xref: /openbmc/linux/arch/arm/mm/context.c (revision 9cfc5c90)
1 /*
2  *  linux/arch/arm/mm/context.c
3  *
4  *  Copyright (C) 2002-2003 Deep Blue Solutions Ltd, all rights reserved.
5  *  Copyright (C) 2012 ARM Limited
6  *
7  *  Author: Will Deacon <will.deacon@arm.com>
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License version 2 as
11  * published by the Free Software Foundation.
12  */
13 #include <linux/init.h>
14 #include <linux/sched.h>
15 #include <linux/mm.h>
16 #include <linux/smp.h>
17 #include <linux/percpu.h>
18 
19 #include <asm/mmu_context.h>
20 #include <asm/smp_plat.h>
21 #include <asm/thread_notify.h>
22 #include <asm/tlbflush.h>
23 #include <asm/proc-fns.h>
24 
25 /*
26  * On ARMv6, we have the following structure in the Context ID:
27  *
28  * 31                         7          0
29  * +-------------------------+-----------+
30  * |      process ID         |   ASID    |
31  * +-------------------------+-----------+
32  * |              context ID             |
33  * +-------------------------------------+
34  *
35  * The ASID is used to tag entries in the CPU caches and TLBs.
36  * The context ID is used by debuggers and trace logic, and
37  * should be unique within all running processes.
38  *
39  * In big endian operation, the two 32 bit words are swapped if accessed
40  * by non-64-bit operations.
41  */
42 #define ASID_FIRST_VERSION	(1ULL << ASID_BITS)
43 #define NUM_USER_ASIDS		ASID_FIRST_VERSION
44 
45 static DEFINE_RAW_SPINLOCK(cpu_asid_lock);
46 static atomic64_t asid_generation = ATOMIC64_INIT(ASID_FIRST_VERSION);
47 static DECLARE_BITMAP(asid_map, NUM_USER_ASIDS);
48 
49 static DEFINE_PER_CPU(atomic64_t, active_asids);
50 static DEFINE_PER_CPU(u64, reserved_asids);
51 static cpumask_t tlb_flush_pending;
52 
53 #ifdef CONFIG_ARM_ERRATA_798181
54 void a15_erratum_get_cpumask(int this_cpu, struct mm_struct *mm,
55 			     cpumask_t *mask)
56 {
57 	int cpu;
58 	unsigned long flags;
59 	u64 context_id, asid;
60 
61 	raw_spin_lock_irqsave(&cpu_asid_lock, flags);
62 	context_id = mm->context.id.counter;
63 	for_each_online_cpu(cpu) {
64 		if (cpu == this_cpu)
65 			continue;
66 		/*
67 		 * We only need to send an IPI if the other CPUs are
68 		 * running the same ASID as the one being invalidated.
69 		 */
70 		asid = per_cpu(active_asids, cpu).counter;
71 		if (asid == 0)
72 			asid = per_cpu(reserved_asids, cpu);
73 		if (context_id == asid)
74 			cpumask_set_cpu(cpu, mask);
75 	}
76 	raw_spin_unlock_irqrestore(&cpu_asid_lock, flags);
77 }
78 #endif
79 
80 #ifdef CONFIG_ARM_LPAE
81 /*
82  * With LPAE, the ASID and page tables are updated atomicly, so there is
83  * no need for a reserved set of tables (the active ASID tracking prevents
84  * any issues across a rollover).
85  */
86 #define cpu_set_reserved_ttbr0()
87 #else
88 static void cpu_set_reserved_ttbr0(void)
89 {
90 	u32 ttb;
91 	/*
92 	 * Copy TTBR1 into TTBR0.
93 	 * This points at swapper_pg_dir, which contains only global
94 	 * entries so any speculative walks are perfectly safe.
95 	 */
96 	asm volatile(
97 	"	mrc	p15, 0, %0, c2, c0, 1		@ read TTBR1\n"
98 	"	mcr	p15, 0, %0, c2, c0, 0		@ set TTBR0\n"
99 	: "=r" (ttb));
100 	isb();
101 }
102 #endif
103 
104 #ifdef CONFIG_PID_IN_CONTEXTIDR
105 static int contextidr_notifier(struct notifier_block *unused, unsigned long cmd,
106 			       void *t)
107 {
108 	u32 contextidr;
109 	pid_t pid;
110 	struct thread_info *thread = t;
111 
112 	if (cmd != THREAD_NOTIFY_SWITCH)
113 		return NOTIFY_DONE;
114 
115 	pid = task_pid_nr(thread->task) << ASID_BITS;
116 	asm volatile(
117 	"	mrc	p15, 0, %0, c13, c0, 1\n"
118 	"	and	%0, %0, %2\n"
119 	"	orr	%0, %0, %1\n"
120 	"	mcr	p15, 0, %0, c13, c0, 1\n"
121 	: "=r" (contextidr), "+r" (pid)
122 	: "I" (~ASID_MASK));
123 	isb();
124 
125 	return NOTIFY_OK;
126 }
127 
128 static struct notifier_block contextidr_notifier_block = {
129 	.notifier_call = contextidr_notifier,
130 };
131 
132 static int __init contextidr_notifier_init(void)
133 {
134 	return thread_register_notifier(&contextidr_notifier_block);
135 }
136 arch_initcall(contextidr_notifier_init);
137 #endif
138 
139 static void flush_context(unsigned int cpu)
140 {
141 	int i;
142 	u64 asid;
143 
144 	/* Update the list of reserved ASIDs and the ASID bitmap. */
145 	bitmap_clear(asid_map, 0, NUM_USER_ASIDS);
146 	for_each_possible_cpu(i) {
147 		asid = atomic64_xchg(&per_cpu(active_asids, i), 0);
148 		/*
149 		 * If this CPU has already been through a
150 		 * rollover, but hasn't run another task in
151 		 * the meantime, we must preserve its reserved
152 		 * ASID, as this is the only trace we have of
153 		 * the process it is still running.
154 		 */
155 		if (asid == 0)
156 			asid = per_cpu(reserved_asids, i);
157 		__set_bit(asid & ~ASID_MASK, asid_map);
158 		per_cpu(reserved_asids, i) = asid;
159 	}
160 
161 	/* Queue a TLB invalidate and flush the I-cache if necessary. */
162 	cpumask_setall(&tlb_flush_pending);
163 
164 	if (icache_is_vivt_asid_tagged())
165 		__flush_icache_all();
166 }
167 
168 static int is_reserved_asid(u64 asid)
169 {
170 	int cpu;
171 	for_each_possible_cpu(cpu)
172 		if (per_cpu(reserved_asids, cpu) == asid)
173 			return 1;
174 	return 0;
175 }
176 
177 static u64 new_context(struct mm_struct *mm, unsigned int cpu)
178 {
179 	static u32 cur_idx = 1;
180 	u64 asid = atomic64_read(&mm->context.id);
181 	u64 generation = atomic64_read(&asid_generation);
182 
183 	if (asid != 0) {
184 		/*
185 		 * If our current ASID was active during a rollover, we
186 		 * can continue to use it and this was just a false alarm.
187 		 */
188 		if (is_reserved_asid(asid))
189 			return generation | (asid & ~ASID_MASK);
190 
191 		/*
192 		 * We had a valid ASID in a previous life, so try to re-use
193 		 * it if possible.,
194 		 */
195 		asid &= ~ASID_MASK;
196 		if (!__test_and_set_bit(asid, asid_map))
197 			goto bump_gen;
198 	}
199 
200 	/*
201 	 * Allocate a free ASID. If we can't find one, take a note of the
202 	 * currently active ASIDs and mark the TLBs as requiring flushes.
203 	 * We always count from ASID #1, as we reserve ASID #0 to switch
204 	 * via TTBR0 and to avoid speculative page table walks from hitting
205 	 * in any partial walk caches, which could be populated from
206 	 * overlapping level-1 descriptors used to map both the module
207 	 * area and the userspace stack.
208 	 */
209 	asid = find_next_zero_bit(asid_map, NUM_USER_ASIDS, cur_idx);
210 	if (asid == NUM_USER_ASIDS) {
211 		generation = atomic64_add_return(ASID_FIRST_VERSION,
212 						 &asid_generation);
213 		flush_context(cpu);
214 		asid = find_next_zero_bit(asid_map, NUM_USER_ASIDS, 1);
215 	}
216 
217 	__set_bit(asid, asid_map);
218 	cur_idx = asid;
219 
220 bump_gen:
221 	asid |= generation;
222 	cpumask_clear(mm_cpumask(mm));
223 	return asid;
224 }
225 
226 void check_and_switch_context(struct mm_struct *mm, struct task_struct *tsk)
227 {
228 	unsigned long flags;
229 	unsigned int cpu = smp_processor_id();
230 	u64 asid;
231 
232 	if (unlikely(mm->context.vmalloc_seq != init_mm.context.vmalloc_seq))
233 		__check_vmalloc_seq(mm);
234 
235 	/*
236 	 * We cannot update the pgd and the ASID atomicly with classic
237 	 * MMU, so switch exclusively to global mappings to avoid
238 	 * speculative page table walking with the wrong TTBR.
239 	 */
240 	cpu_set_reserved_ttbr0();
241 
242 	asid = atomic64_read(&mm->context.id);
243 	if (!((asid ^ atomic64_read(&asid_generation)) >> ASID_BITS)
244 	    && atomic64_xchg(&per_cpu(active_asids, cpu), asid))
245 		goto switch_mm_fastpath;
246 
247 	raw_spin_lock_irqsave(&cpu_asid_lock, flags);
248 	/* Check that our ASID belongs to the current generation. */
249 	asid = atomic64_read(&mm->context.id);
250 	if ((asid ^ atomic64_read(&asid_generation)) >> ASID_BITS) {
251 		asid = new_context(mm, cpu);
252 		atomic64_set(&mm->context.id, asid);
253 	}
254 
255 	if (cpumask_test_and_clear_cpu(cpu, &tlb_flush_pending)) {
256 		local_flush_bp_all();
257 		local_flush_tlb_all();
258 	}
259 
260 	atomic64_set(&per_cpu(active_asids, cpu), asid);
261 	cpumask_set_cpu(cpu, mm_cpumask(mm));
262 	raw_spin_unlock_irqrestore(&cpu_asid_lock, flags);
263 
264 switch_mm_fastpath:
265 	cpu_switch_mm(mm->pgd, mm);
266 }
267