xref: /linux/drivers/hv/hv.c (revision f61389a9cd26b424485acade726ccfff96c749de)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2009, Microsoft Corporation.
4  *
5  * Authors:
6  *   Haiyang Zhang <haiyangz@microsoft.com>
7  *   Hank Janssen  <hjanssen@microsoft.com>
8  */
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/io.h>
12 #include <linux/kernel.h>
13 #include <linux/mm.h>
14 #include <linux/slab.h>
15 #include <linux/vmalloc.h>
16 #include <linux/hyperv.h>
17 #include <linux/random.h>
18 #include <linux/clockchips.h>
19 #include <linux/delay.h>
20 #include <linux/interrupt.h>
21 #include <clocksource/hyperv_timer.h>
22 #include <asm/mshyperv.h>
23 #include <linux/set_memory.h>
24 #include "hyperv_vmbus.h"
25 
26 /* The one and only */
27 struct hv_context hv_context;
28 
29 /*
30  * hv_init - Main initialization routine.
31  *
32  * This routine must be called before any other routines in here are called
33  */
hv_init(void)34 int hv_init(void)
35 {
36 	hv_context.cpu_context = alloc_percpu(struct hv_per_cpu_context);
37 	if (!hv_context.cpu_context)
38 		return -ENOMEM;
39 	return 0;
40 }
41 
42 /*
43  * hv_post_message - Post a message using the hypervisor message IPC.
44  *
45  * This involves a hypercall.
46  */
hv_post_message(union hv_connection_id connection_id,enum hv_message_type message_type,void * payload,size_t payload_size)47 int hv_post_message(union hv_connection_id connection_id,
48 			enum hv_message_type message_type,
49 			void *payload, size_t payload_size)
50 {
51 	struct hv_input_post_message *aligned_msg;
52 	unsigned long flags;
53 	u64 status;
54 
55 	if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
56 		return -EMSGSIZE;
57 
58 	local_irq_save(flags);
59 
60 	/*
61 	 * A TDX VM with the paravisor must use the decrypted post_msg_page: see
62 	 * the comment in struct hv_per_cpu_context. A SNP VM with the paravisor
63 	 * can use the encrypted hyperv_pcpu_input_arg because it copies the
64 	 * input into the GHCB page, which has been decrypted by the paravisor.
65 	 */
66 	if (hv_isolation_type_tdx() && ms_hyperv.paravisor_present)
67 		aligned_msg = this_cpu_ptr(hv_context.cpu_context)->post_msg_page;
68 	else
69 		aligned_msg = *this_cpu_ptr(hyperv_pcpu_input_arg);
70 
71 	aligned_msg->connectionid = connection_id;
72 	aligned_msg->reserved = 0;
73 	aligned_msg->message_type = message_type;
74 	aligned_msg->payload_size = payload_size;
75 	memcpy((void *)aligned_msg->payload, payload, payload_size);
76 
77 	if (ms_hyperv.paravisor_present) {
78 		if (hv_isolation_type_tdx())
79 			status = hv_tdx_hypercall(HVCALL_POST_MESSAGE,
80 						  virt_to_phys(aligned_msg), 0);
81 		else if (hv_isolation_type_snp())
82 			status = hv_ghcb_hypercall(HVCALL_POST_MESSAGE,
83 						   aligned_msg, NULL,
84 						   sizeof(*aligned_msg));
85 		else
86 			status = HV_STATUS_INVALID_PARAMETER;
87 	} else {
88 		u64 control = HVCALL_POST_MESSAGE;
89 
90 		control |= hv_nested ? HV_HYPERCALL_NESTED : 0;
91 		status = hv_do_hypercall(control, aligned_msg, NULL);
92 	}
93 
94 	local_irq_restore(flags);
95 
96 	return hv_result(status);
97 }
98 
hv_synic_alloc(void)99 int hv_synic_alloc(void)
100 {
101 	int cpu, ret = -ENOMEM;
102 	struct hv_per_cpu_context *hv_cpu;
103 
104 	/*
105 	 * First, zero all per-cpu memory areas so hv_synic_free() can
106 	 * detect what memory has been allocated and cleanup properly
107 	 * after any failures.
108 	 */
109 	for_each_present_cpu(cpu) {
110 		hv_cpu = per_cpu_ptr(hv_context.cpu_context, cpu);
111 		memset(hv_cpu, 0, sizeof(*hv_cpu));
112 	}
113 
114 	hv_context.hv_numa_map = kcalloc(nr_node_ids, sizeof(struct cpumask),
115 					 GFP_KERNEL);
116 	if (!hv_context.hv_numa_map) {
117 		pr_err("Unable to allocate NUMA map\n");
118 		goto err;
119 	}
120 
121 	for_each_present_cpu(cpu) {
122 		hv_cpu = per_cpu_ptr(hv_context.cpu_context, cpu);
123 
124 		tasklet_init(&hv_cpu->msg_dpc,
125 			     vmbus_on_msg_dpc, (unsigned long)hv_cpu);
126 
127 		if (ms_hyperv.paravisor_present && hv_isolation_type_tdx()) {
128 			hv_cpu->post_msg_page = (void *)get_zeroed_page(GFP_ATOMIC);
129 			if (!hv_cpu->post_msg_page) {
130 				pr_err("Unable to allocate post msg page\n");
131 				goto err;
132 			}
133 
134 			ret = set_memory_decrypted((unsigned long)hv_cpu->post_msg_page, 1);
135 			if (ret) {
136 				pr_err("Failed to decrypt post msg page: %d\n", ret);
137 				/* Just leak the page, as it's unsafe to free the page. */
138 				hv_cpu->post_msg_page = NULL;
139 				goto err;
140 			}
141 
142 			memset(hv_cpu->post_msg_page, 0, PAGE_SIZE);
143 		}
144 
145 		/*
146 		 * Synic message and event pages are allocated by paravisor.
147 		 * Skip these pages allocation here.
148 		 */
149 		if (!ms_hyperv.paravisor_present && !hv_root_partition()) {
150 			hv_cpu->synic_message_page =
151 				(void *)get_zeroed_page(GFP_ATOMIC);
152 			if (!hv_cpu->synic_message_page) {
153 				pr_err("Unable to allocate SYNIC message page\n");
154 				goto err;
155 			}
156 
157 			hv_cpu->synic_event_page =
158 				(void *)get_zeroed_page(GFP_ATOMIC);
159 			if (!hv_cpu->synic_event_page) {
160 				pr_err("Unable to allocate SYNIC event page\n");
161 
162 				free_page((unsigned long)hv_cpu->synic_message_page);
163 				hv_cpu->synic_message_page = NULL;
164 				goto err;
165 			}
166 		}
167 
168 		if (!ms_hyperv.paravisor_present &&
169 		    (hv_isolation_type_snp() || hv_isolation_type_tdx())) {
170 			ret = set_memory_decrypted((unsigned long)
171 				hv_cpu->synic_message_page, 1);
172 			if (ret) {
173 				pr_err("Failed to decrypt SYNIC msg page: %d\n", ret);
174 				hv_cpu->synic_message_page = NULL;
175 
176 				/*
177 				 * Free the event page here so that hv_synic_free()
178 				 * won't later try to re-encrypt it.
179 				 */
180 				free_page((unsigned long)hv_cpu->synic_event_page);
181 				hv_cpu->synic_event_page = NULL;
182 				goto err;
183 			}
184 
185 			ret = set_memory_decrypted((unsigned long)
186 				hv_cpu->synic_event_page, 1);
187 			if (ret) {
188 				pr_err("Failed to decrypt SYNIC event page: %d\n", ret);
189 				hv_cpu->synic_event_page = NULL;
190 				goto err;
191 			}
192 
193 			memset(hv_cpu->synic_message_page, 0, PAGE_SIZE);
194 			memset(hv_cpu->synic_event_page, 0, PAGE_SIZE);
195 		}
196 	}
197 
198 	return 0;
199 
200 err:
201 	/*
202 	 * Any memory allocations that succeeded will be freed when
203 	 * the caller cleans up by calling hv_synic_free()
204 	 */
205 	return ret;
206 }
207 
hv_synic_free(void)208 void hv_synic_free(void)
209 {
210 	int cpu, ret;
211 
212 	for_each_present_cpu(cpu) {
213 		struct hv_per_cpu_context *hv_cpu =
214 			per_cpu_ptr(hv_context.cpu_context, cpu);
215 
216 		/* It's better to leak the page if the encryption fails. */
217 		if (ms_hyperv.paravisor_present && hv_isolation_type_tdx()) {
218 			if (hv_cpu->post_msg_page) {
219 				ret = set_memory_encrypted((unsigned long)
220 					hv_cpu->post_msg_page, 1);
221 				if (ret) {
222 					pr_err("Failed to encrypt post msg page: %d\n", ret);
223 					hv_cpu->post_msg_page = NULL;
224 				}
225 			}
226 		}
227 
228 		if (!ms_hyperv.paravisor_present &&
229 		    (hv_isolation_type_snp() || hv_isolation_type_tdx())) {
230 			if (hv_cpu->synic_message_page) {
231 				ret = set_memory_encrypted((unsigned long)
232 					hv_cpu->synic_message_page, 1);
233 				if (ret) {
234 					pr_err("Failed to encrypt SYNIC msg page: %d\n", ret);
235 					hv_cpu->synic_message_page = NULL;
236 				}
237 			}
238 
239 			if (hv_cpu->synic_event_page) {
240 				ret = set_memory_encrypted((unsigned long)
241 					hv_cpu->synic_event_page, 1);
242 				if (ret) {
243 					pr_err("Failed to encrypt SYNIC event page: %d\n", ret);
244 					hv_cpu->synic_event_page = NULL;
245 				}
246 			}
247 		}
248 
249 		free_page((unsigned long)hv_cpu->post_msg_page);
250 		free_page((unsigned long)hv_cpu->synic_event_page);
251 		free_page((unsigned long)hv_cpu->synic_message_page);
252 	}
253 
254 	kfree(hv_context.hv_numa_map);
255 }
256 
257 /*
258  * hv_synic_init - Initialize the Synthetic Interrupt Controller.
259  *
260  * If it is already initialized by another entity (ie x2v shim), we need to
261  * retrieve the initialized message and event pages.  Otherwise, we create and
262  * initialize the message and event pages.
263  */
hv_synic_enable_regs(unsigned int cpu)264 void hv_synic_enable_regs(unsigned int cpu)
265 {
266 	struct hv_per_cpu_context *hv_cpu =
267 		per_cpu_ptr(hv_context.cpu_context, cpu);
268 	union hv_synic_simp simp;
269 	union hv_synic_siefp siefp;
270 	union hv_synic_sint shared_sint;
271 	union hv_synic_scontrol sctrl;
272 
273 	/* Setup the Synic's message page */
274 	simp.as_uint64 = hv_get_msr(HV_MSR_SIMP);
275 	simp.simp_enabled = 1;
276 
277 	if (ms_hyperv.paravisor_present || hv_root_partition()) {
278 		/* Mask out vTOM bit. ioremap_cache() maps decrypted */
279 		u64 base = (simp.base_simp_gpa << HV_HYP_PAGE_SHIFT) &
280 				~ms_hyperv.shared_gpa_boundary;
281 		hv_cpu->synic_message_page =
282 			(void *)ioremap_cache(base, HV_HYP_PAGE_SIZE);
283 		if (!hv_cpu->synic_message_page)
284 			pr_err("Fail to map synic message page.\n");
285 	} else {
286 		simp.base_simp_gpa = virt_to_phys(hv_cpu->synic_message_page)
287 			>> HV_HYP_PAGE_SHIFT;
288 	}
289 
290 	hv_set_msr(HV_MSR_SIMP, simp.as_uint64);
291 
292 	/* Setup the Synic's event page */
293 	siefp.as_uint64 = hv_get_msr(HV_MSR_SIEFP);
294 	siefp.siefp_enabled = 1;
295 
296 	if (ms_hyperv.paravisor_present || hv_root_partition()) {
297 		/* Mask out vTOM bit. ioremap_cache() maps decrypted */
298 		u64 base = (siefp.base_siefp_gpa << HV_HYP_PAGE_SHIFT) &
299 				~ms_hyperv.shared_gpa_boundary;
300 		hv_cpu->synic_event_page =
301 			(void *)ioremap_cache(base, HV_HYP_PAGE_SIZE);
302 		if (!hv_cpu->synic_event_page)
303 			pr_err("Fail to map synic event page.\n");
304 	} else {
305 		siefp.base_siefp_gpa = virt_to_phys(hv_cpu->synic_event_page)
306 			>> HV_HYP_PAGE_SHIFT;
307 	}
308 
309 	hv_set_msr(HV_MSR_SIEFP, siefp.as_uint64);
310 
311 	/* Setup the shared SINT. */
312 	if (vmbus_irq != -1)
313 		enable_percpu_irq(vmbus_irq, 0);
314 	shared_sint.as_uint64 = hv_get_msr(HV_MSR_SINT0 + VMBUS_MESSAGE_SINT);
315 
316 	shared_sint.vector = vmbus_interrupt;
317 	shared_sint.masked = false;
318 	shared_sint.auto_eoi = hv_recommend_using_aeoi();
319 	hv_set_msr(HV_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
320 
321 	/* Enable the global synic bit */
322 	sctrl.as_uint64 = hv_get_msr(HV_MSR_SCONTROL);
323 	sctrl.enable = 1;
324 
325 	hv_set_msr(HV_MSR_SCONTROL, sctrl.as_uint64);
326 }
327 
hv_synic_init(unsigned int cpu)328 int hv_synic_init(unsigned int cpu)
329 {
330 	hv_synic_enable_regs(cpu);
331 
332 	hv_stimer_legacy_init(cpu, VMBUS_MESSAGE_SINT);
333 
334 	return 0;
335 }
336 
hv_synic_disable_regs(unsigned int cpu)337 void hv_synic_disable_regs(unsigned int cpu)
338 {
339 	struct hv_per_cpu_context *hv_cpu =
340 		per_cpu_ptr(hv_context.cpu_context, cpu);
341 	union hv_synic_sint shared_sint;
342 	union hv_synic_simp simp;
343 	union hv_synic_siefp siefp;
344 	union hv_synic_scontrol sctrl;
345 
346 	shared_sint.as_uint64 = hv_get_msr(HV_MSR_SINT0 + VMBUS_MESSAGE_SINT);
347 
348 	shared_sint.masked = 1;
349 
350 	/* Need to correctly cleanup in the case of SMP!!! */
351 	/* Disable the interrupt */
352 	hv_set_msr(HV_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
353 
354 	simp.as_uint64 = hv_get_msr(HV_MSR_SIMP);
355 	/*
356 	 * In Isolation VM, sim and sief pages are allocated by
357 	 * paravisor. These pages also will be used by kdump
358 	 * kernel. So just reset enable bit here and keep page
359 	 * addresses.
360 	 */
361 	simp.simp_enabled = 0;
362 	if (ms_hyperv.paravisor_present || hv_root_partition()) {
363 		iounmap(hv_cpu->synic_message_page);
364 		hv_cpu->synic_message_page = NULL;
365 	} else {
366 		simp.base_simp_gpa = 0;
367 	}
368 
369 	hv_set_msr(HV_MSR_SIMP, simp.as_uint64);
370 
371 	siefp.as_uint64 = hv_get_msr(HV_MSR_SIEFP);
372 	siefp.siefp_enabled = 0;
373 
374 	if (ms_hyperv.paravisor_present || hv_root_partition()) {
375 		iounmap(hv_cpu->synic_event_page);
376 		hv_cpu->synic_event_page = NULL;
377 	} else {
378 		siefp.base_siefp_gpa = 0;
379 	}
380 
381 	hv_set_msr(HV_MSR_SIEFP, siefp.as_uint64);
382 
383 	/* Disable the global synic bit */
384 	sctrl.as_uint64 = hv_get_msr(HV_MSR_SCONTROL);
385 	sctrl.enable = 0;
386 	hv_set_msr(HV_MSR_SCONTROL, sctrl.as_uint64);
387 
388 	if (vmbus_irq != -1)
389 		disable_percpu_irq(vmbus_irq);
390 }
391 
392 #define HV_MAX_TRIES 3
393 /*
394  * Scan the event flags page of 'this' CPU looking for any bit that is set.  If we find one
395  * bit set, then wait for a few milliseconds.  Repeat these steps for a maximum of 3 times.
396  * Return 'true', if there is still any set bit after this operation; 'false', otherwise.
397  *
398  * If a bit is set, that means there is a pending channel interrupt.  The expectation is
399  * that the normal interrupt handling mechanism will find and process the channel interrupt
400  * "very soon", and in the process clear the bit.
401  */
hv_synic_event_pending(void)402 static bool hv_synic_event_pending(void)
403 {
404 	struct hv_per_cpu_context *hv_cpu = this_cpu_ptr(hv_context.cpu_context);
405 	union hv_synic_event_flags *event =
406 		(union hv_synic_event_flags *)hv_cpu->synic_event_page + VMBUS_MESSAGE_SINT;
407 	unsigned long *recv_int_page = event->flags; /* assumes VMBus version >= VERSION_WIN8 */
408 	bool pending;
409 	u32 relid;
410 	int tries = 0;
411 
412 retry:
413 	pending = false;
414 	for_each_set_bit(relid, recv_int_page, HV_EVENT_FLAGS_COUNT) {
415 		/* Special case - VMBus channel protocol messages */
416 		if (relid == 0)
417 			continue;
418 		pending = true;
419 		break;
420 	}
421 	if (pending && tries++ < HV_MAX_TRIES) {
422 		usleep_range(10000, 20000);
423 		goto retry;
424 	}
425 	return pending;
426 }
427 
hv_pick_new_cpu(struct vmbus_channel * channel)428 static int hv_pick_new_cpu(struct vmbus_channel *channel)
429 {
430 	int ret = -EBUSY;
431 	int start;
432 	int cpu;
433 
434 	lockdep_assert_cpus_held();
435 	lockdep_assert_held(&vmbus_connection.channel_mutex);
436 
437 	/*
438 	 * We can't assume that the relevant interrupts will be sent before
439 	 * the cpu is offlined on older versions of hyperv.
440 	 */
441 	if (vmbus_proto_version < VERSION_WIN10_V5_3)
442 		return -EBUSY;
443 
444 	start = get_random_u32_below(nr_cpu_ids);
445 
446 	for_each_cpu_wrap(cpu, cpu_online_mask, start) {
447 		if (channel->target_cpu == cpu ||
448 		    channel->target_cpu == VMBUS_CONNECT_CPU)
449 			continue;
450 
451 		ret = vmbus_channel_set_cpu(channel, cpu);
452 		if (!ret)
453 			break;
454 	}
455 
456 	if (ret)
457 		ret = vmbus_channel_set_cpu(channel, VMBUS_CONNECT_CPU);
458 
459 	return ret;
460 }
461 
462 /*
463  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
464  */
hv_synic_cleanup(unsigned int cpu)465 int hv_synic_cleanup(unsigned int cpu)
466 {
467 	struct vmbus_channel *channel, *sc;
468 	int ret = 0;
469 
470 	if (vmbus_connection.conn_state != CONNECTED)
471 		goto always_cleanup;
472 
473 	/*
474 	 * Hyper-V does not provide a way to change the connect CPU once
475 	 * it is set; we must prevent the connect CPU from going offline
476 	 * while the VM is running normally. But in the panic or kexec()
477 	 * path where the vmbus is already disconnected, the CPU must be
478 	 * allowed to shut down.
479 	 */
480 	if (cpu == VMBUS_CONNECT_CPU)
481 		return -EBUSY;
482 
483 	/*
484 	 * Search for channels which are bound to the CPU we're about to
485 	 * cleanup.
486 	 */
487 	mutex_lock(&vmbus_connection.channel_mutex);
488 	list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
489 		if (channel->target_cpu == cpu) {
490 			ret = hv_pick_new_cpu(channel);
491 			if (ret) {
492 				mutex_unlock(&vmbus_connection.channel_mutex);
493 				return ret;
494 			}
495 		}
496 		list_for_each_entry(sc, &channel->sc_list, sc_list) {
497 			if (sc->target_cpu == cpu) {
498 				ret = hv_pick_new_cpu(sc);
499 				if (ret) {
500 					mutex_unlock(&vmbus_connection.channel_mutex);
501 					return ret;
502 				}
503 			}
504 		}
505 	}
506 	mutex_unlock(&vmbus_connection.channel_mutex);
507 
508 	/*
509 	 * Scan the event flags page looking for bits that are set and waiting
510 	 * with a timeout for vmbus_chan_sched() to process such bits. If bits
511 	 * are still set after this operation and VMBus is connected, fail the
512 	 * CPU offlining operation.
513 	 */
514 	if (vmbus_proto_version >= VERSION_WIN10_V4_1 && hv_synic_event_pending())
515 		return -EBUSY;
516 
517 always_cleanup:
518 	hv_stimer_legacy_cleanup(cpu);
519 
520 	hv_synic_disable_regs(cpu);
521 
522 	return ret;
523 }
524