xref: /linux/arch/x86/mm/tlb.c (revision 60e13231561b3a4c5269bfa1ef6c0569ad6f28ec)
1 #include <linux/init.h>
2 
3 #include <linux/mm.h>
4 #include <linux/spinlock.h>
5 #include <linux/smp.h>
6 #include <linux/interrupt.h>
7 #include <linux/module.h>
8 #include <linux/cpu.h>
9 
10 #include <asm/tlbflush.h>
11 #include <asm/mmu_context.h>
12 #include <asm/cache.h>
13 #include <asm/apic.h>
14 #include <asm/uv/uv.h>
15 
16 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate)
17 			= { &init_mm, 0, };
18 
19 /*
20  *	Smarter SMP flushing macros.
21  *		c/o Linus Torvalds.
22  *
23  *	These mean you can really definitely utterly forget about
24  *	writing to user space from interrupts. (Its not allowed anyway).
25  *
26  *	Optimizations Manfred Spraul <manfred@colorfullife.com>
27  *
28  *	More scalable flush, from Andi Kleen
29  *
30  *	To avoid global state use 8 different call vectors.
31  *	Each CPU uses a specific vector to trigger flushes on other
32  *	CPUs. Depending on the received vector the target CPUs look into
33  *	the right array slot for the flush data.
34  *
35  *	With more than 8 CPUs they are hashed to the 8 available
36  *	vectors. The limited global vector space forces us to this right now.
37  *	In future when interrupts are split into per CPU domains this could be
38  *	fixed, at the cost of triggering multiple IPIs in some cases.
39  */
40 
41 union smp_flush_state {
42 	struct {
43 		struct mm_struct *flush_mm;
44 		unsigned long flush_va;
45 		raw_spinlock_t tlbstate_lock;
46 		DECLARE_BITMAP(flush_cpumask, NR_CPUS);
47 	};
48 	char pad[INTERNODE_CACHE_BYTES];
49 } ____cacheline_internodealigned_in_smp;
50 
51 /* State is put into the per CPU data section, but padded
52    to a full cache line because other CPUs can access it and we don't
53    want false sharing in the per cpu data segment. */
54 static union smp_flush_state flush_state[NUM_INVALIDATE_TLB_VECTORS];
55 
56 static DEFINE_PER_CPU_READ_MOSTLY(int, tlb_vector_offset);
57 
58 /*
59  * We cannot call mmdrop() because we are in interrupt context,
60  * instead update mm->cpu_vm_mask.
61  */
62 void leave_mm(int cpu)
63 {
64 	if (percpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
65 		BUG();
66 	cpumask_clear_cpu(cpu,
67 			  mm_cpumask(percpu_read(cpu_tlbstate.active_mm)));
68 	load_cr3(swapper_pg_dir);
69 }
70 EXPORT_SYMBOL_GPL(leave_mm);
71 
72 /*
73  *
74  * The flush IPI assumes that a thread switch happens in this order:
75  * [cpu0: the cpu that switches]
76  * 1) switch_mm() either 1a) or 1b)
77  * 1a) thread switch to a different mm
78  * 1a1) cpu_clear(cpu, old_mm->cpu_vm_mask);
79  *	Stop ipi delivery for the old mm. This is not synchronized with
80  *	the other cpus, but smp_invalidate_interrupt ignore flush ipis
81  *	for the wrong mm, and in the worst case we perform a superfluous
82  *	tlb flush.
83  * 1a2) set cpu mmu_state to TLBSTATE_OK
84  *	Now the smp_invalidate_interrupt won't call leave_mm if cpu0
85  *	was in lazy tlb mode.
86  * 1a3) update cpu active_mm
87  *	Now cpu0 accepts tlb flushes for the new mm.
88  * 1a4) cpu_set(cpu, new_mm->cpu_vm_mask);
89  *	Now the other cpus will send tlb flush ipis.
90  * 1a4) change cr3.
91  * 1b) thread switch without mm change
92  *	cpu active_mm is correct, cpu0 already handles
93  *	flush ipis.
94  * 1b1) set cpu mmu_state to TLBSTATE_OK
95  * 1b2) test_and_set the cpu bit in cpu_vm_mask.
96  *	Atomically set the bit [other cpus will start sending flush ipis],
97  *	and test the bit.
98  * 1b3) if the bit was 0: leave_mm was called, flush the tlb.
99  * 2) switch %%esp, ie current
100  *
101  * The interrupt must handle 2 special cases:
102  * - cr3 is changed before %%esp, ie. it cannot use current->{active_,}mm.
103  * - the cpu performs speculative tlb reads, i.e. even if the cpu only
104  *   runs in kernel space, the cpu could load tlb entries for user space
105  *   pages.
106  *
107  * The good news is that cpu mmu_state is local to each cpu, no
108  * write/read ordering problems.
109  */
110 
111 /*
112  * TLB flush IPI:
113  *
114  * 1) Flush the tlb entries if the cpu uses the mm that's being flushed.
115  * 2) Leave the mm if we are in the lazy tlb mode.
116  *
117  * Interrupts are disabled.
118  */
119 
120 /*
121  * FIXME: use of asmlinkage is not consistent.  On x86_64 it's noop
122  * but still used for documentation purpose but the usage is slightly
123  * inconsistent.  On x86_32, asmlinkage is regparm(0) but interrupt
124  * entry calls in with the first parameter in %eax.  Maybe define
125  * intrlinkage?
126  */
127 #ifdef CONFIG_X86_64
128 asmlinkage
129 #endif
130 void smp_invalidate_interrupt(struct pt_regs *regs)
131 {
132 	unsigned int cpu;
133 	unsigned int sender;
134 	union smp_flush_state *f;
135 
136 	cpu = smp_processor_id();
137 	/*
138 	 * orig_rax contains the negated interrupt vector.
139 	 * Use that to determine where the sender put the data.
140 	 */
141 	sender = ~regs->orig_ax - INVALIDATE_TLB_VECTOR_START;
142 	f = &flush_state[sender];
143 
144 	if (!cpumask_test_cpu(cpu, to_cpumask(f->flush_cpumask)))
145 		goto out;
146 		/*
147 		 * This was a BUG() but until someone can quote me the
148 		 * line from the intel manual that guarantees an IPI to
149 		 * multiple CPUs is retried _only_ on the erroring CPUs
150 		 * its staying as a return
151 		 *
152 		 * BUG();
153 		 */
154 
155 	if (f->flush_mm == percpu_read(cpu_tlbstate.active_mm)) {
156 		if (percpu_read(cpu_tlbstate.state) == TLBSTATE_OK) {
157 			if (f->flush_va == TLB_FLUSH_ALL)
158 				local_flush_tlb();
159 			else
160 				__flush_tlb_one(f->flush_va);
161 		} else
162 			leave_mm(cpu);
163 	}
164 out:
165 	ack_APIC_irq();
166 	smp_mb__before_clear_bit();
167 	cpumask_clear_cpu(cpu, to_cpumask(f->flush_cpumask));
168 	smp_mb__after_clear_bit();
169 	inc_irq_stat(irq_tlb_count);
170 }
171 
172 static void flush_tlb_others_ipi(const struct cpumask *cpumask,
173 				 struct mm_struct *mm, unsigned long va)
174 {
175 	unsigned int sender;
176 	union smp_flush_state *f;
177 
178 	/* Caller has disabled preemption */
179 	sender = this_cpu_read(tlb_vector_offset);
180 	f = &flush_state[sender];
181 
182 	if (nr_cpu_ids > NUM_INVALIDATE_TLB_VECTORS)
183 		raw_spin_lock(&f->tlbstate_lock);
184 
185 	f->flush_mm = mm;
186 	f->flush_va = va;
187 	if (cpumask_andnot(to_cpumask(f->flush_cpumask), cpumask, cpumask_of(smp_processor_id()))) {
188 		/*
189 		 * We have to send the IPI only to
190 		 * CPUs affected.
191 		 */
192 		apic->send_IPI_mask(to_cpumask(f->flush_cpumask),
193 			      INVALIDATE_TLB_VECTOR_START + sender);
194 
195 		while (!cpumask_empty(to_cpumask(f->flush_cpumask)))
196 			cpu_relax();
197 	}
198 
199 	f->flush_mm = NULL;
200 	f->flush_va = 0;
201 	if (nr_cpu_ids > NUM_INVALIDATE_TLB_VECTORS)
202 		raw_spin_unlock(&f->tlbstate_lock);
203 }
204 
205 void native_flush_tlb_others(const struct cpumask *cpumask,
206 			     struct mm_struct *mm, unsigned long va)
207 {
208 	if (is_uv_system()) {
209 		unsigned int cpu;
210 
211 		cpu = smp_processor_id();
212 		cpumask = uv_flush_tlb_others(cpumask, mm, va, cpu);
213 		if (cpumask)
214 			flush_tlb_others_ipi(cpumask, mm, va);
215 		return;
216 	}
217 	flush_tlb_others_ipi(cpumask, mm, va);
218 }
219 
220 static void __cpuinit calculate_tlb_offset(void)
221 {
222 	int cpu, node, nr_node_vecs, idx = 0;
223 	/*
224 	 * we are changing tlb_vector_offset for each CPU in runtime, but this
225 	 * will not cause inconsistency, as the write is atomic under X86. we
226 	 * might see more lock contentions in a short time, but after all CPU's
227 	 * tlb_vector_offset are changed, everything should go normal
228 	 *
229 	 * Note: if NUM_INVALIDATE_TLB_VECTORS % nr_online_nodes !=0, we might
230 	 * waste some vectors.
231 	 **/
232 	if (nr_online_nodes > NUM_INVALIDATE_TLB_VECTORS)
233 		nr_node_vecs = 1;
234 	else
235 		nr_node_vecs = NUM_INVALIDATE_TLB_VECTORS/nr_online_nodes;
236 
237 	for_each_online_node(node) {
238 		int node_offset = (idx % NUM_INVALIDATE_TLB_VECTORS) *
239 			nr_node_vecs;
240 		int cpu_offset = 0;
241 		for_each_cpu(cpu, cpumask_of_node(node)) {
242 			per_cpu(tlb_vector_offset, cpu) = node_offset +
243 				cpu_offset;
244 			cpu_offset++;
245 			cpu_offset = cpu_offset % nr_node_vecs;
246 		}
247 		idx++;
248 	}
249 }
250 
251 static int __cpuinit tlb_cpuhp_notify(struct notifier_block *n,
252 		unsigned long action, void *hcpu)
253 {
254 	switch (action & 0xf) {
255 	case CPU_ONLINE:
256 	case CPU_DEAD:
257 		calculate_tlb_offset();
258 	}
259 	return NOTIFY_OK;
260 }
261 
262 static int __cpuinit init_smp_flush(void)
263 {
264 	int i;
265 
266 	for (i = 0; i < ARRAY_SIZE(flush_state); i++)
267 		raw_spin_lock_init(&flush_state[i].tlbstate_lock);
268 
269 	calculate_tlb_offset();
270 	hotcpu_notifier(tlb_cpuhp_notify, 0);
271 	return 0;
272 }
273 core_initcall(init_smp_flush);
274 
275 void flush_tlb_current_task(void)
276 {
277 	struct mm_struct *mm = current->mm;
278 
279 	preempt_disable();
280 
281 	local_flush_tlb();
282 	if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
283 		flush_tlb_others(mm_cpumask(mm), mm, TLB_FLUSH_ALL);
284 	preempt_enable();
285 }
286 
287 void flush_tlb_mm(struct mm_struct *mm)
288 {
289 	preempt_disable();
290 
291 	if (current->active_mm == mm) {
292 		if (current->mm)
293 			local_flush_tlb();
294 		else
295 			leave_mm(smp_processor_id());
296 	}
297 	if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
298 		flush_tlb_others(mm_cpumask(mm), mm, TLB_FLUSH_ALL);
299 
300 	preempt_enable();
301 }
302 
303 void flush_tlb_page(struct vm_area_struct *vma, unsigned long va)
304 {
305 	struct mm_struct *mm = vma->vm_mm;
306 
307 	preempt_disable();
308 
309 	if (current->active_mm == mm) {
310 		if (current->mm)
311 			__flush_tlb_one(va);
312 		else
313 			leave_mm(smp_processor_id());
314 	}
315 
316 	if (cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids)
317 		flush_tlb_others(mm_cpumask(mm), mm, va);
318 
319 	preempt_enable();
320 }
321 
322 static void do_flush_tlb_all(void *info)
323 {
324 	__flush_tlb_all();
325 	if (percpu_read(cpu_tlbstate.state) == TLBSTATE_LAZY)
326 		leave_mm(smp_processor_id());
327 }
328 
329 void flush_tlb_all(void)
330 {
331 	on_each_cpu(do_flush_tlb_all, NULL, 1);
332 }
333