Merge tag 'timers-core-2022-10-05' of git://git.kernel.org/pub/scm/linux/kernel/git...
[platform/kernel/linux-starfive.git] / drivers / hv / vmbus_drv.c
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  *   K. Y. Srinivasan <kys@microsoft.com>
9  */
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/init.h>
13 #include <linux/module.h>
14 #include <linux/device.h>
15 #include <linux/interrupt.h>
16 #include <linux/sysctl.h>
17 #include <linux/slab.h>
18 #include <linux/acpi.h>
19 #include <linux/completion.h>
20 #include <linux/hyperv.h>
21 #include <linux/kernel_stat.h>
22 #include <linux/clockchips.h>
23 #include <linux/cpu.h>
24 #include <linux/sched/isolation.h>
25 #include <linux/sched/task_stack.h>
26
27 #include <linux/delay.h>
28 #include <linux/notifier.h>
29 #include <linux/panic_notifier.h>
30 #include <linux/ptrace.h>
31 #include <linux/screen_info.h>
32 #include <linux/kdebug.h>
33 #include <linux/efi.h>
34 #include <linux/random.h>
35 #include <linux/kernel.h>
36 #include <linux/syscore_ops.h>
37 #include <linux/dma-map-ops.h>
38 #include <linux/pci.h>
39 #include <clocksource/hyperv_timer.h>
40 #include "hyperv_vmbus.h"
41
42 struct vmbus_dynid {
43         struct list_head node;
44         struct hv_vmbus_device_id id;
45 };
46
47 static struct acpi_device  *hv_acpi_dev;
48
49 static struct completion probe_event;
50
51 static int hyperv_cpuhp_online;
52
53 static void *hv_panic_page;
54
55 static long __percpu *vmbus_evt;
56
57 /* Values parsed from ACPI DSDT */
58 int vmbus_irq;
59 int vmbus_interrupt;
60
61 /*
62  * Boolean to control whether to report panic messages over Hyper-V.
63  *
64  * It can be set via /proc/sys/kernel/hyperv_record_panic_msg
65  */
66 static int sysctl_record_panic_msg = 1;
67
68 static int hyperv_report_reg(void)
69 {
70         return !sysctl_record_panic_msg || !hv_panic_page;
71 }
72
73 static int hyperv_panic_event(struct notifier_block *nb, unsigned long val,
74                               void *args)
75 {
76         struct pt_regs *regs;
77
78         vmbus_initiate_unload(true);
79
80         /*
81          * Hyper-V should be notified only once about a panic.  If we will be
82          * doing hv_kmsg_dump() with kmsg data later, don't do the notification
83          * here.
84          */
85         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE
86             && hyperv_report_reg()) {
87                 regs = current_pt_regs();
88                 hyperv_report_panic(regs, val, false);
89         }
90         return NOTIFY_DONE;
91 }
92
93 static int hyperv_die_event(struct notifier_block *nb, unsigned long val,
94                             void *args)
95 {
96         struct die_args *die = args;
97         struct pt_regs *regs = die->regs;
98
99         /* Don't notify Hyper-V if the die event is other than oops */
100         if (val != DIE_OOPS)
101                 return NOTIFY_DONE;
102
103         /*
104          * Hyper-V should be notified only once about a panic.  If we will be
105          * doing hv_kmsg_dump() with kmsg data later, don't do the notification
106          * here.
107          */
108         if (hyperv_report_reg())
109                 hyperv_report_panic(regs, val, true);
110         return NOTIFY_DONE;
111 }
112
113 static struct notifier_block hyperv_die_block = {
114         .notifier_call = hyperv_die_event,
115 };
116 static struct notifier_block hyperv_panic_block = {
117         .notifier_call = hyperv_panic_event,
118 };
119
120 static const char *fb_mmio_name = "fb_range";
121 static struct resource *fb_mmio;
122 static struct resource *hyperv_mmio;
123 static DEFINE_MUTEX(hyperv_mmio_lock);
124
125 static int vmbus_exists(void)
126 {
127         if (hv_acpi_dev == NULL)
128                 return -ENODEV;
129
130         return 0;
131 }
132
133 static u8 channel_monitor_group(const struct vmbus_channel *channel)
134 {
135         return (u8)channel->offermsg.monitorid / 32;
136 }
137
138 static u8 channel_monitor_offset(const struct vmbus_channel *channel)
139 {
140         return (u8)channel->offermsg.monitorid % 32;
141 }
142
143 static u32 channel_pending(const struct vmbus_channel *channel,
144                            const struct hv_monitor_page *monitor_page)
145 {
146         u8 monitor_group = channel_monitor_group(channel);
147
148         return monitor_page->trigger_group[monitor_group].pending;
149 }
150
151 static u32 channel_latency(const struct vmbus_channel *channel,
152                            const struct hv_monitor_page *monitor_page)
153 {
154         u8 monitor_group = channel_monitor_group(channel);
155         u8 monitor_offset = channel_monitor_offset(channel);
156
157         return monitor_page->latency[monitor_group][monitor_offset];
158 }
159
160 static u32 channel_conn_id(struct vmbus_channel *channel,
161                            struct hv_monitor_page *monitor_page)
162 {
163         u8 monitor_group = channel_monitor_group(channel);
164         u8 monitor_offset = channel_monitor_offset(channel);
165
166         return monitor_page->parameter[monitor_group][monitor_offset].connectionid.u.id;
167 }
168
169 static ssize_t id_show(struct device *dev, struct device_attribute *dev_attr,
170                        char *buf)
171 {
172         struct hv_device *hv_dev = device_to_hv_device(dev);
173
174         if (!hv_dev->channel)
175                 return -ENODEV;
176         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.child_relid);
177 }
178 static DEVICE_ATTR_RO(id);
179
180 static ssize_t state_show(struct device *dev, struct device_attribute *dev_attr,
181                           char *buf)
182 {
183         struct hv_device *hv_dev = device_to_hv_device(dev);
184
185         if (!hv_dev->channel)
186                 return -ENODEV;
187         return sprintf(buf, "%d\n", hv_dev->channel->state);
188 }
189 static DEVICE_ATTR_RO(state);
190
191 static ssize_t monitor_id_show(struct device *dev,
192                                struct device_attribute *dev_attr, char *buf)
193 {
194         struct hv_device *hv_dev = device_to_hv_device(dev);
195
196         if (!hv_dev->channel)
197                 return -ENODEV;
198         return sprintf(buf, "%d\n", hv_dev->channel->offermsg.monitorid);
199 }
200 static DEVICE_ATTR_RO(monitor_id);
201
202 static ssize_t class_id_show(struct device *dev,
203                                struct device_attribute *dev_attr, char *buf)
204 {
205         struct hv_device *hv_dev = device_to_hv_device(dev);
206
207         if (!hv_dev->channel)
208                 return -ENODEV;
209         return sprintf(buf, "{%pUl}\n",
210                        &hv_dev->channel->offermsg.offer.if_type);
211 }
212 static DEVICE_ATTR_RO(class_id);
213
214 static ssize_t device_id_show(struct device *dev,
215                               struct device_attribute *dev_attr, char *buf)
216 {
217         struct hv_device *hv_dev = device_to_hv_device(dev);
218
219         if (!hv_dev->channel)
220                 return -ENODEV;
221         return sprintf(buf, "{%pUl}\n",
222                        &hv_dev->channel->offermsg.offer.if_instance);
223 }
224 static DEVICE_ATTR_RO(device_id);
225
226 static ssize_t modalias_show(struct device *dev,
227                              struct device_attribute *dev_attr, char *buf)
228 {
229         struct hv_device *hv_dev = device_to_hv_device(dev);
230
231         return sprintf(buf, "vmbus:%*phN\n", UUID_SIZE, &hv_dev->dev_type);
232 }
233 static DEVICE_ATTR_RO(modalias);
234
235 #ifdef CONFIG_NUMA
236 static ssize_t numa_node_show(struct device *dev,
237                               struct device_attribute *attr, char *buf)
238 {
239         struct hv_device *hv_dev = device_to_hv_device(dev);
240
241         if (!hv_dev->channel)
242                 return -ENODEV;
243
244         return sprintf(buf, "%d\n", cpu_to_node(hv_dev->channel->target_cpu));
245 }
246 static DEVICE_ATTR_RO(numa_node);
247 #endif
248
249 static ssize_t server_monitor_pending_show(struct device *dev,
250                                            struct device_attribute *dev_attr,
251                                            char *buf)
252 {
253         struct hv_device *hv_dev = device_to_hv_device(dev);
254
255         if (!hv_dev->channel)
256                 return -ENODEV;
257         return sprintf(buf, "%d\n",
258                        channel_pending(hv_dev->channel,
259                                        vmbus_connection.monitor_pages[0]));
260 }
261 static DEVICE_ATTR_RO(server_monitor_pending);
262
263 static ssize_t client_monitor_pending_show(struct device *dev,
264                                            struct device_attribute *dev_attr,
265                                            char *buf)
266 {
267         struct hv_device *hv_dev = device_to_hv_device(dev);
268
269         if (!hv_dev->channel)
270                 return -ENODEV;
271         return sprintf(buf, "%d\n",
272                        channel_pending(hv_dev->channel,
273                                        vmbus_connection.monitor_pages[1]));
274 }
275 static DEVICE_ATTR_RO(client_monitor_pending);
276
277 static ssize_t server_monitor_latency_show(struct device *dev,
278                                            struct device_attribute *dev_attr,
279                                            char *buf)
280 {
281         struct hv_device *hv_dev = device_to_hv_device(dev);
282
283         if (!hv_dev->channel)
284                 return -ENODEV;
285         return sprintf(buf, "%d\n",
286                        channel_latency(hv_dev->channel,
287                                        vmbus_connection.monitor_pages[0]));
288 }
289 static DEVICE_ATTR_RO(server_monitor_latency);
290
291 static ssize_t client_monitor_latency_show(struct device *dev,
292                                            struct device_attribute *dev_attr,
293                                            char *buf)
294 {
295         struct hv_device *hv_dev = device_to_hv_device(dev);
296
297         if (!hv_dev->channel)
298                 return -ENODEV;
299         return sprintf(buf, "%d\n",
300                        channel_latency(hv_dev->channel,
301                                        vmbus_connection.monitor_pages[1]));
302 }
303 static DEVICE_ATTR_RO(client_monitor_latency);
304
305 static ssize_t server_monitor_conn_id_show(struct device *dev,
306                                            struct device_attribute *dev_attr,
307                                            char *buf)
308 {
309         struct hv_device *hv_dev = device_to_hv_device(dev);
310
311         if (!hv_dev->channel)
312                 return -ENODEV;
313         return sprintf(buf, "%d\n",
314                        channel_conn_id(hv_dev->channel,
315                                        vmbus_connection.monitor_pages[0]));
316 }
317 static DEVICE_ATTR_RO(server_monitor_conn_id);
318
319 static ssize_t client_monitor_conn_id_show(struct device *dev,
320                                            struct device_attribute *dev_attr,
321                                            char *buf)
322 {
323         struct hv_device *hv_dev = device_to_hv_device(dev);
324
325         if (!hv_dev->channel)
326                 return -ENODEV;
327         return sprintf(buf, "%d\n",
328                        channel_conn_id(hv_dev->channel,
329                                        vmbus_connection.monitor_pages[1]));
330 }
331 static DEVICE_ATTR_RO(client_monitor_conn_id);
332
333 static ssize_t out_intr_mask_show(struct device *dev,
334                                   struct device_attribute *dev_attr, char *buf)
335 {
336         struct hv_device *hv_dev = device_to_hv_device(dev);
337         struct hv_ring_buffer_debug_info outbound;
338         int ret;
339
340         if (!hv_dev->channel)
341                 return -ENODEV;
342
343         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
344                                           &outbound);
345         if (ret < 0)
346                 return ret;
347
348         return sprintf(buf, "%d\n", outbound.current_interrupt_mask);
349 }
350 static DEVICE_ATTR_RO(out_intr_mask);
351
352 static ssize_t out_read_index_show(struct device *dev,
353                                    struct device_attribute *dev_attr, char *buf)
354 {
355         struct hv_device *hv_dev = device_to_hv_device(dev);
356         struct hv_ring_buffer_debug_info outbound;
357         int ret;
358
359         if (!hv_dev->channel)
360                 return -ENODEV;
361
362         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
363                                           &outbound);
364         if (ret < 0)
365                 return ret;
366         return sprintf(buf, "%d\n", outbound.current_read_index);
367 }
368 static DEVICE_ATTR_RO(out_read_index);
369
370 static ssize_t out_write_index_show(struct device *dev,
371                                     struct device_attribute *dev_attr,
372                                     char *buf)
373 {
374         struct hv_device *hv_dev = device_to_hv_device(dev);
375         struct hv_ring_buffer_debug_info outbound;
376         int ret;
377
378         if (!hv_dev->channel)
379                 return -ENODEV;
380
381         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
382                                           &outbound);
383         if (ret < 0)
384                 return ret;
385         return sprintf(buf, "%d\n", outbound.current_write_index);
386 }
387 static DEVICE_ATTR_RO(out_write_index);
388
389 static ssize_t out_read_bytes_avail_show(struct device *dev,
390                                          struct device_attribute *dev_attr,
391                                          char *buf)
392 {
393         struct hv_device *hv_dev = device_to_hv_device(dev);
394         struct hv_ring_buffer_debug_info outbound;
395         int ret;
396
397         if (!hv_dev->channel)
398                 return -ENODEV;
399
400         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
401                                           &outbound);
402         if (ret < 0)
403                 return ret;
404         return sprintf(buf, "%d\n", outbound.bytes_avail_toread);
405 }
406 static DEVICE_ATTR_RO(out_read_bytes_avail);
407
408 static ssize_t out_write_bytes_avail_show(struct device *dev,
409                                           struct device_attribute *dev_attr,
410                                           char *buf)
411 {
412         struct hv_device *hv_dev = device_to_hv_device(dev);
413         struct hv_ring_buffer_debug_info outbound;
414         int ret;
415
416         if (!hv_dev->channel)
417                 return -ENODEV;
418
419         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->outbound,
420                                           &outbound);
421         if (ret < 0)
422                 return ret;
423         return sprintf(buf, "%d\n", outbound.bytes_avail_towrite);
424 }
425 static DEVICE_ATTR_RO(out_write_bytes_avail);
426
427 static ssize_t in_intr_mask_show(struct device *dev,
428                                  struct device_attribute *dev_attr, char *buf)
429 {
430         struct hv_device *hv_dev = device_to_hv_device(dev);
431         struct hv_ring_buffer_debug_info inbound;
432         int ret;
433
434         if (!hv_dev->channel)
435                 return -ENODEV;
436
437         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
438         if (ret < 0)
439                 return ret;
440
441         return sprintf(buf, "%d\n", inbound.current_interrupt_mask);
442 }
443 static DEVICE_ATTR_RO(in_intr_mask);
444
445 static ssize_t in_read_index_show(struct device *dev,
446                                   struct device_attribute *dev_attr, char *buf)
447 {
448         struct hv_device *hv_dev = device_to_hv_device(dev);
449         struct hv_ring_buffer_debug_info inbound;
450         int ret;
451
452         if (!hv_dev->channel)
453                 return -ENODEV;
454
455         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
456         if (ret < 0)
457                 return ret;
458
459         return sprintf(buf, "%d\n", inbound.current_read_index);
460 }
461 static DEVICE_ATTR_RO(in_read_index);
462
463 static ssize_t in_write_index_show(struct device *dev,
464                                    struct device_attribute *dev_attr, char *buf)
465 {
466         struct hv_device *hv_dev = device_to_hv_device(dev);
467         struct hv_ring_buffer_debug_info inbound;
468         int ret;
469
470         if (!hv_dev->channel)
471                 return -ENODEV;
472
473         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
474         if (ret < 0)
475                 return ret;
476
477         return sprintf(buf, "%d\n", inbound.current_write_index);
478 }
479 static DEVICE_ATTR_RO(in_write_index);
480
481 static ssize_t in_read_bytes_avail_show(struct device *dev,
482                                         struct device_attribute *dev_attr,
483                                         char *buf)
484 {
485         struct hv_device *hv_dev = device_to_hv_device(dev);
486         struct hv_ring_buffer_debug_info inbound;
487         int ret;
488
489         if (!hv_dev->channel)
490                 return -ENODEV;
491
492         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
493         if (ret < 0)
494                 return ret;
495
496         return sprintf(buf, "%d\n", inbound.bytes_avail_toread);
497 }
498 static DEVICE_ATTR_RO(in_read_bytes_avail);
499
500 static ssize_t in_write_bytes_avail_show(struct device *dev,
501                                          struct device_attribute *dev_attr,
502                                          char *buf)
503 {
504         struct hv_device *hv_dev = device_to_hv_device(dev);
505         struct hv_ring_buffer_debug_info inbound;
506         int ret;
507
508         if (!hv_dev->channel)
509                 return -ENODEV;
510
511         ret = hv_ringbuffer_get_debuginfo(&hv_dev->channel->inbound, &inbound);
512         if (ret < 0)
513                 return ret;
514
515         return sprintf(buf, "%d\n", inbound.bytes_avail_towrite);
516 }
517 static DEVICE_ATTR_RO(in_write_bytes_avail);
518
519 static ssize_t channel_vp_mapping_show(struct device *dev,
520                                        struct device_attribute *dev_attr,
521                                        char *buf)
522 {
523         struct hv_device *hv_dev = device_to_hv_device(dev);
524         struct vmbus_channel *channel = hv_dev->channel, *cur_sc;
525         int buf_size = PAGE_SIZE, n_written, tot_written;
526         struct list_head *cur;
527
528         if (!channel)
529                 return -ENODEV;
530
531         mutex_lock(&vmbus_connection.channel_mutex);
532
533         tot_written = snprintf(buf, buf_size, "%u:%u\n",
534                 channel->offermsg.child_relid, channel->target_cpu);
535
536         list_for_each(cur, &channel->sc_list) {
537                 if (tot_written >= buf_size - 1)
538                         break;
539
540                 cur_sc = list_entry(cur, struct vmbus_channel, sc_list);
541                 n_written = scnprintf(buf + tot_written,
542                                      buf_size - tot_written,
543                                      "%u:%u\n",
544                                      cur_sc->offermsg.child_relid,
545                                      cur_sc->target_cpu);
546                 tot_written += n_written;
547         }
548
549         mutex_unlock(&vmbus_connection.channel_mutex);
550
551         return tot_written;
552 }
553 static DEVICE_ATTR_RO(channel_vp_mapping);
554
555 static ssize_t vendor_show(struct device *dev,
556                            struct device_attribute *dev_attr,
557                            char *buf)
558 {
559         struct hv_device *hv_dev = device_to_hv_device(dev);
560
561         return sprintf(buf, "0x%x\n", hv_dev->vendor_id);
562 }
563 static DEVICE_ATTR_RO(vendor);
564
565 static ssize_t device_show(struct device *dev,
566                            struct device_attribute *dev_attr,
567                            char *buf)
568 {
569         struct hv_device *hv_dev = device_to_hv_device(dev);
570
571         return sprintf(buf, "0x%x\n", hv_dev->device_id);
572 }
573 static DEVICE_ATTR_RO(device);
574
575 static ssize_t driver_override_store(struct device *dev,
576                                      struct device_attribute *attr,
577                                      const char *buf, size_t count)
578 {
579         struct hv_device *hv_dev = device_to_hv_device(dev);
580         int ret;
581
582         ret = driver_set_override(dev, &hv_dev->driver_override, buf, count);
583         if (ret)
584                 return ret;
585
586         return count;
587 }
588
589 static ssize_t driver_override_show(struct device *dev,
590                                     struct device_attribute *attr, char *buf)
591 {
592         struct hv_device *hv_dev = device_to_hv_device(dev);
593         ssize_t len;
594
595         device_lock(dev);
596         len = snprintf(buf, PAGE_SIZE, "%s\n", hv_dev->driver_override);
597         device_unlock(dev);
598
599         return len;
600 }
601 static DEVICE_ATTR_RW(driver_override);
602
603 /* Set up per device attributes in /sys/bus/vmbus/devices/<bus device> */
604 static struct attribute *vmbus_dev_attrs[] = {
605         &dev_attr_id.attr,
606         &dev_attr_state.attr,
607         &dev_attr_monitor_id.attr,
608         &dev_attr_class_id.attr,
609         &dev_attr_device_id.attr,
610         &dev_attr_modalias.attr,
611 #ifdef CONFIG_NUMA
612         &dev_attr_numa_node.attr,
613 #endif
614         &dev_attr_server_monitor_pending.attr,
615         &dev_attr_client_monitor_pending.attr,
616         &dev_attr_server_monitor_latency.attr,
617         &dev_attr_client_monitor_latency.attr,
618         &dev_attr_server_monitor_conn_id.attr,
619         &dev_attr_client_monitor_conn_id.attr,
620         &dev_attr_out_intr_mask.attr,
621         &dev_attr_out_read_index.attr,
622         &dev_attr_out_write_index.attr,
623         &dev_attr_out_read_bytes_avail.attr,
624         &dev_attr_out_write_bytes_avail.attr,
625         &dev_attr_in_intr_mask.attr,
626         &dev_attr_in_read_index.attr,
627         &dev_attr_in_write_index.attr,
628         &dev_attr_in_read_bytes_avail.attr,
629         &dev_attr_in_write_bytes_avail.attr,
630         &dev_attr_channel_vp_mapping.attr,
631         &dev_attr_vendor.attr,
632         &dev_attr_device.attr,
633         &dev_attr_driver_override.attr,
634         NULL,
635 };
636
637 /*
638  * Device-level attribute_group callback function. Returns the permission for
639  * each attribute, and returns 0 if an attribute is not visible.
640  */
641 static umode_t vmbus_dev_attr_is_visible(struct kobject *kobj,
642                                          struct attribute *attr, int idx)
643 {
644         struct device *dev = kobj_to_dev(kobj);
645         const struct hv_device *hv_dev = device_to_hv_device(dev);
646
647         /* Hide the monitor attributes if the monitor mechanism is not used. */
648         if (!hv_dev->channel->offermsg.monitor_allocated &&
649             (attr == &dev_attr_monitor_id.attr ||
650              attr == &dev_attr_server_monitor_pending.attr ||
651              attr == &dev_attr_client_monitor_pending.attr ||
652              attr == &dev_attr_server_monitor_latency.attr ||
653              attr == &dev_attr_client_monitor_latency.attr ||
654              attr == &dev_attr_server_monitor_conn_id.attr ||
655              attr == &dev_attr_client_monitor_conn_id.attr))
656                 return 0;
657
658         return attr->mode;
659 }
660
661 static const struct attribute_group vmbus_dev_group = {
662         .attrs = vmbus_dev_attrs,
663         .is_visible = vmbus_dev_attr_is_visible
664 };
665 __ATTRIBUTE_GROUPS(vmbus_dev);
666
667 /* Set up the attribute for /sys/bus/vmbus/hibernation */
668 static ssize_t hibernation_show(struct bus_type *bus, char *buf)
669 {
670         return sprintf(buf, "%d\n", !!hv_is_hibernation_supported());
671 }
672
673 static BUS_ATTR_RO(hibernation);
674
675 static struct attribute *vmbus_bus_attrs[] = {
676         &bus_attr_hibernation.attr,
677         NULL,
678 };
679 static const struct attribute_group vmbus_bus_group = {
680         .attrs = vmbus_bus_attrs,
681 };
682 __ATTRIBUTE_GROUPS(vmbus_bus);
683
684 /*
685  * vmbus_uevent - add uevent for our device
686  *
687  * This routine is invoked when a device is added or removed on the vmbus to
688  * generate a uevent to udev in the userspace. The udev will then look at its
689  * rule and the uevent generated here to load the appropriate driver
690  *
691  * The alias string will be of the form vmbus:guid where guid is the string
692  * representation of the device guid (each byte of the guid will be
693  * represented with two hex characters.
694  */
695 static int vmbus_uevent(struct device *device, struct kobj_uevent_env *env)
696 {
697         struct hv_device *dev = device_to_hv_device(device);
698         const char *format = "MODALIAS=vmbus:%*phN";
699
700         return add_uevent_var(env, format, UUID_SIZE, &dev->dev_type);
701 }
702
703 static const struct hv_vmbus_device_id *
704 hv_vmbus_dev_match(const struct hv_vmbus_device_id *id, const guid_t *guid)
705 {
706         if (id == NULL)
707                 return NULL; /* empty device table */
708
709         for (; !guid_is_null(&id->guid); id++)
710                 if (guid_equal(&id->guid, guid))
711                         return id;
712
713         return NULL;
714 }
715
716 static const struct hv_vmbus_device_id *
717 hv_vmbus_dynid_match(struct hv_driver *drv, const guid_t *guid)
718 {
719         const struct hv_vmbus_device_id *id = NULL;
720         struct vmbus_dynid *dynid;
721
722         spin_lock(&drv->dynids.lock);
723         list_for_each_entry(dynid, &drv->dynids.list, node) {
724                 if (guid_equal(&dynid->id.guid, guid)) {
725                         id = &dynid->id;
726                         break;
727                 }
728         }
729         spin_unlock(&drv->dynids.lock);
730
731         return id;
732 }
733
734 static const struct hv_vmbus_device_id vmbus_device_null;
735
736 /*
737  * Return a matching hv_vmbus_device_id pointer.
738  * If there is no match, return NULL.
739  */
740 static const struct hv_vmbus_device_id *hv_vmbus_get_id(struct hv_driver *drv,
741                                                         struct hv_device *dev)
742 {
743         const guid_t *guid = &dev->dev_type;
744         const struct hv_vmbus_device_id *id;
745
746         /* When driver_override is set, only bind to the matching driver */
747         if (dev->driver_override && strcmp(dev->driver_override, drv->name))
748                 return NULL;
749
750         /* Look at the dynamic ids first, before the static ones */
751         id = hv_vmbus_dynid_match(drv, guid);
752         if (!id)
753                 id = hv_vmbus_dev_match(drv->id_table, guid);
754
755         /* driver_override will always match, send a dummy id */
756         if (!id && dev->driver_override)
757                 id = &vmbus_device_null;
758
759         return id;
760 }
761
762 /* vmbus_add_dynid - add a new device ID to this driver and re-probe devices */
763 static int vmbus_add_dynid(struct hv_driver *drv, guid_t *guid)
764 {
765         struct vmbus_dynid *dynid;
766
767         dynid = kzalloc(sizeof(*dynid), GFP_KERNEL);
768         if (!dynid)
769                 return -ENOMEM;
770
771         dynid->id.guid = *guid;
772
773         spin_lock(&drv->dynids.lock);
774         list_add_tail(&dynid->node, &drv->dynids.list);
775         spin_unlock(&drv->dynids.lock);
776
777         return driver_attach(&drv->driver);
778 }
779
780 static void vmbus_free_dynids(struct hv_driver *drv)
781 {
782         struct vmbus_dynid *dynid, *n;
783
784         spin_lock(&drv->dynids.lock);
785         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
786                 list_del(&dynid->node);
787                 kfree(dynid);
788         }
789         spin_unlock(&drv->dynids.lock);
790 }
791
792 /*
793  * store_new_id - sysfs frontend to vmbus_add_dynid()
794  *
795  * Allow GUIDs to be added to an existing driver via sysfs.
796  */
797 static ssize_t new_id_store(struct device_driver *driver, const char *buf,
798                             size_t count)
799 {
800         struct hv_driver *drv = drv_to_hv_drv(driver);
801         guid_t guid;
802         ssize_t retval;
803
804         retval = guid_parse(buf, &guid);
805         if (retval)
806                 return retval;
807
808         if (hv_vmbus_dynid_match(drv, &guid))
809                 return -EEXIST;
810
811         retval = vmbus_add_dynid(drv, &guid);
812         if (retval)
813                 return retval;
814         return count;
815 }
816 static DRIVER_ATTR_WO(new_id);
817
818 /*
819  * store_remove_id - remove a PCI device ID from this driver
820  *
821  * Removes a dynamic pci device ID to this driver.
822  */
823 static ssize_t remove_id_store(struct device_driver *driver, const char *buf,
824                                size_t count)
825 {
826         struct hv_driver *drv = drv_to_hv_drv(driver);
827         struct vmbus_dynid *dynid, *n;
828         guid_t guid;
829         ssize_t retval;
830
831         retval = guid_parse(buf, &guid);
832         if (retval)
833                 return retval;
834
835         retval = -ENODEV;
836         spin_lock(&drv->dynids.lock);
837         list_for_each_entry_safe(dynid, n, &drv->dynids.list, node) {
838                 struct hv_vmbus_device_id *id = &dynid->id;
839
840                 if (guid_equal(&id->guid, &guid)) {
841                         list_del(&dynid->node);
842                         kfree(dynid);
843                         retval = count;
844                         break;
845                 }
846         }
847         spin_unlock(&drv->dynids.lock);
848
849         return retval;
850 }
851 static DRIVER_ATTR_WO(remove_id);
852
853 static struct attribute *vmbus_drv_attrs[] = {
854         &driver_attr_new_id.attr,
855         &driver_attr_remove_id.attr,
856         NULL,
857 };
858 ATTRIBUTE_GROUPS(vmbus_drv);
859
860
861 /*
862  * vmbus_match - Attempt to match the specified device to the specified driver
863  */
864 static int vmbus_match(struct device *device, struct device_driver *driver)
865 {
866         struct hv_driver *drv = drv_to_hv_drv(driver);
867         struct hv_device *hv_dev = device_to_hv_device(device);
868
869         /* The hv_sock driver handles all hv_sock offers. */
870         if (is_hvsock_channel(hv_dev->channel))
871                 return drv->hvsock;
872
873         if (hv_vmbus_get_id(drv, hv_dev))
874                 return 1;
875
876         return 0;
877 }
878
879 /*
880  * vmbus_probe - Add the new vmbus's child device
881  */
882 static int vmbus_probe(struct device *child_device)
883 {
884         int ret = 0;
885         struct hv_driver *drv =
886                         drv_to_hv_drv(child_device->driver);
887         struct hv_device *dev = device_to_hv_device(child_device);
888         const struct hv_vmbus_device_id *dev_id;
889
890         dev_id = hv_vmbus_get_id(drv, dev);
891         if (drv->probe) {
892                 ret = drv->probe(dev, dev_id);
893                 if (ret != 0)
894                         pr_err("probe failed for device %s (%d)\n",
895                                dev_name(child_device), ret);
896
897         } else {
898                 pr_err("probe not set for driver %s\n",
899                        dev_name(child_device));
900                 ret = -ENODEV;
901         }
902         return ret;
903 }
904
905 /*
906  * vmbus_dma_configure -- Configure DMA coherence for VMbus device
907  */
908 static int vmbus_dma_configure(struct device *child_device)
909 {
910         /*
911          * On ARM64, propagate the DMA coherence setting from the top level
912          * VMbus ACPI device to the child VMbus device being added here.
913          * On x86/x64 coherence is assumed and these calls have no effect.
914          */
915         hv_setup_dma_ops(child_device,
916                 device_get_dma_attr(&hv_acpi_dev->dev) == DEV_DMA_COHERENT);
917         return 0;
918 }
919
920 /*
921  * vmbus_remove - Remove a vmbus device
922  */
923 static void vmbus_remove(struct device *child_device)
924 {
925         struct hv_driver *drv;
926         struct hv_device *dev = device_to_hv_device(child_device);
927
928         if (child_device->driver) {
929                 drv = drv_to_hv_drv(child_device->driver);
930                 if (drv->remove)
931                         drv->remove(dev);
932         }
933 }
934
935 /*
936  * vmbus_shutdown - Shutdown a vmbus device
937  */
938 static void vmbus_shutdown(struct device *child_device)
939 {
940         struct hv_driver *drv;
941         struct hv_device *dev = device_to_hv_device(child_device);
942
943
944         /* The device may not be attached yet */
945         if (!child_device->driver)
946                 return;
947
948         drv = drv_to_hv_drv(child_device->driver);
949
950         if (drv->shutdown)
951                 drv->shutdown(dev);
952 }
953
954 #ifdef CONFIG_PM_SLEEP
955 /*
956  * vmbus_suspend - Suspend a vmbus device
957  */
958 static int vmbus_suspend(struct device *child_device)
959 {
960         struct hv_driver *drv;
961         struct hv_device *dev = device_to_hv_device(child_device);
962
963         /* The device may not be attached yet */
964         if (!child_device->driver)
965                 return 0;
966
967         drv = drv_to_hv_drv(child_device->driver);
968         if (!drv->suspend)
969                 return -EOPNOTSUPP;
970
971         return drv->suspend(dev);
972 }
973
974 /*
975  * vmbus_resume - Resume a vmbus device
976  */
977 static int vmbus_resume(struct device *child_device)
978 {
979         struct hv_driver *drv;
980         struct hv_device *dev = device_to_hv_device(child_device);
981
982         /* The device may not be attached yet */
983         if (!child_device->driver)
984                 return 0;
985
986         drv = drv_to_hv_drv(child_device->driver);
987         if (!drv->resume)
988                 return -EOPNOTSUPP;
989
990         return drv->resume(dev);
991 }
992 #else
993 #define vmbus_suspend NULL
994 #define vmbus_resume NULL
995 #endif /* CONFIG_PM_SLEEP */
996
997 /*
998  * vmbus_device_release - Final callback release of the vmbus child device
999  */
1000 static void vmbus_device_release(struct device *device)
1001 {
1002         struct hv_device *hv_dev = device_to_hv_device(device);
1003         struct vmbus_channel *channel = hv_dev->channel;
1004
1005         hv_debug_rm_dev_dir(hv_dev);
1006
1007         mutex_lock(&vmbus_connection.channel_mutex);
1008         hv_process_channel_removal(channel);
1009         mutex_unlock(&vmbus_connection.channel_mutex);
1010         kfree(hv_dev);
1011 }
1012
1013 /*
1014  * Note: we must use the "noirq" ops: see the comment before vmbus_bus_pm.
1015  *
1016  * suspend_noirq/resume_noirq are set to NULL to support Suspend-to-Idle: we
1017  * shouldn't suspend the vmbus devices upon Suspend-to-Idle, otherwise there
1018  * is no way to wake up a Generation-2 VM.
1019  *
1020  * The other 4 ops are for hibernation.
1021  */
1022
1023 static const struct dev_pm_ops vmbus_pm = {
1024         .suspend_noirq  = NULL,
1025         .resume_noirq   = NULL,
1026         .freeze_noirq   = vmbus_suspend,
1027         .thaw_noirq     = vmbus_resume,
1028         .poweroff_noirq = vmbus_suspend,
1029         .restore_noirq  = vmbus_resume,
1030 };
1031
1032 /* The one and only one */
1033 static struct bus_type  hv_bus = {
1034         .name =         "vmbus",
1035         .match =                vmbus_match,
1036         .shutdown =             vmbus_shutdown,
1037         .remove =               vmbus_remove,
1038         .probe =                vmbus_probe,
1039         .uevent =               vmbus_uevent,
1040         .dma_configure =        vmbus_dma_configure,
1041         .dev_groups =           vmbus_dev_groups,
1042         .drv_groups =           vmbus_drv_groups,
1043         .bus_groups =           vmbus_bus_groups,
1044         .pm =                   &vmbus_pm,
1045 };
1046
1047 struct onmessage_work_context {
1048         struct work_struct work;
1049         struct {
1050                 struct hv_message_header header;
1051                 u8 payload[];
1052         } msg;
1053 };
1054
1055 static void vmbus_onmessage_work(struct work_struct *work)
1056 {
1057         struct onmessage_work_context *ctx;
1058
1059         /* Do not process messages if we're in DISCONNECTED state */
1060         if (vmbus_connection.conn_state == DISCONNECTED)
1061                 return;
1062
1063         ctx = container_of(work, struct onmessage_work_context,
1064                            work);
1065         vmbus_onmessage((struct vmbus_channel_message_header *)
1066                         &ctx->msg.payload);
1067         kfree(ctx);
1068 }
1069
1070 void vmbus_on_msg_dpc(unsigned long data)
1071 {
1072         struct hv_per_cpu_context *hv_cpu = (void *)data;
1073         void *page_addr = hv_cpu->synic_message_page;
1074         struct hv_message msg_copy, *msg = (struct hv_message *)page_addr +
1075                                   VMBUS_MESSAGE_SINT;
1076         struct vmbus_channel_message_header *hdr;
1077         enum vmbus_channel_message_type msgtype;
1078         const struct vmbus_channel_message_table_entry *entry;
1079         struct onmessage_work_context *ctx;
1080         __u8 payload_size;
1081         u32 message_type;
1082
1083         /*
1084          * 'enum vmbus_channel_message_type' is supposed to always be 'u32' as
1085          * it is being used in 'struct vmbus_channel_message_header' definition
1086          * which is supposed to match hypervisor ABI.
1087          */
1088         BUILD_BUG_ON(sizeof(enum vmbus_channel_message_type) != sizeof(u32));
1089
1090         /*
1091          * Since the message is in memory shared with the host, an erroneous or
1092          * malicious Hyper-V could modify the message while vmbus_on_msg_dpc()
1093          * or individual message handlers are executing; to prevent this, copy
1094          * the message into private memory.
1095          */
1096         memcpy(&msg_copy, msg, sizeof(struct hv_message));
1097
1098         message_type = msg_copy.header.message_type;
1099         if (message_type == HVMSG_NONE)
1100                 /* no msg */
1101                 return;
1102
1103         hdr = (struct vmbus_channel_message_header *)msg_copy.u.payload;
1104         msgtype = hdr->msgtype;
1105
1106         trace_vmbus_on_msg_dpc(hdr);
1107
1108         if (msgtype >= CHANNELMSG_COUNT) {
1109                 WARN_ONCE(1, "unknown msgtype=%d\n", msgtype);
1110                 goto msg_handled;
1111         }
1112
1113         payload_size = msg_copy.header.payload_size;
1114         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT) {
1115                 WARN_ONCE(1, "payload size is too large (%d)\n", payload_size);
1116                 goto msg_handled;
1117         }
1118
1119         entry = &channel_message_table[msgtype];
1120
1121         if (!entry->message_handler)
1122                 goto msg_handled;
1123
1124         if (payload_size < entry->min_payload_len) {
1125                 WARN_ONCE(1, "message too short: msgtype=%d len=%d\n", msgtype, payload_size);
1126                 goto msg_handled;
1127         }
1128
1129         if (entry->handler_type == VMHT_BLOCKING) {
1130                 ctx = kmalloc(struct_size(ctx, msg.payload, payload_size), GFP_ATOMIC);
1131                 if (ctx == NULL)
1132                         return;
1133
1134                 INIT_WORK(&ctx->work, vmbus_onmessage_work);
1135                 memcpy(&ctx->msg, &msg_copy, sizeof(msg->header) + payload_size);
1136
1137                 /*
1138                  * The host can generate a rescind message while we
1139                  * may still be handling the original offer. We deal with
1140                  * this condition by relying on the synchronization provided
1141                  * by offer_in_progress and by channel_mutex.  See also the
1142                  * inline comments in vmbus_onoffer_rescind().
1143                  */
1144                 switch (msgtype) {
1145                 case CHANNELMSG_RESCIND_CHANNELOFFER:
1146                         /*
1147                          * If we are handling the rescind message;
1148                          * schedule the work on the global work queue.
1149                          *
1150                          * The OFFER message and the RESCIND message should
1151                          * not be handled by the same serialized work queue,
1152                          * because the OFFER handler may call vmbus_open(),
1153                          * which tries to open the channel by sending an
1154                          * OPEN_CHANNEL message to the host and waits for
1155                          * the host's response; however, if the host has
1156                          * rescinded the channel before it receives the
1157                          * OPEN_CHANNEL message, the host just silently
1158                          * ignores the OPEN_CHANNEL message; as a result,
1159                          * the guest's OFFER handler hangs for ever, if we
1160                          * handle the RESCIND message in the same serialized
1161                          * work queue: the RESCIND handler can not start to
1162                          * run before the OFFER handler finishes.
1163                          */
1164                         if (vmbus_connection.ignore_any_offer_msg)
1165                                 break;
1166                         queue_work(vmbus_connection.rescind_work_queue, &ctx->work);
1167                         break;
1168
1169                 case CHANNELMSG_OFFERCHANNEL:
1170                         /*
1171                          * The host sends the offer message of a given channel
1172                          * before sending the rescind message of the same
1173                          * channel.  These messages are sent to the guest's
1174                          * connect CPU; the guest then starts processing them
1175                          * in the tasklet handler on this CPU:
1176                          *
1177                          * VMBUS_CONNECT_CPU
1178                          *
1179                          * [vmbus_on_msg_dpc()]
1180                          * atomic_inc()  // CHANNELMSG_OFFERCHANNEL
1181                          * queue_work()
1182                          * ...
1183                          * [vmbus_on_msg_dpc()]
1184                          * schedule_work()  // CHANNELMSG_RESCIND_CHANNELOFFER
1185                          *
1186                          * We rely on the memory-ordering properties of the
1187                          * queue_work() and schedule_work() primitives, which
1188                          * guarantee that the atomic increment will be visible
1189                          * to the CPUs which will execute the offer & rescind
1190                          * works by the time these works will start execution.
1191                          */
1192                         if (vmbus_connection.ignore_any_offer_msg)
1193                                 break;
1194                         atomic_inc(&vmbus_connection.offer_in_progress);
1195                         fallthrough;
1196
1197                 default:
1198                         queue_work(vmbus_connection.work_queue, &ctx->work);
1199                 }
1200         } else
1201                 entry->message_handler(hdr);
1202
1203 msg_handled:
1204         vmbus_signal_eom(msg, message_type);
1205 }
1206
1207 #ifdef CONFIG_PM_SLEEP
1208 /*
1209  * Fake RESCIND_CHANNEL messages to clean up hv_sock channels by force for
1210  * hibernation, because hv_sock connections can not persist across hibernation.
1211  */
1212 static void vmbus_force_channel_rescinded(struct vmbus_channel *channel)
1213 {
1214         struct onmessage_work_context *ctx;
1215         struct vmbus_channel_rescind_offer *rescind;
1216
1217         WARN_ON(!is_hvsock_channel(channel));
1218
1219         /*
1220          * Allocation size is small and the allocation should really not fail,
1221          * otherwise the state of the hv_sock connections ends up in limbo.
1222          */
1223         ctx = kzalloc(sizeof(*ctx) + sizeof(*rescind),
1224                       GFP_KERNEL | __GFP_NOFAIL);
1225
1226         /*
1227          * So far, these are not really used by Linux. Just set them to the
1228          * reasonable values conforming to the definitions of the fields.
1229          */
1230         ctx->msg.header.message_type = 1;
1231         ctx->msg.header.payload_size = sizeof(*rescind);
1232
1233         /* These values are actually used by Linux. */
1234         rescind = (struct vmbus_channel_rescind_offer *)ctx->msg.payload;
1235         rescind->header.msgtype = CHANNELMSG_RESCIND_CHANNELOFFER;
1236         rescind->child_relid = channel->offermsg.child_relid;
1237
1238         INIT_WORK(&ctx->work, vmbus_onmessage_work);
1239
1240         queue_work(vmbus_connection.work_queue, &ctx->work);
1241 }
1242 #endif /* CONFIG_PM_SLEEP */
1243
1244 /*
1245  * Schedule all channels with events pending
1246  */
1247 static void vmbus_chan_sched(struct hv_per_cpu_context *hv_cpu)
1248 {
1249         unsigned long *recv_int_page;
1250         u32 maxbits, relid;
1251
1252         /*
1253          * The event page can be directly checked to get the id of
1254          * the channel that has the interrupt pending.
1255          */
1256         void *page_addr = hv_cpu->synic_event_page;
1257         union hv_synic_event_flags *event
1258                 = (union hv_synic_event_flags *)page_addr +
1259                                          VMBUS_MESSAGE_SINT;
1260
1261         maxbits = HV_EVENT_FLAGS_COUNT;
1262         recv_int_page = event->flags;
1263
1264         if (unlikely(!recv_int_page))
1265                 return;
1266
1267         for_each_set_bit(relid, recv_int_page, maxbits) {
1268                 void (*callback_fn)(void *context);
1269                 struct vmbus_channel *channel;
1270
1271                 if (!sync_test_and_clear_bit(relid, recv_int_page))
1272                         continue;
1273
1274                 /* Special case - vmbus channel protocol msg */
1275                 if (relid == 0)
1276                         continue;
1277
1278                 /*
1279                  * Pairs with the kfree_rcu() in vmbus_chan_release().
1280                  * Guarantees that the channel data structure doesn't
1281                  * get freed while the channel pointer below is being
1282                  * dereferenced.
1283                  */
1284                 rcu_read_lock();
1285
1286                 /* Find channel based on relid */
1287                 channel = relid2channel(relid);
1288                 if (channel == NULL)
1289                         goto sched_unlock_rcu;
1290
1291                 if (channel->rescind)
1292                         goto sched_unlock_rcu;
1293
1294                 /*
1295                  * Make sure that the ring buffer data structure doesn't get
1296                  * freed while we dereference the ring buffer pointer.  Test
1297                  * for the channel's onchannel_callback being NULL within a
1298                  * sched_lock critical section.  See also the inline comments
1299                  * in vmbus_reset_channel_cb().
1300                  */
1301                 spin_lock(&channel->sched_lock);
1302
1303                 callback_fn = channel->onchannel_callback;
1304                 if (unlikely(callback_fn == NULL))
1305                         goto sched_unlock;
1306
1307                 trace_vmbus_chan_sched(channel);
1308
1309                 ++channel->interrupts;
1310
1311                 switch (channel->callback_mode) {
1312                 case HV_CALL_ISR:
1313                         (*callback_fn)(channel->channel_callback_context);
1314                         break;
1315
1316                 case HV_CALL_BATCHED:
1317                         hv_begin_read(&channel->inbound);
1318                         fallthrough;
1319                 case HV_CALL_DIRECT:
1320                         tasklet_schedule(&channel->callback_event);
1321                 }
1322
1323 sched_unlock:
1324                 spin_unlock(&channel->sched_lock);
1325 sched_unlock_rcu:
1326                 rcu_read_unlock();
1327         }
1328 }
1329
1330 static void vmbus_isr(void)
1331 {
1332         struct hv_per_cpu_context *hv_cpu
1333                 = this_cpu_ptr(hv_context.cpu_context);
1334         void *page_addr;
1335         struct hv_message *msg;
1336
1337         vmbus_chan_sched(hv_cpu);
1338
1339         page_addr = hv_cpu->synic_message_page;
1340         msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
1341
1342         /* Check if there are actual msgs to be processed */
1343         if (msg->header.message_type != HVMSG_NONE) {
1344                 if (msg->header.message_type == HVMSG_TIMER_EXPIRED) {
1345                         hv_stimer0_isr();
1346                         vmbus_signal_eom(msg, HVMSG_TIMER_EXPIRED);
1347                 } else
1348                         tasklet_schedule(&hv_cpu->msg_dpc);
1349         }
1350
1351         add_interrupt_randomness(vmbus_interrupt);
1352 }
1353
1354 static irqreturn_t vmbus_percpu_isr(int irq, void *dev_id)
1355 {
1356         vmbus_isr();
1357         return IRQ_HANDLED;
1358 }
1359
1360 /*
1361  * Callback from kmsg_dump. Grab as much as possible from the end of the kmsg
1362  * buffer and call into Hyper-V to transfer the data.
1363  */
1364 static void hv_kmsg_dump(struct kmsg_dumper *dumper,
1365                          enum kmsg_dump_reason reason)
1366 {
1367         struct kmsg_dump_iter iter;
1368         size_t bytes_written;
1369
1370         /* We are only interested in panics. */
1371         if ((reason != KMSG_DUMP_PANIC) || (!sysctl_record_panic_msg))
1372                 return;
1373
1374         /*
1375          * Write dump contents to the page. No need to synchronize; panic should
1376          * be single-threaded.
1377          */
1378         kmsg_dump_rewind(&iter);
1379         kmsg_dump_get_buffer(&iter, false, hv_panic_page, HV_HYP_PAGE_SIZE,
1380                              &bytes_written);
1381         if (!bytes_written)
1382                 return;
1383         /*
1384          * P3 to contain the physical address of the panic page & P4 to
1385          * contain the size of the panic data in that page. Rest of the
1386          * registers are no-op when the NOTIFY_MSG flag is set.
1387          */
1388         hv_set_register(HV_REGISTER_CRASH_P0, 0);
1389         hv_set_register(HV_REGISTER_CRASH_P1, 0);
1390         hv_set_register(HV_REGISTER_CRASH_P2, 0);
1391         hv_set_register(HV_REGISTER_CRASH_P3, virt_to_phys(hv_panic_page));
1392         hv_set_register(HV_REGISTER_CRASH_P4, bytes_written);
1393
1394         /*
1395          * Let Hyper-V know there is crash data available along with
1396          * the panic message.
1397          */
1398         hv_set_register(HV_REGISTER_CRASH_CTL,
1399                (HV_CRASH_CTL_CRASH_NOTIFY | HV_CRASH_CTL_CRASH_NOTIFY_MSG));
1400 }
1401
1402 static struct kmsg_dumper hv_kmsg_dumper = {
1403         .dump = hv_kmsg_dump,
1404 };
1405
1406 static void hv_kmsg_dump_register(void)
1407 {
1408         int ret;
1409
1410         hv_panic_page = hv_alloc_hyperv_zeroed_page();
1411         if (!hv_panic_page) {
1412                 pr_err("Hyper-V: panic message page memory allocation failed\n");
1413                 return;
1414         }
1415
1416         ret = kmsg_dump_register(&hv_kmsg_dumper);
1417         if (ret) {
1418                 pr_err("Hyper-V: kmsg dump register error 0x%x\n", ret);
1419                 hv_free_hyperv_page((unsigned long)hv_panic_page);
1420                 hv_panic_page = NULL;
1421         }
1422 }
1423
1424 static struct ctl_table_header *hv_ctl_table_hdr;
1425
1426 /*
1427  * sysctl option to allow the user to control whether kmsg data should be
1428  * reported to Hyper-V on panic.
1429  */
1430 static struct ctl_table hv_ctl_table[] = {
1431         {
1432                 .procname       = "hyperv_record_panic_msg",
1433                 .data           = &sysctl_record_panic_msg,
1434                 .maxlen         = sizeof(int),
1435                 .mode           = 0644,
1436                 .proc_handler   = proc_dointvec_minmax,
1437                 .extra1         = SYSCTL_ZERO,
1438                 .extra2         = SYSCTL_ONE
1439         },
1440         {}
1441 };
1442
1443 static struct ctl_table hv_root_table[] = {
1444         {
1445                 .procname       = "kernel",
1446                 .mode           = 0555,
1447                 .child          = hv_ctl_table
1448         },
1449         {}
1450 };
1451
1452 /*
1453  * vmbus_bus_init -Main vmbus driver initialization routine.
1454  *
1455  * Here, we
1456  *      - initialize the vmbus driver context
1457  *      - invoke the vmbus hv main init routine
1458  *      - retrieve the channel offers
1459  */
1460 static int vmbus_bus_init(void)
1461 {
1462         int ret;
1463
1464         ret = hv_init();
1465         if (ret != 0) {
1466                 pr_err("Unable to initialize the hypervisor - 0x%x\n", ret);
1467                 return ret;
1468         }
1469
1470         ret = bus_register(&hv_bus);
1471         if (ret)
1472                 return ret;
1473
1474         /*
1475          * VMbus interrupts are best modeled as per-cpu interrupts. If
1476          * on an architecture with support for per-cpu IRQs (e.g. ARM64),
1477          * allocate a per-cpu IRQ using standard Linux kernel functionality.
1478          * If not on such an architecture (e.g., x86/x64), then rely on
1479          * code in the arch-specific portion of the code tree to connect
1480          * the VMbus interrupt handler.
1481          */
1482
1483         if (vmbus_irq == -1) {
1484                 hv_setup_vmbus_handler(vmbus_isr);
1485         } else {
1486                 vmbus_evt = alloc_percpu(long);
1487                 ret = request_percpu_irq(vmbus_irq, vmbus_percpu_isr,
1488                                 "Hyper-V VMbus", vmbus_evt);
1489                 if (ret) {
1490                         pr_err("Can't request Hyper-V VMbus IRQ %d, Err %d",
1491                                         vmbus_irq, ret);
1492                         free_percpu(vmbus_evt);
1493                         goto err_setup;
1494                 }
1495         }
1496
1497         ret = hv_synic_alloc();
1498         if (ret)
1499                 goto err_alloc;
1500
1501         /*
1502          * Initialize the per-cpu interrupt state and stimer state.
1503          * Then connect to the host.
1504          */
1505         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "hyperv/vmbus:online",
1506                                 hv_synic_init, hv_synic_cleanup);
1507         if (ret < 0)
1508                 goto err_cpuhp;
1509         hyperv_cpuhp_online = ret;
1510
1511         ret = vmbus_connect();
1512         if (ret)
1513                 goto err_connect;
1514
1515         if (hv_is_isolation_supported())
1516                 sysctl_record_panic_msg = 0;
1517
1518         /*
1519          * Only register if the crash MSRs are available
1520          */
1521         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
1522                 u64 hyperv_crash_ctl;
1523                 /*
1524                  * Panic message recording (sysctl_record_panic_msg)
1525                  * is enabled by default in non-isolated guests and
1526                  * disabled by default in isolated guests; the panic
1527                  * message recording won't be available in isolated
1528                  * guests should the following registration fail.
1529                  */
1530                 hv_ctl_table_hdr = register_sysctl_table(hv_root_table);
1531                 if (!hv_ctl_table_hdr)
1532                         pr_err("Hyper-V: sysctl table register error");
1533
1534                 /*
1535                  * Register for panic kmsg callback only if the right
1536                  * capability is supported by the hypervisor.
1537                  */
1538                 hyperv_crash_ctl = hv_get_register(HV_REGISTER_CRASH_CTL);
1539                 if (hyperv_crash_ctl & HV_CRASH_CTL_CRASH_NOTIFY_MSG)
1540                         hv_kmsg_dump_register();
1541
1542                 register_die_notifier(&hyperv_die_block);
1543         }
1544
1545         /*
1546          * Always register the panic notifier because we need to unload
1547          * the VMbus channel connection to prevent any VMbus
1548          * activity after the VM panics.
1549          */
1550         atomic_notifier_chain_register(&panic_notifier_list,
1551                                &hyperv_panic_block);
1552
1553         vmbus_request_offers();
1554
1555         return 0;
1556
1557 err_connect:
1558         cpuhp_remove_state(hyperv_cpuhp_online);
1559 err_cpuhp:
1560         hv_synic_free();
1561 err_alloc:
1562         if (vmbus_irq == -1) {
1563                 hv_remove_vmbus_handler();
1564         } else {
1565                 free_percpu_irq(vmbus_irq, vmbus_evt);
1566                 free_percpu(vmbus_evt);
1567         }
1568 err_setup:
1569         bus_unregister(&hv_bus);
1570         unregister_sysctl_table(hv_ctl_table_hdr);
1571         hv_ctl_table_hdr = NULL;
1572         return ret;
1573 }
1574
1575 /**
1576  * __vmbus_child_driver_register() - Register a vmbus's driver
1577  * @hv_driver: Pointer to driver structure you want to register
1578  * @owner: owner module of the drv
1579  * @mod_name: module name string
1580  *
1581  * Registers the given driver with Linux through the 'driver_register()' call
1582  * and sets up the hyper-v vmbus handling for this driver.
1583  * It will return the state of the 'driver_register()' call.
1584  *
1585  */
1586 int __vmbus_driver_register(struct hv_driver *hv_driver, struct module *owner, const char *mod_name)
1587 {
1588         int ret;
1589
1590         pr_info("registering driver %s\n", hv_driver->name);
1591
1592         ret = vmbus_exists();
1593         if (ret < 0)
1594                 return ret;
1595
1596         hv_driver->driver.name = hv_driver->name;
1597         hv_driver->driver.owner = owner;
1598         hv_driver->driver.mod_name = mod_name;
1599         hv_driver->driver.bus = &hv_bus;
1600
1601         spin_lock_init(&hv_driver->dynids.lock);
1602         INIT_LIST_HEAD(&hv_driver->dynids.list);
1603
1604         ret = driver_register(&hv_driver->driver);
1605
1606         return ret;
1607 }
1608 EXPORT_SYMBOL_GPL(__vmbus_driver_register);
1609
1610 /**
1611  * vmbus_driver_unregister() - Unregister a vmbus's driver
1612  * @hv_driver: Pointer to driver structure you want to
1613  *             un-register
1614  *
1615  * Un-register the given driver that was previous registered with a call to
1616  * vmbus_driver_register()
1617  */
1618 void vmbus_driver_unregister(struct hv_driver *hv_driver)
1619 {
1620         pr_info("unregistering driver %s\n", hv_driver->name);
1621
1622         if (!vmbus_exists()) {
1623                 driver_unregister(&hv_driver->driver);
1624                 vmbus_free_dynids(hv_driver);
1625         }
1626 }
1627 EXPORT_SYMBOL_GPL(vmbus_driver_unregister);
1628
1629
1630 /*
1631  * Called when last reference to channel is gone.
1632  */
1633 static void vmbus_chan_release(struct kobject *kobj)
1634 {
1635         struct vmbus_channel *channel
1636                 = container_of(kobj, struct vmbus_channel, kobj);
1637
1638         kfree_rcu(channel, rcu);
1639 }
1640
1641 struct vmbus_chan_attribute {
1642         struct attribute attr;
1643         ssize_t (*show)(struct vmbus_channel *chan, char *buf);
1644         ssize_t (*store)(struct vmbus_channel *chan,
1645                          const char *buf, size_t count);
1646 };
1647 #define VMBUS_CHAN_ATTR(_name, _mode, _show, _store) \
1648         struct vmbus_chan_attribute chan_attr_##_name \
1649                 = __ATTR(_name, _mode, _show, _store)
1650 #define VMBUS_CHAN_ATTR_RW(_name) \
1651         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RW(_name)
1652 #define VMBUS_CHAN_ATTR_RO(_name) \
1653         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_RO(_name)
1654 #define VMBUS_CHAN_ATTR_WO(_name) \
1655         struct vmbus_chan_attribute chan_attr_##_name = __ATTR_WO(_name)
1656
1657 static ssize_t vmbus_chan_attr_show(struct kobject *kobj,
1658                                     struct attribute *attr, char *buf)
1659 {
1660         const struct vmbus_chan_attribute *attribute
1661                 = container_of(attr, struct vmbus_chan_attribute, attr);
1662         struct vmbus_channel *chan
1663                 = container_of(kobj, struct vmbus_channel, kobj);
1664
1665         if (!attribute->show)
1666                 return -EIO;
1667
1668         return attribute->show(chan, buf);
1669 }
1670
1671 static ssize_t vmbus_chan_attr_store(struct kobject *kobj,
1672                                      struct attribute *attr, const char *buf,
1673                                      size_t count)
1674 {
1675         const struct vmbus_chan_attribute *attribute
1676                 = container_of(attr, struct vmbus_chan_attribute, attr);
1677         struct vmbus_channel *chan
1678                 = container_of(kobj, struct vmbus_channel, kobj);
1679
1680         if (!attribute->store)
1681                 return -EIO;
1682
1683         return attribute->store(chan, buf, count);
1684 }
1685
1686 static const struct sysfs_ops vmbus_chan_sysfs_ops = {
1687         .show = vmbus_chan_attr_show,
1688         .store = vmbus_chan_attr_store,
1689 };
1690
1691 static ssize_t out_mask_show(struct vmbus_channel *channel, char *buf)
1692 {
1693         struct hv_ring_buffer_info *rbi = &channel->outbound;
1694         ssize_t ret;
1695
1696         mutex_lock(&rbi->ring_buffer_mutex);
1697         if (!rbi->ring_buffer) {
1698                 mutex_unlock(&rbi->ring_buffer_mutex);
1699                 return -EINVAL;
1700         }
1701
1702         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1703         mutex_unlock(&rbi->ring_buffer_mutex);
1704         return ret;
1705 }
1706 static VMBUS_CHAN_ATTR_RO(out_mask);
1707
1708 static ssize_t in_mask_show(struct vmbus_channel *channel, char *buf)
1709 {
1710         struct hv_ring_buffer_info *rbi = &channel->inbound;
1711         ssize_t ret;
1712
1713         mutex_lock(&rbi->ring_buffer_mutex);
1714         if (!rbi->ring_buffer) {
1715                 mutex_unlock(&rbi->ring_buffer_mutex);
1716                 return -EINVAL;
1717         }
1718
1719         ret = sprintf(buf, "%u\n", rbi->ring_buffer->interrupt_mask);
1720         mutex_unlock(&rbi->ring_buffer_mutex);
1721         return ret;
1722 }
1723 static VMBUS_CHAN_ATTR_RO(in_mask);
1724
1725 static ssize_t read_avail_show(struct vmbus_channel *channel, char *buf)
1726 {
1727         struct hv_ring_buffer_info *rbi = &channel->inbound;
1728         ssize_t ret;
1729
1730         mutex_lock(&rbi->ring_buffer_mutex);
1731         if (!rbi->ring_buffer) {
1732                 mutex_unlock(&rbi->ring_buffer_mutex);
1733                 return -EINVAL;
1734         }
1735
1736         ret = sprintf(buf, "%u\n", hv_get_bytes_to_read(rbi));
1737         mutex_unlock(&rbi->ring_buffer_mutex);
1738         return ret;
1739 }
1740 static VMBUS_CHAN_ATTR_RO(read_avail);
1741
1742 static ssize_t write_avail_show(struct vmbus_channel *channel, char *buf)
1743 {
1744         struct hv_ring_buffer_info *rbi = &channel->outbound;
1745         ssize_t ret;
1746
1747         mutex_lock(&rbi->ring_buffer_mutex);
1748         if (!rbi->ring_buffer) {
1749                 mutex_unlock(&rbi->ring_buffer_mutex);
1750                 return -EINVAL;
1751         }
1752
1753         ret = sprintf(buf, "%u\n", hv_get_bytes_to_write(rbi));
1754         mutex_unlock(&rbi->ring_buffer_mutex);
1755         return ret;
1756 }
1757 static VMBUS_CHAN_ATTR_RO(write_avail);
1758
1759 static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
1760 {
1761         return sprintf(buf, "%u\n", channel->target_cpu);
1762 }
1763 static ssize_t target_cpu_store(struct vmbus_channel *channel,
1764                                 const char *buf, size_t count)
1765 {
1766         u32 target_cpu, origin_cpu;
1767         ssize_t ret = count;
1768
1769         if (vmbus_proto_version < VERSION_WIN10_V4_1)
1770                 return -EIO;
1771
1772         if (sscanf(buf, "%uu", &target_cpu) != 1)
1773                 return -EIO;
1774
1775         /* Validate target_cpu for the cpumask_test_cpu() operation below. */
1776         if (target_cpu >= nr_cpumask_bits)
1777                 return -EINVAL;
1778
1779         if (!cpumask_test_cpu(target_cpu, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
1780                 return -EINVAL;
1781
1782         /* No CPUs should come up or down during this. */
1783         cpus_read_lock();
1784
1785         if (!cpu_online(target_cpu)) {
1786                 cpus_read_unlock();
1787                 return -EINVAL;
1788         }
1789
1790         /*
1791          * Synchronizes target_cpu_store() and channel closure:
1792          *
1793          * { Initially: state = CHANNEL_OPENED }
1794          *
1795          * CPU1                         CPU2
1796          *
1797          * [target_cpu_store()]         [vmbus_disconnect_ring()]
1798          *
1799          * LOCK channel_mutex           LOCK channel_mutex
1800          * LOAD r1 = state              LOAD r2 = state
1801          * IF (r1 == CHANNEL_OPENED)    IF (r2 == CHANNEL_OPENED)
1802          *   SEND MODIFYCHANNEL           STORE state = CHANNEL_OPEN
1803          *   [...]                        SEND CLOSECHANNEL
1804          * UNLOCK channel_mutex         UNLOCK channel_mutex
1805          *
1806          * Forbids: r1 == r2 == CHANNEL_OPENED (i.e., CPU1's LOCK precedes
1807          *              CPU2's LOCK) && CPU2's SEND precedes CPU1's SEND
1808          *
1809          * Note.  The host processes the channel messages "sequentially", in
1810          * the order in which they are received on a per-partition basis.
1811          */
1812         mutex_lock(&vmbus_connection.channel_mutex);
1813
1814         /*
1815          * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
1816          * avoid sending the message and fail here for such channels.
1817          */
1818         if (channel->state != CHANNEL_OPENED_STATE) {
1819                 ret = -EIO;
1820                 goto cpu_store_unlock;
1821         }
1822
1823         origin_cpu = channel->target_cpu;
1824         if (target_cpu == origin_cpu)
1825                 goto cpu_store_unlock;
1826
1827         if (vmbus_send_modifychannel(channel,
1828                                      hv_cpu_number_to_vp_number(target_cpu))) {
1829                 ret = -EIO;
1830                 goto cpu_store_unlock;
1831         }
1832
1833         /*
1834          * For version before VERSION_WIN10_V5_3, the following warning holds:
1835          *
1836          * Warning.  At this point, there is *no* guarantee that the host will
1837          * have successfully processed the vmbus_send_modifychannel() request.
1838          * See the header comment of vmbus_send_modifychannel() for more info.
1839          *
1840          * Lags in the processing of the above vmbus_send_modifychannel() can
1841          * result in missed interrupts if the "old" target CPU is taken offline
1842          * before Hyper-V starts sending interrupts to the "new" target CPU.
1843          * But apart from this offlining scenario, the code tolerates such
1844          * lags.  It will function correctly even if a channel interrupt comes
1845          * in on a CPU that is different from the channel target_cpu value.
1846          */
1847
1848         channel->target_cpu = target_cpu;
1849
1850         /* See init_vp_index(). */
1851         if (hv_is_perf_channel(channel))
1852                 hv_update_allocated_cpus(origin_cpu, target_cpu);
1853
1854         /* Currently set only for storvsc channels. */
1855         if (channel->change_target_cpu_callback) {
1856                 (*channel->change_target_cpu_callback)(channel,
1857                                 origin_cpu, target_cpu);
1858         }
1859
1860 cpu_store_unlock:
1861         mutex_unlock(&vmbus_connection.channel_mutex);
1862         cpus_read_unlock();
1863         return ret;
1864 }
1865 static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
1866
1867 static ssize_t channel_pending_show(struct vmbus_channel *channel,
1868                                     char *buf)
1869 {
1870         return sprintf(buf, "%d\n",
1871                        channel_pending(channel,
1872                                        vmbus_connection.monitor_pages[1]));
1873 }
1874 static VMBUS_CHAN_ATTR(pending, 0444, channel_pending_show, NULL);
1875
1876 static ssize_t channel_latency_show(struct vmbus_channel *channel,
1877                                     char *buf)
1878 {
1879         return sprintf(buf, "%d\n",
1880                        channel_latency(channel,
1881                                        vmbus_connection.monitor_pages[1]));
1882 }
1883 static VMBUS_CHAN_ATTR(latency, 0444, channel_latency_show, NULL);
1884
1885 static ssize_t channel_interrupts_show(struct vmbus_channel *channel, char *buf)
1886 {
1887         return sprintf(buf, "%llu\n", channel->interrupts);
1888 }
1889 static VMBUS_CHAN_ATTR(interrupts, 0444, channel_interrupts_show, NULL);
1890
1891 static ssize_t channel_events_show(struct vmbus_channel *channel, char *buf)
1892 {
1893         return sprintf(buf, "%llu\n", channel->sig_events);
1894 }
1895 static VMBUS_CHAN_ATTR(events, 0444, channel_events_show, NULL);
1896
1897 static ssize_t channel_intr_in_full_show(struct vmbus_channel *channel,
1898                                          char *buf)
1899 {
1900         return sprintf(buf, "%llu\n",
1901                        (unsigned long long)channel->intr_in_full);
1902 }
1903 static VMBUS_CHAN_ATTR(intr_in_full, 0444, channel_intr_in_full_show, NULL);
1904
1905 static ssize_t channel_intr_out_empty_show(struct vmbus_channel *channel,
1906                                            char *buf)
1907 {
1908         return sprintf(buf, "%llu\n",
1909                        (unsigned long long)channel->intr_out_empty);
1910 }
1911 static VMBUS_CHAN_ATTR(intr_out_empty, 0444, channel_intr_out_empty_show, NULL);
1912
1913 static ssize_t channel_out_full_first_show(struct vmbus_channel *channel,
1914                                            char *buf)
1915 {
1916         return sprintf(buf, "%llu\n",
1917                        (unsigned long long)channel->out_full_first);
1918 }
1919 static VMBUS_CHAN_ATTR(out_full_first, 0444, channel_out_full_first_show, NULL);
1920
1921 static ssize_t channel_out_full_total_show(struct vmbus_channel *channel,
1922                                            char *buf)
1923 {
1924         return sprintf(buf, "%llu\n",
1925                        (unsigned long long)channel->out_full_total);
1926 }
1927 static VMBUS_CHAN_ATTR(out_full_total, 0444, channel_out_full_total_show, NULL);
1928
1929 static ssize_t subchannel_monitor_id_show(struct vmbus_channel *channel,
1930                                           char *buf)
1931 {
1932         return sprintf(buf, "%u\n", channel->offermsg.monitorid);
1933 }
1934 static VMBUS_CHAN_ATTR(monitor_id, 0444, subchannel_monitor_id_show, NULL);
1935
1936 static ssize_t subchannel_id_show(struct vmbus_channel *channel,
1937                                   char *buf)
1938 {
1939         return sprintf(buf, "%u\n",
1940                        channel->offermsg.offer.sub_channel_index);
1941 }
1942 static VMBUS_CHAN_ATTR_RO(subchannel_id);
1943
1944 static struct attribute *vmbus_chan_attrs[] = {
1945         &chan_attr_out_mask.attr,
1946         &chan_attr_in_mask.attr,
1947         &chan_attr_read_avail.attr,
1948         &chan_attr_write_avail.attr,
1949         &chan_attr_cpu.attr,
1950         &chan_attr_pending.attr,
1951         &chan_attr_latency.attr,
1952         &chan_attr_interrupts.attr,
1953         &chan_attr_events.attr,
1954         &chan_attr_intr_in_full.attr,
1955         &chan_attr_intr_out_empty.attr,
1956         &chan_attr_out_full_first.attr,
1957         &chan_attr_out_full_total.attr,
1958         &chan_attr_monitor_id.attr,
1959         &chan_attr_subchannel_id.attr,
1960         NULL
1961 };
1962
1963 /*
1964  * Channel-level attribute_group callback function. Returns the permission for
1965  * each attribute, and returns 0 if an attribute is not visible.
1966  */
1967 static umode_t vmbus_chan_attr_is_visible(struct kobject *kobj,
1968                                           struct attribute *attr, int idx)
1969 {
1970         const struct vmbus_channel *channel =
1971                 container_of(kobj, struct vmbus_channel, kobj);
1972
1973         /* Hide the monitor attributes if the monitor mechanism is not used. */
1974         if (!channel->offermsg.monitor_allocated &&
1975             (attr == &chan_attr_pending.attr ||
1976              attr == &chan_attr_latency.attr ||
1977              attr == &chan_attr_monitor_id.attr))
1978                 return 0;
1979
1980         return attr->mode;
1981 }
1982
1983 static struct attribute_group vmbus_chan_group = {
1984         .attrs = vmbus_chan_attrs,
1985         .is_visible = vmbus_chan_attr_is_visible
1986 };
1987
1988 static struct kobj_type vmbus_chan_ktype = {
1989         .sysfs_ops = &vmbus_chan_sysfs_ops,
1990         .release = vmbus_chan_release,
1991 };
1992
1993 /*
1994  * vmbus_add_channel_kobj - setup a sub-directory under device/channels
1995  */
1996 int vmbus_add_channel_kobj(struct hv_device *dev, struct vmbus_channel *channel)
1997 {
1998         const struct device *device = &dev->device;
1999         struct kobject *kobj = &channel->kobj;
2000         u32 relid = channel->offermsg.child_relid;
2001         int ret;
2002
2003         kobj->kset = dev->channels_kset;
2004         ret = kobject_init_and_add(kobj, &vmbus_chan_ktype, NULL,
2005                                    "%u", relid);
2006         if (ret) {
2007                 kobject_put(kobj);
2008                 return ret;
2009         }
2010
2011         ret = sysfs_create_group(kobj, &vmbus_chan_group);
2012
2013         if (ret) {
2014                 /*
2015                  * The calling functions' error handling paths will cleanup the
2016                  * empty channel directory.
2017                  */
2018                 kobject_put(kobj);
2019                 dev_err(device, "Unable to set up channel sysfs files\n");
2020                 return ret;
2021         }
2022
2023         kobject_uevent(kobj, KOBJ_ADD);
2024
2025         return 0;
2026 }
2027
2028 /*
2029  * vmbus_remove_channel_attr_group - remove the channel's attribute group
2030  */
2031 void vmbus_remove_channel_attr_group(struct vmbus_channel *channel)
2032 {
2033         sysfs_remove_group(&channel->kobj, &vmbus_chan_group);
2034 }
2035
2036 /*
2037  * vmbus_device_create - Creates and registers a new child device
2038  * on the vmbus.
2039  */
2040 struct hv_device *vmbus_device_create(const guid_t *type,
2041                                       const guid_t *instance,
2042                                       struct vmbus_channel *channel)
2043 {
2044         struct hv_device *child_device_obj;
2045
2046         child_device_obj = kzalloc(sizeof(struct hv_device), GFP_KERNEL);
2047         if (!child_device_obj) {
2048                 pr_err("Unable to allocate device object for child device\n");
2049                 return NULL;
2050         }
2051
2052         child_device_obj->channel = channel;
2053         guid_copy(&child_device_obj->dev_type, type);
2054         guid_copy(&child_device_obj->dev_instance, instance);
2055         child_device_obj->vendor_id = 0x1414; /* MSFT vendor ID */
2056
2057         return child_device_obj;
2058 }
2059
2060 /*
2061  * vmbus_device_register - Register the child device
2062  */
2063 int vmbus_device_register(struct hv_device *child_device_obj)
2064 {
2065         struct kobject *kobj = &child_device_obj->device.kobj;
2066         int ret;
2067
2068         dev_set_name(&child_device_obj->device, "%pUl",
2069                      &child_device_obj->channel->offermsg.offer.if_instance);
2070
2071         child_device_obj->device.bus = &hv_bus;
2072         child_device_obj->device.parent = &hv_acpi_dev->dev;
2073         child_device_obj->device.release = vmbus_device_release;
2074
2075         child_device_obj->device.dma_parms = &child_device_obj->dma_parms;
2076         child_device_obj->device.dma_mask = &child_device_obj->dma_mask;
2077         dma_set_mask(&child_device_obj->device, DMA_BIT_MASK(64));
2078
2079         /*
2080          * Register with the LDM. This will kick off the driver/device
2081          * binding...which will eventually call vmbus_match() and vmbus_probe()
2082          */
2083         ret = device_register(&child_device_obj->device);
2084         if (ret) {
2085                 pr_err("Unable to register child device\n");
2086                 return ret;
2087         }
2088
2089         child_device_obj->channels_kset = kset_create_and_add("channels",
2090                                                               NULL, kobj);
2091         if (!child_device_obj->channels_kset) {
2092                 ret = -ENOMEM;
2093                 goto err_dev_unregister;
2094         }
2095
2096         ret = vmbus_add_channel_kobj(child_device_obj,
2097                                      child_device_obj->channel);
2098         if (ret) {
2099                 pr_err("Unable to register primary channeln");
2100                 goto err_kset_unregister;
2101         }
2102         hv_debug_add_dev_dir(child_device_obj);
2103
2104         return 0;
2105
2106 err_kset_unregister:
2107         kset_unregister(child_device_obj->channels_kset);
2108
2109 err_dev_unregister:
2110         device_unregister(&child_device_obj->device);
2111         return ret;
2112 }
2113
2114 /*
2115  * vmbus_device_unregister - Remove the specified child device
2116  * from the vmbus.
2117  */
2118 void vmbus_device_unregister(struct hv_device *device_obj)
2119 {
2120         pr_debug("child device %s unregistered\n",
2121                 dev_name(&device_obj->device));
2122
2123         kset_unregister(device_obj->channels_kset);
2124
2125         /*
2126          * Kick off the process of unregistering the device.
2127          * This will call vmbus_remove() and eventually vmbus_device_release()
2128          */
2129         device_unregister(&device_obj->device);
2130 }
2131
2132
2133 /*
2134  * VMBUS is an acpi enumerated device. Get the information we
2135  * need from DSDT.
2136  */
2137 #define VTPM_BASE_ADDRESS 0xfed40000
2138 static acpi_status vmbus_walk_resources(struct acpi_resource *res, void *ctx)
2139 {
2140         resource_size_t start = 0;
2141         resource_size_t end = 0;
2142         struct resource *new_res;
2143         struct resource **old_res = &hyperv_mmio;
2144         struct resource **prev_res = NULL;
2145         struct resource r;
2146
2147         switch (res->type) {
2148
2149         /*
2150          * "Address" descriptors are for bus windows. Ignore
2151          * "memory" descriptors, which are for registers on
2152          * devices.
2153          */
2154         case ACPI_RESOURCE_TYPE_ADDRESS32:
2155                 start = res->data.address32.address.minimum;
2156                 end = res->data.address32.address.maximum;
2157                 break;
2158
2159         case ACPI_RESOURCE_TYPE_ADDRESS64:
2160                 start = res->data.address64.address.minimum;
2161                 end = res->data.address64.address.maximum;
2162                 break;
2163
2164         /*
2165          * The IRQ information is needed only on ARM64, which Hyper-V
2166          * sets up in the extended format. IRQ information is present
2167          * on x86/x64 in the non-extended format but it is not used by
2168          * Linux. So don't bother checking for the non-extended format.
2169          */
2170         case ACPI_RESOURCE_TYPE_EXTENDED_IRQ:
2171                 if (!acpi_dev_resource_interrupt(res, 0, &r)) {
2172                         pr_err("Unable to parse Hyper-V ACPI interrupt\n");
2173                         return AE_ERROR;
2174                 }
2175                 /* ARM64 INTID for VMbus */
2176                 vmbus_interrupt = res->data.extended_irq.interrupts[0];
2177                 /* Linux IRQ number */
2178                 vmbus_irq = r.start;
2179                 return AE_OK;
2180
2181         default:
2182                 /* Unused resource type */
2183                 return AE_OK;
2184
2185         }
2186         /*
2187          * Ignore ranges that are below 1MB, as they're not
2188          * necessary or useful here.
2189          */
2190         if (end < 0x100000)
2191                 return AE_OK;
2192
2193         new_res = kzalloc(sizeof(*new_res), GFP_ATOMIC);
2194         if (!new_res)
2195                 return AE_NO_MEMORY;
2196
2197         /* If this range overlaps the virtual TPM, truncate it. */
2198         if (end > VTPM_BASE_ADDRESS && start < VTPM_BASE_ADDRESS)
2199                 end = VTPM_BASE_ADDRESS;
2200
2201         new_res->name = "hyperv mmio";
2202         new_res->flags = IORESOURCE_MEM;
2203         new_res->start = start;
2204         new_res->end = end;
2205
2206         /*
2207          * If two ranges are adjacent, merge them.
2208          */
2209         do {
2210                 if (!*old_res) {
2211                         *old_res = new_res;
2212                         break;
2213                 }
2214
2215                 if (((*old_res)->end + 1) == new_res->start) {
2216                         (*old_res)->end = new_res->end;
2217                         kfree(new_res);
2218                         break;
2219                 }
2220
2221                 if ((*old_res)->start == new_res->end + 1) {
2222                         (*old_res)->start = new_res->start;
2223                         kfree(new_res);
2224                         break;
2225                 }
2226
2227                 if ((*old_res)->start > new_res->end) {
2228                         new_res->sibling = *old_res;
2229                         if (prev_res)
2230                                 (*prev_res)->sibling = new_res;
2231                         *old_res = new_res;
2232                         break;
2233                 }
2234
2235                 prev_res = old_res;
2236                 old_res = &(*old_res)->sibling;
2237
2238         } while (1);
2239
2240         return AE_OK;
2241 }
2242
2243 static int vmbus_acpi_remove(struct acpi_device *device)
2244 {
2245         struct resource *cur_res;
2246         struct resource *next_res;
2247
2248         if (hyperv_mmio) {
2249                 if (fb_mmio) {
2250                         __release_region(hyperv_mmio, fb_mmio->start,
2251                                          resource_size(fb_mmio));
2252                         fb_mmio = NULL;
2253                 }
2254
2255                 for (cur_res = hyperv_mmio; cur_res; cur_res = next_res) {
2256                         next_res = cur_res->sibling;
2257                         kfree(cur_res);
2258                 }
2259         }
2260
2261         return 0;
2262 }
2263
2264 static void vmbus_reserve_fb(void)
2265 {
2266         resource_size_t start = 0, size;
2267         struct pci_dev *pdev;
2268
2269         if (efi_enabled(EFI_BOOT)) {
2270                 /* Gen2 VM: get FB base from EFI framebuffer */
2271                 start = screen_info.lfb_base;
2272                 size = max_t(__u32, screen_info.lfb_size, 0x800000);
2273         } else {
2274                 /* Gen1 VM: get FB base from PCI */
2275                 pdev = pci_get_device(PCI_VENDOR_ID_MICROSOFT,
2276                                       PCI_DEVICE_ID_HYPERV_VIDEO, NULL);
2277                 if (!pdev)
2278                         return;
2279
2280                 if (pdev->resource[0].flags & IORESOURCE_MEM) {
2281                         start = pci_resource_start(pdev, 0);
2282                         size = pci_resource_len(pdev, 0);
2283                 }
2284
2285                 /*
2286                  * Release the PCI device so hyperv_drm or hyperv_fb driver can
2287                  * grab it later.
2288                  */
2289                 pci_dev_put(pdev);
2290         }
2291
2292         if (!start)
2293                 return;
2294
2295         /*
2296          * Make a claim for the frame buffer in the resource tree under the
2297          * first node, which will be the one below 4GB.  The length seems to
2298          * be underreported, particularly in a Generation 1 VM.  So start out
2299          * reserving a larger area and make it smaller until it succeeds.
2300          */
2301         for (; !fb_mmio && (size >= 0x100000); size >>= 1)
2302                 fb_mmio = __request_region(hyperv_mmio, start, size, fb_mmio_name, 0);
2303 }
2304
2305 /**
2306  * vmbus_allocate_mmio() - Pick a memory-mapped I/O range.
2307  * @new:                If successful, supplied a pointer to the
2308  *                      allocated MMIO space.
2309  * @device_obj:         Identifies the caller
2310  * @min:                Minimum guest physical address of the
2311  *                      allocation
2312  * @max:                Maximum guest physical address
2313  * @size:               Size of the range to be allocated
2314  * @align:              Alignment of the range to be allocated
2315  * @fb_overlap_ok:      Whether this allocation can be allowed
2316  *                      to overlap the video frame buffer.
2317  *
2318  * This function walks the resources granted to VMBus by the
2319  * _CRS object in the ACPI namespace underneath the parent
2320  * "bridge" whether that's a root PCI bus in the Generation 1
2321  * case or a Module Device in the Generation 2 case.  It then
2322  * attempts to allocate from the global MMIO pool in a way that
2323  * matches the constraints supplied in these parameters and by
2324  * that _CRS.
2325  *
2326  * Return: 0 on success, -errno on failure
2327  */
2328 int vmbus_allocate_mmio(struct resource **new, struct hv_device *device_obj,
2329                         resource_size_t min, resource_size_t max,
2330                         resource_size_t size, resource_size_t align,
2331                         bool fb_overlap_ok)
2332 {
2333         struct resource *iter, *shadow;
2334         resource_size_t range_min, range_max, start, end;
2335         const char *dev_n = dev_name(&device_obj->device);
2336         int retval;
2337
2338         retval = -ENXIO;
2339         mutex_lock(&hyperv_mmio_lock);
2340
2341         /*
2342          * If overlaps with frame buffers are allowed, then first attempt to
2343          * make the allocation from within the reserved region.  Because it
2344          * is already reserved, no shadow allocation is necessary.
2345          */
2346         if (fb_overlap_ok && fb_mmio && !(min > fb_mmio->end) &&
2347             !(max < fb_mmio->start)) {
2348
2349                 range_min = fb_mmio->start;
2350                 range_max = fb_mmio->end;
2351                 start = (range_min + align - 1) & ~(align - 1);
2352                 for (; start + size - 1 <= range_max; start += align) {
2353                         *new = request_mem_region_exclusive(start, size, dev_n);
2354                         if (*new) {
2355                                 retval = 0;
2356                                 goto exit;
2357                         }
2358                 }
2359         }
2360
2361         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2362                 if ((iter->start >= max) || (iter->end <= min))
2363                         continue;
2364
2365                 range_min = iter->start;
2366                 range_max = iter->end;
2367                 start = (range_min + align - 1) & ~(align - 1);
2368                 for (; start + size - 1 <= range_max; start += align) {
2369                         end = start + size - 1;
2370
2371                         /* Skip the whole fb_mmio region if not fb_overlap_ok */
2372                         if (!fb_overlap_ok && fb_mmio &&
2373                             (((start >= fb_mmio->start) && (start <= fb_mmio->end)) ||
2374                              ((end >= fb_mmio->start) && (end <= fb_mmio->end))))
2375                                 continue;
2376
2377                         shadow = __request_region(iter, start, size, NULL,
2378                                                   IORESOURCE_BUSY);
2379                         if (!shadow)
2380                                 continue;
2381
2382                         *new = request_mem_region_exclusive(start, size, dev_n);
2383                         if (*new) {
2384                                 shadow->name = (char *)*new;
2385                                 retval = 0;
2386                                 goto exit;
2387                         }
2388
2389                         __release_region(iter, start, size);
2390                 }
2391         }
2392
2393 exit:
2394         mutex_unlock(&hyperv_mmio_lock);
2395         return retval;
2396 }
2397 EXPORT_SYMBOL_GPL(vmbus_allocate_mmio);
2398
2399 /**
2400  * vmbus_free_mmio() - Free a memory-mapped I/O range.
2401  * @start:              Base address of region to release.
2402  * @size:               Size of the range to be allocated
2403  *
2404  * This function releases anything requested by
2405  * vmbus_mmio_allocate().
2406  */
2407 void vmbus_free_mmio(resource_size_t start, resource_size_t size)
2408 {
2409         struct resource *iter;
2410
2411         mutex_lock(&hyperv_mmio_lock);
2412         for (iter = hyperv_mmio; iter; iter = iter->sibling) {
2413                 if ((iter->start >= start + size) || (iter->end <= start))
2414                         continue;
2415
2416                 __release_region(iter, start, size);
2417         }
2418         release_mem_region(start, size);
2419         mutex_unlock(&hyperv_mmio_lock);
2420
2421 }
2422 EXPORT_SYMBOL_GPL(vmbus_free_mmio);
2423
2424 static int vmbus_acpi_add(struct acpi_device *device)
2425 {
2426         acpi_status result;
2427         int ret_val = -ENODEV;
2428         struct acpi_device *ancestor;
2429
2430         hv_acpi_dev = device;
2431
2432         /*
2433          * Older versions of Hyper-V for ARM64 fail to include the _CCA
2434          * method on the top level VMbus device in the DSDT. But devices
2435          * are hardware coherent in all current Hyper-V use cases, so fix
2436          * up the ACPI device to behave as if _CCA is present and indicates
2437          * hardware coherence.
2438          */
2439         ACPI_COMPANION_SET(&device->dev, device);
2440         if (IS_ENABLED(CONFIG_ACPI_CCA_REQUIRED) &&
2441             device_get_dma_attr(&device->dev) == DEV_DMA_NOT_SUPPORTED) {
2442                 pr_info("No ACPI _CCA found; assuming coherent device I/O\n");
2443                 device->flags.cca_seen = true;
2444                 device->flags.coherent_dma = true;
2445         }
2446
2447         result = acpi_walk_resources(device->handle, METHOD_NAME__CRS,
2448                                         vmbus_walk_resources, NULL);
2449
2450         if (ACPI_FAILURE(result))
2451                 goto acpi_walk_err;
2452         /*
2453          * Some ancestor of the vmbus acpi device (Gen1 or Gen2
2454          * firmware) is the VMOD that has the mmio ranges. Get that.
2455          */
2456         for (ancestor = acpi_dev_parent(device); ancestor;
2457              ancestor = acpi_dev_parent(ancestor)) {
2458                 result = acpi_walk_resources(ancestor->handle, METHOD_NAME__CRS,
2459                                              vmbus_walk_resources, NULL);
2460
2461                 if (ACPI_FAILURE(result))
2462                         continue;
2463                 if (hyperv_mmio) {
2464                         vmbus_reserve_fb();
2465                         break;
2466                 }
2467         }
2468         ret_val = 0;
2469
2470 acpi_walk_err:
2471         complete(&probe_event);
2472         if (ret_val)
2473                 vmbus_acpi_remove(device);
2474         return ret_val;
2475 }
2476
2477 #ifdef CONFIG_PM_SLEEP
2478 static int vmbus_bus_suspend(struct device *dev)
2479 {
2480         struct hv_per_cpu_context *hv_cpu = per_cpu_ptr(
2481                         hv_context.cpu_context, VMBUS_CONNECT_CPU);
2482         struct vmbus_channel *channel, *sc;
2483
2484         tasklet_disable(&hv_cpu->msg_dpc);
2485         vmbus_connection.ignore_any_offer_msg = true;
2486         /* The tasklet_enable() takes care of providing a memory barrier */
2487         tasklet_enable(&hv_cpu->msg_dpc);
2488
2489         /* Drain all the workqueues as we are in suspend */
2490         drain_workqueue(vmbus_connection.rescind_work_queue);
2491         drain_workqueue(vmbus_connection.work_queue);
2492         drain_workqueue(vmbus_connection.handle_primary_chan_wq);
2493         drain_workqueue(vmbus_connection.handle_sub_chan_wq);
2494
2495         mutex_lock(&vmbus_connection.channel_mutex);
2496         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2497                 if (!is_hvsock_channel(channel))
2498                         continue;
2499
2500                 vmbus_force_channel_rescinded(channel);
2501         }
2502         mutex_unlock(&vmbus_connection.channel_mutex);
2503
2504         /*
2505          * Wait until all the sub-channels and hv_sock channels have been
2506          * cleaned up. Sub-channels should be destroyed upon suspend, otherwise
2507          * they would conflict with the new sub-channels that will be created
2508          * in the resume path. hv_sock channels should also be destroyed, but
2509          * a hv_sock channel of an established hv_sock connection can not be
2510          * really destroyed since it may still be referenced by the userspace
2511          * application, so we just force the hv_sock channel to be rescinded
2512          * by vmbus_force_channel_rescinded(), and the userspace application
2513          * will thoroughly destroy the channel after hibernation.
2514          *
2515          * Note: the counter nr_chan_close_on_suspend may never go above 0 if
2516          * the VM has no sub-channel and hv_sock channel, e.g. a 1-vCPU VM.
2517          */
2518         if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
2519                 wait_for_completion(&vmbus_connection.ready_for_suspend_event);
2520
2521         if (atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0) {
2522                 pr_err("Can not suspend due to a previous failed resuming\n");
2523                 return -EBUSY;
2524         }
2525
2526         mutex_lock(&vmbus_connection.channel_mutex);
2527
2528         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
2529                 /*
2530                  * Remove the channel from the array of channels and invalidate
2531                  * the channel's relid.  Upon resume, vmbus_onoffer() will fix
2532                  * up the relid (and other fields, if necessary) and add the
2533                  * channel back to the array.
2534                  */
2535                 vmbus_channel_unmap_relid(channel);
2536                 channel->offermsg.child_relid = INVALID_RELID;
2537
2538                 if (is_hvsock_channel(channel)) {
2539                         if (!channel->rescind) {
2540                                 pr_err("hv_sock channel not rescinded!\n");
2541                                 WARN_ON_ONCE(1);
2542                         }
2543                         continue;
2544                 }
2545
2546                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
2547                         pr_err("Sub-channel not deleted!\n");
2548                         WARN_ON_ONCE(1);
2549                 }
2550
2551                 atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
2552         }
2553
2554         mutex_unlock(&vmbus_connection.channel_mutex);
2555
2556         vmbus_initiate_unload(false);
2557
2558         /* Reset the event for the next resume. */
2559         reinit_completion(&vmbus_connection.ready_for_resume_event);
2560
2561         return 0;
2562 }
2563
2564 static int vmbus_bus_resume(struct device *dev)
2565 {
2566         struct vmbus_channel_msginfo *msginfo;
2567         size_t msgsize;
2568         int ret;
2569
2570         vmbus_connection.ignore_any_offer_msg = false;
2571
2572         /*
2573          * We only use the 'vmbus_proto_version', which was in use before
2574          * hibernation, to re-negotiate with the host.
2575          */
2576         if (!vmbus_proto_version) {
2577                 pr_err("Invalid proto version = 0x%x\n", vmbus_proto_version);
2578                 return -EINVAL;
2579         }
2580
2581         msgsize = sizeof(*msginfo) +
2582                   sizeof(struct vmbus_channel_initiate_contact);
2583
2584         msginfo = kzalloc(msgsize, GFP_KERNEL);
2585
2586         if (msginfo == NULL)
2587                 return -ENOMEM;
2588
2589         ret = vmbus_negotiate_version(msginfo, vmbus_proto_version);
2590
2591         kfree(msginfo);
2592
2593         if (ret != 0)
2594                 return ret;
2595
2596         WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
2597
2598         vmbus_request_offers();
2599
2600         if (wait_for_completion_timeout(
2601                 &vmbus_connection.ready_for_resume_event, 10 * HZ) == 0)
2602                 pr_err("Some vmbus device is missing after suspending?\n");
2603
2604         /* Reset the event for the next suspend. */
2605         reinit_completion(&vmbus_connection.ready_for_suspend_event);
2606
2607         return 0;
2608 }
2609 #else
2610 #define vmbus_bus_suspend NULL
2611 #define vmbus_bus_resume NULL
2612 #endif /* CONFIG_PM_SLEEP */
2613
2614 static const struct acpi_device_id vmbus_acpi_device_ids[] = {
2615         {"VMBUS", 0},
2616         {"VMBus", 0},
2617         {"", 0},
2618 };
2619 MODULE_DEVICE_TABLE(acpi, vmbus_acpi_device_ids);
2620
2621 /*
2622  * Note: we must use the "no_irq" ops, otherwise hibernation can not work with
2623  * PCI device assignment, because "pci_dev_pm_ops" uses the "noirq" ops: in
2624  * the resume path, the pci "noirq" restore op runs before "non-noirq" op (see
2625  * resume_target_kernel() -> dpm_resume_start(), and hibernation_restore() ->
2626  * dpm_resume_end()). This means vmbus_bus_resume() and the pci-hyperv's
2627  * resume callback must also run via the "noirq" ops.
2628  *
2629  * Set suspend_noirq/resume_noirq to NULL for Suspend-to-Idle: see the comment
2630  * earlier in this file before vmbus_pm.
2631  */
2632
2633 static const struct dev_pm_ops vmbus_bus_pm = {
2634         .suspend_noirq  = NULL,
2635         .resume_noirq   = NULL,
2636         .freeze_noirq   = vmbus_bus_suspend,
2637         .thaw_noirq     = vmbus_bus_resume,
2638         .poweroff_noirq = vmbus_bus_suspend,
2639         .restore_noirq  = vmbus_bus_resume
2640 };
2641
2642 static struct acpi_driver vmbus_acpi_driver = {
2643         .name = "vmbus",
2644         .ids = vmbus_acpi_device_ids,
2645         .ops = {
2646                 .add = vmbus_acpi_add,
2647                 .remove = vmbus_acpi_remove,
2648         },
2649         .drv.pm = &vmbus_bus_pm,
2650 };
2651
2652 static void hv_kexec_handler(void)
2653 {
2654         hv_stimer_global_cleanup();
2655         vmbus_initiate_unload(false);
2656         /* Make sure conn_state is set as hv_synic_cleanup checks for it */
2657         mb();
2658         cpuhp_remove_state(hyperv_cpuhp_online);
2659 };
2660
2661 static void hv_crash_handler(struct pt_regs *regs)
2662 {
2663         int cpu;
2664
2665         vmbus_initiate_unload(true);
2666         /*
2667          * In crash handler we can't schedule synic cleanup for all CPUs,
2668          * doing the cleanup for current CPU only. This should be sufficient
2669          * for kdump.
2670          */
2671         cpu = smp_processor_id();
2672         hv_stimer_cleanup(cpu);
2673         hv_synic_disable_regs(cpu);
2674 };
2675
2676 static int hv_synic_suspend(void)
2677 {
2678         /*
2679          * When we reach here, all the non-boot CPUs have been offlined.
2680          * If we're in a legacy configuration where stimer Direct Mode is
2681          * not enabled, the stimers on the non-boot CPUs have been unbound
2682          * in hv_synic_cleanup() -> hv_stimer_legacy_cleanup() ->
2683          * hv_stimer_cleanup() -> clockevents_unbind_device().
2684          *
2685          * hv_synic_suspend() only runs on CPU0 with interrupts disabled.
2686          * Here we do not call hv_stimer_legacy_cleanup() on CPU0 because:
2687          * 1) it's unnecessary as interrupts remain disabled between
2688          * syscore_suspend() and syscore_resume(): see create_image() and
2689          * resume_target_kernel()
2690          * 2) the stimer on CPU0 is automatically disabled later by
2691          * syscore_suspend() -> timekeeping_suspend() -> tick_suspend() -> ...
2692          * -> clockevents_shutdown() -> ... -> hv_ce_shutdown()
2693          * 3) a warning would be triggered if we call
2694          * clockevents_unbind_device(), which may sleep, in an
2695          * interrupts-disabled context.
2696          */
2697
2698         hv_synic_disable_regs(0);
2699
2700         return 0;
2701 }
2702
2703 static void hv_synic_resume(void)
2704 {
2705         hv_synic_enable_regs(0);
2706
2707         /*
2708          * Note: we don't need to call hv_stimer_init(0), because the timer
2709          * on CPU0 is not unbound in hv_synic_suspend(), and the timer is
2710          * automatically re-enabled in timekeeping_resume().
2711          */
2712 }
2713
2714 /* The callbacks run only on CPU0, with irqs_disabled. */
2715 static struct syscore_ops hv_synic_syscore_ops = {
2716         .suspend = hv_synic_suspend,
2717         .resume = hv_synic_resume,
2718 };
2719
2720 static int __init hv_acpi_init(void)
2721 {
2722         int ret, t;
2723
2724         if (!hv_is_hyperv_initialized())
2725                 return -ENODEV;
2726
2727         if (hv_root_partition)
2728                 return 0;
2729
2730         init_completion(&probe_event);
2731
2732         /*
2733          * Get ACPI resources first.
2734          */
2735         ret = acpi_bus_register_driver(&vmbus_acpi_driver);
2736
2737         if (ret)
2738                 return ret;
2739
2740         t = wait_for_completion_timeout(&probe_event, 5*HZ);
2741         if (t == 0) {
2742                 ret = -ETIMEDOUT;
2743                 goto cleanup;
2744         }
2745
2746         /*
2747          * If we're on an architecture with a hardcoded hypervisor
2748          * vector (i.e. x86/x64), override the VMbus interrupt found
2749          * in the ACPI tables. Ensure vmbus_irq is not set since the
2750          * normal Linux IRQ mechanism is not used in this case.
2751          */
2752 #ifdef HYPERVISOR_CALLBACK_VECTOR
2753         vmbus_interrupt = HYPERVISOR_CALLBACK_VECTOR;
2754         vmbus_irq = -1;
2755 #endif
2756
2757         hv_debug_init();
2758
2759         ret = vmbus_bus_init();
2760         if (ret)
2761                 goto cleanup;
2762
2763         hv_setup_kexec_handler(hv_kexec_handler);
2764         hv_setup_crash_handler(hv_crash_handler);
2765
2766         register_syscore_ops(&hv_synic_syscore_ops);
2767
2768         return 0;
2769
2770 cleanup:
2771         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2772         hv_acpi_dev = NULL;
2773         return ret;
2774 }
2775
2776 static void __exit vmbus_exit(void)
2777 {
2778         int cpu;
2779
2780         unregister_syscore_ops(&hv_synic_syscore_ops);
2781
2782         hv_remove_kexec_handler();
2783         hv_remove_crash_handler();
2784         vmbus_connection.conn_state = DISCONNECTED;
2785         hv_stimer_global_cleanup();
2786         vmbus_disconnect();
2787         if (vmbus_irq == -1) {
2788                 hv_remove_vmbus_handler();
2789         } else {
2790                 free_percpu_irq(vmbus_irq, vmbus_evt);
2791                 free_percpu(vmbus_evt);
2792         }
2793         for_each_online_cpu(cpu) {
2794                 struct hv_per_cpu_context *hv_cpu
2795                         = per_cpu_ptr(hv_context.cpu_context, cpu);
2796
2797                 tasklet_kill(&hv_cpu->msg_dpc);
2798         }
2799         hv_debug_rm_all_dir();
2800
2801         vmbus_free_channels();
2802         kfree(vmbus_connection.channels);
2803
2804         if (ms_hyperv.misc_features & HV_FEATURE_GUEST_CRASH_MSR_AVAILABLE) {
2805                 kmsg_dump_unregister(&hv_kmsg_dumper);
2806                 unregister_die_notifier(&hyperv_die_block);
2807         }
2808
2809         /*
2810          * The panic notifier is always registered, hence we should
2811          * also unconditionally unregister it here as well.
2812          */
2813         atomic_notifier_chain_unregister(&panic_notifier_list,
2814                                          &hyperv_panic_block);
2815
2816         free_page((unsigned long)hv_panic_page);
2817         unregister_sysctl_table(hv_ctl_table_hdr);
2818         hv_ctl_table_hdr = NULL;
2819         bus_unregister(&hv_bus);
2820
2821         cpuhp_remove_state(hyperv_cpuhp_online);
2822         hv_synic_free();
2823         acpi_bus_unregister_driver(&vmbus_acpi_driver);
2824 }
2825
2826
2827 MODULE_LICENSE("GPL");
2828 MODULE_DESCRIPTION("Microsoft Hyper-V VMBus Driver");
2829
2830 subsys_initcall(hv_acpi_init);
2831 module_exit(vmbus_exit);