1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3 * Copyright (c) 2023, Microsoft Corporation.
4 */
5
6 #ifndef _MSHV_ROOT_H_
7 #define _MSHV_ROOT_H_
8
9 #include <linux/spinlock.h>
10 #include <linux/mutex.h>
11 #include <linux/semaphore.h>
12 #include <linux/sched.h>
13 #include <linux/srcu.h>
14 #include <linux/wait.h>
15 #include <linux/hashtable.h>
16 #include <linux/dev_printk.h>
17 #include <linux/build_bug.h>
18 #include <linux/mmu_notifier.h>
19 #include <uapi/linux/mshv.h>
20 #include "mshv_trace.h"
21
22 /*
23 * Hypervisor must be between these version numbers (inclusive)
24 * to guarantee compatibility
25 */
26 #define MSHV_HV_MIN_VERSION (27744)
27 #define MSHV_HV_MAX_VERSION (27751)
28
29 static_assert(HV_HYP_PAGE_SIZE == MSHV_HV_PAGE_SIZE);
30
31 #define MSHV_MAX_VPS 256
32
33 #define MSHV_PARTITIONS_HASH_BITS 9
34
35 #define MSHV_PIN_PAGES_BATCH_SIZE (0x10000000ULL / HV_HYP_PAGE_SIZE)
36
37 struct mshv_vp {
38 u32 vp_index;
39 struct mshv_partition *vp_partition;
40 struct mutex vp_mutex;
41 struct hv_vp_register_page *vp_register_page;
42 struct hv_message *vp_intercept_msg_page;
43 void *vp_ghcb_page;
44 struct hv_stats_page *vp_stats_pages[2];
45 struct {
46 atomic64_t vp_signaled_count;
47 struct {
48 u64 intercept_suspend: 1;
49 u64 root_sched_blocked: 1; /* root scheduler only */
50 u64 root_sched_dispatched: 1; /* root scheduler only */
51 u64 reserved: 61;
52 } flags;
53 unsigned int kicked_by_hv;
54 wait_queue_head_t vp_suspend_queue;
55 } run;
56 #if IS_ENABLED(CONFIG_DEBUG_FS)
57 struct dentry *vp_stats_dentry;
58 #endif
59 };
60
61 #define vp_fmt(fmt) "p%lluvp%u: " fmt
62 #define vp_devprintk(level, v, fmt, ...) \
63 do { \
64 const struct mshv_vp *__vp = (v); \
65 const struct mshv_partition *__pt = __vp->vp_partition; \
66 dev_##level(__pt->pt_module_dev, vp_fmt(fmt), __pt->pt_id, \
67 __vp->vp_index, ##__VA_ARGS__); \
68 } while (0)
69 #define vp_emerg(v, fmt, ...) vp_devprintk(emerg, v, fmt, ##__VA_ARGS__)
70 #define vp_crit(v, fmt, ...) vp_devprintk(crit, v, fmt, ##__VA_ARGS__)
71 #define vp_alert(v, fmt, ...) vp_devprintk(alert, v, fmt, ##__VA_ARGS__)
72 #define vp_err(v, fmt, ...) vp_devprintk(err, v, fmt, ##__VA_ARGS__)
73 #define vp_warn(v, fmt, ...) vp_devprintk(warn, v, fmt, ##__VA_ARGS__)
74 #define vp_notice(v, fmt, ...) vp_devprintk(notice, v, fmt, ##__VA_ARGS__)
75 #define vp_info(v, fmt, ...) vp_devprintk(info, v, fmt, ##__VA_ARGS__)
76 #define vp_dbg(v, fmt, ...) vp_devprintk(dbg, v, fmt, ##__VA_ARGS__)
77
78 enum mshv_region_type {
79 MSHV_REGION_TYPE_MEM_PINNED,
80 MSHV_REGION_TYPE_MEM_MOVABLE,
81 MSHV_REGION_TYPE_MMIO
82 };
83
84 struct mshv_mem_region {
85 struct hlist_node hnode;
86 struct kref mreg_refcount;
87 u64 nr_pages;
88 u64 start_gfn;
89 u64 start_uaddr;
90 u32 hv_map_flags;
91 struct mshv_partition *partition;
92 enum mshv_region_type mreg_type;
93 struct mmu_interval_notifier mreg_mni;
94 struct mutex mreg_mutex; /* protects region pages remapping */
95 struct page *mreg_pages[];
96 };
97
98 struct mshv_irq_ack_notifier {
99 struct hlist_node link;
100 unsigned int irq_ack_gsi;
101 void (*irq_acked)(struct mshv_irq_ack_notifier *mian);
102 };
103
104 struct mshv_partition {
105 struct device *pt_module_dev;
106
107 struct hlist_node pt_hnode;
108 u64 pt_id;
109 refcount_t pt_ref_count;
110 struct mutex pt_mutex;
111
112 spinlock_t pt_mem_regions_lock;
113 struct hlist_head pt_mem_regions; // not ordered
114
115 u32 pt_vp_count;
116 struct mshv_vp *pt_vp_array[MSHV_MAX_VPS];
117
118 struct mutex pt_irq_lock;
119 struct srcu_struct pt_irq_srcu;
120 struct hlist_head irq_ack_notifier_list;
121
122 struct hlist_head pt_devices;
123
124 /*
125 * MSHV does not support more than one async hypercall in flight
126 * for a single partition. Thus, it is okay to define per partition
127 * async hypercall status.
128 */
129 struct completion async_hypercall;
130 u64 async_hypercall_status;
131
132 spinlock_t pt_irqfds_lock;
133 struct hlist_head pt_irqfds_list;
134 struct mutex irqfds_resampler_lock;
135 struct hlist_head irqfds_resampler_list;
136
137 struct hlist_head ioeventfds_list;
138
139 struct mshv_girq_routing_table __rcu *pt_girq_tbl;
140 u64 isolation_type;
141 bool import_completed;
142 bool pt_initialized;
143 #if IS_ENABLED(CONFIG_DEBUG_FS)
144 struct dentry *pt_stats_dentry;
145 struct dentry *pt_vp_dentry;
146 #endif
147 };
148
149 #define pt_fmt(fmt) "p%llu: " fmt
150 #define pt_devprintk(level, p, fmt, ...) \
151 do { \
152 const struct mshv_partition *__pt = (p); \
153 dev_##level(__pt->pt_module_dev, pt_fmt(fmt), __pt->pt_id, \
154 ##__VA_ARGS__); \
155 } while (0)
156 #define pt_emerg(p, fmt, ...) pt_devprintk(emerg, p, fmt, ##__VA_ARGS__)
157 #define pt_crit(p, fmt, ...) pt_devprintk(crit, p, fmt, ##__VA_ARGS__)
158 #define pt_alert(p, fmt, ...) pt_devprintk(alert, p, fmt, ##__VA_ARGS__)
159 #define pt_err(p, fmt, ...) pt_devprintk(err, p, fmt, ##__VA_ARGS__)
160 #define pt_warn(p, fmt, ...) pt_devprintk(warn, p, fmt, ##__VA_ARGS__)
161 #define pt_notice(p, fmt, ...) pt_devprintk(notice, p, fmt, ##__VA_ARGS__)
162 #define pt_info(p, fmt, ...) pt_devprintk(info, p, fmt, ##__VA_ARGS__)
163 #define pt_dbg(p, fmt, ...) pt_devprintk(dbg, p, fmt, ##__VA_ARGS__)
164
165 struct mshv_lapic_irq {
166 u32 lapic_vector;
167 u64 lapic_apic_id;
168 union hv_interrupt_control lapic_control;
169 };
170
171 #define MSHV_MAX_GUEST_IRQS 4096
172
173 /* representation of one guest irq entry, either msi or legacy */
174 struct mshv_guest_irq_ent {
175 u32 girq_entry_valid; /* vfio looks at this */
176 u32 guest_irq_num; /* a unique number for each irq */
177 u32 girq_addr_lo; /* guest irq msi address info */
178 u32 girq_addr_hi;
179 u32 girq_irq_data; /* idt vector in some cases */
180 };
181
182 struct mshv_girq_routing_table {
183 u32 num_rt_entries;
184 struct mshv_guest_irq_ent mshv_girq_info_tbl[];
185 };
186
187 struct hv_synic_pages {
188 struct hv_message_page *hyp_synic_message_page;
189 struct hv_synic_event_flags_page *synic_event_flags_page;
190 struct hv_synic_event_ring_page *synic_event_ring_page;
191 };
192
193 struct mshv_root {
194 spinlock_t pt_ht_lock;
195 DECLARE_HASHTABLE(pt_htable, MSHV_PARTITIONS_HASH_BITS);
196 struct hv_partition_property_vmm_capabilities vmm_caps;
197 };
198
199 /*
200 * Callback for doorbell events.
201 * NOTE: This is called in interrupt context. Callback
202 * should defer slow and sleeping logic to later.
203 */
204 typedef void (*doorbell_cb_t) (int doorbell_id, void *);
205
206 /*
207 * port table information
208 */
209 struct port_table_info {
210 struct rcu_head portbl_rcu;
211 enum hv_port_type hv_port_type;
212 union {
213 struct {
214 u64 reserved[2];
215 } hv_port_message;
216 struct {
217 u64 reserved[2];
218 } hv_port_event;
219 struct {
220 u64 reserved[2];
221 } hv_port_monitor;
222 struct {
223 doorbell_cb_t doorbell_cb;
224 void *data;
225 } hv_port_doorbell;
226 };
227 };
228
229 int mshv_update_routing_table(struct mshv_partition *partition,
230 const struct mshv_user_irq_entry *entries,
231 unsigned int numents);
232 void mshv_free_routing_table(struct mshv_partition *partition);
233
234 struct mshv_guest_irq_ent mshv_ret_girq_entry(struct mshv_partition *partition,
235 u32 irq_num);
236
237 void mshv_copy_girq_info(struct mshv_guest_irq_ent *src_irq,
238 struct mshv_lapic_irq *dest_irq);
239
240 void mshv_irqfd_routing_update(struct mshv_partition *partition);
241
242 void mshv_port_table_fini(void);
243 int mshv_portid_alloc(struct port_table_info *info);
244 int mshv_portid_lookup(int port_id, struct port_table_info *info);
245 void mshv_portid_free(int port_id);
246
247 int mshv_register_doorbell(u64 partition_id, doorbell_cb_t doorbell_cb,
248 void *data, u64 gpa, u64 val, u64 flags);
249 void mshv_unregister_doorbell(u64 partition_id, int doorbell_portid);
250
251 void mshv_isr(void);
252 int mshv_synic_init(struct device *dev);
253 void mshv_synic_exit(void);
254
mshv_partition_encrypted(struct mshv_partition * partition)255 static inline bool mshv_partition_encrypted(struct mshv_partition *partition)
256 {
257 return partition->isolation_type == HV_PARTITION_ISOLATION_TYPE_SNP;
258 }
259
260 struct mshv_partition *mshv_partition_get(struct mshv_partition *partition);
261 void mshv_partition_put(struct mshv_partition *partition);
262 struct mshv_partition *mshv_partition_find(u64 partition_id) __must_hold(RCU);
263
is_l1vh_parent(u64 partition_id)264 static inline bool is_l1vh_parent(u64 partition_id)
265 {
266 return hv_l1vh_partition() && (partition_id == HV_PARTITION_ID_SELF);
267 }
268
269 int mshv_vp_stats_map(u64 partition_id, u32 vp_index,
270 struct hv_stats_page **stats_pages);
271 void mshv_vp_stats_unmap(u64 partition_id, u32 vp_index,
272 struct hv_stats_page **stats_pages);
273
274 /* hypercalls */
275
276 int hv_call_withdraw_memory(u64 count, int node, u64 partition_id);
277 int hv_call_create_partition(u64 flags,
278 struct hv_partition_creation_properties creation_properties,
279 union hv_partition_isolation_properties isolation_properties,
280 u64 *partition_id);
281 int hv_call_initialize_partition(u64 partition_id);
282 int hv_call_finalize_partition(u64 partition_id);
283 int hv_call_delete_partition(u64 partition_id);
284 int hv_call_map_mmio_pages(u64 partition_id, u64 gfn, u64 mmio_spa, u64 numpgs);
285 int hv_call_map_gpa_pages(u64 partition_id, u64 gpa_target, u64 page_count,
286 u32 flags, struct page **pages);
287 int hv_call_unmap_gpa_pages(u64 partition_id, u64 gpa_target, u64 page_count,
288 u32 flags);
289 int hv_call_delete_vp(u64 partition_id, u32 vp_index);
290 int hv_call_assert_virtual_interrupt(u64 partition_id, u32 vector,
291 u64 dest_addr,
292 union hv_interrupt_control control);
293 int hv_call_clear_virtual_interrupt(u64 partition_id);
294 int hv_call_get_gpa_access_states(u64 partition_id, u32 count, u64 gpa_base_pfn,
295 union hv_gpa_page_access_state_flags state_flags,
296 int *written_total,
297 union hv_gpa_page_access_state *states);
298 int hv_call_get_vp_state(u32 vp_index, u64 partition_id,
299 struct hv_vp_state_data state_data,
300 /* Choose between pages and ret_output */
301 u64 page_count, struct page **pages,
302 union hv_output_get_vp_state *ret_output);
303 int hv_call_set_vp_state(u32 vp_index, u64 partition_id,
304 /* Choose between pages and bytes */
305 struct hv_vp_state_data state_data, u64 page_count,
306 struct page **pages, u32 num_bytes, u8 *bytes);
307 int hv_map_vp_state_page(u64 partition_id, u32 vp_index, u32 type,
308 union hv_input_vtl input_vtl,
309 struct page **state_page);
310 int hv_unmap_vp_state_page(u64 partition_id, u32 vp_index, u32 type,
311 struct page *state_page,
312 union hv_input_vtl input_vtl);
313 int hv_call_create_port(u64 port_partition_id, union hv_port_id port_id,
314 u64 connection_partition_id, struct hv_port_info *port_info,
315 u8 port_vtl, u8 min_connection_vtl, int node);
316 int hv_call_delete_port(u64 port_partition_id, union hv_port_id port_id);
317 int hv_call_connect_port(u64 port_partition_id, union hv_port_id port_id,
318 u64 connection_partition_id,
319 union hv_connection_id connection_id,
320 struct hv_connection_info *connection_info,
321 u8 connection_vtl, int node);
322 int hv_call_disconnect_port(u64 connection_partition_id,
323 union hv_connection_id connection_id);
324 int hv_call_notify_port_ring_empty(u32 sint_index);
325 int hv_map_stats_page(enum hv_stats_object_type type,
326 const union hv_stats_object_identity *identity,
327 struct hv_stats_page **addr);
328 int hv_unmap_stats_page(enum hv_stats_object_type type,
329 struct hv_stats_page *page_addr,
330 const union hv_stats_object_identity *identity);
331 int hv_call_modify_spa_host_access(u64 partition_id, struct page **pages,
332 u64 page_struct_count, u32 host_access,
333 u32 flags, u8 acquire);
334 int hv_call_get_partition_property_ex(u64 partition_id, u64 property_code, u64 arg,
335 void *property_value, size_t property_value_sz);
336
337 #if IS_ENABLED(CONFIG_DEBUG_FS)
338 int __init mshv_debugfs_init(void);
339 void mshv_debugfs_exit(void);
340
341 int mshv_debugfs_partition_create(struct mshv_partition *partition);
342 void mshv_debugfs_partition_remove(struct mshv_partition *partition);
343 int mshv_debugfs_vp_create(struct mshv_vp *vp);
344 void mshv_debugfs_vp_remove(struct mshv_vp *vp);
345 #else
mshv_debugfs_init(void)346 static inline int __init mshv_debugfs_init(void)
347 {
348 return 0;
349 }
mshv_debugfs_exit(void)350 static inline void mshv_debugfs_exit(void) { }
351
mshv_debugfs_partition_create(struct mshv_partition * partition)352 static inline int mshv_debugfs_partition_create(struct mshv_partition *partition)
353 {
354 return 0;
355 }
mshv_debugfs_partition_remove(struct mshv_partition * partition)356 static inline void mshv_debugfs_partition_remove(struct mshv_partition *partition) { }
mshv_debugfs_vp_create(struct mshv_vp * vp)357 static inline int mshv_debugfs_vp_create(struct mshv_vp *vp)
358 {
359 return 0;
360 }
mshv_debugfs_vp_remove(struct mshv_vp * vp)361 static inline void mshv_debugfs_vp_remove(struct mshv_vp *vp) { }
362 #endif
363
364 extern struct mshv_root mshv_root;
365 extern enum hv_scheduler_type hv_scheduler_type;
366 extern u8 * __percpu *hv_synic_eventring_tail;
367
368 struct mshv_mem_region *mshv_region_create(u64 guest_pfn, u64 nr_pages,
369 u64 uaddr, u32 flags);
370 int mshv_region_share(struct mshv_mem_region *region);
371 int mshv_region_unshare(struct mshv_mem_region *region);
372 int mshv_region_map(struct mshv_mem_region *region);
373 void mshv_region_invalidate(struct mshv_mem_region *region);
374 int mshv_region_pin(struct mshv_mem_region *region);
375 void mshv_region_put(struct mshv_mem_region *region);
376 int mshv_region_get(struct mshv_mem_region *region);
377 bool mshv_region_handle_gfn_fault(struct mshv_mem_region *region, u64 gfn);
378 void mshv_region_movable_fini(struct mshv_mem_region *region);
379 bool mshv_region_movable_init(struct mshv_mem_region *region);
380
381 #endif /* _MSHV_ROOT_H_ */
382