xref: /linux/drivers/xen/events/events_2l.c (revision 06ed6aa56ffac9241e03a24649e8d048f8f1b10c)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Xen event channels (2-level ABI)
4  *
5  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
6  */
7 
8 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
9 
10 #include <linux/linkage.h>
11 #include <linux/interrupt.h>
12 #include <linux/irq.h>
13 
14 #include <asm/sync_bitops.h>
15 #include <asm/xen/hypercall.h>
16 #include <asm/xen/hypervisor.h>
17 
18 #include <xen/xen.h>
19 #include <xen/xen-ops.h>
20 #include <xen/events.h>
21 #include <xen/interface/xen.h>
22 #include <xen/interface/event_channel.h>
23 
24 #include "events_internal.h"
25 
26 /*
27  * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be
28  * careful to only use bitops which allow for this (e.g
29  * test_bit/find_first_bit and friends but not __ffs) and to pass
30  * BITS_PER_EVTCHN_WORD as the bitmask length.
31  */
32 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8)
33 /*
34  * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t
35  * array. Primarily to avoid long lines (hence the terse name).
36  */
37 #define BM(x) (unsigned long *)(x)
38 /* Find the first set bit in a evtchn mask */
39 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD)
40 
41 #define EVTCHN_MASK_SIZE (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD)
42 
43 static DEFINE_PER_CPU(xen_ulong_t [EVTCHN_MASK_SIZE], cpu_evtchn_mask);
44 
45 static unsigned evtchn_2l_max_channels(void)
46 {
47 	return EVTCHN_2L_NR_CHANNELS;
48 }
49 
50 static void evtchn_2l_bind_to_cpu(struct irq_info *info, unsigned cpu)
51 {
52 	clear_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, info->cpu)));
53 	set_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, cpu)));
54 }
55 
56 static void evtchn_2l_clear_pending(evtchn_port_t port)
57 {
58 	struct shared_info *s = HYPERVISOR_shared_info;
59 	sync_clear_bit(port, BM(&s->evtchn_pending[0]));
60 }
61 
62 static void evtchn_2l_set_pending(evtchn_port_t port)
63 {
64 	struct shared_info *s = HYPERVISOR_shared_info;
65 	sync_set_bit(port, BM(&s->evtchn_pending[0]));
66 }
67 
68 static bool evtchn_2l_is_pending(evtchn_port_t port)
69 {
70 	struct shared_info *s = HYPERVISOR_shared_info;
71 	return sync_test_bit(port, BM(&s->evtchn_pending[0]));
72 }
73 
74 static bool evtchn_2l_test_and_set_mask(evtchn_port_t port)
75 {
76 	struct shared_info *s = HYPERVISOR_shared_info;
77 	return sync_test_and_set_bit(port, BM(&s->evtchn_mask[0]));
78 }
79 
80 static void evtchn_2l_mask(evtchn_port_t port)
81 {
82 	struct shared_info *s = HYPERVISOR_shared_info;
83 	sync_set_bit(port, BM(&s->evtchn_mask[0]));
84 }
85 
86 static void evtchn_2l_unmask(evtchn_port_t port)
87 {
88 	struct shared_info *s = HYPERVISOR_shared_info;
89 	unsigned int cpu = get_cpu();
90 	int do_hypercall = 0, evtchn_pending = 0;
91 
92 	BUG_ON(!irqs_disabled());
93 
94 	if (unlikely((cpu != cpu_from_evtchn(port))))
95 		do_hypercall = 1;
96 	else {
97 		/*
98 		 * Need to clear the mask before checking pending to
99 		 * avoid a race with an event becoming pending.
100 		 *
101 		 * EVTCHNOP_unmask will only trigger an upcall if the
102 		 * mask bit was set, so if a hypercall is needed
103 		 * remask the event.
104 		 */
105 		sync_clear_bit(port, BM(&s->evtchn_mask[0]));
106 		evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
107 
108 		if (unlikely(evtchn_pending && xen_hvm_domain())) {
109 			sync_set_bit(port, BM(&s->evtchn_mask[0]));
110 			do_hypercall = 1;
111 		}
112 	}
113 
114 	/* Slow path (hypercall) if this is a non-local port or if this is
115 	 * an hvm domain and an event is pending (hvm domains don't have
116 	 * their own implementation of irq_enable). */
117 	if (do_hypercall) {
118 		struct evtchn_unmask unmask = { .port = port };
119 		(void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
120 	} else {
121 		struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
122 
123 		/*
124 		 * The following is basically the equivalent of
125 		 * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
126 		 * the interrupt edge' if the channel is masked.
127 		 */
128 		if (evtchn_pending &&
129 		    !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
130 					   BM(&vcpu_info->evtchn_pending_sel)))
131 			vcpu_info->evtchn_upcall_pending = 1;
132 	}
133 
134 	put_cpu();
135 }
136 
137 static DEFINE_PER_CPU(unsigned int, current_word_idx);
138 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
139 
140 /*
141  * Mask out the i least significant bits of w
142  */
143 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i))
144 
145 static inline xen_ulong_t active_evtchns(unsigned int cpu,
146 					 struct shared_info *sh,
147 					 unsigned int idx)
148 {
149 	return sh->evtchn_pending[idx] &
150 		per_cpu(cpu_evtchn_mask, cpu)[idx] &
151 		~sh->evtchn_mask[idx];
152 }
153 
154 /*
155  * Search the CPU's pending events bitmasks.  For each one found, map
156  * the event number to an irq, and feed it into do_IRQ() for handling.
157  *
158  * Xen uses a two-level bitmap to speed searching.  The first level is
159  * a bitset of words which contain pending event bits.  The second
160  * level is a bitset of pending events themselves.
161  */
162 static void evtchn_2l_handle_events(unsigned cpu)
163 {
164 	int irq;
165 	xen_ulong_t pending_words;
166 	xen_ulong_t pending_bits;
167 	int start_word_idx, start_bit_idx;
168 	int word_idx, bit_idx;
169 	int i;
170 	struct shared_info *s = HYPERVISOR_shared_info;
171 	struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
172 
173 	/* Timer interrupt has highest priority. */
174 	irq = irq_from_virq(cpu, VIRQ_TIMER);
175 	if (irq != -1) {
176 		evtchn_port_t evtchn = evtchn_from_irq(irq);
177 		word_idx = evtchn / BITS_PER_LONG;
178 		bit_idx = evtchn % BITS_PER_LONG;
179 		if (active_evtchns(cpu, s, word_idx) & (1ULL << bit_idx))
180 			generic_handle_irq(irq);
181 	}
182 
183 	/*
184 	 * Master flag must be cleared /before/ clearing
185 	 * selector flag. xchg_xen_ulong must contain an
186 	 * appropriate barrier.
187 	 */
188 	pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0);
189 
190 	start_word_idx = __this_cpu_read(current_word_idx);
191 	start_bit_idx = __this_cpu_read(current_bit_idx);
192 
193 	word_idx = start_word_idx;
194 
195 	for (i = 0; pending_words != 0; i++) {
196 		xen_ulong_t words;
197 
198 		words = MASK_LSBS(pending_words, word_idx);
199 
200 		/*
201 		 * If we masked out all events, wrap to beginning.
202 		 */
203 		if (words == 0) {
204 			word_idx = 0;
205 			bit_idx = 0;
206 			continue;
207 		}
208 		word_idx = EVTCHN_FIRST_BIT(words);
209 
210 		pending_bits = active_evtchns(cpu, s, word_idx);
211 		bit_idx = 0; /* usually scan entire word from start */
212 		/*
213 		 * We scan the starting word in two parts.
214 		 *
215 		 * 1st time: start in the middle, scanning the
216 		 * upper bits.
217 		 *
218 		 * 2nd time: scan the whole word (not just the
219 		 * parts skipped in the first pass) -- if an
220 		 * event in the previously scanned bits is
221 		 * pending again it would just be scanned on
222 		 * the next loop anyway.
223 		 */
224 		if (word_idx == start_word_idx) {
225 			if (i == 0)
226 				bit_idx = start_bit_idx;
227 		}
228 
229 		do {
230 			xen_ulong_t bits;
231 			evtchn_port_t port;
232 
233 			bits = MASK_LSBS(pending_bits, bit_idx);
234 
235 			/* If we masked out all events, move on. */
236 			if (bits == 0)
237 				break;
238 
239 			bit_idx = EVTCHN_FIRST_BIT(bits);
240 
241 			/* Process port. */
242 			port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
243 			irq = get_evtchn_to_irq(port);
244 
245 			if (irq != -1)
246 				generic_handle_irq(irq);
247 
248 			bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
249 
250 			/* Next caller starts at last processed + 1 */
251 			__this_cpu_write(current_word_idx,
252 					 bit_idx ? word_idx :
253 					 (word_idx+1) % BITS_PER_EVTCHN_WORD);
254 			__this_cpu_write(current_bit_idx, bit_idx);
255 		} while (bit_idx != 0);
256 
257 		/* Scan start_l1i twice; all others once. */
258 		if ((word_idx != start_word_idx) || (i != 0))
259 			pending_words &= ~(1UL << word_idx);
260 
261 		word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
262 	}
263 }
264 
265 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
266 {
267 	struct shared_info *sh = HYPERVISOR_shared_info;
268 	int cpu = smp_processor_id();
269 	xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
270 	int i;
271 	unsigned long flags;
272 	static DEFINE_SPINLOCK(debug_lock);
273 	struct vcpu_info *v;
274 
275 	spin_lock_irqsave(&debug_lock, flags);
276 
277 	printk("\nvcpu %d\n  ", cpu);
278 
279 	for_each_online_cpu(i) {
280 		int pending;
281 		v = per_cpu(xen_vcpu, i);
282 		pending = (get_irq_regs() && i == cpu)
283 			? xen_irqs_disabled(get_irq_regs())
284 			: v->evtchn_upcall_mask;
285 		printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n  ", i,
286 		       pending, v->evtchn_upcall_pending,
287 		       (int)(sizeof(v->evtchn_pending_sel)*2),
288 		       v->evtchn_pending_sel);
289 	}
290 	v = per_cpu(xen_vcpu, cpu);
291 
292 	printk("\npending:\n   ");
293 	for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
294 		printk("%0*"PRI_xen_ulong"%s",
295 		       (int)sizeof(sh->evtchn_pending[0])*2,
296 		       sh->evtchn_pending[i],
297 		       i % 8 == 0 ? "\n   " : " ");
298 	printk("\nglobal mask:\n   ");
299 	for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
300 		printk("%0*"PRI_xen_ulong"%s",
301 		       (int)(sizeof(sh->evtchn_mask[0])*2),
302 		       sh->evtchn_mask[i],
303 		       i % 8 == 0 ? "\n   " : " ");
304 
305 	printk("\nglobally unmasked:\n   ");
306 	for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
307 		printk("%0*"PRI_xen_ulong"%s",
308 		       (int)(sizeof(sh->evtchn_mask[0])*2),
309 		       sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
310 		       i % 8 == 0 ? "\n   " : " ");
311 
312 	printk("\nlocal cpu%d mask:\n   ", cpu);
313 	for (i = (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--)
314 		printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2),
315 		       cpu_evtchn[i],
316 		       i % 8 == 0 ? "\n   " : " ");
317 
318 	printk("\nlocally unmasked:\n   ");
319 	for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
320 		xen_ulong_t pending = sh->evtchn_pending[i]
321 			& ~sh->evtchn_mask[i]
322 			& cpu_evtchn[i];
323 		printk("%0*"PRI_xen_ulong"%s",
324 		       (int)(sizeof(sh->evtchn_mask[0])*2),
325 		       pending, i % 8 == 0 ? "\n   " : " ");
326 	}
327 
328 	printk("\npending list:\n");
329 	for (i = 0; i < EVTCHN_2L_NR_CHANNELS; i++) {
330 		if (sync_test_bit(i, BM(sh->evtchn_pending))) {
331 			int word_idx = i / BITS_PER_EVTCHN_WORD;
332 			printk("  %d: event %d -> irq %d%s%s%s\n",
333 			       cpu_from_evtchn(i), i,
334 			       get_evtchn_to_irq(i),
335 			       sync_test_bit(word_idx, BM(&v->evtchn_pending_sel))
336 			       ? "" : " l2-clear",
337 			       !sync_test_bit(i, BM(sh->evtchn_mask))
338 			       ? "" : " globally-masked",
339 			       sync_test_bit(i, BM(cpu_evtchn))
340 			       ? "" : " locally-masked");
341 		}
342 	}
343 
344 	spin_unlock_irqrestore(&debug_lock, flags);
345 
346 	return IRQ_HANDLED;
347 }
348 
349 static void evtchn_2l_resume(void)
350 {
351 	int i;
352 
353 	for_each_online_cpu(i)
354 		memset(per_cpu(cpu_evtchn_mask, i), 0, sizeof(xen_ulong_t) *
355 				EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD);
356 }
357 
358 static const struct evtchn_ops evtchn_ops_2l = {
359 	.max_channels      = evtchn_2l_max_channels,
360 	.nr_channels       = evtchn_2l_max_channels,
361 	.bind_to_cpu       = evtchn_2l_bind_to_cpu,
362 	.clear_pending     = evtchn_2l_clear_pending,
363 	.set_pending       = evtchn_2l_set_pending,
364 	.is_pending        = evtchn_2l_is_pending,
365 	.test_and_set_mask = evtchn_2l_test_and_set_mask,
366 	.mask              = evtchn_2l_mask,
367 	.unmask            = evtchn_2l_unmask,
368 	.handle_events     = evtchn_2l_handle_events,
369 	.resume	           = evtchn_2l_resume,
370 };
371 
372 void __init xen_evtchn_2l_init(void)
373 {
374 	pr_info("Using 2-level ABI\n");
375 	evtchn_ops = &evtchn_ops_2l;
376 }
377