xref: /linux/drivers/hv/hv.c (revision b68fc09be48edbc47de1a0f3d42ef8adf6c0ac55)
1 /*
2  * Copyright (c) 2009, Microsoft Corporation.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms and conditions of the GNU General Public License,
6  * version 2, as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
11  * more details.
12  *
13  * You should have received a copy of the GNU General Public License along with
14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
15  * Place - Suite 330, Boston, MA 02111-1307 USA.
16  *
17  * Authors:
18  *   Haiyang Zhang <haiyangz@microsoft.com>
19  *   Hank Janssen  <hjanssen@microsoft.com>
20  *
21  */
22 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
23 
24 #include <linux/kernel.h>
25 #include <linux/mm.h>
26 #include <linux/slab.h>
27 #include <linux/vmalloc.h>
28 #include <linux/hyperv.h>
29 #include <linux/version.h>
30 #include <linux/random.h>
31 #include <linux/clockchips.h>
32 #include <asm/mshyperv.h>
33 #include "hyperv_vmbus.h"
34 
35 /* The one and only */
36 struct hv_context hv_context = {
37 	.synic_initialized	= false,
38 };
39 
40 /*
41  * If false, we're using the old mechanism for stimer0 interrupts
42  * where it sends a VMbus message when it expires. The old
43  * mechanism is used when running on older versions of Hyper-V
44  * that don't support Direct Mode. While Hyper-V provides
45  * four stimer's per CPU, Linux uses only stimer0.
46  */
47 static bool direct_mode_enabled;
48 static int stimer0_irq;
49 static int stimer0_vector;
50 
51 #define HV_TIMER_FREQUENCY (10 * 1000 * 1000) /* 100ns period */
52 #define HV_MAX_MAX_DELTA_TICKS 0xffffffff
53 #define HV_MIN_DELTA_TICKS 1
54 
55 /*
56  * hv_init - Main initialization routine.
57  *
58  * This routine must be called before any other routines in here are called
59  */
60 int hv_init(void)
61 {
62 	hv_context.cpu_context = alloc_percpu(struct hv_per_cpu_context);
63 	if (!hv_context.cpu_context)
64 		return -ENOMEM;
65 
66 	direct_mode_enabled = ms_hyperv.misc_features &
67 			HV_STIMER_DIRECT_MODE_AVAILABLE;
68 	return 0;
69 }
70 
71 /*
72  * hv_post_message - Post a message using the hypervisor message IPC.
73  *
74  * This involves a hypercall.
75  */
76 int hv_post_message(union hv_connection_id connection_id,
77 		  enum hv_message_type message_type,
78 		  void *payload, size_t payload_size)
79 {
80 	struct hv_input_post_message *aligned_msg;
81 	struct hv_per_cpu_context *hv_cpu;
82 	u64 status;
83 
84 	if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
85 		return -EMSGSIZE;
86 
87 	hv_cpu = get_cpu_ptr(hv_context.cpu_context);
88 	aligned_msg = hv_cpu->post_msg_page;
89 	aligned_msg->connectionid = connection_id;
90 	aligned_msg->reserved = 0;
91 	aligned_msg->message_type = message_type;
92 	aligned_msg->payload_size = payload_size;
93 	memcpy((void *)aligned_msg->payload, payload, payload_size);
94 
95 	status = hv_do_hypercall(HVCALL_POST_MESSAGE, aligned_msg, NULL);
96 
97 	/* Preemption must remain disabled until after the hypercall
98 	 * so some other thread can't get scheduled onto this cpu and
99 	 * corrupt the per-cpu post_msg_page
100 	 */
101 	put_cpu_ptr(hv_cpu);
102 
103 	return status & 0xFFFF;
104 }
105 
106 /*
107  * ISR for when stimer0 is operating in Direct Mode.  Direct Mode
108  * does not use VMbus or any VMbus messages, so process here and not
109  * in the VMbus driver code.
110  */
111 
112 static void hv_stimer0_isr(void)
113 {
114 	struct hv_per_cpu_context *hv_cpu;
115 
116 	hv_cpu = this_cpu_ptr(hv_context.cpu_context);
117 	hv_cpu->clk_evt->event_handler(hv_cpu->clk_evt);
118 	add_interrupt_randomness(stimer0_vector, 0);
119 }
120 
121 static int hv_ce_set_next_event(unsigned long delta,
122 				struct clock_event_device *evt)
123 {
124 	u64 current_tick;
125 
126 	WARN_ON(!clockevent_state_oneshot(evt));
127 
128 	current_tick = hyperv_cs->read(NULL);
129 	current_tick += delta;
130 	hv_init_timer(0, current_tick);
131 	return 0;
132 }
133 
134 static int hv_ce_shutdown(struct clock_event_device *evt)
135 {
136 	hv_init_timer(0, 0);
137 	hv_init_timer_config(0, 0);
138 	if (direct_mode_enabled)
139 		hv_disable_stimer0_percpu_irq(stimer0_irq);
140 
141 	return 0;
142 }
143 
144 static int hv_ce_set_oneshot(struct clock_event_device *evt)
145 {
146 	union hv_timer_config timer_cfg;
147 
148 	timer_cfg.as_uint64 = 0;
149 	timer_cfg.enable = 1;
150 	timer_cfg.auto_enable = 1;
151 	if (direct_mode_enabled) {
152 		/*
153 		 * When it expires, the timer will directly interrupt
154 		 * on the specified hardware vector/IRQ.
155 		 */
156 		timer_cfg.direct_mode = 1;
157 		timer_cfg.apic_vector = stimer0_vector;
158 		hv_enable_stimer0_percpu_irq(stimer0_irq);
159 	} else {
160 		/*
161 		 * When it expires, the timer will generate a VMbus message,
162 		 * to be handled by the normal VMbus interrupt handler.
163 		 */
164 		timer_cfg.direct_mode = 0;
165 		timer_cfg.sintx = VMBUS_MESSAGE_SINT;
166 	}
167 	hv_init_timer_config(0, timer_cfg.as_uint64);
168 	return 0;
169 }
170 
171 static void hv_init_clockevent_device(struct clock_event_device *dev, int cpu)
172 {
173 	dev->name = "Hyper-V clockevent";
174 	dev->features = CLOCK_EVT_FEAT_ONESHOT;
175 	dev->cpumask = cpumask_of(cpu);
176 	dev->rating = 1000;
177 	/*
178 	 * Avoid settint dev->owner = THIS_MODULE deliberately as doing so will
179 	 * result in clockevents_config_and_register() taking additional
180 	 * references to the hv_vmbus module making it impossible to unload.
181 	 */
182 
183 	dev->set_state_shutdown = hv_ce_shutdown;
184 	dev->set_state_oneshot = hv_ce_set_oneshot;
185 	dev->set_next_event = hv_ce_set_next_event;
186 }
187 
188 
189 int hv_synic_alloc(void)
190 {
191 	int cpu;
192 
193 	hv_context.hv_numa_map = kcalloc(nr_node_ids, sizeof(struct cpumask),
194 					 GFP_KERNEL);
195 	if (hv_context.hv_numa_map == NULL) {
196 		pr_err("Unable to allocate NUMA map\n");
197 		goto err;
198 	}
199 
200 	for_each_present_cpu(cpu) {
201 		struct hv_per_cpu_context *hv_cpu
202 			= per_cpu_ptr(hv_context.cpu_context, cpu);
203 
204 		memset(hv_cpu, 0, sizeof(*hv_cpu));
205 		tasklet_init(&hv_cpu->msg_dpc,
206 			     vmbus_on_msg_dpc, (unsigned long) hv_cpu);
207 
208 		hv_cpu->clk_evt = kzalloc(sizeof(struct clock_event_device),
209 					  GFP_KERNEL);
210 		if (hv_cpu->clk_evt == NULL) {
211 			pr_err("Unable to allocate clock event device\n");
212 			goto err;
213 		}
214 		hv_init_clockevent_device(hv_cpu->clk_evt, cpu);
215 
216 		hv_cpu->synic_message_page =
217 			(void *)get_zeroed_page(GFP_ATOMIC);
218 		if (hv_cpu->synic_message_page == NULL) {
219 			pr_err("Unable to allocate SYNIC message page\n");
220 			goto err;
221 		}
222 
223 		hv_cpu->synic_event_page = (void *)get_zeroed_page(GFP_ATOMIC);
224 		if (hv_cpu->synic_event_page == NULL) {
225 			pr_err("Unable to allocate SYNIC event page\n");
226 			goto err;
227 		}
228 
229 		hv_cpu->post_msg_page = (void *)get_zeroed_page(GFP_ATOMIC);
230 		if (hv_cpu->post_msg_page == NULL) {
231 			pr_err("Unable to allocate post msg page\n");
232 			goto err;
233 		}
234 
235 		INIT_LIST_HEAD(&hv_cpu->chan_list);
236 	}
237 
238 	if (direct_mode_enabled &&
239 	    hv_setup_stimer0_irq(&stimer0_irq, &stimer0_vector,
240 				hv_stimer0_isr))
241 		goto err;
242 
243 	return 0;
244 err:
245 	/*
246 	 * Any memory allocations that succeeded will be freed when
247 	 * the caller cleans up by calling hv_synic_free()
248 	 */
249 	return -ENOMEM;
250 }
251 
252 
253 void hv_synic_free(void)
254 {
255 	int cpu;
256 
257 	for_each_present_cpu(cpu) {
258 		struct hv_per_cpu_context *hv_cpu
259 			= per_cpu_ptr(hv_context.cpu_context, cpu);
260 
261 		kfree(hv_cpu->clk_evt);
262 		free_page((unsigned long)hv_cpu->synic_event_page);
263 		free_page((unsigned long)hv_cpu->synic_message_page);
264 		free_page((unsigned long)hv_cpu->post_msg_page);
265 	}
266 
267 	kfree(hv_context.hv_numa_map);
268 }
269 
270 /*
271  * hv_synic_init - Initialize the Synthetic Interrupt Controller.
272  *
273  * If it is already initialized by another entity (ie x2v shim), we need to
274  * retrieve the initialized message and event pages.  Otherwise, we create and
275  * initialize the message and event pages.
276  */
277 int hv_synic_init(unsigned int cpu)
278 {
279 	struct hv_per_cpu_context *hv_cpu
280 		= per_cpu_ptr(hv_context.cpu_context, cpu);
281 	union hv_synic_simp simp;
282 	union hv_synic_siefp siefp;
283 	union hv_synic_sint shared_sint;
284 	union hv_synic_scontrol sctrl;
285 
286 	/* Setup the Synic's message page */
287 	hv_get_simp(simp.as_uint64);
288 	simp.simp_enabled = 1;
289 	simp.base_simp_gpa = virt_to_phys(hv_cpu->synic_message_page)
290 		>> PAGE_SHIFT;
291 
292 	hv_set_simp(simp.as_uint64);
293 
294 	/* Setup the Synic's event page */
295 	hv_get_siefp(siefp.as_uint64);
296 	siefp.siefp_enabled = 1;
297 	siefp.base_siefp_gpa = virt_to_phys(hv_cpu->synic_event_page)
298 		>> PAGE_SHIFT;
299 
300 	hv_set_siefp(siefp.as_uint64);
301 
302 	/* Setup the shared SINT. */
303 	hv_get_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
304 
305 	shared_sint.vector = HYPERVISOR_CALLBACK_VECTOR;
306 	shared_sint.masked = false;
307 	if (ms_hyperv.hints & HV_DEPRECATING_AEOI_RECOMMENDED)
308 		shared_sint.auto_eoi = false;
309 	else
310 		shared_sint.auto_eoi = true;
311 
312 	hv_set_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
313 
314 	/* Enable the global synic bit */
315 	hv_get_synic_state(sctrl.as_uint64);
316 	sctrl.enable = 1;
317 
318 	hv_set_synic_state(sctrl.as_uint64);
319 
320 	hv_context.synic_initialized = true;
321 
322 	/*
323 	 * Register the per-cpu clockevent source.
324 	 */
325 	if (ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE)
326 		clockevents_config_and_register(hv_cpu->clk_evt,
327 						HV_TIMER_FREQUENCY,
328 						HV_MIN_DELTA_TICKS,
329 						HV_MAX_MAX_DELTA_TICKS);
330 	return 0;
331 }
332 
333 /*
334  * hv_synic_clockevents_cleanup - Cleanup clockevent devices
335  */
336 void hv_synic_clockevents_cleanup(void)
337 {
338 	int cpu;
339 
340 	if (!(ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE))
341 		return;
342 
343 	if (direct_mode_enabled)
344 		hv_remove_stimer0_irq(stimer0_irq);
345 
346 	for_each_present_cpu(cpu) {
347 		struct hv_per_cpu_context *hv_cpu
348 			= per_cpu_ptr(hv_context.cpu_context, cpu);
349 
350 		clockevents_unbind_device(hv_cpu->clk_evt, cpu);
351 	}
352 }
353 
354 /*
355  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
356  */
357 int hv_synic_cleanup(unsigned int cpu)
358 {
359 	union hv_synic_sint shared_sint;
360 	union hv_synic_simp simp;
361 	union hv_synic_siefp siefp;
362 	union hv_synic_scontrol sctrl;
363 	struct vmbus_channel *channel, *sc;
364 	bool channel_found = false;
365 	unsigned long flags;
366 
367 	if (!hv_context.synic_initialized)
368 		return -EFAULT;
369 
370 	/*
371 	 * Search for channels which are bound to the CPU we're about to
372 	 * cleanup. In case we find one and vmbus is still connected we need to
373 	 * fail, this will effectively prevent CPU offlining. There is no way
374 	 * we can re-bind channels to different CPUs for now.
375 	 */
376 	mutex_lock(&vmbus_connection.channel_mutex);
377 	list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
378 		if (channel->target_cpu == cpu) {
379 			channel_found = true;
380 			break;
381 		}
382 		spin_lock_irqsave(&channel->lock, flags);
383 		list_for_each_entry(sc, &channel->sc_list, sc_list) {
384 			if (sc->target_cpu == cpu) {
385 				channel_found = true;
386 				break;
387 			}
388 		}
389 		spin_unlock_irqrestore(&channel->lock, flags);
390 		if (channel_found)
391 			break;
392 	}
393 	mutex_unlock(&vmbus_connection.channel_mutex);
394 
395 	if (channel_found && vmbus_connection.conn_state == CONNECTED)
396 		return -EBUSY;
397 
398 	/* Turn off clockevent device */
399 	if (ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE) {
400 		struct hv_per_cpu_context *hv_cpu
401 			= this_cpu_ptr(hv_context.cpu_context);
402 
403 		clockevents_unbind_device(hv_cpu->clk_evt, cpu);
404 		hv_ce_shutdown(hv_cpu->clk_evt);
405 		put_cpu_ptr(hv_cpu);
406 	}
407 
408 	hv_get_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
409 
410 	shared_sint.masked = 1;
411 
412 	/* Need to correctly cleanup in the case of SMP!!! */
413 	/* Disable the interrupt */
414 	hv_set_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
415 
416 	hv_get_simp(simp.as_uint64);
417 	simp.simp_enabled = 0;
418 	simp.base_simp_gpa = 0;
419 
420 	hv_set_simp(simp.as_uint64);
421 
422 	hv_get_siefp(siefp.as_uint64);
423 	siefp.siefp_enabled = 0;
424 	siefp.base_siefp_gpa = 0;
425 
426 	hv_set_siefp(siefp.as_uint64);
427 
428 	/* Disable the global synic bit */
429 	hv_get_synic_state(sctrl.as_uint64);
430 	sctrl.enable = 0;
431 	hv_set_synic_state(sctrl.as_uint64);
432 
433 	return 0;
434 }
435