/* allocated from bootmem */
328 			if (page_size < PAGE_SIZE) {
329 				/*
330 				 * this shouldn't happen, but if it is
331 				 * the case, leave the memory there
332 				 */
333 				WARN_ON_ONCE(1);
334 			} else {
335 				while (nr_pages--)
336 					free_reserved_page(page++);
337 			}
338 		} else {
339 			free_pages((unsigned long)(__va(addr)), page_order);
340 		}
341 
342 		vmemmap_remove_mapping(start, page_size);
343 	}
344 }
345 #endif
346 void register_page_bootmem_memmap(unsigned long section_nr,
347 				  struct page *start_page, unsigned long size)
348 {
349 }
350 
351 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
352 
353 #ifdef CONFIG_PPC_BOOK3S_64
354 static bool disable_radix = !IS_ENABLED(CONFIG_PPC_RADIX_MMU_DEFAULT);
355 
356 static int __init parse_disable_radix(char *p)
357 {
358 	bool val;
359 
360 	if (!p)
361 		val = true;
362 	else if (kstrtobool(p, &val))
363 		return -EINVAL;
364 
365 	disable_radix = val;
366 
367 	return 0;
368 }
369 early_param("disable_radix", parse_disable_radix);
370 
371 /*
372  * If we're running under a hypervisor, we need to check the contents of
373  * /chosen/ibm,architecture-vec-5 to see if the hypervisor is willing to do
374  * radix.  If not, we clear the radix feature bit so we fall back to hash.
375  */
376 static void __init early_check_vec5(void)
377 {
378 	unsigned long root, chosen;
379 	int size;
380 	const u8 *vec5;
381 	u8 mmu_supported;
382 
383 	root = of_get_flat_dt_root();
384 	chosen = of_get_flat_dt_subnode_by_name(root, "chosen");
385 	if (chosen == -FDT_ERR_NOTFOUND) {
386 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
387 		return;
388 	}
389 	vec5 = of_get_flat_dt_prop(chosen, "ibm,architecture-vec-5", &size);
390 	if (!vec5) {
391 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
392 		return;
393 	}
394 	if (size <= OV5_INDX(OV5_MMU_SUPPORT)) {
395 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
396 		return;
397 	}
398 
399 	/* Check for supported configuration */
400 	mmu_supported = vec5[OV5_INDX(OV5_MMU_SUPPORT)] &
401 			OV5_FEAT(OV5_MMU_SUPPORT);
402 	if (mmu_supported == OV5_FEAT(OV5_MMU_RADIX)) {
403 		/* Hypervisor only supports radix - check enabled && GTSE */
404 		if (!early_radix_enabled()) {
405 			pr_warn("WARNING: Ignoring cmdline option disable_radix\n");
406 		}
407 		if (!(vec5[OV5_INDX(OV5_RADIX_GTSE)] &
408 						OV5_FEAT(OV5_RADIX_GTSE))) {
409 			pr_warn("WARNING: Hypervisor doesn't support RADIX with GTSE\n");
410 		}
411 		/* Do radix anyway - the hypervisor said we had to */
412 		cur_cpu_spec->mmu_features |= MMU_FTR_TYPE_RADIX;
413 	} else if (mmu_supported == OV5_FEAT(OV5_MMU_HASH)) {
414 		/* Hypervisor only supports hash - disable radix */
415 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
416 	}
417 }
418 
419 void __init mmu_early_init_devtree(void)
420 {
421 	/* Disable radix mode based on kernel command line. */
422 	if (disable_radix)
423 		cur_cpu_spec->mmu_features &= ~MMU_FTR_TYPE_RADIX;
424 
425 	/*
426 	 * Check /chosen/ibm,architecture-vec-5 if running as a guest.
427 	 * When running bare-metal, we can use radix if we like
428 	 * even though the ibm,architecture-vec-5 property created by
429 	 * skiboot doesn't have the necessary bits set.
430 	 */
431 	if (!(mfmsr() & MSR_HV))
432 		early_check_vec5();
433 
434 	if (early_radix_enabled())
435 		radix__early_init_devtree();
436 	else
437 		hash__early_init_devtree();
438 }
439 #endif /* CONFIG_PPC_BOOK3S_64 */
440