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/string_helpers.h>
17 #include <linux/stop_machine.h>
18 
19 #include <asm/pgtable.h>
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/trace.h>
29 #include <asm/uaccess.h>
30 
31 #include <trace/events/thp.h>
32 
33 unsigned int mmu_pid_bits;
34 unsigned int mmu_base_pid;
35 
36 static int native_register_process_table(unsigned long base, unsigned long pg_sz,
37 					 unsigned long table_size)
38 {
39 	unsigned long patb0, patb1;
40 
41 	patb0 = be64_to_cpu(partition_tb[0].patb0);
42 	patb1 = base | table_size | PATB_GR;
43 
44 	mmu_partition_table_set_entry(0, patb0, patb1);
45 
46 	return 0;
47 }
48 
49 static __ref void *early_alloc_pgtable(unsigned long size, int nid,
50 			unsigned long region_start, unsigned long region_end)
51 {
52 	phys_addr_t min_addr = MEMBLOCK_LOW_LIMIT;
53 	phys_addr_t max_addr = MEMBLOCK_ALLOC_ANYWHERE;
54 	void *ptr;
55 
56 	if (region_start)
57 		min_addr = region_start;
58 	if (region_end)
59 		max_addr = region_end;
60 
61 	ptr = memblock_alloc_try_nid(size, size, min_addr, max_addr, nid);
62 
63 	if (!ptr)
64 		panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa max_addr=%pa\n",
65 		      __func__, size, size, nid, &min_addr, &max_addr);
66 
67 	return ptr;
68 }
69 
70 static int early_map_kernel_page(unsigned long ea, unsigned long pa,
71 			  pgprot_t flags,
72 			  unsigned int map_page_size,
73 			  int nid,
74 			  unsigned long region_start, unsigned long region_end)
75 {
76 	unsigned long pfn = pa >> PAGE_SHIFT;
77 	pgd_t *pgdp;
78 	pud_t *pudp;
79 	pmd_t *pmdp;
80 	pte_t *ptep;
81 
82 	pgdp = pgd_offset_k(ea);
83 	if (pgd_none(*pgdp)) {
84 		pudp = early_alloc_pgtable(PUD_TABLE_SIZE, nid,
85 						region_start, region_end);
86 		pgd_populate(&init_mm, pgdp, pudp);
87 	}
88 	pudp = pud_offset(pgdp, ea);
89 	if (map_page_size == PUD_SIZE) {
90 		ptep = (pte_t *)pudp;
91 		goto set_the_pte;
92 	}
93 	if (pud_none(*pudp)) {
94 		pmdp = early_alloc_pgtable(PMD_TABLE_SIZE, nid,
95 						region_start, region_end);
96 		pud_populate(&init_mm, pudp, pmdp);
97 	}
98 	pmdp = pmd_offset(pudp, ea);
99 	if (map_page_size == PMD_SIZE) {
100 		ptep = pmdp_ptep(pmdp);
101 		goto set_the_pte;
102 	}
103 	if (!pmd_present(*pmdp)) {
104 		ptep = early_alloc_pgtable(PAGE_SIZE, nid,
105 						region_start, region_end);
106 		pmd_populate_kernel(&init_mm, pmdp, ptep);
107 	}
108 	ptep = pte_offset_kernel(pmdp, ea);
109 
110 set_the_pte:
111 	set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
112 	smp_wmb();
113 	return 0;
114 }
115 
116 /*
117  * nid, region_start, and region_end are hints to try to place the page
118  * table memory in the same node or region.
119  */
120 static int __map_kernel_page(unsigned long ea, unsigned long pa,
121 			  pgprot_t flags,
122 			  unsigned int map_page_size,
123 			  int nid,
124 			  unsigned long region_start, unsigned long region_end)
125 {
126 	unsigned long pfn = pa >> PAGE_SHIFT;
127 	pgd_t *pgdp;
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 	pudp = pud_alloc(&init_mm, pgdp, ea);
151 	if (!pudp)
152 		return -ENOMEM;
153 	if (map_page_size == PUD_SIZE) {
154 		ptep = (pte_t *)pudp;
155 		goto set_the_pte;
156 	}
157 	pmdp = pmd_alloc(&init_mm, pudp, ea);
158 	if (!pmdp)
159 		return -ENOMEM;
160 	if (map_page_size == PMD_SIZE) {
161 		ptep = pmdp_ptep(pmdp);
162 		goto set_the_pte;
163 	}
164 	ptep = pte_alloc_kernel(pmdp, ea);
165 	if (!ptep)
166 		return -ENOMEM;
167 
168 set_the_pte:
169 	set_pte_at(&init_mm, ea, ptep, pfn_pte(pfn, flags));
170 	smp_wmb();
171 	return 0;
172 }
173 
174 int radix__map_kernel_page(unsigned long ea, unsigned long pa,
175 			  pgprot_t flags,
176 			  unsigned int map_page_size)
177 {
178 	return __map_kernel_page(ea, pa, flags, map_page_size, -1, 0, 0);
179 }
180 
181 #ifdef CONFIG_STRICT_KERNEL_RWX
182 void radix__change_memory_range(unsigned long start, unsigned long end,
183 				unsigned long clear)
184 {
185 	unsigned long idx;
186 	pgd_t *pgdp;
187 	pud_t *pudp;
188 	pmd_t *pmdp;
189 	pte_t *ptep;
190 
191 	start = ALIGN_DOWN(start, PAGE_SIZE);
192 	end = PAGE_ALIGN(end); // aligns up
193 
194 	pr_debug("Changing flags on range %lx-%lx removing 0x%lx\n",
195 		 start, end, clear);
196 
197 	for (idx = start; idx < end; idx += PAGE_SIZE) {
198 		pgdp = pgd_offset_k(idx);
199 		pudp = pud_alloc(&init_mm, pgdp, idx);
200 		if (!pudp)
201 			continue;
202 		if (pud_is_leaf(*pudp)) {
203 			ptep = (pte_t *)pudp;
204 			goto update_the_pte;
205 		}
206 		pmdp = pmd_alloc(&init_mm, pudp, idx);
207 		if (!pmdp)
208 			continue;
209 		if (pmd_is_leaf(*pmdp)) {
210 			ptep = pmdp_ptep(pmdp);
211 			goto update_the_pte;
212 		}
213 		ptep = pte_alloc_kernel(pmdp, idx);
214 		if (!ptep)
215 			continue;
216 update_the_pte:
217 		radix__pte_update(&init_mm, idx, ptep, clear, 0, 0);
218 	}
219 
220 	radix__flush_tlb_kernel_range(start, end);
221 }
222 
223 void radix__mark_rodata_ro(void)
224 {
225 	unsigned long start, end;
226 
227 	start = (unsigned long)_stext;
228 	end = (unsigned long)__init_begin;
229 
230 	radix__change_memory_range(start, end, _PAGE_WRITE);
231 }
232 
233 void radix__mark_initmem_nx(void)
234 {
235 	unsigned long start = (unsigned long)__init_begin;
236 	unsigned long end = (unsigned long)__init_end;
237 
238 	radix__change_memory_range(start, end, _PAGE_EXEC);
239 }
240 #endif /* CONFIG_STRICT_KERNEL_RWX */
241 
242 static inline void __meminit
243 print_mapping(unsigned long start, unsigned long end, unsigned long size, bool exec)
244 {
245 	char buf[10];
246 
247 	if (end <= start)
248 		return;
249 
250 	string_get_size(size, 1, STRING_UNITS_2, buf, sizeof(buf));
251 
252 	pr_info("Mapped 0x%016lx-0x%016lx with %s pages%s\n", start, end, buf,
253 		exec ? " (exec)" : "");
254 }
255 
256 static unsigned long next_boundary(unsigned long addr, unsigned long end)
257 {
258 #ifdef CONFIG_STRICT_KERNEL_RWX
259 	if (addr < __pa_symbol(__init_begin))
260 		return __pa_symbol(__init_begin);
261 #endif
262 	return end;
263 }
264 
265 static int __meminit create_physical_mapping(unsigned long start,
266 					     unsigned long end,
267 					     int nid)
268 {
269 	unsigned long vaddr, addr, mapping_size = 0;
270 	bool prev_exec, exec = false;
271 	pgprot_t prot;
272 	int psize;
273 
274 	start = _ALIGN_UP(start, PAGE_SIZE);
275 	for (addr = start; addr < end; addr += mapping_size) {
276 		unsigned long gap, previous_size;
277 		int rc;
278 
279 		gap = next_boundary(addr, end) - addr;
280 		previous_size = mapping_size;
281 		prev_exec = exec;
282 
283 		if (IS_ALIGNED(addr, PUD_SIZE) && gap >= PUD_SIZE &&
284 		    mmu_psize_defs[MMU_PAGE_1G].shift) {
285 			mapping_size = PUD_SIZE;
286 			psize = MMU_PAGE_1G;
287 		} else if (IS_ALIGNED(addr, PMD_SIZE) && gap >= PMD_SIZE &&
288 			   mmu_psize_defs[MMU_PAGE_2M].shift) {
289 			mapping_size = PMD_SIZE;
290 			psize = MMU_PAGE_2M;
291 		} else {
292 			mapping_size = PAGE_SIZE;
293 			psize = mmu_virtual_psize;
294 		}
295 
296 		vaddr = (unsigned long)__va(addr);
297 
298 		if (overlaps_kernel_text(vaddr, vaddr + mapping_size) ||
299 		    overlaps_interrupt_vector_text(vaddr, vaddr + mapping_size)) {
300 			prot = PAGE_KERNEL_X;
301 			exec = true;
302 		} else {
303 			prot = PAGE_KERNEL;
304 			exec = false;
305 		}
306 
307 		if (mapping_size != previous_size || exec != prev_exec) {
308 			print_mapping(start, addr, previous_size, prev_exec);
309 			start = addr;
310 		}
311 
312 		rc = __map_kernel_page(vaddr, addr, prot, mapping_size, nid, start, end);
313 		if (rc)
314 			return rc;
315 
316 		update_page_count(psize, 1);
317 	}
318 
319 	print_mapping(start, addr, mapping_size, exec);
320 	return 0;
321 }
322 
323 static void __init radix_init_pgtable(void)
324 {
325 	unsigned long rts_field;
326 	struct memblock_region *reg;
327 
328 	/* We don't support slb for radix */
329 	mmu_slb_size = 0;
330 	/*
331 	 * Create the linear mapping, using standard page size for now
332 	 */
333 	for_each_memblock(memory, reg) {
334 		/*
335 		 * The memblock allocator  is up at this point, so the
336 		 * page tables will be allocated within the range. No
337 		 * need or a node (which we don't have yet).
338 		 */
339 
340 		if ((reg->base + reg->size) >= RADIX_VMALLOC_START) {
341 			pr_warn("Outside the supported range\n");
342 			continue;
343 		}
344 
345 		WARN_ON(create_physical_mapping(reg->base,
346 						reg->base + reg->size,
347 						-1));
348 	}
349 
350 	/* Find out how many PID bits are supported */
351 	if (cpu_has_feature(CPU_FTR_HVMODE)) {
352 		if (!mmu_pid_bits)
353 			mmu_pid_bits = 20;
354 #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
355 		/*
356 		 * When KVM is possible, we only use the top half of the
357 		 * PID space to avoid collisions between host and guest PIDs
358 		 * which can cause problems due to prefetch when exiting the
359 		 * guest with AIL=3
360 		 */
361 		mmu_base_pid = 1 << (mmu_pid_bits - 1);
362 #else
363 		mmu_base_pid = 1;
364 #endif
365 	} else {
366 		/* The guest uses the bottom half of the PID space */
367 		if (!mmu_pid_bits)
368 			mmu_pid_bits = 19;
369 		mmu_base_pid = 1;
370 	}
371 
372 	/*
373 	 * Allocate Partition table and process table for the
374 	 * host.
375 	 */
376 	BUG_ON(PRTB_SIZE_SHIFT > 36);
377 	process_tb = early_alloc_pgtable(1UL << PRTB_SIZE_SHIFT, -1, 0, 0);
378 	/*
379 	 * Fill in the process table.
380 	 */
381 	rts_field = radix__get_tree_size();
382 	process_tb->prtb0 = cpu_to_be64(rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE);
383 	/*
384 	 * Fill in the partition table. We are suppose to use effective address
385 	 * of process table here. But our linear mapping also enable us to use
386 	 * physical address here.
387 	 */
388 	register_process_table(__pa(process_tb), 0, PRTB_SIZE_SHIFT - 12);
389 	pr_info("Process table %p and radix root for kernel: %p\n", process_tb, init_mm.pgd);
390 	asm volatile("ptesync" : : : "memory");
391 	asm volatile(PPC_TLBIE_5(%0,%1,2,1,1) : :
392 		     "r" (TLBIEL_INVAL_SET_LPID), "r" (0));
393 	asm volatile("eieio; tlbsync; ptesync" : : : "memory");
394 	trace_tlbie(0, 0, TLBIEL_INVAL_SET_LPID, 0, 2, 1, 1);
395 
396 	/*
397 	 * The init_mm context is given the first available (non-zero) PID,
398 	 * which is the "guard PID" and contains no page table. PIDR should
399 	 * never be set to zero because that duplicates the kernel address
400 	 * space at the 0x0... offset (quadrant 0)!
401 	 *
402 	 * An arbitrary PID that may later be allocated by the PID allocator
403 	 * for userspace processes must not be used either, because that
404 	 * would cause stale user mappings for that PID on CPUs outside of
405 	 * the TLB invalidation scheme (because it won't be in mm_cpumask).
406 	 *
407 	 * So permanently carve out one PID for the purpose of a guard PID.
408 	 */
409 	init_mm.context.id = mmu_base_pid;
410 	mmu_base_pid++;
411 }
412 
413 static void __init radix_init_partition_table(void)
414 {
415 	unsigned long rts_field, dw0;
416 
417 	mmu_partition_table_init();
418 	rts_field = radix__get_tree_size();
419 	dw0 = rts_field | __pa(init_mm.pgd) | RADIX_PGD_INDEX_SIZE | PATB_HR;
420 	mmu_partition_table_set_entry(0, dw0, 0);
421 
422 	pr_info("Initializing Radix MMU\n");
423 	pr_info("Partition table %p\n", partition_tb);
424 }
425 
426 void __init radix_init_native(void)
427 {
428 	register_process_table = native_register_process_table;
429 }
430 
431 static int __init get_idx_from_shift(unsigned int shift)
432 {
433 	int idx = -1;
434 
435 	switch (shift) {
436 	case 0xc:
437 		idx = MMU_PAGE_4K;
438 		break;
439 	case 0x10:
440 		idx = MMU_PAGE_64K;
441 		break;
442 	case 0x15:
443 		idx = MMU_PAGE_2M;
444 		break;
445 	case 0x1e:
446 		idx = MMU_PAGE_1G;
447 		break;
448 	}
449 	return idx;
450 }
451 
452 static int __init radix_dt_scan_page_sizes(unsigned long node,
453 					   const char *uname, int depth,
454 					   void *data)
455 {
456 	int size = 0;
457 	int shift, idx;
458 	unsigned int ap;
459 	const __be32 *prop;
460 	const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
461 
462 	/* We are scanning "cpu" nodes only */
463 	if (type == NULL || strcmp(type, "cpu") != 0)
464 		return 0;
465 
466 	/* Find MMU PID size */
467 	prop = of_get_flat_dt_prop(node, "ibm,mmu-pid-bits", &size);
468 	if (prop && size == 4)
469 		mmu_pid_bits = be32_to_cpup(prop);
470 
471 	/* Grab page size encodings */
472 	prop = of_get_flat_dt_prop(node, "ibm,processor-radix-AP-encodings", &size);
473 	if (!prop)
474 		return 0;
475 
476 	pr_info("Page sizes from device-tree:\n");
477 	for (; size >= 4; size -= 4, ++prop) {
478 
479 		struct mmu_psize_def *def;
480 
481 		/* top 3 bit is AP encoding */
482 		shift = be32_to_cpu(prop[0]) & ~(0xe << 28);
483 		ap = be32_to_cpu(prop[0]) >> 29;
484 		pr_info("Page size shift = %d AP=0x%x\n", shift, ap);
485 
486 		idx = get_idx_from_shift(shift);
487 		if (idx < 0)
488 			continue;
489 
490 		def = &mmu_psize_defs[idx];
491 		def->shift = shift;
492 		def->ap  = ap;
493 	}
494 
495 	/* needed ? */
496 	cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
497 	return 1;
498 }
499 
500 void __init radix__early_init_devtree(void)
501 {
502 	int rc;
503 
504 	/*
505 	 * Try to find the available page sizes in the device-tree
506 	 */
507 	rc = of_scan_flat_dt(radix_dt_scan_page_sizes, NULL);
508 	if (rc != 0)  /* Found */
509 		goto found;
510 	/*
511 	 * let's assume we have page 4k and 64k support
512 	 */
513 	mmu_psize_defs[MMU_PAGE_4K].shift = 12;
514 	mmu_psize_defs[MMU_PAGE_4K].ap = 0x0;
515 
516 	mmu_psize_defs[MMU_PAGE_64K].shift = 16;
517 	mmu_psize_defs[MMU_PAGE_64K].ap = 0x5;
518 found:
519 	return;
520 }
521 
522 static void radix_init_amor(void)
523 {
524 	/*
525 	* In HV mode, we init AMOR (Authority Mask Override Register) so that
526 	* the hypervisor and guest can setup IAMR (Instruction Authority Mask
527 	* Register), enable key 0 and set it to 1.
528 	*
529 	* AMOR = 0b1100 .... 0000 (Mask for key 0 is 11)
530 	*/
531 	mtspr(SPRN_AMOR, (3ul << 62));
532 }
533 
534 #ifdef CONFIG_PPC_KUEP
535 void setup_kuep(bool disabled)
536 {
537 	if (disabled || !early_radix_enabled())
538 		return;
539 
540 	if (smp_processor_id() == boot_cpuid)
541 		pr_info("Activating Kernel Userspace Execution Prevention\n");
542 
543 	/*
544 	 * Radix always uses key0 of the IAMR to determine if an access is
545 	 * allowed. We set bit 0 (IBM bit 1) of key0, to prevent instruction
546 	 * fetch.
547 	 */
548 	mtspr(SPRN_IAMR, (1ul << 62));
549 }
550 #endif
551 
552 #ifdef CONFIG_PPC_KUAP
553 void setup_kuap(bool disabled)
554 {
555 	if (disabled || !early_radix_enabled())
556 		return;
557 
558 	if (smp_processor_id() == boot_cpuid) {
559 		pr_info("Activating Kernel Userspace Access Prevention\n");
560 		cur_cpu_spec->mmu_features |= MMU_FTR_RADIX_KUAP;
561 	}
562 
563 	/* Make sure userspace can't change the AMR */
564 	mtspr(SPRN_UAMOR, 0);
565 	mtspr(SPRN_AMR, AMR_KUAP_BLOCKED);
566 	isync();
567 }
568 #endif
569 
570 void __init radix__early_init_mmu(void)
571 {
572 	unsigned long lpcr;
573 
574 #ifdef CONFIG_PPC_64K_PAGES
575 	/* PAGE_SIZE mappings */
576 	mmu_virtual_psize = MMU_PAGE_64K;
577 #else
578 	mmu_virtual_psize = MMU_PAGE_4K;
579 #endif
580 
581 #ifdef CONFIG_SPARSEMEM_VMEMMAP
582 	/* vmemmap mapping */
583 	if (mmu_psize_defs[MMU_PAGE_2M].shift) {
584 		/*
585 		 * map vmemmap using 2M if available
586 		 */
587 		mmu_vmemmap_psize = MMU_PAGE_2M;
588 	} else
589 		mmu_vmemmap_psize = mmu_virtual_psize;
590 #endif
591 	/*
592 	 * initialize page table size
593 	 */
594 	__pte_index_size = RADIX_PTE_INDEX_SIZE;
595 	__pmd_index_size = RADIX_PMD_INDEX_SIZE;
596 	__pud_index_size = RADIX_PUD_INDEX_SIZE;
597 	__pgd_index_size = RADIX_PGD_INDEX_SIZE;
598 	__pud_cache_index = RADIX_PUD_INDEX_SIZE;
599 	__pte_table_size = RADIX_PTE_TABLE_SIZE;
600 	__pmd_table_size = RADIX_PMD_TABLE_SIZE;
601 	__pud_table_size = RADIX_PUD_TABLE_SIZE;
602 	__pgd_table_size = RADIX_PGD_TABLE_SIZE;
603 
604 	__pmd_val_bits = RADIX_PMD_VAL_BITS;
605 	__pud_val_bits = RADIX_PUD_VAL_BITS;
606 	__pgd_val_bits = RADIX_PGD_VAL_BITS;
607 
608 	__kernel_virt_start = RADIX_KERN_VIRT_START;
609 	__vmalloc_start = RADIX_VMALLOC_START;
610 	__vmalloc_end = RADIX_VMALLOC_END;
611 	__kernel_io_start = RADIX_KERN_IO_START;
612 	__kernel_io_end = RADIX_KERN_IO_END;
613 	vmemmap = (struct page *)RADIX_VMEMMAP_START;
614 	ioremap_bot = IOREMAP_BASE;
615 
616 #ifdef CONFIG_PCI
617 	pci_io_base = ISA_IO_BASE;
618 #endif
619 	__pte_frag_nr = RADIX_PTE_FRAG_NR;
620 	__pte_frag_size_shift = RADIX_PTE_FRAG_SIZE_SHIFT;
621 	__pmd_frag_nr = RADIX_PMD_FRAG_NR;
622 	__pmd_frag_size_shift = RADIX_PMD_FRAG_SIZE_SHIFT;
623 
624 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
625 		radix_init_native();
626 		lpcr = mfspr(SPRN_LPCR);
627 		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
628 		radix_init_partition_table();
629 		radix_init_amor();
630 	} else {
631 		radix_init_pseries();
632 	}
633 
634 	memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
635 
636 	radix_init_pgtable();
637 	/* Switch to the guard PID before turning on MMU */
638 	radix__switch_mmu_context(NULL, &init_mm);
639 	if (cpu_has_feature(CPU_FTR_HVMODE))
640 		tlbiel_all();
641 }
642 
643 void radix__early_init_mmu_secondary(void)
644 {
645 	unsigned long lpcr;
646 	/*
647 	 * update partition table control register and UPRT
648 	 */
649 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
650 		lpcr = mfspr(SPRN_LPCR);
651 		mtspr(SPRN_LPCR, lpcr | LPCR_UPRT | LPCR_HR);
652 
653 		mtspr(SPRN_PTCR,
654 		      __pa(partition_tb) | (PATB_SIZE_SHIFT - 12));
655 		radix_init_amor();
656 	}
657 
658 	radix__switch_mmu_context(NULL, &init_mm);
659 	if (cpu_has_feature(CPU_FTR_HVMODE))
660 		tlbiel_all();
661 }
662 
663 void radix__mmu_cleanup_all(void)
664 {
665 	unsigned long lpcr;
666 
667 	if (!firmware_has_feature(FW_FEATURE_LPAR)) {
668 		lpcr = mfspr(SPRN_LPCR);
669 		mtspr(SPRN_LPCR, lpcr & ~LPCR_UPRT);
670 		mtspr(SPRN_PTCR, 0);
671 		powernv_set_nmmu_ptcr(0);
672 		radix__flush_tlb_all();
673 	}
674 }
675 
676 void radix__setup_initial_memory_limit(phys_addr_t first_memblock_base,
677 				phys_addr_t first_memblock_size)
678 {
679 	/*
680 	 * We don't currently support the first MEMBLOCK not mapping 0
681 	 * physical on those processors
682 	 */
683 	BUG_ON(first_memblock_base != 0);
684 
685 	/*
686 	 * Radix mode is not limited by RMA / VRMA addressing.
687 	 */
688 	ppc64_rma_size = ULONG_MAX;
689 }
690 
691 #ifdef CONFIG_MEMORY_HOTPLUG
692 static void free_pte_table(pte_t *pte_start, pmd_t *pmd)
693 {
694 	pte_t *pte;
695 	int i;
696 
697 	for (i = 0; i < PTRS_PER_PTE; i++) {
698 		pte = pte_start + i;
699 		if (!pte_none(*pte))
700 			return;
701 	}
702 
703 	pte_free_kernel(&init_mm, pte_start);
704 	pmd_clear(pmd);
705 }
706 
707 static void free_pmd_table(pmd_t *pmd_start, pud_t *pud)
708 {
709 	pmd_t *pmd;
710 	int i;
711 
712 	for (i = 0; i < PTRS_PER_PMD; i++) {
713 		pmd = pmd_start + i;
714 		if (!pmd_none(*pmd))
715 			return;
716 	}
717 
718 	pmd_free(&init_mm, pmd_start);
719 	pud_clear(pud);
720 }
721 
722 struct change_mapping_params {
723 	pte_t *pte;
724 	unsigned long start;
725 	unsigned long end;
726 	unsigned long aligned_start;
727 	unsigned long aligned_end;
728 };
729 
730 static int __meminit stop_machine_change_mapping(void *data)
731 {
732 	struct change_mapping_params *params =
733 			(struct change_mapping_params *)data;
734 
735 	if (!data)
736 		return -1;
737 
738 	spin_unlock(&init_mm.page_table_lock);
739 	pte_clear(&init_mm, params->aligned_start, params->pte);
740 	create_physical_mapping(params->aligned_start, params->start, -1);
741 	create_physical_mapping(params->end, params->aligned_end, -1);
742 	spin_lock(&init_mm.page_table_lock);
743 	return 0;
744 }
745 
746 static void remove_pte_table(pte_t *pte_start, unsigned long addr,
747 			     unsigned long end)
748 {
749 	unsigned long next;
750 	pte_t *pte;
751 
752 	pte = pte_start + pte_index(addr);
753 	for (; addr < end; addr = next, pte++) {
754 		next = (addr + PAGE_SIZE) & PAGE_MASK;
755 		if (next > end)
756 			next = end;
757 
758 		if (!pte_present(*pte))
759 			continue;
760 
761 		if (!PAGE_ALIGNED(addr) || !PAGE_ALIGNED(next)) {
762 			/*
763 			 * The vmemmap_free() and remove_section_mapping()
764 			 * codepaths call us with aligned addresses.
765 			 */
766 			WARN_ONCE(1, "%s: unaligned range\n", __func__);
767 			continue;
768 		}
769 
770 		pte_clear(&init_mm, addr, pte);
771 	}
772 }
773 
774 /*
775  * clear the pte and potentially split the mapping helper
776  */
777 static void __meminit split_kernel_mapping(unsigned long addr, unsigned long end,
778 				unsigned long size, pte_t *pte)
779 {
780 	unsigned long mask = ~(size - 1);
781 	unsigned long aligned_start = addr & mask;
782 	unsigned long aligned_end = addr + size;
783 	struct change_mapping_params params;
784 	bool split_region = false;
785 
786 	if ((end - addr) < size) {
787 		/*
788 		 * We're going to clear the PTE, but not flushed
789 		 * the mapping, time to remap and flush. The
790 		 * effects if visible outside the processor or
791 		 * if we are running in code close to the
792 		 * mapping we cleared, we are in trouble.
793 		 */
794 		if (overlaps_kernel_text(aligned_start, addr) ||
795 			overlaps_kernel_text(end, aligned_end)) {
796 			/*
797 			 * Hack, just return, don't pte_clear
798 			 */
799 			WARN_ONCE(1, "Linear mapping %lx->%lx overlaps kernel "
800 				  "text, not splitting\n", addr, end);
801 			return;
802 		}
803 		split_region = true;
804 	}
805 
806 	if (split_region) {
807 		params.pte = pte;
808 		params.start = addr;
809 		params.end = end;
810 		params.aligned_start = addr & ~(size - 1);
811 		params.aligned_end = min_t(unsigned long, aligned_end,
812 				(unsigned long)__va(memblock_end_of_DRAM()));
813 		stop_machine(stop_machine_change_mapping, &params, NULL);
814 		return;
815 	}
816 
817 	pte_clear(&init_mm, addr, pte);
818 }
819 
820 static void remove_pmd_table(pmd_t *pmd_start, unsigned long addr,
821 			     unsigned long end)
822 {
823 	unsigned long next;
824 	pte_t *pte_base;
825 	pmd_t *pmd;
826 
827 	pmd = pmd_start + pmd_index(addr);
828 	for (; addr < end; addr = next, pmd++) {
829 		next = pmd_addr_end(addr, end);
830 
831 		if (!pmd_present(*pmd))
832 			continue;
833 
834 		if (pmd_is_leaf(*pmd)) {
835 			split_kernel_mapping(addr, end, PMD_SIZE, (pte_t *)pmd);
836 			continue;
837 		}
838 
839 		pte_base = (pte_t *)pmd_page_vaddr(*pmd);
840 		remove_pte_table(pte_base, addr, next);
841 		free_pte_table(pte_base, pmd);
842 	}
843 }
844 
845 static void remove_pud_table(pud_t *pud_start, unsigned long addr,
846 			     unsigned long end)
847 {
848 	unsigned long next;
849 	pmd_t *pmd_base;
850 	pud_t *pud;
851 
852 	pud = pud_start + pud_index(addr);
853 	for (; addr < end; addr = next, pud++) {
854 		next = pud_addr_end(addr, end);
855 
856 		if (!pud_present(*pud))
857 			continue;
858 
859 		if (pud_is_leaf(*pud)) {
860 			split_kernel_mapping(addr, end, PUD_SIZE, (pte_t *)pud);
861 			continue;
862 		}
863 
864 		pmd_base = (pmd_t *)pud_page_vaddr(*pud);
865 		remove_pmd_table(pmd_base, addr, next);
866 		free_pmd_table(pmd_base, pud);
867 	}
868 }
869 
870 static void __meminit remove_pagetable(unsigned long start, unsigned long end)
871 {
872 	unsigned long addr, next;
873 	pud_t *pud_base;
874 	pgd_t *pgd;
875 
876 	spin_lock(&init_mm.page_table_lock);
877 
878 	for (addr = start; addr < end; addr = next) {
879 		next = pgd_addr_end(addr, end);
880 
881 		pgd = pgd_offset_k(addr);
882 		if (!pgd_present(*pgd))
883 			continue;
884 
885 		if (pgd_is_leaf(*pgd)) {
886 			split_kernel_mapping(addr, end, PGDIR_SIZE, (pte_t *)pgd);
887 			continue;
888 		}
889 
890 		pud_base = (pud_t *)pgd_page_vaddr(*pgd);
891 		remove_pud_table(pud_base, addr, next);
892 	}
893 
894 	spin_unlock(&init_mm.page_table_lock);
895 	radix__flush_tlb_kernel_range(start, end);
896 }
897 
898 int __meminit radix__create_section_mapping(unsigned long start, unsigned long end, int nid)
899 {
900 	if (end >= RADIX_VMALLOC_START) {
901 		pr_warn("Outside the supported range\n");
902 		return -1;
903 	}
904 
905 	return create_physical_mapping(start, end, nid);
906 }
907 
908 int __meminit radix__remove_section_mapping(unsigned long start, unsigned long end)
909 {
910 	remove_pagetable(start, end);
911 	return 0;
912 }
913 #endif /* CONFIG_MEMORY_HOTPLUG */
914 
915 #ifdef CONFIG_SPARSEMEM_VMEMMAP
916 static int __map_kernel_page_nid(unsigned long ea, unsigned long pa,
917 				 pgprot_t flags, unsigned int map_page_size,
918 				 int nid)
919 {
920 	return __map_kernel_page(ea, pa, flags, map_page_size, nid, 0, 0);
921 }
922 
923 int __meminit radix__vmemmap_create_mapping(unsigned long start,
924 				      unsigned long page_size,
925 				      unsigned long phys)
926 {
927 	/* Create a PTE encoding */
928 	unsigned long flags = _PAGE_PRESENT | _PAGE_ACCESSED | _PAGE_KERNEL_RW;
929 	int nid = early_pfn_to_nid(phys >> PAGE_SHIFT);
930 	int ret;
931 
932 	if ((start + page_size) >= RADIX_VMEMMAP_END) {
933 		pr_warn("Outside the supported range\n");
934 		return -1;
935 	}
936 
937 	ret = __map_kernel_page_nid(start, phys, __pgprot(flags), page_size, nid);
938 	BUG_ON(ret);
939 
940 	return 0;
941 }
942 
943 #ifdef CONFIG_MEMORY_HOTPLUG
944 void __meminit radix__vmemmap_remove_mapping(unsigned long start, unsigned long page_size)
945 {
946 	remove_pagetable(start, start + page_size);
947 }
948 #endif
949 #endif
950 
951 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
952 
953 unsigned long radix__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
954 				  pmd_t *pmdp, unsigned long clr,
955 				  unsigned long set)
956 {
957 	unsigned long old;
958 
959 #ifdef CONFIG_DEBUG_VM
960 	WARN_ON(!radix__pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));
961 	assert_spin_locked(pmd_lockptr(mm, pmdp));
962 #endif
963 
964 	old = radix__pte_update(mm, addr, (pte_t *)pmdp, clr, set, 1);
965 	trace_hugepage_update(addr, old, clr, set);
966 
967 	return old;
968 }
969 
970 pmd_t radix__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address,
971 			pmd_t *pmdp)
972 
973 {
974 	pmd_t pmd;
975 
976 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
977 	VM_BUG_ON(radix__pmd_trans_huge(*pmdp));
978 	VM_BUG_ON(pmd_devmap(*pmdp));
979 	/*
980 	 * khugepaged calls this for normal pmd
981 	 */
982 	pmd = *pmdp;
983 	pmd_clear(pmdp);
984 
985 	/*FIXME!!  Verify whether we need this kick below */
986 	serialize_against_pte_lookup(vma->vm_mm);
987 
988 	radix__flush_tlb_collapsed_pmd(vma->vm_mm, address);
989 
990 	return pmd;
991 }
992 
993 /*
994  * For us pgtable_t is pte_t *. Inorder to save the deposisted
995  * page table, we consider the allocated page table as a list
996  * head. On withdraw we need to make sure we zero out the used
997  * list_head memory area.
998  */
999 void radix__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
1000 				 pgtable_t pgtable)
1001 {
1002 	struct list_head *lh = (struct list_head *) pgtable;
1003 
1004 	assert_spin_locked(pmd_lockptr(mm, pmdp));
1005 
1006 	/* FIFO */
1007 	if (!pmd_huge_pte(mm, pmdp))
1008 		INIT_LIST_HEAD(lh);
1009 	else
1010 		list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
1011 	pmd_huge_pte(mm, pmdp) = pgtable;
1012 }
1013 
1014 pgtable_t radix__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
1015 {
1016 	pte_t *ptep;
1017 	pgtable_t pgtable;
1018 	struct list_head *lh;
1019 
1020 	assert_spin_locked(pmd_lockptr(mm, pmdp));
1021 
1022 	/* FIFO */
1023 	pgtable = pmd_huge_pte(mm, pmdp);
1024 	lh = (struct list_head *) pgtable;
1025 	if (list_empty(lh))
1026 		pmd_huge_pte(mm, pmdp) = NULL;
1027 	else {
1028 		pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
1029 		list_del(lh);
1030 	}
1031 	ptep = (pte_t *) pgtable;
1032 	*ptep = __pte(0);
1033 	ptep++;
1034 	*ptep = __pte(0);
1035 	return pgtable;
1036 }
1037 
1038 pmd_t radix__pmdp_huge_get_and_clear(struct mm_struct *mm,
1039 				     unsigned long addr, pmd_t *pmdp)
1040 {
1041 	pmd_t old_pmd;
1042 	unsigned long old;
1043 
1044 	old = radix__pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
1045 	old_pmd = __pmd(old);
1046 	/*
1047 	 * Serialize against find_current_mm_pte which does lock-less
1048 	 * lookup in page tables with local interrupts disabled. For huge pages
1049 	 * it casts pmd_t to pte_t. Since format of pte_t is different from
1050 	 * pmd_t we want to prevent transit from pmd pointing to page table
1051 	 * to pmd pointing to huge page (and back) while interrupts are disabled.
1052 	 * We clear pmd to possibly replace it with page table pointer in
1053 	 * different code paths. So make sure we wait for the parallel
1054 	 * find_current_mm_pte to finish.
1055 	 */
1056 	serialize_against_pte_lookup(mm);
1057 	return old_pmd;
1058 }
1059 
1060 int radix__has_transparent_hugepage(void)
1061 {
1062 	/* For radix 2M at PMD level means thp */
1063 	if (mmu_psize_defs[MMU_PAGE_2M].shift == PMD_SHIFT)
1064 		return 1;
1065 	return 0;
1066 }
1067 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1068 
1069 void radix__ptep_set_access_flags(struct vm_area_struct *vma, pte_t *ptep,
1070 				  pte_t entry, unsigned long address, int psize)
1071 {
1072 	struct mm_struct *mm = vma->vm_mm;
1073 	unsigned long set = pte_val(entry) & (_PAGE_DIRTY | _PAGE_ACCESSED |
1074 					      _PAGE_RW | _PAGE_EXEC);
1075 
1076 	unsigned long change = pte_val(entry) ^ pte_val(*ptep);
1077 	/*
1078 	 * To avoid NMMU hang while relaxing access, we need mark
1079 	 * the pte invalid in between.
1080 	 */
1081 	if ((change & _PAGE_RW) && atomic_read(&mm->context.copros) > 0) {
1082 		unsigned long old_pte, new_pte;
1083 
1084 		old_pte = __radix_pte_update(ptep, _PAGE_PRESENT, _PAGE_INVALID);
1085 		/*
1086 		 * new value of pte
1087 		 */
1088 		new_pte = old_pte | set;
1089 		radix__flush_tlb_page_psize(mm, address, psize);
1090 		__radix_pte_update(ptep, _PAGE_INVALID, new_pte);
1091 	} else {
1092 		__radix_pte_update(ptep, 0, set);
1093 		/*
1094 		 * Book3S does not require a TLB flush when relaxing access
1095 		 * restrictions when the address space is not attached to a
1096 		 * NMMU, because the core MMU will reload the pte after taking
1097 		 * an access fault, which is defined by the architectue.
1098 		 */
1099 	}
1100 	/* See ptesync comment in radix__set_pte_at */
1101 }
1102 
1103 void radix__ptep_modify_prot_commit(struct vm_area_struct *vma,
1104 				    unsigned long addr, pte_t *ptep,
1105 				    pte_t old_pte, pte_t pte)
1106 {
1107 	struct mm_struct *mm = vma->vm_mm;
1108 
1109 	/*
1110 	 * To avoid NMMU hang while relaxing access we need to flush the tlb before
1111 	 * we set the new value. We need to do this only for radix, because hash
1112 	 * translation does flush when updating the linux pte.
1113 	 */
1114 	if (is_pte_rw_upgrade(pte_val(old_pte), pte_val(pte)) &&
1115 	    (atomic_read(&mm->context.copros) > 0))
1116 		radix__flush_tlb_page(vma, addr);
1117 
1118 	set_pte_at(mm, addr, ptep, pte);
1119 }
1120 
1121 int __init arch_ioremap_pud_supported(void)
1122 {
1123 	/* HPT does not cope with large pages in the vmalloc area */
1124 	return radix_enabled();
1125 }
1126 
1127 int __init arch_ioremap_pmd_supported(void)
1128 {
1129 	return radix_enabled();
1130 }
1131 
1132 int p4d_free_pud_page(p4d_t *p4d, unsigned long addr)
1133 {
1134 	return 0;
1135 }
1136 
1137 int pud_set_huge(pud_t *pud, phys_addr_t addr, pgprot_t prot)
1138 {
1139 	pte_t *ptep = (pte_t *)pud;
1140 	pte_t new_pud = pfn_pte(__phys_to_pfn(addr), prot);
1141 
1142 	if (!radix_enabled())
1143 		return 0;
1144 
1145 	set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pud);
1146 
1147 	return 1;
1148 }
1149 
1150 int pud_clear_huge(pud_t *pud)
1151 {
1152 	if (pud_huge(*pud)) {
1153 		pud_clear(pud);
1154 		return 1;
1155 	}
1156 
1157 	return 0;
1158 }
1159 
1160 int pud_free_pmd_page(pud_t *pud, unsigned long addr)
1161 {
1162 	pmd_t *pmd;
1163 	int i;
1164 
1165 	pmd = (pmd_t *)pud_page_vaddr(*pud);
1166 	pud_clear(pud);
1167 
1168 	flush_tlb_kernel_range(addr, addr + PUD_SIZE);
1169 
1170 	for (i = 0; i < PTRS_PER_PMD; i++) {
1171 		if (!pmd_none(pmd[i])) {
1172 			pte_t *pte;
1173 			pte = (pte_t *)pmd_page_vaddr(pmd[i]);
1174 
1175 			pte_free_kernel(&init_mm, pte);
1176 		}
1177 	}
1178 
1179 	pmd_free(&init_mm, pmd);
1180 
1181 	return 1;
1182 }
1183 
1184 int pmd_set_huge(pmd_t *pmd, phys_addr_t addr, pgprot_t prot)
1185 {
1186 	pte_t *ptep = (pte_t *)pmd;
1187 	pte_t new_pmd = pfn_pte(__phys_to_pfn(addr), prot);
1188 
1189 	if (!radix_enabled())
1190 		return 0;
1191 
1192 	set_pte_at(&init_mm, 0 /* radix unused */, ptep, new_pmd);
1193 
1194 	return 1;
1195 }
1196 
1197 int pmd_clear_huge(pmd_t *pmd)
1198 {
1199 	if (pmd_huge(*pmd)) {
1200 		pmd_clear(pmd);
1201 		return 1;
1202 	}
1203 
1204 	return 0;
1205 }
1206 
1207 int pmd_free_pte_page(pmd_t *pmd, unsigned long addr)
1208 {
1209 	pte_t *pte;
1210 
1211 	pte = (pte_t *)pmd_page_vaddr(*pmd);
1212 	pmd_clear(pmd);
1213 
1214 	flush_tlb_kernel_range(addr, addr + PMD_SIZE);
1215 
1216 	pte_free_kernel(&init_mm, pte);
1217 
1218 	return 1;
1219 }
1220 
1221 int radix__ioremap_range(unsigned long ea, phys_addr_t pa, unsigned long size,
1222 			pgprot_t prot, int nid)
1223 {
1224 	if (likely(slab_is_available())) {
1225 		int err = ioremap_page_range(ea, ea + size, pa, prot);
1226 		if (err)
1227 			unmap_kernel_range(ea, size);
1228 		return err;
1229 	} else {
1230 		unsigned long i;
1231 
1232 		for (i = 0; i < size; i += PAGE_SIZE) {
1233 			int err = map_kernel_page(ea + i, pa + i, prot);
1234 			if (WARN_ON_ONCE(err)) /* Should clean up */
1235 				return err;
1236 		}
1237 		return 0;
1238 	}
1239 }
1240 
1241 int __init arch_ioremap_p4d_supported(void)
1242 {
1243 	return 0;
1244 }
1245