1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2009, Microsoft Corporation.
6 * Haiyang Zhang <haiyangz@microsoft.com>
7 * Hank Janssen <hjanssen@microsoft.com>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/kernel.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 "hyperv_vmbus.h"
25 /* The one and only */
26 struct hv_context hv_context;
29 * hv_init - Main initialization routine.
31 * This routine must be called before any other routines in here are called
35 hv_context.cpu_context = alloc_percpu(struct hv_per_cpu_context);
36 if (!hv_context.cpu_context)
42 * Functions for allocating and freeing memory with size and
43 * alignment HV_HYP_PAGE_SIZE. These functions are needed because
44 * the guest page size may not be the same as the Hyper-V page
45 * size. We depend upon kmalloc() aligning power-of-two size
46 * allocations to the allocation size boundary, so that the
47 * allocated memory appears to Hyper-V as a page of the size
51 void *hv_alloc_hyperv_page(void)
53 BUILD_BUG_ON(PAGE_SIZE < HV_HYP_PAGE_SIZE);
55 if (PAGE_SIZE == HV_HYP_PAGE_SIZE)
56 return (void *)__get_free_page(GFP_KERNEL);
58 return kmalloc(HV_HYP_PAGE_SIZE, GFP_KERNEL);
61 void *hv_alloc_hyperv_zeroed_page(void)
63 if (PAGE_SIZE == HV_HYP_PAGE_SIZE)
64 return (void *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
66 return kzalloc(HV_HYP_PAGE_SIZE, GFP_KERNEL);
69 void hv_free_hyperv_page(unsigned long addr)
71 if (PAGE_SIZE == HV_HYP_PAGE_SIZE)
78 * hv_post_message - Post a message using the hypervisor message IPC.
80 * This involves a hypercall.
82 int hv_post_message(union hv_connection_id connection_id,
83 enum hv_message_type message_type,
84 void *payload, size_t payload_size)
86 struct hv_input_post_message *aligned_msg;
87 struct hv_per_cpu_context *hv_cpu;
90 if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
93 hv_cpu = get_cpu_ptr(hv_context.cpu_context);
94 aligned_msg = hv_cpu->post_msg_page;
95 aligned_msg->connectionid = connection_id;
96 aligned_msg->reserved = 0;
97 aligned_msg->message_type = message_type;
98 aligned_msg->payload_size = payload_size;
99 memcpy((void *)aligned_msg->payload, payload, payload_size);
101 if (hv_isolation_type_snp())
102 status = hv_ghcb_hypercall(HVCALL_POST_MESSAGE,
103 (void *)aligned_msg, NULL,
104 sizeof(*aligned_msg));
106 status = hv_do_hypercall(HVCALL_POST_MESSAGE,
109 /* Preemption must remain disabled until after the hypercall
110 * so some other thread can't get scheduled onto this cpu and
111 * corrupt the per-cpu post_msg_page
115 return hv_result(status);
118 int hv_synic_alloc(void)
121 struct hv_per_cpu_context *hv_cpu;
124 * First, zero all per-cpu memory areas so hv_synic_free() can
125 * detect what memory has been allocated and cleanup properly
126 * after any failures.
128 for_each_present_cpu(cpu) {
129 hv_cpu = per_cpu_ptr(hv_context.cpu_context, cpu);
130 memset(hv_cpu, 0, sizeof(*hv_cpu));
133 hv_context.hv_numa_map = kcalloc(nr_node_ids, sizeof(struct cpumask),
135 if (hv_context.hv_numa_map == NULL) {
136 pr_err("Unable to allocate NUMA map\n");
140 for_each_present_cpu(cpu) {
141 hv_cpu = per_cpu_ptr(hv_context.cpu_context, cpu);
143 tasklet_init(&hv_cpu->msg_dpc,
144 vmbus_on_msg_dpc, (unsigned long) hv_cpu);
147 * Synic message and event pages are allocated by paravisor.
148 * Skip these pages allocation here.
150 if (!hv_isolation_type_snp()) {
151 hv_cpu->synic_message_page =
152 (void *)get_zeroed_page(GFP_ATOMIC);
153 if (hv_cpu->synic_message_page == NULL) {
154 pr_err("Unable to allocate SYNIC message page\n");
158 hv_cpu->synic_event_page =
159 (void *)get_zeroed_page(GFP_ATOMIC);
160 if (hv_cpu->synic_event_page == NULL) {
161 pr_err("Unable to allocate SYNIC event page\n");
166 hv_cpu->post_msg_page = (void *)get_zeroed_page(GFP_ATOMIC);
167 if (hv_cpu->post_msg_page == NULL) {
168 pr_err("Unable to allocate post msg page\n");
176 * Any memory allocations that succeeded will be freed when
177 * the caller cleans up by calling hv_synic_free()
183 void hv_synic_free(void)
187 for_each_present_cpu(cpu) {
188 struct hv_per_cpu_context *hv_cpu
189 = per_cpu_ptr(hv_context.cpu_context, cpu);
191 free_page((unsigned long)hv_cpu->synic_event_page);
192 free_page((unsigned long)hv_cpu->synic_message_page);
193 free_page((unsigned long)hv_cpu->post_msg_page);
196 kfree(hv_context.hv_numa_map);
200 * hv_synic_init - Initialize the Synthetic Interrupt Controller.
202 * If it is already initialized by another entity (ie x2v shim), we need to
203 * retrieve the initialized message and event pages. Otherwise, we create and
204 * initialize the message and event pages.
206 void hv_synic_enable_regs(unsigned int cpu)
208 struct hv_per_cpu_context *hv_cpu
209 = per_cpu_ptr(hv_context.cpu_context, cpu);
210 union hv_synic_simp simp;
211 union hv_synic_siefp siefp;
212 union hv_synic_sint shared_sint;
213 union hv_synic_scontrol sctrl;
215 /* Setup the Synic's message page */
216 simp.as_uint64 = hv_get_register(HV_REGISTER_SIMP);
217 simp.simp_enabled = 1;
219 if (hv_isolation_type_snp()) {
220 hv_cpu->synic_message_page
221 = memremap(simp.base_simp_gpa << HV_HYP_PAGE_SHIFT,
222 HV_HYP_PAGE_SIZE, MEMREMAP_WB);
223 if (!hv_cpu->synic_message_page)
224 pr_err("Fail to map syinc message page.\n");
226 simp.base_simp_gpa = virt_to_phys(hv_cpu->synic_message_page)
227 >> HV_HYP_PAGE_SHIFT;
230 hv_set_register(HV_REGISTER_SIMP, simp.as_uint64);
232 /* Setup the Synic's event page */
233 siefp.as_uint64 = hv_get_register(HV_REGISTER_SIEFP);
234 siefp.siefp_enabled = 1;
236 if (hv_isolation_type_snp()) {
237 hv_cpu->synic_event_page =
238 memremap(siefp.base_siefp_gpa << HV_HYP_PAGE_SHIFT,
239 HV_HYP_PAGE_SIZE, MEMREMAP_WB);
241 if (!hv_cpu->synic_event_page)
242 pr_err("Fail to map syinc event page.\n");
244 siefp.base_siefp_gpa = virt_to_phys(hv_cpu->synic_event_page)
245 >> HV_HYP_PAGE_SHIFT;
248 hv_set_register(HV_REGISTER_SIEFP, siefp.as_uint64);
250 /* Setup the shared SINT. */
252 enable_percpu_irq(vmbus_irq, 0);
253 shared_sint.as_uint64 = hv_get_register(HV_REGISTER_SINT0 +
256 shared_sint.vector = vmbus_interrupt;
257 shared_sint.masked = false;
260 * On architectures where Hyper-V doesn't support AEOI (e.g., ARM64),
261 * it doesn't provide a recommendation flag and AEOI must be disabled.
263 #ifdef HV_DEPRECATING_AEOI_RECOMMENDED
264 shared_sint.auto_eoi =
265 !(ms_hyperv.hints & HV_DEPRECATING_AEOI_RECOMMENDED);
267 shared_sint.auto_eoi = 0;
269 hv_set_register(HV_REGISTER_SINT0 + VMBUS_MESSAGE_SINT,
270 shared_sint.as_uint64);
272 /* Enable the global synic bit */
273 sctrl.as_uint64 = hv_get_register(HV_REGISTER_SCONTROL);
276 hv_set_register(HV_REGISTER_SCONTROL, sctrl.as_uint64);
279 int hv_synic_init(unsigned int cpu)
281 hv_synic_enable_regs(cpu);
283 hv_stimer_legacy_init(cpu, VMBUS_MESSAGE_SINT);
289 * hv_synic_cleanup - Cleanup routine for hv_synic_init().
291 void hv_synic_disable_regs(unsigned int cpu)
293 struct hv_per_cpu_context *hv_cpu
294 = per_cpu_ptr(hv_context.cpu_context, cpu);
295 union hv_synic_sint shared_sint;
296 union hv_synic_simp simp;
297 union hv_synic_siefp siefp;
298 union hv_synic_scontrol sctrl;
300 shared_sint.as_uint64 = hv_get_register(HV_REGISTER_SINT0 +
303 shared_sint.masked = 1;
305 /* Need to correctly cleanup in the case of SMP!!! */
306 /* Disable the interrupt */
307 hv_set_register(HV_REGISTER_SINT0 + VMBUS_MESSAGE_SINT,
308 shared_sint.as_uint64);
310 simp.as_uint64 = hv_get_register(HV_REGISTER_SIMP);
312 * In Isolation VM, sim and sief pages are allocated by
313 * paravisor. These pages also will be used by kdump
314 * kernel. So just reset enable bit here and keep page
317 simp.simp_enabled = 0;
318 if (hv_isolation_type_snp())
319 memunmap(hv_cpu->synic_message_page);
321 simp.base_simp_gpa = 0;
323 hv_set_register(HV_REGISTER_SIMP, simp.as_uint64);
325 siefp.as_uint64 = hv_get_register(HV_REGISTER_SIEFP);
326 siefp.siefp_enabled = 0;
328 if (hv_isolation_type_snp())
329 memunmap(hv_cpu->synic_event_page);
331 siefp.base_siefp_gpa = 0;
333 hv_set_register(HV_REGISTER_SIEFP, siefp.as_uint64);
335 /* Disable the global synic bit */
336 sctrl.as_uint64 = hv_get_register(HV_REGISTER_SCONTROL);
338 hv_set_register(HV_REGISTER_SCONTROL, sctrl.as_uint64);
341 disable_percpu_irq(vmbus_irq);
344 #define HV_MAX_TRIES 3
346 * Scan the event flags page of 'this' CPU looking for any bit that is set. If we find one
347 * bit set, then wait for a few milliseconds. Repeat these steps for a maximum of 3 times.
348 * Return 'true', if there is still any set bit after this operation; 'false', otherwise.
350 * If a bit is set, that means there is a pending channel interrupt. The expectation is
351 * that the normal interrupt handling mechanism will find and process the channel interrupt
352 * "very soon", and in the process clear the bit.
354 static bool hv_synic_event_pending(void)
356 struct hv_per_cpu_context *hv_cpu = this_cpu_ptr(hv_context.cpu_context);
357 union hv_synic_event_flags *event =
358 (union hv_synic_event_flags *)hv_cpu->synic_event_page + VMBUS_MESSAGE_SINT;
359 unsigned long *recv_int_page = event->flags; /* assumes VMBus version >= VERSION_WIN8 */
366 for_each_set_bit(relid, recv_int_page, HV_EVENT_FLAGS_COUNT) {
367 /* Special case - VMBus channel protocol messages */
373 if (pending && tries++ < HV_MAX_TRIES) {
374 usleep_range(10000, 20000);
380 int hv_synic_cleanup(unsigned int cpu)
382 struct vmbus_channel *channel, *sc;
383 bool channel_found = false;
385 if (vmbus_connection.conn_state != CONNECTED)
389 * Hyper-V does not provide a way to change the connect CPU once
390 * it is set; we must prevent the connect CPU from going offline
391 * while the VM is running normally. But in the panic or kexec()
392 * path where the vmbus is already disconnected, the CPU must be
393 * allowed to shut down.
395 if (cpu == VMBUS_CONNECT_CPU)
399 * Search for channels which are bound to the CPU we're about to
400 * cleanup. In case we find one and vmbus is still connected, we
401 * fail; this will effectively prevent CPU offlining.
403 * TODO: Re-bind the channels to different CPUs.
405 mutex_lock(&vmbus_connection.channel_mutex);
406 list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
407 if (channel->target_cpu == cpu) {
408 channel_found = true;
411 list_for_each_entry(sc, &channel->sc_list, sc_list) {
412 if (sc->target_cpu == cpu) {
413 channel_found = true;
420 mutex_unlock(&vmbus_connection.channel_mutex);
426 * channel_found == false means that any channels that were previously
427 * assigned to the CPU have been reassigned elsewhere with a call of
428 * vmbus_send_modifychannel(). Scan the event flags page looking for
429 * bits that are set and waiting with a timeout for vmbus_chan_sched()
430 * to process such bits. If bits are still set after this operation
431 * and VMBus is connected, fail the CPU offlining operation.
433 if (vmbus_proto_version >= VERSION_WIN10_V4_1 && hv_synic_event_pending())
437 hv_stimer_legacy_cleanup(cpu);
439 hv_synic_disable_regs(cpu);