hyperlight-host 0.19.0

A lightweight Virtual Machine Manager that can be hosted in an application to safely run untrusted or code within a VM partition with very low latency and overhead.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
// SPDX-License-Identifier: Apache-2.0
// Copyright 2026 The Hyperlight Authors.

//! Snapshot memory model:
//!
//! ```text
//! Snapshot
//! `-- memory: Arc<SnapshotMemory>
//!     |-- layers: Box<[Arc<SnapshotBlob>]>
//!     |   |-- SnapshotBlob 0
//!     |   |   |-- memory: ReadonlySharedMemory   [data | page tables]
//!     |   |   |-- gpa_start: u64
//!     |   |   `-- page_table_len: usize
//!     |   `-- SnapshotBlob 1
//!     |       `-- ...
//!     `-- page_tables: Arc<SnapshotBlob>
//! ```

use std::ops::Range;
use std::sync::Arc;

use hyperlight_common::vmem::PAGE_SIZE;

use crate::mem::layout::SandboxMemoryLayout;
use crate::mem::shared_mem::{ReadonlySharedMemory, SharedMemory};
use crate::{Result, new_error};

mod backing;
pub(crate) use backing::SnapshotMemoryBacking;

// Arbitrary cap on the layers of one snapshot. Each layer is one mapping, which
// costs a map/unmap when a sandbox switches snapshots, so this bounds restore
// latency.
pub(crate) const MAX_SNAPSHOT_LAYERS: usize = 30;

/// Memory state assembled from immutable snapshot layers.
///
/// Layer ranges never overlap.
#[derive(Debug)]
pub(crate) struct SnapshotMemory {
    /// Blobs with guest data, sorted by start GPA, which lookups binary
    /// search.
    layers: Box<[Arc<SnapshotBlob>]>,
    /// The blob whose page tables this snapshot restores into scratch memory.
    /// It is one of `layers` when it also holds guest data.
    page_tables: Arc<SnapshotBlob>,
}

/// Immutable memory one snapshot creates: guest data followed by the page
/// tables of that snapshot. The VM maps the data whole at `gpa_start`,
/// including pages a later layer replaced. The page tables are never mapped.
/// Snapshots that reuse the data share the blob.
#[derive(Debug)]
pub(crate) struct SnapshotBlob {
    /// Its `guest_mapped_size` is the data length.
    memory: ReadonlySharedMemory,
    gpa_start: u64,
    page_table_len: usize,
}

impl SnapshotBlob {
    /// `gpa_start` is unused when the blob has no data.
    pub(crate) fn new(
        memory: ReadonlySharedMemory,
        gpa_start: u64,
        page_table_len: usize,
        scratch_base_gpa: u64,
    ) -> Result<Self> {
        let len = memory.mem_size();
        let data_len = memory.guest_mapped_size();
        let expected_len = data_len
            .checked_add(page_table_len)
            .and_then(|len| len.checked_next_multiple_of(host_page_size()));
        if !is_page_aligned(data_len as u64)
            || !page_table_len.is_multiple_of(PAGE_SIZE)
            || len == 0
            || expected_len != Some(len)
        {
            return Err(new_error!("snapshot layer size is invalid"));
        }
        if data_len > 0 {
            if !is_page_aligned(gpa_start) {
                return Err(new_error!("snapshot layer GPA is not aligned"));
            }
            if gpa_start < SandboxMemoryLayout::BASE_ADDRESS as u64 {
                return Err(new_error!("snapshot layer starts below BASE_ADDRESS"));
            }
            let gpa_end = gpa_start
                .checked_add(data_len as u64)
                .ok_or_else(|| new_error!("snapshot layer GPA range overflows"))?;
            if gpa_end > scratch_base_gpa {
                return Err(new_error!("snapshot layer data overlaps scratch memory"));
            }
        }
        Ok(Self {
            memory,
            gpa_start,
            page_table_len,
        })
    }

    pub(crate) fn memory(&self) -> &ReadonlySharedMemory {
        &self.memory
    }

    pub(crate) fn gpa_start(&self) -> u64 {
        self.gpa_start
    }

    /// Length of the guest data, which the VM maps.
    pub(crate) fn data_len(&self) -> usize {
        self.memory.guest_mapped_size()
    }

    pub(crate) fn page_table_len(&self) -> usize {
        self.page_table_len
    }

    pub(crate) fn gpa_range(&self) -> Range<u64> {
        self.gpa_start..self.gpa_start + self.data_len() as u64
    }

    fn page_table_bytes(&self) -> &[u8] {
        let data_len = self.data_len();
        &self.memory.as_slice()[data_len..data_len + self.page_table_len]
    }
}

impl SnapshotMemory {
    pub(crate) fn new(
        mut layers: Box<[Arc<SnapshotBlob>]>,
        page_tables: Arc<SnapshotBlob>,
    ) -> Result<Self> {
        if layers.is_empty() {
            return Err(new_error!("snapshot has no layers"));
        }
        if layers.len() > MAX_SNAPSHOT_LAYERS {
            return Err(new_error!(
                "snapshot layer count {} exceeds {}",
                layers.len(),
                MAX_SNAPSHOT_LAYERS
            ));
        }
        if layers.iter().any(|layer| layer.data_len() == 0) {
            return Err(new_error!("snapshot layer has no data"));
        }
        layers.sort_by_key(|layer| layer.gpa_start);
        // Layers sit between BASE_ADDRESS and scratch, so disjoint layers also
        // bound the total layer size. Overlap also rejects a blob referenced
        // twice.
        if layers
            .windows(2)
            .any(|pair| pair[0].gpa_range().end > pair[1].gpa_start)
        {
            return Err(new_error!("snapshot blob GPA ranges overlap"));
        }
        if page_tables.page_table_len == 0 {
            return Err(new_error!("snapshot page tables are empty"));
        }
        if page_tables.data_len() > 0
            && !layers.iter().any(|layer| Arc::ptr_eq(layer, &page_tables))
        {
            return Err(new_error!("snapshot page-table blob data is not a layer"));
        }

        Ok(Self {
            layers,
            page_tables,
        })
    }

    /// Builds a one-layer snapshot from `memory`, which holds guest data
    /// followed by `page_table_len` bytes of page tables.
    pub(crate) fn from_flat(
        memory: ReadonlySharedMemory,
        gpa_start: u64,
        page_table_len: usize,
        scratch_base_gpa: u64,
    ) -> Result<Self> {
        let blob = Arc::new(SnapshotBlob::new(
            memory,
            gpa_start,
            page_table_len,
            scratch_base_gpa,
        )?);
        Self::new(Box::new([blob.clone()]), blob)
    }

    pub(crate) fn layers(&self) -> &[Arc<SnapshotBlob>] {
        &self.layers
    }

    /// Every blob this snapshot holds: the layers, then the page-table blob
    /// when it holds no data.
    pub(crate) fn blobs(&self) -> impl Iterator<Item = &Arc<SnapshotBlob>> {
        let page_table_only = (self.page_tables.data_len() == 0).then_some(&self.page_tables);
        self.layers.iter().chain(page_table_only)
    }

    pub(crate) fn page_table_blob(&self) -> &Arc<SnapshotBlob> {
        &self.page_tables
    }

    pub(crate) fn page_table_bytes(&self) -> &[u8] {
        self.page_tables.page_table_bytes()
    }

    pub(crate) fn gpa_span_len(&self) -> usize {
        // Sorted, disjoint layers put the highest end in the last one.
        self.layers.last().map_or(0, |last| {
            // Blobs end below the scratch base, so the offset fits in usize.
            (last.gpa_start - SandboxMemoryLayout::BASE_ADDRESS as u64) as usize + last.data_len()
        })
    }

    pub(crate) fn page_table_len(&self) -> usize {
        self.page_tables.page_table_len
    }

    pub(crate) fn resolve(&self, gpa: u64, len: usize) -> Option<(usize, usize)> {
        let (index, offset, available) = self.resolve_chunk(gpa)?;
        (len <= available).then_some((index, offset))
    }

    /// Returns the layer index, the offset into its blob, and the data bytes
    /// left in the blob from `gpa`.
    pub(crate) fn resolve_chunk(&self, gpa: u64) -> Option<(usize, usize, usize)> {
        let index = self
            .layers
            .partition_point(|layer| layer.gpa_start <= gpa)
            .checked_sub(1)?;
        let layer = &self.layers[index];
        let offset = usize::try_from(gpa - layer.gpa_start).ok()?;
        let data_len = layer.data_len();
        (offset < data_len).then(|| (index, offset, data_len - offset))
    }

    pub(crate) fn read_page_tables(
        &self,
        pt_gpa_base: u64,
        gpa: u64,
        destination: &mut [u8],
    ) -> Result<()> {
        let source = gpa
            .checked_sub(pt_gpa_base)
            .and_then(|offset| usize::try_from(offset).ok())
            .and_then(|offset| {
                self.page_table_bytes()
                    .get(offset..offset.checked_add(destination.len())?)
            })
            .ok_or_else(|| new_error!("page-table read range is out of bounds"))?;
        destination.copy_from_slice(source);
        Ok(())
    }
}

pub(crate) fn host_page_size() -> usize {
    PAGE_SIZE.max(page_size::get())
}

/// Guest and host pages are powers of two, so the larger one implies both.
fn is_page_aligned(value: u64) -> bool {
    value.is_multiple_of(host_page_size() as u64)
}

#[cfg(test)]
pub(crate) mod tests {
    use super::*;
    use crate::mem::shared_mem::ExclusiveSharedMemory;

    const SCRATCH_BASE_GPA: u64 = 0x1_0000_0000;

    /// Copies guest data and page tables into one blob. The data is padded to a
    /// host page.
    pub(crate) fn blob_bytes(data: &[u8], page_tables: &[u8]) -> Result<ReadonlySharedMemory> {
        let data_len = data.len().next_multiple_of(host_page_size());
        let mut bytes = vec![0u8; data_len + page_tables.len()];
        bytes[..data.len()].copy_from_slice(data);
        bytes[data_len..].copy_from_slice(page_tables);
        Ok(ReadonlySharedMemory::from_bytes(&bytes, data_len)?)
    }

    pub(crate) fn flat_memory(
        data: &[u8],
        page_tables: &[u8],
        gpa_start: u64,
        scratch_base_gpa: u64,
    ) -> SnapshotMemory {
        SnapshotMemory::from_flat(
            blob_bytes(data, page_tables).unwrap(),
            gpa_start,
            page_tables.len(),
            scratch_base_gpa,
        )
        .unwrap()
    }

    pub(crate) fn page_tables_from_bytes(bytes: &[u8]) -> Arc<SnapshotBlob> {
        Arc::new(SnapshotBlob::new(blob_bytes(&[], bytes).unwrap(), 0, bytes.len(), 0).unwrap())
    }

    pub(crate) fn stub_page_tables() -> Arc<SnapshotBlob> {
        page_tables_from_bytes(&vec![0u8; PAGE_SIZE])
    }

    fn blob(data_pages: usize, gpa_start: u64) -> Arc<SnapshotBlob> {
        let data_len = data_pages * page_size::get();
        let storage = ExclusiveSharedMemory::new(data_len)
            .unwrap()
            .freeze(data_len)
            .unwrap();
        Arc::new(SnapshotBlob::new(storage, gpa_start, 0, SCRATCH_BASE_GPA).unwrap())
    }

    #[test]
    fn blob_rejects_overflowing_gpa_range() {
        let host_page_size = page_size::get();
        let storage = ExclusiveSharedMemory::new(host_page_size)
            .unwrap()
            .freeze(host_page_size)
            .unwrap();
        let gpa_start = u64::MAX - host_page_size as u64 + 1;

        assert!(SnapshotBlob::new(storage, gpa_start, 0, u64::MAX).is_err());
    }

    #[test]
    fn blob_rejects_sizes_that_do_not_fill_it() {
        let host_page_size = page_size::get();
        let base = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let storage = |data_len| {
            ExclusiveSharedMemory::new(2 * host_page_size)
                .unwrap()
                .freeze(data_len)
                .unwrap()
        };

        assert!(SnapshotBlob::new(storage(host_page_size), base, 0, SCRATCH_BASE_GPA).is_err());
        assert!(SnapshotBlob::new(storage(0), base, 0, SCRATCH_BASE_GPA).is_err());
        assert!(SnapshotBlob::new(storage(host_page_size), base, 1, SCRATCH_BASE_GPA).is_err());
        SnapshotBlob::new(
            storage(host_page_size),
            base,
            host_page_size,
            SCRATCH_BASE_GPA,
        )
        .unwrap();
    }

    #[test]
    fn page_tables_follow_the_data_and_stay_unmapped() {
        let host_page_size = page_size::get();
        let base = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let memory = SnapshotMemory::from_flat(
            blob_bytes(&vec![0x11; host_page_size], &vec![0x22; PAGE_SIZE]).unwrap(),
            base,
            PAGE_SIZE,
            SCRATCH_BASE_GPA,
        )
        .unwrap();

        assert_eq!(memory.page_table_bytes(), vec![0x22; PAGE_SIZE]);
        assert_eq!(memory.gpa_span_len(), host_page_size);
        assert_eq!(memory.resolve(base + host_page_size as u64, 1), None);
        assert_eq!(memory.blobs().count(), 1);
    }

    #[test]
    fn memory_rejects_page_table_blob_data_outside_the_layers() {
        let host_page_size = page_size::get();
        let base = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let page_tables = Arc::new(
            SnapshotBlob::new(
                blob_bytes(&vec![0; host_page_size], &vec![0; PAGE_SIZE]).unwrap(),
                base + host_page_size as u64,
                PAGE_SIZE,
                SCRATCH_BASE_GPA,
            )
            .unwrap(),
        );

        assert!(SnapshotMemory::new(Box::new([blob(1, base)]), page_tables.clone()).is_err());
        SnapshotMemory::new(Box::new([blob(1, base), page_tables.clone()]), page_tables).unwrap();
    }

    #[test]
    fn reads_page_tables_from_page_table_only_blob() {
        let base = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let page_tables = page_tables_from_bytes(&vec![0x33; PAGE_SIZE]);
        let memory = SnapshotMemory::new(Box::new([blob(1, base)]), page_tables).unwrap();
        let mut byte = [0u8; 1];

        memory.read_page_tables(0x8000, 0x8fff, &mut byte).unwrap();

        assert_eq!(byte, [0x33]);
        assert_eq!(memory.resolve(base, PAGE_SIZE), Some((0, 0)));
    }

    #[test]
    fn memory_rejects_overlapping_blob_gpas() {
        let host_page_size = page_size::get();
        let gpa = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let layers = Box::new([blob(2, gpa), blob(1, gpa + host_page_size as u64)]);

        assert!(SnapshotMemory::new(layers, stub_page_tables()).is_err());
    }

    #[test]
    fn memory_canonicalizes_layers() {
        let host_page_size = page_size::get();
        let base = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let layers = Box::new([blob(1, base + host_page_size as u64), blob(1, base)]);

        let memory = SnapshotMemory::new(layers, stub_page_tables()).unwrap();

        assert_eq!(memory.layers()[0].gpa_start(), base);
    }

    #[test]
    fn resolve_covers_whole_blob_and_rejects_past_end() {
        let host_page_size = page_size::get();
        let gpa = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let memory = SnapshotMemory::new(Box::new([blob(3, gpa)]), stub_page_tables()).unwrap();

        assert_eq!(memory.resolve(gpa, 1), Some((0, 0)));
        assert_eq!(
            memory.resolve(gpa + host_page_size as u64, host_page_size),
            Some((0, host_page_size))
        );
        assert_eq!(
            memory.resolve(gpa + (3 * host_page_size - 1) as u64, 1),
            Some((0, 3 * host_page_size - 1))
        );
        assert_eq!(
            memory.resolve(gpa + (3 * host_page_size - 1) as u64, 2),
            None
        );
        assert_eq!(memory.resolve(gpa - 1, 1), None);
    }

    #[test]
    fn reads_reject_addresses_past_the_blob() {
        let host_page_size = page_size::get();
        let gpa = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let memory = SnapshotMemory::new(Box::new([blob(3, gpa)]), stub_page_tables()).unwrap();
        let (backing, _) = SnapshotMemoryBacking::from_snapshot(Arc::new(memory))
            .unwrap()
            .build();
        let mut byte = [0u8; 1];

        backing
            .read_snapshot_gpa(gpa + host_page_size as u64 + 1, &mut byte)
            .unwrap();
        assert!(
            backing
                .read_snapshot_gpa(gpa + 3 * host_page_size as u64 + 1, &mut byte)
                .is_err()
        );
    }

    #[test]
    fn memory_rejects_excess_layers() {
        let host_page_size = page_size::get() as u64;
        let base = SandboxMemoryLayout::BASE_ADDRESS as u64;
        let layers = (0..=MAX_SNAPSHOT_LAYERS as u64)
            .map(|index| blob(1, base + index * host_page_size))
            .collect();

        assert!(SnapshotMemory::new(layers, stub_page_tables()).is_err());
    }
}