xref: /linux/drivers/hv/hv.c (revision db6d8d5fdf9537641c76ba7f32e02b4bcc600972)
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/interrupt.h>
31 #include <linux/clockchips.h>
32 #include <asm/hyperv.h>
33 #include <asm/mshyperv.h>
34 #include "hyperv_vmbus.h"
35 
36 /* The one and only */
37 struct hv_context hv_context = {
38 	.synic_initialized	= false,
39 	.hypercall_page		= NULL,
40 };
41 
42 #define HV_TIMER_FREQUENCY (10 * 1000 * 1000) /* 100ns period */
43 #define HV_MAX_MAX_DELTA_TICKS 0xffffffff
44 #define HV_MIN_DELTA_TICKS 1
45 
46 /*
47  * query_hypervisor_info - Get version info of the windows hypervisor
48  */
49 unsigned int host_info_eax;
50 unsigned int host_info_ebx;
51 unsigned int host_info_ecx;
52 unsigned int host_info_edx;
53 
54 static int query_hypervisor_info(void)
55 {
56 	unsigned int eax;
57 	unsigned int ebx;
58 	unsigned int ecx;
59 	unsigned int edx;
60 	unsigned int max_leaf;
61 	unsigned int op;
62 
63 	/*
64 	* Its assumed that this is called after confirming that Viridian
65 	* is present. Query id and revision.
66 	*/
67 	eax = 0;
68 	ebx = 0;
69 	ecx = 0;
70 	edx = 0;
71 	op = HVCPUID_VENDOR_MAXFUNCTION;
72 	cpuid(op, &eax, &ebx, &ecx, &edx);
73 
74 	max_leaf = eax;
75 
76 	if (max_leaf >= HVCPUID_VERSION) {
77 		eax = 0;
78 		ebx = 0;
79 		ecx = 0;
80 		edx = 0;
81 		op = HVCPUID_VERSION;
82 		cpuid(op, &eax, &ebx, &ecx, &edx);
83 		host_info_eax = eax;
84 		host_info_ebx = ebx;
85 		host_info_ecx = ecx;
86 		host_info_edx = edx;
87 	}
88 	return max_leaf;
89 }
90 
91 /*
92  * hv_do_hypercall- Invoke the specified hypercall
93  */
94 u64 hv_do_hypercall(u64 control, void *input, void *output)
95 {
96 	u64 input_address = (input) ? virt_to_phys(input) : 0;
97 	u64 output_address = (output) ? virt_to_phys(output) : 0;
98 	void *hypercall_page = hv_context.hypercall_page;
99 #ifdef CONFIG_X86_64
100 	u64 hv_status = 0;
101 
102 	if (!hypercall_page)
103 		return (u64)ULLONG_MAX;
104 
105 	__asm__ __volatile__("mov %0, %%r8" : : "r" (output_address) : "r8");
106 	__asm__ __volatile__("call *%3" : "=a" (hv_status) :
107 			     "c" (control), "d" (input_address),
108 			     "m" (hypercall_page));
109 
110 	return hv_status;
111 
112 #else
113 
114 	u32 control_hi = control >> 32;
115 	u32 control_lo = control & 0xFFFFFFFF;
116 	u32 hv_status_hi = 1;
117 	u32 hv_status_lo = 1;
118 	u32 input_address_hi = input_address >> 32;
119 	u32 input_address_lo = input_address & 0xFFFFFFFF;
120 	u32 output_address_hi = output_address >> 32;
121 	u32 output_address_lo = output_address & 0xFFFFFFFF;
122 
123 	if (!hypercall_page)
124 		return (u64)ULLONG_MAX;
125 
126 	__asm__ __volatile__ ("call *%8" : "=d"(hv_status_hi),
127 			      "=a"(hv_status_lo) : "d" (control_hi),
128 			      "a" (control_lo), "b" (input_address_hi),
129 			      "c" (input_address_lo), "D"(output_address_hi),
130 			      "S"(output_address_lo), "m" (hypercall_page));
131 
132 	return hv_status_lo | ((u64)hv_status_hi << 32);
133 #endif /* !x86_64 */
134 }
135 EXPORT_SYMBOL_GPL(hv_do_hypercall);
136 
137 #ifdef CONFIG_X86_64
138 static cycle_t read_hv_clock_tsc(struct clocksource *arg)
139 {
140 	cycle_t current_tick;
141 	struct ms_hyperv_tsc_page *tsc_pg = hv_context.tsc_page;
142 
143 	if (tsc_pg->tsc_sequence != 0) {
144 		/*
145 		 * Use the tsc page to compute the value.
146 		 */
147 
148 		while (1) {
149 			cycle_t tmp;
150 			u32 sequence = tsc_pg->tsc_sequence;
151 			u64 cur_tsc;
152 			u64 scale = tsc_pg->tsc_scale;
153 			s64 offset = tsc_pg->tsc_offset;
154 
155 			rdtscll(cur_tsc);
156 			/* current_tick = ((cur_tsc *scale) >> 64) + offset */
157 			asm("mulq %3"
158 				: "=d" (current_tick), "=a" (tmp)
159 				: "a" (cur_tsc), "r" (scale));
160 
161 			current_tick += offset;
162 			if (tsc_pg->tsc_sequence == sequence)
163 				return current_tick;
164 
165 			if (tsc_pg->tsc_sequence != 0)
166 				continue;
167 			/*
168 			 * Fallback using MSR method.
169 			 */
170 			break;
171 		}
172 	}
173 	rdmsrl(HV_X64_MSR_TIME_REF_COUNT, current_tick);
174 	return current_tick;
175 }
176 
177 static struct clocksource hyperv_cs_tsc = {
178 		.name           = "hyperv_clocksource_tsc_page",
179 		.rating         = 425,
180 		.read           = read_hv_clock_tsc,
181 		.mask           = CLOCKSOURCE_MASK(64),
182 		.flags          = CLOCK_SOURCE_IS_CONTINUOUS,
183 };
184 #endif
185 
186 
187 /*
188  * hv_init - Main initialization routine.
189  *
190  * This routine must be called before any other routines in here are called
191  */
192 int hv_init(void)
193 {
194 	int max_leaf;
195 	union hv_x64_msr_hypercall_contents hypercall_msr;
196 	void *virtaddr = NULL;
197 
198 	memset(hv_context.synic_event_page, 0, sizeof(void *) * NR_CPUS);
199 	memset(hv_context.synic_message_page, 0,
200 	       sizeof(void *) * NR_CPUS);
201 	memset(hv_context.post_msg_page, 0,
202 	       sizeof(void *) * NR_CPUS);
203 	memset(hv_context.vp_index, 0,
204 	       sizeof(int) * NR_CPUS);
205 	memset(hv_context.event_dpc, 0,
206 	       sizeof(void *) * NR_CPUS);
207 	memset(hv_context.msg_dpc, 0,
208 	       sizeof(void *) * NR_CPUS);
209 	memset(hv_context.clk_evt, 0,
210 	       sizeof(void *) * NR_CPUS);
211 
212 	max_leaf = query_hypervisor_info();
213 
214 	/*
215 	 * Write our OS ID.
216 	 */
217 	hv_context.guestid = generate_guest_id(0, LINUX_VERSION_CODE, 0);
218 	wrmsrl(HV_X64_MSR_GUEST_OS_ID, hv_context.guestid);
219 
220 	/* See if the hypercall page is already set */
221 	rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
222 
223 	virtaddr = __vmalloc(PAGE_SIZE, GFP_KERNEL, PAGE_KERNEL_EXEC);
224 
225 	if (!virtaddr)
226 		goto cleanup;
227 
228 	hypercall_msr.enable = 1;
229 
230 	hypercall_msr.guest_physical_address = vmalloc_to_pfn(virtaddr);
231 	wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
232 
233 	/* Confirm that hypercall page did get setup. */
234 	hypercall_msr.as_uint64 = 0;
235 	rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
236 
237 	if (!hypercall_msr.enable)
238 		goto cleanup;
239 
240 	hv_context.hypercall_page = virtaddr;
241 
242 #ifdef CONFIG_X86_64
243 	if (ms_hyperv.features & HV_X64_MSR_REFERENCE_TSC_AVAILABLE) {
244 		union hv_x64_msr_hypercall_contents tsc_msr;
245 		void *va_tsc;
246 
247 		va_tsc = __vmalloc(PAGE_SIZE, GFP_KERNEL, PAGE_KERNEL);
248 		if (!va_tsc)
249 			goto cleanup;
250 		hv_context.tsc_page = va_tsc;
251 
252 		rdmsrl(HV_X64_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
253 
254 		tsc_msr.enable = 1;
255 		tsc_msr.guest_physical_address = vmalloc_to_pfn(va_tsc);
256 
257 		wrmsrl(HV_X64_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
258 		clocksource_register_hz(&hyperv_cs_tsc, NSEC_PER_SEC/100);
259 	}
260 #endif
261 	return 0;
262 
263 cleanup:
264 	if (virtaddr) {
265 		if (hypercall_msr.enable) {
266 			hypercall_msr.as_uint64 = 0;
267 			wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
268 		}
269 
270 		vfree(virtaddr);
271 	}
272 
273 	return -ENOTSUPP;
274 }
275 
276 /*
277  * hv_cleanup - Cleanup routine.
278  *
279  * This routine is called normally during driver unloading or exiting.
280  */
281 void hv_cleanup(bool crash)
282 {
283 	union hv_x64_msr_hypercall_contents hypercall_msr;
284 
285 	/* Reset our OS id */
286 	wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
287 
288 	if (hv_context.hypercall_page) {
289 		hypercall_msr.as_uint64 = 0;
290 		wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
291 		if (!crash)
292 			vfree(hv_context.hypercall_page);
293 		hv_context.hypercall_page = NULL;
294 	}
295 
296 #ifdef CONFIG_X86_64
297 	/*
298 	 * Cleanup the TSC page based CS.
299 	 */
300 	if (ms_hyperv.features & HV_X64_MSR_REFERENCE_TSC_AVAILABLE) {
301 		/*
302 		 * Crash can happen in an interrupt context and unregistering
303 		 * a clocksource is impossible and redundant in this case.
304 		 */
305 		if (!oops_in_progress) {
306 			clocksource_change_rating(&hyperv_cs_tsc, 10);
307 			clocksource_unregister(&hyperv_cs_tsc);
308 		}
309 
310 		hypercall_msr.as_uint64 = 0;
311 		wrmsrl(HV_X64_MSR_REFERENCE_TSC, hypercall_msr.as_uint64);
312 		if (!crash)
313 			vfree(hv_context.tsc_page);
314 		hv_context.tsc_page = NULL;
315 	}
316 #endif
317 }
318 
319 /*
320  * hv_post_message - Post a message using the hypervisor message IPC.
321  *
322  * This involves a hypercall.
323  */
324 int hv_post_message(union hv_connection_id connection_id,
325 		  enum hv_message_type message_type,
326 		  void *payload, size_t payload_size)
327 {
328 
329 	struct hv_input_post_message *aligned_msg;
330 	u64 status;
331 
332 	if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
333 		return -EMSGSIZE;
334 
335 	aligned_msg = (struct hv_input_post_message *)
336 			hv_context.post_msg_page[get_cpu()];
337 
338 	aligned_msg->connectionid = connection_id;
339 	aligned_msg->reserved = 0;
340 	aligned_msg->message_type = message_type;
341 	aligned_msg->payload_size = payload_size;
342 	memcpy((void *)aligned_msg->payload, payload, payload_size);
343 
344 	status = hv_do_hypercall(HVCALL_POST_MESSAGE, aligned_msg, NULL);
345 
346 	put_cpu();
347 	return status & 0xFFFF;
348 }
349 
350 static int hv_ce_set_next_event(unsigned long delta,
351 				struct clock_event_device *evt)
352 {
353 	cycle_t current_tick;
354 
355 	WARN_ON(!clockevent_state_oneshot(evt));
356 
357 	rdmsrl(HV_X64_MSR_TIME_REF_COUNT, current_tick);
358 	current_tick += delta;
359 	wrmsrl(HV_X64_MSR_STIMER0_COUNT, current_tick);
360 	return 0;
361 }
362 
363 static int hv_ce_shutdown(struct clock_event_device *evt)
364 {
365 	wrmsrl(HV_X64_MSR_STIMER0_COUNT, 0);
366 	wrmsrl(HV_X64_MSR_STIMER0_CONFIG, 0);
367 
368 	return 0;
369 }
370 
371 static int hv_ce_set_oneshot(struct clock_event_device *evt)
372 {
373 	union hv_timer_config timer_cfg;
374 
375 	timer_cfg.enable = 1;
376 	timer_cfg.auto_enable = 1;
377 	timer_cfg.sintx = VMBUS_MESSAGE_SINT;
378 	wrmsrl(HV_X64_MSR_STIMER0_CONFIG, timer_cfg.as_uint64);
379 
380 	return 0;
381 }
382 
383 static void hv_init_clockevent_device(struct clock_event_device *dev, int cpu)
384 {
385 	dev->name = "Hyper-V clockevent";
386 	dev->features = CLOCK_EVT_FEAT_ONESHOT;
387 	dev->cpumask = cpumask_of(cpu);
388 	dev->rating = 1000;
389 	/*
390 	 * Avoid settint dev->owner = THIS_MODULE deliberately as doing so will
391 	 * result in clockevents_config_and_register() taking additional
392 	 * references to the hv_vmbus module making it impossible to unload.
393 	 */
394 
395 	dev->set_state_shutdown = hv_ce_shutdown;
396 	dev->set_state_oneshot = hv_ce_set_oneshot;
397 	dev->set_next_event = hv_ce_set_next_event;
398 }
399 
400 
401 int hv_synic_alloc(void)
402 {
403 	size_t size = sizeof(struct tasklet_struct);
404 	size_t ced_size = sizeof(struct clock_event_device);
405 	int cpu;
406 
407 	hv_context.hv_numa_map = kzalloc(sizeof(struct cpumask) * nr_node_ids,
408 					 GFP_ATOMIC);
409 	if (hv_context.hv_numa_map == NULL) {
410 		pr_err("Unable to allocate NUMA map\n");
411 		goto err;
412 	}
413 
414 	for_each_online_cpu(cpu) {
415 		hv_context.event_dpc[cpu] = kmalloc(size, GFP_ATOMIC);
416 		if (hv_context.event_dpc[cpu] == NULL) {
417 			pr_err("Unable to allocate event dpc\n");
418 			goto err;
419 		}
420 		tasklet_init(hv_context.event_dpc[cpu], vmbus_on_event, cpu);
421 
422 		hv_context.msg_dpc[cpu] = kmalloc(size, GFP_ATOMIC);
423 		if (hv_context.msg_dpc[cpu] == NULL) {
424 			pr_err("Unable to allocate event dpc\n");
425 			goto err;
426 		}
427 		tasklet_init(hv_context.msg_dpc[cpu], vmbus_on_msg_dpc, cpu);
428 
429 		hv_context.clk_evt[cpu] = kzalloc(ced_size, GFP_ATOMIC);
430 		if (hv_context.clk_evt[cpu] == NULL) {
431 			pr_err("Unable to allocate clock event device\n");
432 			goto err;
433 		}
434 
435 		hv_init_clockevent_device(hv_context.clk_evt[cpu], cpu);
436 
437 		hv_context.synic_message_page[cpu] =
438 			(void *)get_zeroed_page(GFP_ATOMIC);
439 
440 		if (hv_context.synic_message_page[cpu] == NULL) {
441 			pr_err("Unable to allocate SYNIC message page\n");
442 			goto err;
443 		}
444 
445 		hv_context.synic_event_page[cpu] =
446 			(void *)get_zeroed_page(GFP_ATOMIC);
447 
448 		if (hv_context.synic_event_page[cpu] == NULL) {
449 			pr_err("Unable to allocate SYNIC event page\n");
450 			goto err;
451 		}
452 
453 		hv_context.post_msg_page[cpu] =
454 			(void *)get_zeroed_page(GFP_ATOMIC);
455 
456 		if (hv_context.post_msg_page[cpu] == NULL) {
457 			pr_err("Unable to allocate post msg page\n");
458 			goto err;
459 		}
460 	}
461 
462 	return 0;
463 err:
464 	return -ENOMEM;
465 }
466 
467 static void hv_synic_free_cpu(int cpu)
468 {
469 	kfree(hv_context.event_dpc[cpu]);
470 	kfree(hv_context.msg_dpc[cpu]);
471 	kfree(hv_context.clk_evt[cpu]);
472 	if (hv_context.synic_event_page[cpu])
473 		free_page((unsigned long)hv_context.synic_event_page[cpu]);
474 	if (hv_context.synic_message_page[cpu])
475 		free_page((unsigned long)hv_context.synic_message_page[cpu]);
476 	if (hv_context.post_msg_page[cpu])
477 		free_page((unsigned long)hv_context.post_msg_page[cpu]);
478 }
479 
480 void hv_synic_free(void)
481 {
482 	int cpu;
483 
484 	kfree(hv_context.hv_numa_map);
485 	for_each_online_cpu(cpu)
486 		hv_synic_free_cpu(cpu);
487 }
488 
489 /*
490  * hv_synic_init - Initialize the Synthethic Interrupt Controller.
491  *
492  * If it is already initialized by another entity (ie x2v shim), we need to
493  * retrieve the initialized message and event pages.  Otherwise, we create and
494  * initialize the message and event pages.
495  */
496 void hv_synic_init(void *arg)
497 {
498 	u64 version;
499 	union hv_synic_simp simp;
500 	union hv_synic_siefp siefp;
501 	union hv_synic_sint shared_sint;
502 	union hv_synic_scontrol sctrl;
503 	u64 vp_index;
504 
505 	int cpu = smp_processor_id();
506 
507 	if (!hv_context.hypercall_page)
508 		return;
509 
510 	/* Check the version */
511 	rdmsrl(HV_X64_MSR_SVERSION, version);
512 
513 	/* Setup the Synic's message page */
514 	rdmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
515 	simp.simp_enabled = 1;
516 	simp.base_simp_gpa = virt_to_phys(hv_context.synic_message_page[cpu])
517 		>> PAGE_SHIFT;
518 
519 	wrmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
520 
521 	/* Setup the Synic's event page */
522 	rdmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
523 	siefp.siefp_enabled = 1;
524 	siefp.base_siefp_gpa = virt_to_phys(hv_context.synic_event_page[cpu])
525 		>> PAGE_SHIFT;
526 
527 	wrmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
528 
529 	/* Setup the shared SINT. */
530 	rdmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
531 
532 	shared_sint.as_uint64 = 0;
533 	shared_sint.vector = HYPERVISOR_CALLBACK_VECTOR;
534 	shared_sint.masked = false;
535 	shared_sint.auto_eoi = true;
536 
537 	wrmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
538 
539 	/* Enable the global synic bit */
540 	rdmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
541 	sctrl.enable = 1;
542 
543 	wrmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
544 
545 	hv_context.synic_initialized = true;
546 
547 	/*
548 	 * Setup the mapping between Hyper-V's notion
549 	 * of cpuid and Linux' notion of cpuid.
550 	 * This array will be indexed using Linux cpuid.
551 	 */
552 	rdmsrl(HV_X64_MSR_VP_INDEX, vp_index);
553 	hv_context.vp_index[cpu] = (u32)vp_index;
554 
555 	INIT_LIST_HEAD(&hv_context.percpu_list[cpu]);
556 
557 	/*
558 	 * Register the per-cpu clockevent source.
559 	 */
560 	if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE)
561 		clockevents_config_and_register(hv_context.clk_evt[cpu],
562 						HV_TIMER_FREQUENCY,
563 						HV_MIN_DELTA_TICKS,
564 						HV_MAX_MAX_DELTA_TICKS);
565 	return;
566 }
567 
568 /*
569  * hv_synic_clockevents_cleanup - Cleanup clockevent devices
570  */
571 void hv_synic_clockevents_cleanup(void)
572 {
573 	int cpu;
574 
575 	if (!(ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE))
576 		return;
577 
578 	for_each_online_cpu(cpu)
579 		clockevents_unbind_device(hv_context.clk_evt[cpu], cpu);
580 }
581 
582 /*
583  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
584  */
585 void hv_synic_cleanup(void *arg)
586 {
587 	union hv_synic_sint shared_sint;
588 	union hv_synic_simp simp;
589 	union hv_synic_siefp siefp;
590 	union hv_synic_scontrol sctrl;
591 	int cpu = smp_processor_id();
592 
593 	if (!hv_context.synic_initialized)
594 		return;
595 
596 	/* Turn off clockevent device */
597 	if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE)
598 		hv_ce_shutdown(hv_context.clk_evt[cpu]);
599 
600 	rdmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
601 
602 	shared_sint.masked = 1;
603 
604 	/* Need to correctly cleanup in the case of SMP!!! */
605 	/* Disable the interrupt */
606 	wrmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
607 
608 	rdmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
609 	simp.simp_enabled = 0;
610 	simp.base_simp_gpa = 0;
611 
612 	wrmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
613 
614 	rdmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
615 	siefp.siefp_enabled = 0;
616 	siefp.base_siefp_gpa = 0;
617 
618 	wrmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
619 
620 	/* Disable the global synic bit */
621 	rdmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
622 	sctrl.enable = 0;
623 	wrmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
624 }
625