1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Page table handling routines for radix page table.
4  *
5  * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation.
6  */
7 
8 #define pr_fmt(fmt) "radix-mmu: " fmt
9 
10 #include <linux/io.h>
11 #include <linux/kernel.h>
12 #include <linux/sched/mm.h>
13 #include <linux/memblock.h>
14 #include <linux/of_fdt.h>
15 #include <linux/mm.h>
16 #include <linux/hugetlb.h>
17 #include <linux/string_helpers.h>
18 #include <linux/memory.h>
19 
20 #include <asm/pgalloc.h>
21 #include <asm/mmu_context.h>
22 #include <asm/dma.h>
23 #include <asm/machdep.h>
24 #include <asm/mmu.h>
25 #include <asm/firmware.h>
26 #include <asm/powernv.h>
27 #include <asm/sections.h>
28 #include <asm/smp.h>
29 #include <asm/trace.h>
30 #include <asm/uaccess.h>
31 #include <asm/ultravisor.h>
32 
33 #include <trace/events/thp.h>
34 
35 unsigned int mmu_pid_bits;
36 unsigned int mmu_base_pid;
37 unsigned long radix_mem_block_size __ro_after_init;
38 
39 static __ref void *early_alloc_pgtable(unsigned long size, int nid,
40 			unsigned long region_start, unsigned long region_end)
41 {
42 	phys_addr_t min_addr = MEMBLOCK_LOW_LIMIT;
43 	phys_addr_t max_addr = MEMBLOCK_ALLOC_ANYWHERE;
44 	void *ptr;
45 
46 	if (region_start)
47 		min_addr = region_start;
48 	if (region_end)
49 		max_addr = region_end;
50 
51 	ptr = memblock_alloc_try_nid(size, size, min_addr, max_addr, nid);
52 
53 	if (!ptr)
54 		panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa max_addr=%pa\n",
55 		      __func__, size, size, nid, &min_addr, &max_addr);
56 
57 	return ptr;
58 }
59 
60 /*
61  * When allocating pud or pmd pointers, we allocate a complete page
62  * of PAGE_SIZE rather than PUD_TABLE_SIZE or PMD_TABLE_SIZE. This
63  * is to ensure that the page obtained from the memblock allocator
64  * can be completely used as page table page and can be freed
65  * correctly when the page table entries are removed.
66  */
67 static int early_map_kernel_page(unsigned long ea, unsigned long pa,
68 			  pgprot_t flags,
69 			  unsigned int map_page_size,
70 			  int nid,
71 			  unsigned long region_start, unsigned long region_end)
72 {
73 	unsigned long pfn = pa >> PAGE_SHIFT;
74 	pgd_t *pgdp;
75 	p4d_t *p4dp;
76 	pud_t *pudp;
77 	pmd_t *pmdp;
78 	pte_t *ptep;
79 
80 	pgdp = pgd_offset_k(ea);
81 	p4dp = p4d_offset(pgdp, ea);
82 	if (p4d_none(*p4dp)) {
83 		pudp = early_alloc_pgtable(PAGE_SIZE, nid,
84 					   region_start, region_end);
85 		p4d_populate(&init_mm, p4dp, pudp);
86 	}
87 	pudp = pud_offset(p4dp, ea);
88 	if (map_page_size == PUD_SIZE) {
89 		ptep = (pte_t *)pudp;
90 		goto set_the_pte;
91 	}
92 	if (pud_none(*pudp)) {
93 		pmdp = early_alloc_pgtable(PAGE_SIZE, nid, region_start,
94 					   region_end);
95 		pud_populate(&init_mm, pudp, pmdp);
96 	}
97 	pmdp = pmd_offset(pudp, ea);
98 	if (map_page_size == PMD_SIZE) {
99 		ptep = pmdp_ptep(pmdp);
100 		goto set_the_pte;
101 	}
102 	if (!pmd_present(*pmdp)) {
103 		ptep = early_alloc_pgtable(PAGE_SIZE, nid,
104 						region_start, region_end);
105 		pmd_populate_kernel(&init_mm, pmdp, ptep);
106 	}
107 	ptep = pte_offset_kernel(pmdp, ea);
108 
109 set_the_pte:
110 	set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
111 	asm volatile("ptesync": : :"memory");
112 	return 0;
113 }
114 
115 /*
116  * nid, region_start, and region_end are hints to try to place the page
117  * table memory in the same node or region.
118  */
119 static int __map_kernel_page(unsigned long ea, unsigned long pa,
120 			  pgprot_t flags,
121 			  unsigned int map_page_size,
122 			  int nid,
123 			  unsigned long region_start, unsigned long region_end)
124 {
125 	unsigned long pfn = pa >> PAGE_SHIFT;
126 	pgd_t *pgdp;
127 	p4d_t *p4dp;
128 	pud_t *pudp;
129 	pmd_t *pmdp;
130 	pte_t *ptep;
131 	/*
132 	 * Make sure task size is correct as per the max adddr
133 	 */
134 	BUILD_BUG_ON(TASK_SIZE_USER64 > RADIX_PGTABLE_RANGE);
135 
136 #ifdef CONFIG_PPC_64K_PAGES
137 	BUILD_BUG_ON(RADIX_KERN_MAP_SIZE != (1UL << MAX_EA_BITS_PER_CONTEXT));
138 #endif
139 
140 	if (unlikely(!slab_is_available()))
141 		return early_map_kernel_page(ea, pa, flags, map_page_size,
142 						nid, region_start, region_end);
143 
144 	/*
145 	 * Should make page table allocation functions be able to take a
146 	 * node, so we can place kernel page tables on the right nodes after
147 	 * boot.
148 	 */
149 	pgdp = pgd_offset_k(ea);
150 	p4dp = p4d_offset(pgdp, ea);
151 	pudp = pud_alloc(&init_mm, p4dp, ea);
152 	if (!pudp)
153 		return -ENOMEM;
154 	if (map_page_size == PUD_SIZE) {
155 		ptep = (pte_t *)pudp;
156 		goto set_the_pte;
157 	}
158 	pmdp = pmd_alloc(&init_mm, pudp, ea);
159 	if (!pmdp)
160 		return -ENOMEM;
161 	if (map_page_size == PMD_SIZE) {
162 		ptep = pmdp_ptep(pmdp);
163 		goto set_the_pte;
164 	}
165 	ptep = pte_alloc_kernel(pmdp, ea);
166 	if (!ptep)
167 		return -ENOMEM;
168 
169 set_the_pte:
170 	set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
171 	asm volatile("ptesync": : :"memory");
172 	return 0;
173 }
174 
175 int radix__map_kernel_page(unsigned long ea, unsigned long pa,
176 			  pgprot_t flags,
177 			  unsigned int map_page_size)
178 {
179 	return __map_kernel_page(ea, pa, flags, map_page_size, -1, 0, 0);
180 }
181 
182 #ifdef CONFIG_STRICT_KERNEL_RWX
183 static void radix__change_memory_range(unsigned long start, unsigned long end,
184 				       unsigned long clear)
185 {
186 	unsigned long idx;
187 	pgd_t *pgdp;
188 	p4d_t *p4dp;
189 	pud_t *pudp;
190 	pmd_t *pmdp;
191 	pte_t *ptep;
192 
193 	start = ALIGN_DOWN(start, PAGE_SIZE);
194 	end = PAGE_ALIGN(end); // aligns up
195 
196 	pr_debug("Changing flags on range %lx-%lx removing 0x%lx\n",
197 		 start, end, clear);
198 
199 	for (idx = start; idx < end; idx += PAGE_SIZE) {
200 		pgdp = pgd_offset_k(idx);
201 		p4dp = p4d_offset(pgdp, idx);
202 		pudp = pud_alloc(&init_mm, p4dp, idx);
203 		if (!pudp)
204 			continue;
205 		if (pud_is_leaf(*pudp)) {
206 			ptep = (pte_t *)pudp;
207 			goto update_the_pte;
208 		}
209 		pmdp = pmd_alloc(&init_mm, pudp, idx);
210 		if (!pmdp)
211 			continue;
212 		if (pmd_is_leaf(*pmdp)) {
213 			ptep = pmdp_ptep(pmdp);
214 			goto update_the_pte;
215 		}
216 		ptep = pte_alloc_kernel(pmdp, idx);
217 		if (!ptep)
218 			continue;
219 update_the_pte:
220 		radix__pte_update(&init_mm, idx, ptep, clear, 0, 0);
221 	}
222 
223 	radix__flush_tlb_kernel_range(start, end);
224 }
225 
226 void radix__mark_rodata_ro(void)
227 {
228 	unsigned long start, end;
229 
230 	start = (unsigned long)_stext;
231 	end = (unsigned long)__init_begin;
232 
233 	radix__change_memory_range(start, end, _PAGE_WRITE);
234 }
235 
236 void radix__mark_initmem_nx(void)
237 {
238 	unsigned long start = (unsigned long)__init_begin;
239 	unsigned long end = (unsigned long)__init_end;
240 
241 	radix__change_memory_range(start, end, _PAGE_EXEC);
242 }
243 #endif /* CONFIG_STRICT_KERNEL_RWX */
244 
245 static inline void __meminit
246 print_mapping(unsigned long start, unsigned long end, unsigned long size, bool exec)
247 {
248 	char buf[10];
249 
250 	if (end <= start)
251 		return;
252 
253 	string_get_size(size, 1, STRING_UNITS_2, buf, sizeof(buf));
254 
255 	pr_info("Mapped 0x%016lx-0x%016lx with %s pages%s\n", start, end, buf,
256 		exec ? " (exec)" : "");
257 }
258 
259 static unsigned long next_boundary(unsigned long addr, unsigned long end)
260 {
261 #ifdef CONFIG_STRICT_KERNEL_RWX
262 	if (addr < __pa_symbol(__init_begin))
263 		return __pa_symbol(__init_begin);
264 #endif
265 	return end;
266 }
267 
268 static int __meminit create_physical_mapping(unsigned long start,
269 					     unsigned long end,
270 					     unsigned long max_mapping_size,
271 					     int nid, pgprot_t _prot)
272 {
273 	unsigned long vaddr, addr, mapping_size = 0;
274 	bool prev_exec, exec = false;
275 	pgprot_t prot;
276 	int psize;
277 
278 	start = ALIGN(start, PAGE_SIZE);
279 	end   = ALIGN_DOWN(end, PAGE_SIZE);
280 	for (addr = start; addr < end; addr += mapping_size) {
281 		unsigned long gap, previous_size;
282 		int rc;
283 
284 		gap = next_boundary(addr, end) - addr;
285 		if (gap > max_mapping_size)
286 			gap = max_mapping_size;
287 		previous_size = mapping_size;
288 		prev_exec = exec;
289 
290 		if (IS_ALIGNED(addr, PUD_SIZE) && gap >= PUD_SIZE &&
291 		    mmu_psize_defs[MMU_PAGE_1G].shift) {
292 			mapping_size = PUD_SIZE;
293 			psize = MMU_PAGE_1G;
294 		} else if (IS_ALIGNED(addr, PMD_SIZE) && gap >= PMD_SIZE &&
295 			   mmu_psize_defs[MMU_PAGE_2M].shift) {
296 			mapping_size = PMD_SIZE;
297 			psize = MMU_PAGE_2M;
298 		} else {
299 			mapping_size = PAGE_SIZE;
300 			psize = mmu_virtual_psize;
301 		}
302 
303 		vaddr = (unsigned long)__va(addr);
304 
305 		if (overlaps_kernel_text(vaddr, vaddr + mapping_size) ||
306 		    overlaps_interrupt_vector_text(vaddr, vaddr + mapping_size)) {
307 			prot = PAGE_KERNEL_X;
308 			exec = true;
309 		} else {
310 			prot = _prot;
311 			exec = false;
312 		}
313 
314 		if (mapping_size != previous_size || exec != prev_exec) {
315 			print_mapping(start, addr, previous_size, prev_exec);
316 			start = addr;
317 		}
318 
319 		rc = __map_kernel_page(vaddr, addr, prot, mapping_size, nid, start, end);
320 		if (rc)
321 			return rc;
322 
323 		update_page_count(psize, 1);
324 	}
325 
326 	print_mapping(start, addr, mapping_size, exec);
327 	return 0;
328 }
329 
330 static void __init radix_init_pgtable(void)
331 {
332 	unsigned long rts_field;
333 	phys_addr_t start, end;
334 	u64 i;
335 
336 	/* We don't support slb for radix */
337 	mmu_slb_size = 0;
338 
339 	/*
340 	 * Create the linear mapping
341 	 */
342 	for_each_mem_range(i, &start, &end) {
343 		/*
344 		 * The memblock allocator  is up at this point, so the
345 		 * page tables will be allocated within the range. No
346 		 * need or a node (which we don't have yet).
347 		 */
348 
349 		if (end >= RADIX_VMALLOC_START) {
350 			pr_warn("Outside the supported range\n");
351 			continue;
352 		}
353 
354 		WARN_ON(create_physical_mapping(start, end,
355 						radix_mem_block_size,
356 						-1, PAGE_KERNEL));
357 	}
358 
359 	/* Find out how many PID bits are supported */
360 	if (!cpu_has_feature(CPU_FTR_HVMODE) &&
361 			cpu_has_feature(CPU_FTR_P9_RADIX_PREFETCH_BUG)) {
362 		/*
363 		 * Older versions of KVM on these machines perfer if the
364 		 * guest only uses the low 19 PID bits.
365 		 */
366 		if (!mmu_pid_bits)
367 			mmu_pid_bits = 19;
368 	} else {
369 		if (!mmu_pid_bits)
370 			mmu_pid_bits = 20;
371 	}
372 	mmu_base_pid = 1;
373 
374 	/*
375 	 * Allocate Partition table and process table for the
376 	 * host.
377 	 */
378 	BUG_ON(PRTB_SIZE_SHIFT > 36);
379 	process_tb = early_alloc_pgtable(1UL << PRTB_SIZE_SHIFT, -1, 0, 0);
380 	/*
381 	 * Fill in the process table.
382 	 */
383 	rts_field = radix__get_tree_size();
384 	process_tb->prtb0 = cpu_to_be64(rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE);
385 
386 	/*
387 	 * The init_mm context is given the first available (non-zero) PID,
388 	 * which is the "guard PID" and contains no page table. PIDR should
389 	 * never be set to zero because that duplicates the kernel address
390 	 * space at the 0x0... offset (quadrant 0)!
391 	 *
392 	 * An arbitrary PID that may later be allocated by the PID allocator
393 	 * for userspace processes must not be used either, because that
394 	 * would cause stale user mappings for that PID on CPUs outside of
395 	 * the TLB invalidation scheme (because it won't be in mm_cpumask).
396 	 *
397 	 * So permanently carve out one PID for the purpose of a guard PID.
398 	 */
399 	init_mm.context.id = mmu_base_pid;
400 	mmu_base_pid++;
401 }
402 
403 static void __init radix_init_partition_table(void)
404 {
405 	unsigned long rts_field, dw0, dw1;
406 
407 	mmu_partition_table_init();
408 	rts_field = radix__get_tree_size();
409 	dw0 = rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE | PATB_HR;
410 	dw1 = __pa(process_tb) | (PRTB_SIZE_SHIFT - 12) | PATB_GR;
411 	mmu_partition_table_set_entry(0, dw0, dw1, false);
412 
413 	pr_info("Initializing Radix MMU\n");
414 }
415 
416 static int __init get_idx_from_shift(unsigned int shift)
417 {
418 	int idx = -1;
419 
420 	switch (shift) {
421 	case 0xc:
422 		idx = MMU_PAGE_4K;
423 		break;
424 	case 0x10:
425 		idx = MMU_PAGE_64K;
426 		break;
427 	case 0x15:
428 		idx = MMU_PAGE_2M;
429 		break;
430 	case 0x1e:
431 		idx = MMU_PAGE_1G;
432 		break;
433 	}
434 	return idx;
435 }
436 
437 static int __init radix_dt_scan_page_sizes(unsigned long node,
438 					   const char *uname, int depth,
439 					   void *data)
440 {
441 	int size = 0;
442 	int shift, idx;
443 	unsigned int ap;
444 	const __be32 *prop;
445 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
446 
447 	/* We are scanning "cpu" nodes only */
448 	if (type == NULL || strcmp(type, "cpu") != 0)
449 		return 0;
450 
451 	/* Find MMU PID size */
452 	prop = of_get_flat_dt_prop(node, "ibm,mmu-pid-bits", &size);
453 	if (prop && size == 4)
454 		mmu_pid_bits = be32_to_cpup(prop);
455 
456 	/* Grab page size encodings */
457 	prop = of_get_flat_dt_prop(node, "ibm,processor-radix-AP-encodings", &size);
458 	if (!prop)
459 		return 0;
460 
461 	pr_info("Page sizes from device-tree:\n");
462 	for (; size >= 4; size -= 4, ++prop) {
463 
464 		struct mmu_psize_def *def;
465 
466 		/* top 3 bit is AP encoding */
467 		shift = be32_to_cpu(prop[0]) & ~(0xe << 28);
468 		ap = be32_to_cpu(prop[0]) >> 29;
469 		pr_info("Page size shift = %d AP=0x%x\n", shift, ap);
470 
471 		idx = get_idx_from_shift(shift);
472 		if (idx < 0)
473 			continue;
474 
475 		def = &mmu_psize_defs[idx];
476 		def->shift = shift;
477 		def->ap  = ap;
478 	}
479 
480 	/* needed ? */
481 	cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
482 	return 1;
483 }
484 
485 #ifdef CONFIG_MEMORY_HOTPLUG
486 static int __init probe_memory_block_size(unsigned long node, const char *uname, int
487 					  depth, void *data)
488 {
489 	unsigned long *mem_block_size = (unsigned long *)data;
490 	const __be32 *prop;
491 	int len;
492 
493 	if (depth != 1)
494 		return 0;
495 
496 	if (strcmp(uname, "ibm,dynamic-reconfiguration-memory"))
497 		return 0;
498 
499 	prop = of_get_flat_dt_prop(node, "ibm,lmb-size", &len);
500 
501 	if (!prop || len < dt_root_size_cells * sizeof(__be32))
502 		/*
503 		 * Nothing in the device tree
504 		 */
505 		*mem_block_size = MIN_MEMORY_BLOCK_SIZE;
506 	else
507 		*mem_block_size = of_read_number(prop, dt_root_size_cells);
508 	return 1;
509 }
510 
511 static unsigned long radix_memory_block_size(void)
512 {
513 	unsigned long mem_block_size = MIN_MEMORY_BLOCK_SIZE;
514 
515 	/*
516 	 * OPAL firmware feature is set by now. Hence we are ok
517 	 * to test OPAL feature.
518 	 */
519 	if (firmware_has_feature(FW_FEATURE_OPAL))
520 		mem_block_size = 1UL * 1024 * 1024 * 1024;
521 	else
522 		of_scan_flat_dt(probe_memory_block_size, &mem_block_size);
523 
524 	return mem_block_size;
525 }
526 
527 #else   /* CONFIG_MEMORY_HOTPLUG */
528 
529 static unsigned long radix_memory_block_size(void)
530 {
531 	return 1UL * 1024 * 1024 * 1024;
532 }
533 
534 #endif /* CONFIG_MEMORY_HOTPLUG */
535 
536 
537 void __init radix__early_init_devtree(void)
538 {
539 	int rc;
540 
541 	/*
542 	 * Try to find the available page sizes in the device-tree
543 	 */
544 	rc = of_scan_flat_dt(radix_dt_scan_page_sizes, NULL);
545 	if (!rc) {
546 		/*
547 		 * No page size details found in device tree.
548 		 * Let's assume we have page 4k and 64k support
549 		 */
550 		mmu_psize_defs[MMU_PAGE_4K].shift = 12;
551 		mmu_psize_defs[MMU_PAGE_4K].ap = 0x0;
552 
553 		mmu_psize_defs[MMU_PAGE_64K].shift = 16;
554 		mmu_psize_defs[MMU_PAGE_64K].ap = 0x5;
555 	}
556 
557 	/*
558 	 * Max mapping size used when mapping pages. We don't use
559 	 * ppc_md.memory_block_size() here because this get called
560 	 * early and we don't have machine probe called yet. Also
561 	 * the pseries implementation only check for ibm,lmb-size.
562 	 * All hypervisor supporting radix do expose that device
563 	 * tree node.
564 	 */
565 	radix_mem_block_size = radix_memory_block_size();
566 	return;
567 }
568 
569 static void radix_init_amor(void)
570 {
571 	/*
572 	* In HV mode, we init AMOR (Authority Mask Override Register) so that
573 	* the hypervisor and guest can setup IAMR (Instruction Authority Mask
574 	* Register), enable key 0 and set it to 1.
575 	*
576 	* AMOR = 0b1100 .... 0000 (Mask for key 0 is 11)
577 	*/
578 	mtspr(SPRN_AMOR, (3ul << 62));
579 }
580 
581 void __init radix__early_init_mmu(void)
582 {
583 	unsigned long lpcr;
584 
585 #ifdef CONFIG_PPC_64K_PAGES
586 	/* PAGE_SIZE mappings */
587 	mmu_virtual_psize = MMU_PAGE_64K;
588 #else
589 	mmu_virtual_psize = MMU_PAGE_4K;
590 #endif
591 
592 #ifdef CONFIG_SPARSEMEM_VMEMMAP
593 	/* vmemmap mapping */
594 	if (mmu_psize_defs[MMU_PAGE_2M].shift) {
595 		/*
596 		 * map vmemmap using 2M if available
597 		 */
598 		mmu_vmemmap_psize = MMU_PAGE_2M;
599 	} else
600 		mmu_vmemmap_psize = mmu_virtual_psize;
601 #endif
602 	/*
603 	 * initialize page table size
604 	 */
605 	__pte_index_size = RADIX_PTE_INDEX_SIZE;
606 	__pmd_index_size = RADIX_PMD_INDEX_SIZE;
607 	__pud_index_size = RADIX_PUD_INDEX_SIZE;
608 	__pgd_index_size = RADIX_PGD_INDEX_SIZE;
609 	__pud_cache_index = RADIX_PUD_INDEX_SIZE;
610 	__pte_table_size = RADIX_PTE_TABLE_SIZE;
611 	__pmd_table_size = RADIX_PMD_TABLE_SIZE;
612 	__pud_table_size = RADIX_PUD_TABLE_SIZE;
613 	__pgd_table_size = RADIX_PGD_TABLE_SIZE;
614 
615 	__pmd_val_bits = RADIX_PMD_VAL_BITS;
616 	__pud_val_bits = RADIX_PUD_VAL_BITS;
617 	__pgd_val_bits = RADIX_PGD_VAL_BITS;
618 
619 	__kernel_virt_start = RADIX_KERN_VIRT_START;
620 	__vmalloc_start = RADIX_VMALLOC_START;
621 	__vmalloc_end = RADIX_VMALLOC_END;
622 	__kernel_io_start = RADIX_KERN_IO_START;
623 	__kernel_io_end = RADIX_KERN_IO_END;
624 	vmemmap = (struct page *)RADIX_VMEMMAP_START;
625 	ioremap_bot = IOREMAP_BASE;
626 
627 #ifdef CONFIG_PCI
628 	pci_io_base = ISA_IO_BASE;
629 #endif
630 	__pte_frag_nr = RADIX_PTE_FRAG_NR;
631 	__pte_frag_size_shift = RADIX_PTE_FRAG_SIZE_SHIFT;
632 	__pmd_frag_nr = RADIX_PMD_FRAG_NR;
633 	__pmd_frag_size_shift = RADIX_PMD_FRAG_SIZE_SHIFT;
634 
635 	radix_init_pgtable();
636 
637 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
638 		lpcr = mfspr(SPRN_LPCR);
639 		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
640 		radix_init_partition_table();
641 		radix_init_amor();
642 	} else {
643 		radix_init_pseries();
644 	}
645 
646 	memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
647 
648 	/* Switch to the guard PID before turning on MMU */
649 	radix__switch_mmu_context(NULL, &init_mm);
650 	tlbiel_all();
651 }
652 
653 void radix__early_init_mmu_secondary(void)
654 {
655 	unsigned long lpcr;
656 	/*
657 	 * update partition table control register and UPRT
658 	 */
659 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
660 		lpcr = mfspr(SPRN_LPCR);
661 		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
662 
663 		set_ptcr_when_no_uv(__pa(partition_tb) |
664 				    (PATB_SIZE_SHIFT - 12));
665 
666 		radix_init_amor();
667 	}
668 
669 	radix__switch_mmu_context(NULL, &init_mm);
670 	tlbiel_all();
671 
672 	/* Make sure userspace can't change the AMR */
673 	mtspr(SPRN_UAMOR, 0);
674 }
675 
676 void radix__mmu_cleanup_all(void)
677 {
678 	unsigned long lpcr;
679 
680 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
681 		lpcr = mfspr(SPRN_LPCR);
682 		mtspr(SPRN_LPCR, lpcr & ~LPCR_UPRT);
683 		set_ptcr_when_no_uv(0);
684 		powernv_set_nmmu_ptcr(0);
685 		radix__flush_tlb_all();
686 	}
687 }
688 
689 #ifdef CONFIG_MEMORY_HOTPLUG
690 static void free_pte_table(pte_t *pte_start, pmd_t *pmd)
691 {
692 	pte_t *pte;
693 	int i;
694 
695 	for (i = 0; i < PTRS_PER_PTE; i++) {
696 		pte = pte_start + i;
697 		if (!pte_none(*pte))
698 			return;
699 	}
700 
701 	pte_free_kernel(&init_mm, pte_start);
702 	pmd_clear(pmd);
703 }
704 
705 static void free_pmd_table(pmd_t *pmd_start, pud_t *pud)
706 {
707 	pmd_t *pmd;
708 	int i;
709 
710 	for (i = 0; i < PTRS_PER_PMD; i++) {
711 		pmd = pmd_start + i;
712 		if (!pmd_none(*pmd))
713 			return;
714 	}
715 
716 	pmd_free(&init_mm, pmd_start);
717 	pud_clear(pud);
718 }
719 
720 static void free_pud_table(pud_t *pud_start, p4d_t *p4d)
721 {
722 	pud_t *pud;
723 	int i;
724 
725 	for (i = 0; i < PTRS_PER_PUD; i++) {
726 		pud = pud_start + i;
727 		if (!pud_none(*pud))
728 			return;
729 	}
730 
731 	pud_free(&init_mm, pud_start);
732 	p4d_clear(p4d);
733 }
734 
735 static void remove_pte_table(pte_t *pte_start, unsigned long addr,
736 			     unsigned long end)
737 {
738 	unsigned long next;
739 	pte_t *pte;
740 
741 	pte = pte_start + pte_index(addr);
742 	for (; addr < end; addr = next, pte++) {
743 		next = (addr + PAGE_SIZE) & PAGE_MASK;
744 		if (next > end)
745 			next = end;
746 
747 		if (!pte_present(*pte))
748 			continue;
749 
750 		if (!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(next)) {
751 			/*
752 			 * The vmemmap_free() and remove_section_mapping()
753 			 * codepaths call us with aligned addresses.
754 			 */
755 			WARN_ONCE(1, "%s: unaligned range\n", __func__);
756 			continue;
757 		}
758 
759 		pte_clear(&init_mm, addr, pte);
760 	}
761 }
762 
763 static void __meminit remove_pmd_table(pmd_t *pmd_start, unsigned long addr,
764 			     unsigned long end)
765 {
766 	unsigned long next;
767 	pte_t *pte_base;
768 	pmd_t *pmd;
769 
770 	pmd = pmd_start + pmd_index(addr);
771 	for (; addr < end; addr = next, pmd++) {
772 		next = pmd_addr_end(addr, end);
773 
774 		if (!pmd_present(*pmd))
775 			continue;
776 
777 		if (pmd_is_leaf(*pmd)) {
778 			if (!IS_ALIGNED(addr, PMD_SIZE) ||
779 			    !IS_ALIGNED(next, PMD_SIZE)) {
780 				WARN_ONCE(1, "%s: unaligned range\n", __func__);
781 				continue;
782 			}
783 			pte_clear(&init_mm, addr, (pte_t *)pmd);
784 			continue;
785 		}
786 
787 		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
788 		remove_pte_table(pte_base, addr, next);
789 		free_pte_table(pte_base, pmd);
790 	}
791 }
792 
793 static void __meminit remove_pud_table(pud_t *pud_start, unsigned long addr,
794 			     unsigned long end)
795 {
796 	unsigned long next;
797 	pmd_t *pmd_base;
798 	pud_t *pud;
799 
800 	pud = pud_start + pud_index(addr);
801 	for (; addr < end; addr = next, pud++) {
802 		next = pud_addr_end(addr, end);
803 
804 		if (!pud_present(*pud))
805 			continue;
806 
807 		if (pud_is_leaf(*pud)) {
808 			if (!IS_ALIGNED(addr, PUD_SIZE) ||
809 			    !IS_ALIGNED(next, PUD_SIZE)) {
810 				WARN_ONCE(1, "%s: unaligned range\n", __func__);
811 				continue;
812 			}
813 			pte_clear(&init_mm, addr, (pte_t *)pud);
814 			continue;
815 		}
816 
817 		pmd_base = (pmd_t *)pud_page_vaddr(*pud);
818 		remove_pmd_table(pmd_base, addr, next);
819 		free_pmd_table(pmd_base, pud);
820 	}
821 }
822 
823 static void __meminit remove_pagetable(unsigned long start, unsigned long end)
824 {
825 	unsigned long addr, next;
826 	pud_t *pud_base;
827 	pgd_t *pgd;
828 	p4d_t *p4d;
829 
830 	spin_lock(&init_mm.page_table_lock);
831 
832 	for (addr = start; addr < end; addr = next) {
833 		next = pgd_addr_end(addr, end);
834 
835 		pgd = pgd_offset_k(addr);
836 		p4d = p4d_offset(pgd, addr);
837 		if (!p4d_present(*p4d))
838 			continue;
839 
840 		if (p4d_is_leaf(*p4d)) {
841 			if (!IS_ALIGNED(addr, P4D_SIZE) ||
842 			    !IS_ALIGNED(next, P4D_SIZE)) {
843 				WARN_ONCE(1, "%s: unaligned range\n", __func__);
844 				continue;
845 			}
846 
847 			pte_clear(&init_mm, addr, (pte_t *)pgd);
848 			continue;
849 		}
850 
851 		pud_base = (pud_t *)p4d_page_vaddr(*p4d);
852 		remove_pud_table(pud_base, addr, next);
853 		free_pud_table(pud_base, p4d);
854 	}
855 
856 	spin_unlock(&init_mm.page_table_lock);
857 	radix__flush_tlb_kernel_range(start, end);
858 }
859 
860 int __meminit radix__create_section_mapping(unsigned long start,
861 					    unsigned long end, int nid,
862 					    pgprot_t prot)
863 {
864 	if (end >= RADIX_VMALLOC_START) {
865 		pr_warn("Outside the supported range\n");
866 		return -1;
867 	}
868 
869 	return create_physical_mapping(__pa(start), __pa(end),
870 				       radix_mem_block_size, nid, prot);
871 }
872 
873 int __meminit radix__remove_section_mapping(unsigned long start, unsigned long end)
874 {
875 	remove_pagetable(start, end);
876 	return 0;
877 }
878 #endif /* CONFIG_MEMORY_HOTPLUG */
879 
880 #ifdef CONFIG_SPARSEMEM_VMEMMAP
881 static int __map_kernel_page_nid(unsigned long ea, unsigned long pa,
882 				 pgprot_t flags, unsigned int map_page_size,
883 				 int nid)
884 {
885 	return __map_kernel_page(ea, pa, flags, map_page_size, nid, 0, 0);
886 }
887 
888 int __meminit radix__vmemmap_create_mapping(unsigned long start,
889 				      unsigned long page_size,
890 				      unsigned long phys)
891 {
892 	/* Create a PTE encoding */
893 	unsigned long flags = _PAGE_PRESENT | _PAGE_ACCESSED | _PAGE_KERNEL_RW;
894 	int nid = early_pfn_to_nid(phys >> PAGE_SHIFT);
895 	int ret;
896 
897 	if ((start + page_size) >= RADIX_VMEMMAP_END) {
898 		pr_warn("Outside the supported range\n");
899 		return -1;
900 	}
901 
902 	ret = __map_kernel_page_nid(start, phys, __pgprot(flags), page_size, nid);
903 	BUG_ON(ret);
904 
905 	return 0;
906 }
907 
908 #ifdef CONFIG_MEMORY_HOTPLUG
909 void __meminit radix__vmemmap_remove_mapping(unsigned long start, unsigned long page_size)
910 {
911 	remove_pagetable(start, start + page_size);
912 }
913 #endif
914 #endif
915 
916 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
917 
918 unsigned long radix__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
919 				  pmd_t *pmdp, unsigned long clr,
920 				  unsigned long set)
921 {
922 	unsigned long old;
923 
924 #ifdef CONFIG_DEBUG_VM
925 	WARN_ON(!radix__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
926 	assert_spin_locked(pmd_lockptr(mm, pmdp));
927 #endif
928 
929 	old = radix__pte_update(mm, addr, (pte_t *)pmdp, clr, set, 1);
930 	trace_hugepage_update(addr, old, clr, set);
931 
932 	return old;
933 }
934 
935 pmd_t radix__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
936 			pmd_t *pmdp)
937 
938 {
939 	pmd_t pmd;
940 
941 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
942 	VM_BUG_ON(radix__pmd_trans_huge(*pmdp));
943 	VM_BUG_ON(pmd_devmap(*pmdp));
944 	/*
945 	 * khugepaged calls this for normal pmd
946 	 */
947 	pmd = *pmdp;
948 	pmd_clear(pmdp);
949 
950 	/*
951 	 * pmdp collapse_flush need to ensure that there are no parallel gup
952 	 * walk after this call. This is needed so that we can have stable
953 	 * page ref count when collapsing a page. We don't allow a collapse page
954 	 * if we have gup taken on the page. We can ensure that by sending IPI
955 	 * because gup walk happens with IRQ disabled.
956 	 */
957 	serialize_against_pte_lookup(vma->vm_mm);
958 
959 	radix__flush_tlb_collapsed_pmd(vma->vm_mm, address);
960 
961 	return pmd;
962 }
963 
964 /*
965  * For us pgtable_t is pte_t *. Inorder to save the deposisted
966  * page table, we consider the allocated page table as a list
967  * head. On withdraw we need to make sure we zero out the used
968  * list_head memory area.
969  */
970 void radix__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
971 				 pgtable_t pgtable)
972 {
973 	struct list_head *lh = (struct list_head *) pgtable;
974 
975 	assert_spin_locked(pmd_lockptr(mm, pmdp));
976 
977 	/* FIFO */
978 	if (!pmd_huge_pte(mm, pmdp))
979 		INIT_LIST_HEAD(lh);
980 	else
981 		list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
982 	pmd_huge_pte(mm, pmdp) = pgtable;
983 }
984 
985 pgtable_t radix__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
986 {
987 	pte_t *ptep;
988 	pgtable_t pgtable;
989 	struct list_head *lh;
990 
991 	assert_spin_locked(pmd_lockptr(mm, pmdp));
992 
993 	/* FIFO */
994 	pgtable = pmd_huge_pte(mm, pmdp);
995 	lh = (struct list_head *) pgtable;
996 	if (list_empty(lh))
997 		pmd_huge_pte(mm, pmdp) = NULL;
998 	else {
999 		pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
1000 		list_del(lh);
1001 	}
1002 	ptep = (pte_t *) pgtable;
1003 	*ptep = __pte(0);
1004 	ptep++;
1005 	*ptep = __pte(0);
1006 	return pgtable;
1007 }
1008 
1009 pmd_t radix__pmdp_huge_get_and_clear(struct mm_struct *mm,
1010 				     unsigned long addr, pmd_t *pmdp)
1011 {
1012 	pmd_t old_pmd;
1013 	unsigned long old;
1014 
1015 	old = radix__pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
1016 	old_pmd = __pmd(old);
1017 	return old_pmd;
1018 }
1019 
1020 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1021 
1022 void radix__ptep_set_access_flags(struct vm_area_struct *vma, pte_t *ptep,
1023 				  pte_t entry, unsigned long address, int psize)
1024 {
1025 	struct mm_struct *mm = vma->vm_mm;
1026 	unsigned long set = pte_val(entry) & (_PAGE_DIRTY | _PAGE_ACCESSED |
1027 					      _PAGE_RW | _PAGE_EXEC);
1028 
1029 	unsigned long change = pte_val(entry) ^ pte_val(*ptep);
1030 	/*
1031 	 * To avoid NMMU hang while relaxing access, we need mark
1032 	 * the pte invalid in between.
1033 	 */
1034 	if ((change & _PAGE_RW) && atomic_read(&mm->context.copros) > 0) {
1035 		unsigned long old_pte, new_pte;
1036 
1037 		old_pte = __radix_pte_update(ptep, _PAGE_PRESENT, _PAGE_INVALID);
1038 		/*
1039 		 * new value of pte
1040 		 */
1041 		new_pte = old_pte | set;
1042 		radix__flush_tlb_page_psize(mm, address, psize);
1043 		__radix_pte_update(ptep, _PAGE_INVALID, new_pte);
1044 	} else {
1045 		__radix_pte_update(ptep, 0, set);
1046 		/*
1047 		 * Book3S does not require a TLB flush when relaxing access
1048 		 * restrictions when the address space is not attached to a
1049 		 * NMMU, because the core MMU will reload the pte after taking
1050 		 * an access fault, which is defined by the architecture.
1051 		 */
1052 	}
1053 	/* See ptesync comment in radix__set_pte_at */
1054 }
1055 
1056 void radix__ptep_modify_prot_commit(struct vm_area_struct *vma,
1057 				    unsigned long addr, pte_t *ptep,
1058 				    pte_t old_pte, pte_t pte)
1059 {
1060 	struct mm_struct *mm = vma->vm_mm;
1061 
1062 	/*
1063 	 * To avoid NMMU hang while relaxing access we need to flush the tlb before
1064 	 * we set the new value. We need to do this only for radix, because hash
1065 	 * translation does flush when updating the linux pte.
1066 	 */
1067 	if (is_pte_rw_upgrade(pte_val(old_pte), pte_val(pte)) &&
1068 	    (atomic_read(&mm->context.copros) > 0))
1069 		radix__flush_tlb_page(vma, addr);
1070 
1071 	set_pte_at(mm, addr, ptep, pte);
1072 }
1073 
1074 int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
1075 {
1076 	pte_t *ptep = (pte_t *)pud;
1077 	pte_t new_pud = pfn_pte(__phys_to_pfn(addr), prot);
1078 
1079 	if (!radix_enabled())
1080 		return 0;
1081 
1082 	set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pud);
1083 
1084 	return 1;
1085 }
1086 
1087 int pud_clear_huge(pud_t *pud)
1088 {
1089 	if (pud_huge(*pud)) {
1090 		pud_clear(pud);
1091 		return 1;
1092 	}
1093 
1094 	return 0;
1095 }
1096 
1097 int pud_free_pmd_page(pud_t *pud, unsigned long addr)
1098 {
1099 	pmd_t *pmd;
1100 	int i;
1101 
1102 	pmd = (pmd_t *)pud_page_vaddr(*pud);
1103 	pud_clear(pud);
1104 
1105 	flush_tlb_kernel_range(addr, addr + PUD_SIZE);
1106 
1107 	for (i = 0; i < PTRS_PER_PMD; i++) {
1108 		if (!pmd_none(pmd[i])) {
1109 			pte_t *pte;
1110 			pte = (pte_t *)pmd_page_vaddr(pmd[i]);
1111 
1112 			pte_free_kernel(&init_mm, pte);
1113 		}
1114 	}
1115 
1116 	pmd_free(&init_mm, pmd);
1117 
1118 	return 1;
1119 }
1120 
1121 int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
1122 {
1123 	pte_t *ptep = (pte_t *)pmd;
1124 	pte_t new_pmd = pfn_pte(__phys_to_pfn(addr), prot);
1125 
1126 	if (!radix_enabled())
1127 		return 0;
1128 
1129 	set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pmd);
1130 
1131 	return 1;
1132 }
1133 
1134 int pmd_clear_huge(pmd_t *pmd)
1135 {
1136 	if (pmd_huge(*pmd)) {
1137 		pmd_clear(pmd);
1138 		return 1;
1139 	}
1140 
1141 	return 0;
1142 }
1143 
1144 int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
1145 {
1146 	pte_t *pte;
1147 
1148 	pte = (pte_t *)pmd_page_vaddr(*pmd);
1149 	pmd_clear(pmd);
1150 
1151 	flush_tlb_kernel_range(addr, addr + PMD_SIZE);
1152 
1153 	pte_free_kernel(&init_mm, pte);
1154 
1155 	return 1;
1156 }
1157