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 	smp_wmb();
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 	smp_wmb();
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 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_P9_RADIX_PREFETCH_BUG)) {
361 		if (!mmu_pid_bits)
362 			mmu_pid_bits = 20;
363 		mmu_base_pid = 1;
364 	} else if (cpu_has_feature(CPU_FTR_HVMODE)) {
365 		if (!mmu_pid_bits)
366 			mmu_pid_bits = 20;
367 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
368 		/*
369 		 * When KVM is possible, we only use the top half of the
370 		 * PID space to avoid collisions between host and guest PIDs
371 		 * which can cause problems due to prefetch when exiting the
372 		 * guest with AIL=3
373 		 */
374 		mmu_base_pid = 1 << (mmu_pid_bits - 1);
375 #else
376 		mmu_base_pid = 1;
377 #endif
378 	} else {
379 		/* The guest uses the bottom half of the PID space */
380 		if (!mmu_pid_bits)
381 			mmu_pid_bits = 19;
382 		mmu_base_pid = 1;
383 	}
384 
385 	/*
386 	 * Allocate Partition table and process table for the
387 	 * host.
388 	 */
389 	BUG_ON(PRTB_SIZE_SHIFT > 36);
390 	process_tb = early_alloc_pgtable(1UL << PRTB_SIZE_SHIFT, -1, 0, 0);
391 	/*
392 	 * Fill in the process table.
393 	 */
394 	rts_field = radix__get_tree_size();
395 	process_tb->prtb0 = cpu_to_be64(rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE);
396 
397 	/*
398 	 * The init_mm context is given the first available (non-zero) PID,
399 	 * which is the "guard PID" and contains no page table. PIDR should
400 	 * never be set to zero because that duplicates the kernel address
401 	 * space at the 0x0... offset (quadrant 0)!
402 	 *
403 	 * An arbitrary PID that may later be allocated by the PID allocator
404 	 * for userspace processes must not be used either, because that
405 	 * would cause stale user mappings for that PID on CPUs outside of
406 	 * the TLB invalidation scheme (because it won't be in mm_cpumask).
407 	 *
408 	 * So permanently carve out one PID for the purpose of a guard PID.
409 	 */
410 	init_mm.context.id = mmu_base_pid;
411 	mmu_base_pid++;
412 }
413 
414 static void __init radix_init_partition_table(void)
415 {
416 	unsigned long rts_field, dw0, dw1;
417 
418 	mmu_partition_table_init();
419 	rts_field = radix__get_tree_size();
420 	dw0 = rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE | PATB_HR;
421 	dw1 = __pa(process_tb) | (PRTB_SIZE_SHIFT - 12) | PATB_GR;
422 	mmu_partition_table_set_entry(0, dw0, dw1, false);
423 
424 	pr_info("Initializing Radix MMU\n");
425 }
426 
427 static int __init get_idx_from_shift(unsigned int shift)
428 {
429 	int idx = -1;
430 
431 	switch (shift) {
432 	case 0xc:
433 		idx = MMU_PAGE_4K;
434 		break;
435 	case 0x10:
436 		idx = MMU_PAGE_64K;
437 		break;
438 	case 0x15:
439 		idx = MMU_PAGE_2M;
440 		break;
441 	case 0x1e:
442 		idx = MMU_PAGE_1G;
443 		break;
444 	}
445 	return idx;
446 }
447 
448 static int __init radix_dt_scan_page_sizes(unsigned long node,
449 					   const char *uname, int depth,
450 					   void *data)
451 {
452 	int size = 0;
453 	int shift, idx;
454 	unsigned int ap;
455 	const __be32 *prop;
456 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
457 
458 	/* We are scanning "cpu" nodes only */
459 	if (type == NULL || strcmp(type, "cpu") != 0)
460 		return 0;
461 
462 	/* Find MMU PID size */
463 	prop = of_get_flat_dt_prop(node, "ibm,mmu-pid-bits", &size);
464 	if (prop && size == 4)
465 		mmu_pid_bits = be32_to_cpup(prop);
466 
467 	/* Grab page size encodings */
468 	prop = of_get_flat_dt_prop(node, "ibm,processor-radix-AP-encodings", &size);
469 	if (!prop)
470 		return 0;
471 
472 	pr_info("Page sizes from device-tree:\n");
473 	for (; size >= 4; size -= 4, ++prop) {
474 
475 		struct mmu_psize_def *def;
476 
477 		/* top 3 bit is AP encoding */
478 		shift = be32_to_cpu(prop[0]) & ~(0xe << 28);
479 		ap = be32_to_cpu(prop[0]) >> 29;
480 		pr_info("Page size shift = %d AP=0x%x\n", shift, ap);
481 
482 		idx = get_idx_from_shift(shift);
483 		if (idx < 0)
484 			continue;
485 
486 		def = &mmu_psize_defs[idx];
487 		def->shift = shift;
488 		def->ap  = ap;
489 	}
490 
491 	/* needed ? */
492 	cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
493 	return 1;
494 }
495 
496 #ifdef CONFIG_MEMORY_HOTPLUG
497 static int __init probe_memory_block_size(unsigned long node, const char *uname, int
498 					  depth, void *data)
499 {
500 	unsigned long *mem_block_size = (unsigned long *)data;
501 	const __be32 *prop;
502 	int len;
503 
504 	if (depth != 1)
505 		return 0;
506 
507 	if (strcmp(uname, "ibm,dynamic-reconfiguration-memory"))
508 		return 0;
509 
510 	prop = of_get_flat_dt_prop(node, "ibm,lmb-size", &len);
511 
512 	if (!prop || len < dt_root_size_cells * sizeof(__be32))
513 		/*
514 		 * Nothing in the device tree
515 		 */
516 		*mem_block_size = MIN_MEMORY_BLOCK_SIZE;
517 	else
518 		*mem_block_size = of_read_number(prop, dt_root_size_cells);
519 	return 1;
520 }
521 
522 static unsigned long radix_memory_block_size(void)
523 {
524 	unsigned long mem_block_size = MIN_MEMORY_BLOCK_SIZE;
525 
526 	/*
527 	 * OPAL firmware feature is set by now. Hence we are ok
528 	 * to test OPAL feature.
529 	 */
530 	if (firmware_has_feature(FW_FEATURE_OPAL))
531 		mem_block_size = 1UL * 1024 * 1024 * 1024;
532 	else
533 		of_scan_flat_dt(probe_memory_block_size, &mem_block_size);
534 
535 	return mem_block_size;
536 }
537 
538 #else   /* CONFIG_MEMORY_HOTPLUG */
539 
540 static unsigned long radix_memory_block_size(void)
541 {
542 	return 1UL * 1024 * 1024 * 1024;
543 }
544 
545 #endif /* CONFIG_MEMORY_HOTPLUG */
546 
547 
548 void __init radix__early_init_devtree(void)
549 {
550 	int rc;
551 
552 	/*
553 	 * Try to find the available page sizes in the device-tree
554 	 */
555 	rc = of_scan_flat_dt(radix_dt_scan_page_sizes, NULL);
556 	if (!rc) {
557 		/*
558 		 * No page size details found in device tree.
559 		 * Let's assume we have page 4k and 64k support
560 		 */
561 		mmu_psize_defs[MMU_PAGE_4K].shift = 12;
562 		mmu_psize_defs[MMU_PAGE_4K].ap = 0x0;
563 
564 		mmu_psize_defs[MMU_PAGE_64K].shift = 16;
565 		mmu_psize_defs[MMU_PAGE_64K].ap = 0x5;
566 	}
567 
568 	/*
569 	 * Max mapping size used when mapping pages. We don't use
570 	 * ppc_md.memory_block_size() here because this get called
571 	 * early and we don't have machine probe called yet. Also
572 	 * the pseries implementation only check for ibm,lmb-size.
573 	 * All hypervisor supporting radix do expose that device
574 	 * tree node.
575 	 */
576 	radix_mem_block_size = radix_memory_block_size();
577 	return;
578 }
579 
580 static void radix_init_amor(void)
581 {
582 	/*
583 	* In HV mode, we init AMOR (Authority Mask Override Register) so that
584 	* the hypervisor and guest can setup IAMR (Instruction Authority Mask
585 	* Register), enable key 0 and set it to 1.
586 	*
587 	* AMOR = 0b1100 .... 0000 (Mask for key 0 is 11)
588 	*/
589 	mtspr(SPRN_AMOR, (3ul << 62));
590 }
591 
592 #ifdef CONFIG_PPC_KUEP
593 void setup_kuep(bool disabled)
594 {
595 	if (disabled || !early_radix_enabled())
596 		return;
597 
598 	if (smp_processor_id() == boot_cpuid) {
599 		pr_info("Activating Kernel Userspace Execution Prevention\n");
600 		cur_cpu_spec->mmu_features |= MMU_FTR_KUEP;
601 	}
602 
603 	/*
604 	 * Radix always uses key0 of the IAMR to determine if an access is
605 	 * allowed. We set bit 0 (IBM bit 1) of key0, to prevent instruction
606 	 * fetch.
607 	 */
608 	mtspr(SPRN_IAMR, (1ul << 62));
609 }
610 #endif
611 
612 #ifdef CONFIG_PPC_KUAP
613 void setup_kuap(bool disabled)
614 {
615 	if (disabled || !early_radix_enabled())
616 		return;
617 
618 	if (smp_processor_id() == boot_cpuid) {
619 		pr_info("Activating Kernel Userspace Access Prevention\n");
620 		cur_cpu_spec->mmu_features |= MMU_FTR_RADIX_KUAP;
621 	}
622 
623 	/* Make sure userspace can't change the AMR */
624 	mtspr(SPRN_UAMOR, 0);
625 
626 	/*
627 	 * Set the default kernel AMR values on all cpus.
628 	 */
629 	mtspr(SPRN_AMR, AMR_KUAP_BLOCKED);
630 	isync();
631 }
632 #endif
633 
634 void __init radix__early_init_mmu(void)
635 {
636 	unsigned long lpcr;
637 
638 #ifdef CONFIG_PPC_64K_PAGES
639 	/* PAGE_SIZE mappings */
640 	mmu_virtual_psize = MMU_PAGE_64K;
641 #else
642 	mmu_virtual_psize = MMU_PAGE_4K;
643 #endif
644 
645 #ifdef CONFIG_SPARSEMEM_VMEMMAP
646 	/* vmemmap mapping */
647 	if (mmu_psize_defs[MMU_PAGE_2M].shift) {
648 		/*
649 		 * map vmemmap using 2M if available
650 		 */
651 		mmu_vmemmap_psize = MMU_PAGE_2M;
652 	} else
653 		mmu_vmemmap_psize = mmu_virtual_psize;
654 #endif
655 	/*
656 	 * initialize page table size
657 	 */
658 	__pte_index_size = RADIX_PTE_INDEX_SIZE;
659 	__pmd_index_size = RADIX_PMD_INDEX_SIZE;
660 	__pud_index_size = RADIX_PUD_INDEX_SIZE;
661 	__pgd_index_size = RADIX_PGD_INDEX_SIZE;
662 	__pud_cache_index = RADIX_PUD_INDEX_SIZE;
663 	__pte_table_size = RADIX_PTE_TABLE_SIZE;
664 	__pmd_table_size = RADIX_PMD_TABLE_SIZE;
665 	__pud_table_size = RADIX_PUD_TABLE_SIZE;
666 	__pgd_table_size = RADIX_PGD_TABLE_SIZE;
667 
668 	__pmd_val_bits = RADIX_PMD_VAL_BITS;
669 	__pud_val_bits = RADIX_PUD_VAL_BITS;
670 	__pgd_val_bits = RADIX_PGD_VAL_BITS;
671 
672 	__kernel_virt_start = RADIX_KERN_VIRT_START;
673 	__vmalloc_start = RADIX_VMALLOC_START;
674 	__vmalloc_end = RADIX_VMALLOC_END;
675 	__kernel_io_start = RADIX_KERN_IO_START;
676 	__kernel_io_end = RADIX_KERN_IO_END;
677 	vmemmap = (struct page *)RADIX_VMEMMAP_START;
678 	ioremap_bot = IOREMAP_BASE;
679 
680 #ifdef CONFIG_PCI
681 	pci_io_base = ISA_IO_BASE;
682 #endif
683 	__pte_frag_nr = RADIX_PTE_FRAG_NR;
684 	__pte_frag_size_shift = RADIX_PTE_FRAG_SIZE_SHIFT;
685 	__pmd_frag_nr = RADIX_PMD_FRAG_NR;
686 	__pmd_frag_size_shift = RADIX_PMD_FRAG_SIZE_SHIFT;
687 
688 	radix_init_pgtable();
689 
690 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
691 		lpcr = mfspr(SPRN_LPCR);
692 		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
693 		radix_init_partition_table();
694 		radix_init_amor();
695 	} else {
696 		radix_init_pseries();
697 	}
698 
699 	memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
700 
701 	/* Switch to the guard PID before turning on MMU */
702 	radix__switch_mmu_context(NULL, &init_mm);
703 	tlbiel_all();
704 }
705 
706 void radix__early_init_mmu_secondary(void)
707 {
708 	unsigned long lpcr;
709 	/*
710 	 * update partition table control register and UPRT
711 	 */
712 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
713 		lpcr = mfspr(SPRN_LPCR);
714 		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
715 
716 		set_ptcr_when_no_uv(__pa(partition_tb) |
717 				    (PATB_SIZE_SHIFT - 12));
718 
719 		radix_init_amor();
720 	}
721 
722 	radix__switch_mmu_context(NULL, &init_mm);
723 	tlbiel_all();
724 }
725 
726 void radix__mmu_cleanup_all(void)
727 {
728 	unsigned long lpcr;
729 
730 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
731 		lpcr = mfspr(SPRN_LPCR);
732 		mtspr(SPRN_LPCR, lpcr & ~LPCR_UPRT);
733 		set_ptcr_when_no_uv(0);
734 		powernv_set_nmmu_ptcr(0);
735 		radix__flush_tlb_all();
736 	}
737 }
738 
739 #ifdef CONFIG_MEMORY_HOTPLUG
740 static void free_pte_table(pte_t *pte_start, pmd_t *pmd)
741 {
742 	pte_t *pte;
743 	int i;
744 
745 	for (i = 0; i < PTRS_PER_PTE; i++) {
746 		pte = pte_start + i;
747 		if (!pte_none(*pte))
748 			return;
749 	}
750 
751 	pte_free_kernel(&init_mm, pte_start);
752 	pmd_clear(pmd);
753 }
754 
755 static void free_pmd_table(pmd_t *pmd_start, pud_t *pud)
756 {
757 	pmd_t *pmd;
758 	int i;
759 
760 	for (i = 0; i < PTRS_PER_PMD; i++) {
761 		pmd = pmd_start + i;
762 		if (!pmd_none(*pmd))
763 			return;
764 	}
765 
766 	pmd_free(&init_mm, pmd_start);
767 	pud_clear(pud);
768 }
769 
770 static void free_pud_table(pud_t *pud_start, p4d_t *p4d)
771 {
772 	pud_t *pud;
773 	int i;
774 
775 	for (i = 0; i < PTRS_PER_PUD; i++) {
776 		pud = pud_start + i;
777 		if (!pud_none(*pud))
778 			return;
779 	}
780 
781 	pud_free(&init_mm, pud_start);
782 	p4d_clear(p4d);
783 }
784 
785 static void remove_pte_table(pte_t *pte_start, unsigned long addr,
786 			     unsigned long end)
787 {
788 	unsigned long next;
789 	pte_t *pte;
790 
791 	pte = pte_start + pte_index(addr);
792 	for (; addr < end; addr = next, pte++) {
793 		next = (addr + PAGE_SIZE) & PAGE_MASK;
794 		if (next > end)
795 			next = end;
796 
797 		if (!pte_present(*pte))
798 			continue;
799 
800 		if (!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(next)) {
801 			/*
802 			 * The vmemmap_free() and remove_section_mapping()
803 			 * codepaths call us with aligned addresses.
804 			 */
805 			WARN_ONCE(1, "%s: unaligned range\n", __func__);
806 			continue;
807 		}
808 
809 		pte_clear(&init_mm, addr, pte);
810 	}
811 }
812 
813 static void __meminit remove_pmd_table(pmd_t *pmd_start, unsigned long addr,
814 			     unsigned long end)
815 {
816 	unsigned long next;
817 	pte_t *pte_base;
818 	pmd_t *pmd;
819 
820 	pmd = pmd_start + pmd_index(addr);
821 	for (; addr < end; addr = next, pmd++) {
822 		next = pmd_addr_end(addr, end);
823 
824 		if (!pmd_present(*pmd))
825 			continue;
826 
827 		if (pmd_is_leaf(*pmd)) {
828 			if (!IS_ALIGNED(addr, PMD_SIZE) ||
829 			    !IS_ALIGNED(next, PMD_SIZE)) {
830 				WARN_ONCE(1, "%s: unaligned range\n", __func__);
831 				continue;
832 			}
833 			pte_clear(&init_mm, addr, (pte_t *)pmd);
834 			continue;
835 		}
836 
837 		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
838 		remove_pte_table(pte_base, addr, next);
839 		free_pte_table(pte_base, pmd);
840 	}
841 }
842 
843 static void __meminit remove_pud_table(pud_t *pud_start, unsigned long addr,
844 			     unsigned long end)
845 {
846 	unsigned long next;
847 	pmd_t *pmd_base;
848 	pud_t *pud;
849 
850 	pud = pud_start + pud_index(addr);
851 	for (; addr < end; addr = next, pud++) {
852 		next = pud_addr_end(addr, end);
853 
854 		if (!pud_present(*pud))
855 			continue;
856 
857 		if (pud_is_leaf(*pud)) {
858 			if (!IS_ALIGNED(addr, PUD_SIZE) ||
859 			    !IS_ALIGNED(next, PUD_SIZE)) {
860 				WARN_ONCE(1, "%s: unaligned range\n", __func__);
861 				continue;
862 			}
863 			pte_clear(&init_mm, addr, (pte_t *)pud);
864 			continue;
865 		}
866 
867 		pmd_base = (pmd_t *)pud_page_vaddr(*pud);
868 		remove_pmd_table(pmd_base, addr, next);
869 		free_pmd_table(pmd_base, pud);
870 	}
871 }
872 
873 static void __meminit remove_pagetable(unsigned long start, unsigned long end)
874 {
875 	unsigned long addr, next;
876 	pud_t *pud_base;
877 	pgd_t *pgd;
878 	p4d_t *p4d;
879 
880 	spin_lock(&init_mm.page_table_lock);
881 
882 	for (addr = start; addr < end; addr = next) {
883 		next = pgd_addr_end(addr, end);
884 
885 		pgd = pgd_offset_k(addr);
886 		p4d = p4d_offset(pgd, addr);
887 		if (!p4d_present(*p4d))
888 			continue;
889 
890 		if (p4d_is_leaf(*p4d)) {
891 			if (!IS_ALIGNED(addr, P4D_SIZE) ||
892 			    !IS_ALIGNED(next, P4D_SIZE)) {
893 				WARN_ONCE(1, "%s: unaligned range\n", __func__);
894 				continue;
895 			}
896 
897 			pte_clear(&init_mm, addr, (pte_t *)pgd);
898 			continue;
899 		}
900 
901 		pud_base = (pud_t *)p4d_page_vaddr(*p4d);
902 		remove_pud_table(pud_base, addr, next);
903 		free_pud_table(pud_base, p4d);
904 	}
905 
906 	spin_unlock(&init_mm.page_table_lock);
907 	radix__flush_tlb_kernel_range(start, end);
908 }
909 
910 int __meminit radix__create_section_mapping(unsigned long start,
911 					    unsigned long end, int nid,
912 					    pgprot_t prot)
913 {
914 	if (end >= RADIX_VMALLOC_START) {
915 		pr_warn("Outside the supported range\n");
916 		return -1;
917 	}
918 
919 	return create_physical_mapping(__pa(start), __pa(end),
920 				       radix_mem_block_size, nid, prot);
921 }
922 
923 int __meminit radix__remove_section_mapping(unsigned long start, unsigned long end)
924 {
925 	remove_pagetable(start, end);
926 	return 0;
927 }
928 #endif /* CONFIG_MEMORY_HOTPLUG */
929 
930 #ifdef CONFIG_SPARSEMEM_VMEMMAP
931 static int __map_kernel_page_nid(unsigned long ea, unsigned long pa,
932 				 pgprot_t flags, unsigned int map_page_size,
933 				 int nid)
934 {
935 	return __map_kernel_page(ea, pa, flags, map_page_size, nid, 0, 0);
936 }
937 
938 int __meminit radix__vmemmap_create_mapping(unsigned long start,
939 				      unsigned long page_size,
940 				      unsigned long phys)
941 {
942 	/* Create a PTE encoding */
943 	unsigned long flags = _PAGE_PRESENT | _PAGE_ACCESSED | _PAGE_KERNEL_RW;
944 	int nid = early_pfn_to_nid(phys >> PAGE_SHIFT);
945 	int ret;
946 
947 	if ((start + page_size) >= RADIX_VMEMMAP_END) {
948 		pr_warn("Outside the supported range\n");
949 		return -1;
950 	}
951 
952 	ret = __map_kernel_page_nid(start, phys, __pgprot(flags), page_size, nid);
953 	BUG_ON(ret);
954 
955 	return 0;
956 }
957 
958 #ifdef CONFIG_MEMORY_HOTPLUG
959 void __meminit radix__vmemmap_remove_mapping(unsigned long start, unsigned long page_size)
960 {
961 	remove_pagetable(start, start + page_size);
962 }
963 #endif
964 #endif
965 
966 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
967 
968 unsigned long radix__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
969 				  pmd_t *pmdp, unsigned long clr,
970 				  unsigned long set)
971 {
972 	unsigned long old;
973 
974 #ifdef CONFIG_DEBUG_VM
975 	WARN_ON(!radix__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
976 	assert_spin_locked(pmd_lockptr(mm, pmdp));
977 #endif
978 
979 	old = radix__pte_update(mm, addr, (pte_t *)pmdp, clr, set, 1);
980 	trace_hugepage_update(addr, old, clr, set);
981 
982 	return old;
983 }
984 
985 pmd_t radix__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
986 			pmd_t *pmdp)
987 
988 {
989 	pmd_t pmd;
990 
991 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
992 	VM_BUG_ON(radix__pmd_trans_huge(*pmdp));
993 	VM_BUG_ON(pmd_devmap(*pmdp));
994 	/*
995 	 * khugepaged calls this for normal pmd
996 	 */
997 	pmd = *pmdp;
998 	pmd_clear(pmdp);
999 
1000 	/*
1001 	 * pmdp collapse_flush need to ensure that there are no parallel gup
1002 	 * walk after this call. This is needed so that we can have stable
1003 	 * page ref count when collapsing a page. We don't allow a collapse page
1004 	 * if we have gup taken on the page. We can ensure that by sending IPI
1005 	 * because gup walk happens with IRQ disabled.
1006 	 */
1007 	serialize_against_pte_lookup(vma->vm_mm);
1008 
1009 	radix__flush_tlb_collapsed_pmd(vma->vm_mm, address);
1010 
1011 	return pmd;
1012 }
1013 
1014 /*
1015  * For us pgtable_t is pte_t *. Inorder to save the deposisted
1016  * page table, we consider the allocated page table as a list
1017  * head. On withdraw we need to make sure we zero out the used
1018  * list_head memory area.
1019  */
1020 void radix__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
1021 				 pgtable_t pgtable)
1022 {
1023 	struct list_head *lh = (struct list_head *) pgtable;
1024 
1025 	assert_spin_locked(pmd_lockptr(mm, pmdp));
1026 
1027 	/* FIFO */
1028 	if (!pmd_huge_pte(mm, pmdp))
1029 		INIT_LIST_HEAD(lh);
1030 	else
1031 		list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
1032 	pmd_huge_pte(mm, pmdp) = pgtable;
1033 }
1034 
1035 pgtable_t radix__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
1036 {
1037 	pte_t *ptep;
1038 	pgtable_t pgtable;
1039 	struct list_head *lh;
1040 
1041 	assert_spin_locked(pmd_lockptr(mm, pmdp));
1042 
1043 	/* FIFO */
1044 	pgtable = pmd_huge_pte(mm, pmdp);
1045 	lh = (struct list_head *) pgtable;
1046 	if (list_empty(lh))
1047 		pmd_huge_pte(mm, pmdp) = NULL;
1048 	else {
1049 		pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
1050 		list_del(lh);
1051 	}
1052 	ptep = (pte_t *) pgtable;
1053 	*ptep = __pte(0);
1054 	ptep++;
1055 	*ptep = __pte(0);
1056 	return pgtable;
1057 }
1058 
1059 pmd_t radix__pmdp_huge_get_and_clear(struct mm_struct *mm,
1060 				     unsigned long addr, pmd_t *pmdp)
1061 {
1062 	pmd_t old_pmd;
1063 	unsigned long old;
1064 
1065 	old = radix__pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
1066 	old_pmd = __pmd(old);
1067 	return old_pmd;
1068 }
1069 
1070 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1071 
1072 void radix__ptep_set_access_flags(struct vm_area_struct *vma, pte_t *ptep,
1073 				  pte_t entry, unsigned long address, int psize)
1074 {
1075 	struct mm_struct *mm = vma->vm_mm;
1076 	unsigned long set = pte_val(entry) & (_PAGE_DIRTY | _PAGE_ACCESSED |
1077 					      _PAGE_RW | _PAGE_EXEC);
1078 
1079 	unsigned long change = pte_val(entry) ^ pte_val(*ptep);
1080 	/*
1081 	 * To avoid NMMU hang while relaxing access, we need mark
1082 	 * the pte invalid in between.
1083 	 */
1084 	if ((change & _PAGE_RW) && atomic_read(&mm->context.copros) > 0) {
1085 		unsigned long old_pte, new_pte;
1086 
1087 		old_pte = __radix_pte_update(ptep, _PAGE_PRESENT, _PAGE_INVALID);
1088 		/*
1089 		 * new value of pte
1090 		 */
1091 		new_pte = old_pte | set;
1092 		radix__flush_tlb_page_psize(mm, address, psize);
1093 		__radix_pte_update(ptep, _PAGE_INVALID, new_pte);
1094 	} else {
1095 		__radix_pte_update(ptep, 0, set);
1096 		/*
1097 		 * Book3S does not require a TLB flush when relaxing access
1098 		 * restrictions when the address space is not attached to a
1099 		 * NMMU, because the core MMU will reload the pte after taking
1100 		 * an access fault, which is defined by the architectue.
1101 		 */
1102 	}
1103 	/* See ptesync comment in radix__set_pte_at */
1104 }
1105 
1106 void radix__ptep_modify_prot_commit(struct vm_area_struct *vma,
1107 				    unsigned long addr, pte_t *ptep,
1108 				    pte_t old_pte, pte_t pte)
1109 {
1110 	struct mm_struct *mm = vma->vm_mm;
1111 
1112 	/*
1113 	 * To avoid NMMU hang while relaxing access we need to flush the tlb before
1114 	 * we set the new value. We need to do this only for radix, because hash
1115 	 * translation does flush when updating the linux pte.
1116 	 */
1117 	if (is_pte_rw_upgrade(pte_val(old_pte), pte_val(pte)) &&
1118 	    (atomic_read(&mm->context.copros) > 0))
1119 		radix__flush_tlb_page(vma, addr);
1120 
1121 	set_pte_at(mm, addr, ptep, pte);
1122 }
1123 
1124 int __init arch_ioremap_pud_supported(void)
1125 {
1126 	/* HPT does not cope with large pages in the vmalloc area */
1127 	return radix_enabled();
1128 }
1129 
1130 int __init arch_ioremap_pmd_supported(void)
1131 {
1132 	return radix_enabled();
1133 }
1134 
1135 int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
1136 {
1137 	return 0;
1138 }
1139 
1140 int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
1141 {
1142 	pte_t *ptep = (pte_t *)pud;
1143 	pte_t new_pud = pfn_pte(__phys_to_pfn(addr), prot);
1144 
1145 	if (!radix_enabled())
1146 		return 0;
1147 
1148 	set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pud);
1149 
1150 	return 1;
1151 }
1152 
1153 int pud_clear_huge(pud_t *pud)
1154 {
1155 	if (pud_huge(*pud)) {
1156 		pud_clear(pud);
1157 		return 1;
1158 	}
1159 
1160 	return 0;
1161 }
1162 
1163 int pud_free_pmd_page(pud_t *pud, unsigned long addr)
1164 {
1165 	pmd_t *pmd;
1166 	int i;
1167 
1168 	pmd = (pmd_t *)pud_page_vaddr(*pud);
1169 	pud_clear(pud);
1170 
1171 	flush_tlb_kernel_range(addr, addr + PUD_SIZE);
1172 
1173 	for (i = 0; i < PTRS_PER_PMD; i++) {
1174 		if (!pmd_none(pmd[i])) {
1175 			pte_t *pte;
1176 			pte = (pte_t *)pmd_page_vaddr(pmd[i]);
1177 
1178 			pte_free_kernel(&init_mm, pte);
1179 		}
1180 	}
1181 
1182 	pmd_free(&init_mm, pmd);
1183 
1184 	return 1;
1185 }
1186 
1187 int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
1188 {
1189 	pte_t *ptep = (pte_t *)pmd;
1190 	pte_t new_pmd = pfn_pte(__phys_to_pfn(addr), prot);
1191 
1192 	if (!radix_enabled())
1193 		return 0;
1194 
1195 	set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pmd);
1196 
1197 	return 1;
1198 }
1199 
1200 int pmd_clear_huge(pmd_t *pmd)
1201 {
1202 	if (pmd_huge(*pmd)) {
1203 		pmd_clear(pmd);
1204 		return 1;
1205 	}
1206 
1207 	return 0;
1208 }
1209 
1210 int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
1211 {
1212 	pte_t *pte;
1213 
1214 	pte = (pte_t *)pmd_page_vaddr(*pmd);
1215 	pmd_clear(pmd);
1216 
1217 	flush_tlb_kernel_range(addr, addr + PMD_SIZE);
1218 
1219 	pte_free_kernel(&init_mm, pte);
1220 
1221 	return 1;
1222 }
1223 
1224 int __init arch_ioremap_p4d_supported(void)
1225 {
1226 	return 0;
1227 }
1228