1 /*
2  * This file contains the routines for flushing entries from the
3  * TLB and MMU hash table.
4  *
5  *  Derived from arch/ppc64/mm/init.c:
6  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
7  *
8  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
9  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
10  *    Copyright (C) 1996 Paul Mackerras
11  *
12  *  Derived from "arch/i386/mm/init.c"
13  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
14  *
15  *  Dave Engebretsen <engebret@us.ibm.com>
16  *      Rework for PPC64 port.
17  *
18  *  This program is free software; you can redistribute it and/or
19  *  modify it under the terms of the GNU General Public License
20  *  as published by the Free Software Foundation; either version
21  *  2 of the License, or (at your option) any later version.
22  */
23 
24 #include <linux/kernel.h>
25 #include <linux/mm.h>
26 #include <linux/percpu.h>
27 #include <linux/hardirq.h>
28 #include <asm/pgalloc.h>
29 #include <asm/tlbflush.h>
30 #include <asm/tlb.h>
31 #include <asm/bug.h>
32 #include <asm/pte-walk.h>
33 
34 
35 #include <trace/events/thp.h>
36 
37 DEFINE_PER_CPU(struct ppc64_tlb_batch, ppc64_tlb_batch);
38 
39 /*
40  * A linux PTE was changed and the corresponding hash table entry
41  * neesd to be flushed. This function will either perform the flush
42  * immediately or will batch it up if the current CPU has an active
43  * batch on it.
44  */
45 void hpte_need_flush(struct mm_struct *mm, unsigned long addr,
46 		     pte_t *ptep, unsigned long pte, int huge)
47 {
48 	unsigned long vpn;
49 	struct ppc64_tlb_batch *batch = &get_cpu_var(ppc64_tlb_batch);
50 	unsigned long vsid;
51 	unsigned int psize;
52 	int ssize;
53 	real_pte_t rpte;
54 	int i, offset;
55 
56 	i = batch->index;
57 
58 	/*
59 	 * Get page size (maybe move back to caller).
60 	 *
61 	 * NOTE: when using special 64K mappings in 4K environment like
62 	 * for SPEs, we obtain the page size from the slice, which thus
63 	 * must still exist (and thus the VMA not reused) at the time
64 	 * of this call
65 	 */
66 	if (huge) {
67 #ifdef CONFIG_HUGETLB_PAGE
68 		psize = get_slice_psize(mm, addr);
69 		/* Mask the address for the correct page size */
70 		addr &= ~((1UL << mmu_psize_defs[psize].shift) - 1);
71 		if (unlikely(psize == MMU_PAGE_16G))
72 			offset = PTRS_PER_PUD;
73 		else
74 			offset = PTRS_PER_PMD;
75 #else
76 		BUG();
77 		psize = pte_pagesize_index(mm, addr, pte); /* shutup gcc */
78 #endif
79 	} else {
80 		psize = pte_pagesize_index(mm, addr, pte);
81 		/*
82 		 * Mask the address for the standard page size.  If we
83 		 * have a 64k page kernel, but the hardware does not
84 		 * support 64k pages, this might be different from the
85 		 * hardware page size encoded in the slice table.
86 		 */
87 		addr &= PAGE_MASK;
88 		offset = PTRS_PER_PTE;
89 	}
90 
91 
92 	/* Build full vaddr */
93 	if (!is_kernel_addr(addr)) {
94 		ssize = user_segment_size(addr);
95 		vsid = get_user_vsid(&mm->context, addr, ssize);
96 	} else {
97 		vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
98 		ssize = mmu_kernel_ssize;
99 	}
100 	WARN_ON(vsid == 0);
101 	vpn = hpt_vpn(addr, vsid, ssize);
102 	rpte = __real_pte(__pte(pte), ptep, offset);
103 
104 	/*
105 	 * Check if we have an active batch on this CPU. If not, just
106 	 * flush now and return.
107 	 */
108 	if (!batch->active) {
109 		flush_hash_page(vpn, rpte, psize, ssize, mm_is_thread_local(mm));
110 		put_cpu_var(ppc64_tlb_batch);
111 		return;
112 	}
113 
114 	/*
115 	 * This can happen when we are in the middle of a TLB batch and
116 	 * we encounter memory pressure (eg copy_page_range when it tries
117 	 * to allocate a new pte). If we have to reclaim memory and end
118 	 * up scanning and resetting referenced bits then our batch context
119 	 * will change mid stream.
120 	 *
121 	 * We also need to ensure only one page size is present in a given
122 	 * batch
123 	 */
124 	if (i != 0 && (mm != batch->mm || batch->psize != psize ||
125 		       batch->ssize != ssize)) {
126 		__flush_tlb_pending(batch);
127 		i = 0;
128 	}
129 	if (i == 0) {
130 		batch->mm = mm;
131 		batch->psize = psize;
132 		batch->ssize = ssize;
133 	}
134 	batch->pte[i] = rpte;
135 	batch->vpn[i] = vpn;
136 	batch->index = ++i;
137 	if (i >= PPC64_TLB_BATCH_NR)
138 		__flush_tlb_pending(batch);
139 	put_cpu_var(ppc64_tlb_batch);
140 }
141 
142 /*
143  * This function is called when terminating an mmu batch or when a batch
144  * is full. It will perform the flush of all the entries currently stored
145  * in a batch.
146  *
147  * Must be called from within some kind of spinlock/non-preempt region...
148  */
149 void __flush_tlb_pending(struct ppc64_tlb_batch *batch)
150 {
151 	int i, local;
152 
153 	i = batch->index;
154 	local = mm_is_thread_local(batch->mm);
155 	if (i == 1)
156 		flush_hash_page(batch->vpn[0], batch->pte[0],
157 				batch->psize, batch->ssize, local);
158 	else
159 		flush_hash_range(i, local);
160 	batch->index = 0;
161 }
162 
163 void hash__tlb_flush(struct mmu_gather *tlb)
164 {
165 	struct ppc64_tlb_batch *tlbbatch = &get_cpu_var(ppc64_tlb_batch);
166 
167 	/*
168 	 * If there's a TLB batch pending, then we must flush it because the
169 	 * pages are going to be freed and we really don't want to have a CPU
170 	 * access a freed page because it has a stale TLB
171 	 */
172 	if (tlbbatch->index)
173 		__flush_tlb_pending(tlbbatch);
174 
175 	put_cpu_var(ppc64_tlb_batch);
176 }
177 
178 /**
179  * __flush_hash_table_range - Flush all HPTEs for a given address range
180  *                            from the hash table (and the TLB). But keeps
181  *                            the linux PTEs intact.
182  *
183  * @mm		: mm_struct of the target address space (generally init_mm)
184  * @start	: starting address
185  * @end         : ending address (not included in the flush)
186  *
187  * This function is mostly to be used by some IO hotplug code in order
188  * to remove all hash entries from a given address range used to map IO
189  * space on a removed PCI-PCI bidge without tearing down the full mapping
190  * since 64K pages may overlap with other bridges when using 64K pages
191  * with 4K HW pages on IO space.
192  *
193  * Because of that usage pattern, it is implemented for small size rather
194  * than speed.
195  */
196 void __flush_hash_table_range(struct mm_struct *mm, unsigned long start,
197 			      unsigned long end)
198 {
199 	bool is_thp;
200 	int hugepage_shift;
201 	unsigned long flags;
202 
203 	start = _ALIGN_DOWN(start, PAGE_SIZE);
204 	end = _ALIGN_UP(end, PAGE_SIZE);
205 
206 	BUG_ON(!mm->pgd);
207 
208 	/*
209 	 * Note: Normally, we should only ever use a batch within a
210 	 * PTE locked section. This violates the rule, but will work
211 	 * since we don't actually modify the PTEs, we just flush the
212 	 * hash while leaving the PTEs intact (including their reference
213 	 * to being hashed). This is not the most performance oriented
214 	 * way to do things but is fine for our needs here.
215 	 */
216 	local_irq_save(flags);
217 	arch_enter_lazy_mmu_mode();
218 	for (; start < end; start += PAGE_SIZE) {
219 		pte_t *ptep = find_current_mm_pte(mm->pgd, start, &is_thp,
220 						  &hugepage_shift);
221 		unsigned long pte;
222 
223 		if (ptep == NULL)
224 			continue;
225 		pte = pte_val(*ptep);
226 		if (is_thp)
227 			trace_hugepage_invalidate(start, pte);
228 		if (!(pte & H_PAGE_HASHPTE))
229 			continue;
230 		if (unlikely(is_thp))
231 			hpte_do_hugepage_flush(mm, start, (pmd_t *)ptep, pte);
232 		else
233 			hpte_need_flush(mm, start, ptep, pte, hugepage_shift);
234 	}
235 	arch_leave_lazy_mmu_mode();
236 	local_irq_restore(flags);
237 }
238 
239 void flush_tlb_pmd_range(struct mm_struct *mm, pmd_t *pmd, unsigned long addr)
240 {
241 	pte_t *pte;
242 	pte_t *start_pte;
243 	unsigned long flags;
244 
245 	addr = _ALIGN_DOWN(addr, PMD_SIZE);
246 	/*
247 	 * Note: Normally, we should only ever use a batch within a
248 	 * PTE locked section. This violates the rule, but will work
249 	 * since we don't actually modify the PTEs, we just flush the
250 	 * hash while leaving the PTEs intact (including their reference
251 	 * to being hashed). This is not the most performance oriented
252 	 * way to do things but is fine for our needs here.
253 	 */
254 	local_irq_save(flags);
255 	arch_enter_lazy_mmu_mode();
256 	start_pte = pte_offset_map(pmd, addr);
257 	for (pte = start_pte; pte < start_pte + PTRS_PER_PTE; pte++) {
258 		unsigned long pteval = pte_val(*pte);
259 		if (pteval & H_PAGE_HASHPTE)
260 			hpte_need_flush(mm, addr, pte, pteval, 0);
261 		addr += PAGE_SIZE;
262 	}
263 	arch_leave_lazy_mmu_mode();
264 	local_irq_restore(flags);
265 }
266