Properly CoW page tables to the scratch region #1198

syntactically · 2026-01-28T02:36:10Z

This combines a few commits from the original guest-CoW PR, all of which involve properly CoW'ing the page tables into the scratch region, dealing with this in snapshots, etc.

Previously, this was inconsistent. Signed-off-by: Lucy Menon <[email protected]>

Signed-off-by: Lucy Menon <[email protected]>

whp can now map any region as long as it contains a file handle (which, presently, only snapshot and scratch regions do---not user-provided regions). Signed-off-by: Lucy Menon <[email protected]>

This fixes an error introduced by a merge conflict in a previous PR. Signed-off-by: Lucy Menon <[email protected]>

Previously, in 7ca0a78, MAP_SHARED was replaced with MAP_PRIVATE, primarily to bring the Miri and non-Miri code paths closer together, since it did not appear to have any detrimental effect. However, this seems to cause issues with the scratch mapping on mshv on Linux. It is not entirely clear what mshv is doing here; possibly something like "mapping a page which has never been written to" causes the problem. Signed-off-by: Lucy Menon <[email protected]>

Signed-off-by: Lucy Menon <[email protected]>

As per the comments preexisting in src/hyperlight_common/src/layout.rs. This avoids some fairly cumbersome proliferation of #[cfg(feature = "init-paging")], and may be developed further in the future. Signed-off-by: Lucy Menon <[email protected]>

Instead of updating page tables in place, and using the heap region memory for new page tables, this changes the hyperlight_guest virtual memory subsystem to treat snapshot page tables as copy-on-write and to allocate all new page tables in the scratch region. Signed-off-by: Lucy Menon <[email protected]>

These were already being CoW'd. Signed-off-by: Lucy Menon <[email protected]>

Now that some important things are being written to the scratch region, this makes the snapshot process aware of it. In particular, rather than just copying the current snapshot region into a new snapshot, the snapshot process now traverses the guest page tables and picks all pages that are mapped into guest virtual memory, placing them into the new snapshot along with new page tables which account for all physical memory having moved to the snapshot region. Several tests are temporarily disabled in this commit. One, `snapshot_restore_handles_remapping_correctly`, will be re-enabled in the next commit; its failure is due to an issue with state in the guest page mapping process not being reset on snapshot. Several others, related to the stack guard page, are also disabled, because the snapshot process loses the context, kept in the host, of which physical page is the stack guard page. These will be re-enabled in an upcoming patch series, which moves stack management fully into the guest. This commit also disables an optimisation around restoring to a just-taken snapshot in certain cases; see the comment at the head of `MultiUseSandbox::restore` for details. Note that the optimisation of taking a new snapshot immediately after a restore remains intact. Signed-off-by: Lucy Menon <[email protected]>

Previously, the guest tried to infer the addresses of the snapshot page tables from the values of cr3 that it saw at various times, but this was brittle and irritating to do without writing to the wrong address at times. This replaces that mechanism with an extra word in the metadata portion of the scratch region that explicitly provides this information. Signed-off-by: Lucy Menon <[email protected]>

src/hyperlight_common/src/layout.rs

src/hyperlight_host/src/mem/mgr.rs

Remove the unused constant `SCRATCH_TOP_USED_OFFSET`. Signed-off-by: Lucy Menon <[email protected]>

src/hyperlight_common/src/arch/amd64/vmem.rs

andreiltd

Great improvement! I'm looking forward to removing more special regions from shared memory. Having an extra pair of eyes on the changes would be great also.

Justfile

src/hyperlight_host/src/sandbox/snapshot.rs

dblnz

Great work! I left some small comments

dblnz · 2026-01-29T09:23:47Z

src/hyperlight_common/src/vmem.rs

-pub use arch::{PAGE_SIZE, PAGE_TABLE_SIZE, PageTableEntry, PhysAddr, VirtAddr};
+/// This is always the page size that the /guest/ is being compiled
+/// for, which may or may not be the same as the host page size.
+pub use arch::PAGE_SIZE;


What happens if the page size differs? Does the host signal it somehow?

The goal is to use each of these in the correct places (although I'm not sure whether or not the host page size is important to the guest---maybe in the future as a hint for zerocopy buffer alignment?), but as with some of the other portability stuff that's incidental to this series it's more aspirational/laying the groundwork for the future than "all issues related to this fixed".

src/hyperlight_guest/src/arch/amd64/layout.rs

src/hyperlight_host/src/hypervisor/virtual_machine/mod.rs

dblnz · 2026-01-29T09:55:13Z

src/hyperlight_host/src/hypervisor/hyperlight_vm.rs

-            regions.extend(self.get_mapped_regions().cloned());
+            // Include dynamically mapped regions
+            // TODO: include the snapshot and scratch regions
+            let regions: Vec<MemoryRegion> = self.get_mapped_regions().cloned().collect();


So for the crashdump to work, we need to generate the MemoryRegions for snapshot and scratch.
We'd have to walk the page tables for that, right?

I'm not sure whether you would want the crash dump to be virtually indexed or physically indexed. If the former, yeah, this should be replaced by code that does the same page table walk as snapshot.rs does right now, I think.

dblnz · 2026-01-29T10:02:12Z

src/hyperlight_host/src/mem/mgr.rs

+        #[allow(clippy::unwrap_used)]
+        self.scratch_mem
+            .write::<u64>(size_offset, scratch_size as u64)
+            .unwrap();


Maybe just return an error and Poison the sandbox?

I tend to agree that we should not crash the host if we can absolutely avoid it, maybe we need a new error type besides Poisoned that means something along the lines of "something very bad happened and you should consider restarting your process because we'd really like to panic here"

I don't think I agree about this. Things that are a clearly impossible should panic because who knows what is wrong / whether something else in either hyperlight or the host has been corrupted. See also this thread on the previous PR: #1182 (comment).

Even in the case where this should never happen, I get nervous about being in the failure path as a library that could cause down time. If we poison the sandbox and return an error that is specific to things that should not happen then we provide the consumer the ability to handle it fit to their particular needs

I think this PR is not the right place to have the discussion about panicking on invariant violations in general, so while I still maintain my position that we should, I have for now added fallibility to this function in 54cd065.

A close inspection of the code has however convinced me that there is one caller (SandboxMemoryManager<ExclusiveSharedMemory>::build) for which panicking (at least from our code) is more completely impossible. I have added a slightly more detailed comment to that function arguing why it is morally infallible, but have added a Result type to its return for now anyway in the interests of expediency.

I'm on Lucy's side on this one in general. The tricky part here is that one may modify write function in the future without realizing it is breaking upstream assumption. If we decide to keep an error I think having debug_assert is a good way to document violation of invariant.

src/hyperlight_host/src/sandbox/snapshot.rs

dblnz · 2026-01-29T10:20:25Z

src/hyperlight_host/src/sandbox/snapshot.rs

+    }
+    let (mem, off) = access_gpa(snap, scratch, scratch_size, gpa);
+    if off + PAGE_SIZE <= mem.as_slice().len() {
+        Some(&mem.as_slice()[off..off + PAGE_SIZE])


Can the observation in the comment above happen here also?

Not sure which comment above you are referring to---if it's the one about page-unaligned accesses reading beyond memory, then I don't think so since the if condition checks for that here?

simongdavies

A fwe questions/comments, I think there are a few places where there are types that are pub that could be pub (crate) , didn't look for them all, but possibly worth a pass of the whole PR for this.

src/hyperlight_common/src/arch/amd64/vmem.rs

src/hyperlight_host/src/sandbox/snapshot.rs

simongdavies · 2026-01-29T10:07:40Z

src/hyperlight_guest/src/arch/amd64/layout.rs

+
+pub const MAIN_STACK_TOP_GVA: usize = 0xffff_feff_ffff_f000;
+
+pub fn scratch_size() -> u64 {


should this and the following functions be pub(crate) ?

The following two functions are definitely used in hyperlight_guest_bin. This one isn't currently but I see little downside to making it match.

src/hyperlight_guest/src/prim_alloc.rs

src/hyperlight_host/src/mem/mgr.rs

simongdavies · 2026-01-29T11:20:21Z

src/hyperlight_host/src/mem/mgr.rs

+        #[allow(clippy::unwrap_used)]
+        self.scratch_mem
+            .write::<u64>(size_offset, scratch_size as u64)
+            .unwrap();


I tend to agree that we should not crash the host if we can absolutely avoid it, maybe we need a new error type besides Poisoned that means something along the lines of "something very bad happened and you should consider restarting your process because we'd really like to panic here"

src/hyperlight_host/src/sandbox/initialized_multi_use.rs

simongdavies · 2026-01-29T11:28:39Z

src/hyperlight_host/src/sandbox/snapshot.rs

+    // impossible unless an extremely significant invariant violation
+    // has occurred.
+    #[allow(clippy::panic)]
+    unsafe fn alloc_table(&self) -> u64 {


Can ew split TableOps into a TableOpsRead and TableOpsWrite to avoid this?

We probably could, but I didn't think it was worth the effort at the moment. It is pretty clear semantically that these implementations could not possibly be r/w and I can't see any natural way in which host code would want to use them that way.

I would prefer compile time checks to make this impossible but OK.

Is there a downside to splitting the trait into read and write?

No, and I agree that it is the correct thing to do, I just wasn't sure if it was worth the hassle. Since it seems like there is more interest in getting right, I went ahead and did it in a2ece56.

simongdavies · 2026-01-29T11:29:57Z

src/hyperlight_host/src/sandbox/snapshot.rs

+    // should possibly ever call an update_parent.  Therefore, this is
+    // impossible unless an extremely significant invariant violation
+    // has occurred.
+    #[allow(clippy::panic)]


As above can we split up the TableOps trait into read/write and then avoid this entirely?

Make virt_to_phys use the aligned and un-canonicalised version of its argument for traversal. This does not actually matter, because only certain bits are actually taken from vmin by the (chained) `modify_ptes` iterator(s). However, since it doesn't matter, this is probably better for clarity. Signed-off-by: Lucy Menon <[email protected]>

Fix bug that resulted in the first page of scratch memory being unconditionally copied into the snapshot. Signed-off-by: Lucy Menon <[email protected]>

Be slightly more careful about handling offsets from the guest in snapshot.rs: previously, a maliciously-constructed sandbox could possibly cause underflow panics, and another function could be misused in a way that would result in a slice out-of-bounds panic (although it was not currently being misused in such a way). Signed-off-by: Lucy Menon <[email protected]>

syntactically · 2026-01-29T15:20:37Z

Thanks for the comments everyone! I have added three fixups that address some of them. Sorry for the commit message summaries---the first line is used by autosquash so has to just be "which commit this is amending", although I did add a few words tying each of them to an issue identified in the review.

I think there are a few places where there are types that are pub that could be pub (crate) , didn't look for them all, but possibly worth a pass of the whole PR for this.

I did a quick grep for added lines starting with pub and pretty much everything I saw was guest library exports for memory layout or paging, which I think will be useful in other guests (e.g. hyperlight-wasm, which has its own page-table manipulations to implement).

ludfjig

Looks ok to me :D

ludfjig · 2026-01-29T19:08:33Z

src/hyperlight_host/src/sandbox/initialized_multi_use.rs

-        if let Some(snap) = &self.snapshot
-            && snap.as_ref() == snapshot.as_ref()
-        {
-            // If the snapshot is already the current one, no need to restore
-            return Ok(());
-        }


Can we remove the self.snapshot or is it used somewhere else? (Maybe in future PR if deemed OK?)

It's still used for the optimisation of taking a snapshot immediately after a restore

…tch region Signed-off-by: Lucy Menon <[email protected]>

This splits `hyperlight_common::vmem::TableOps` into `TableReadOps` and `TableOps`, so that implementations like the one in snapshot.rs that cannot support writes don't have to provide the (extremely partial!) write functions. This is slightly involved, because we need to keep to track of the right information in the type system about whether writes are possible, and, if so, whether they can move existing page tables. The implementation of of writing to cr3 to update page tables is also moved into `hyperlight_guest_bin::paging`, where it belongs, which is a simple change. Signed-off-by: Lucy Menon <[email protected]>

syntactically added 3 commits January 28, 2026 02:28

Update guest Cargo.lock

312862f

Previously, this was inconsistent. Signed-off-by: Lucy Menon <[email protected]>

[nix] include miri for the nightly toolchain

58d5506

Signed-off-by: Lucy Menon <[email protected]>

[nix] hack to make libgit2-sys builds more reliable

0839184

Signed-off-by: Lucy Menon <[email protected]>

syntactically requested review from danbugs, dblnz, devigned, jprendes, ludfjig, marosset and simongdavies as code owners January 28, 2026 02:36

syntactically added kind/enhancement For PRs adding features, improving functionality, docs, tests, etc. Guest-COW PRs that form part of the Guest-COW change labels Jan 28, 2026

syntactically mentioned this pull request Jan 28, 2026

[DO NOT MERGE] Implement Copy On Write in the hyperlight guest #1091

Closed

12 tasks

syntactically force-pushed the lm/pt-scratch branch from 08e43a5 to 7e3f9a5 Compare January 28, 2026 03:04

syntactically added 4 commits January 28, 2026 03:13

[just] Bring just like-ci in line with CI changes

22b433a

Signed-off-by: Lucy Menon <[email protected]>

[kvm] Better document the cpuid leaf used to set GPA width

415823f

Signed-off-by: Lucy Menon <[email protected]>

[whp] Remove initial_memory_setup_done, which is no longer used

665a86f

whp can now map any region as long as it contains a file handle (which, presently, only snapshot and scratch regions do---not user-provided regions). Signed-off-by: Lucy Menon <[email protected]>

Make scratch region writable again

2cc3e55

This fixes an error introduced by a merge conflict in a previous PR. Signed-off-by: Lucy Menon <[email protected]>

syntactically force-pushed the lm/pt-scratch branch 2 times, most recently from 1b10f54 to 22173e0 Compare January 28, 2026 06:53

syntactically added 7 commits January 28, 2026 06:54

Modify guest physical page allocator to allocate from the scratch region

a9355b1

Signed-off-by: Lucy Menon <[email protected]>

i686: add layout, vmem, etc stubs

b6b839d

As per the comments preexisting in src/hyperlight_common/src/layout.rs. This avoids some fairly cumbersome proliferation of #[cfg(feature = "init-paging")], and may be developed further in the future. Signed-off-by: Lucy Menon <[email protected]>

Make the snapshot page table mapping readonly

5bbb393

These were already being CoW'd. Signed-off-by: Lucy Menon <[email protected]>

syntactically force-pushed the lm/pt-scratch branch from 22173e0 to 350a605 Compare January 28, 2026 06:54

andreiltd reviewed Jan 28, 2026

View reviewed changes

src/hyperlight_common/src/layout.rs Show resolved Hide resolved

andreiltd reviewed Jan 28, 2026

View reviewed changes

src/hyperlight_host/src/mem/mgr.rs Outdated Show resolved Hide resolved

fixup! Explicitly specify the physical addresses of snapshot page tables

b1274bd

Remove the unused constant `SCRATCH_TOP_USED_OFFSET`. Signed-off-by: Lucy Menon <[email protected]>

andreiltd reviewed Jan 28, 2026

View reviewed changes

src/hyperlight_common/src/arch/amd64/vmem.rs Outdated Show resolved Hide resolved

andreiltd previously approved these changes Jan 28, 2026

View reviewed changes

jsturtevant mentioned this pull request Jan 28, 2026

Guest Copy on Write #1200

Open

17 tasks

jsturtevant reviewed Jan 28, 2026

View reviewed changes

Justfile Show resolved Hide resolved

jsturtevant reviewed Jan 29, 2026

View reviewed changes

src/hyperlight_host/src/sandbox/snapshot.rs Outdated Show resolved Hide resolved

jsturtevant reviewed Jan 29, 2026

View reviewed changes

src/hyperlight_host/src/sandbox/snapshot.rs Outdated Show resolved Hide resolved

jsturtevant reviewed Jan 29, 2026

View reviewed changes

src/hyperlight_host/src/sandbox/snapshot.rs Show resolved Hide resolved

simongdavies mentioned this pull request Jan 29, 2026

Update CI to use just commands rather than have 2 sets of steps that have to be in sync #1201

Open

dblnz reviewed Jan 29, 2026

View reviewed changes

simongdavies reviewed Jan 29, 2026

View reviewed changes

syntactically dismissed andreiltd’s stale review via 1fa74ff January 29, 2026 15:13

syntactically added 3 commits January 29, 2026 15:13

fixup! snapshot: handle the scratch region correctly

638bcb7

Fix bug that resulted in the first page of scratch memory being unconditionally copied into the snapshot. Signed-off-by: Lucy Menon <[email protected]>

syntactically force-pushed the lm/pt-scratch branch from 1fa74ff to 2a2b444 Compare January 29, 2026 15:13

ludfjig previously approved these changes Jan 29, 2026

View reviewed changes

andreiltd previously approved these changes Jan 29, 2026

View reviewed changes

syntactically added 2 commits January 30, 2026 02:06

fixup! Modify guest physical page allocator to allocate from the scra…

54cd065

…tch region Signed-off-by: Lucy Menon <[email protected]>

syntactically dismissed stale reviews from andreiltd and ludfjig via a2ece56 January 30, 2026 02:12

andreiltd approved these changes Jan 30, 2026

View reviewed changes


		pub const MAIN_STACK_TOP_GVA: usize = 0xffff_feff_ffff_f000;

		pub fn scratch_size() -> u64 {

Properly CoW page tables to the scratch region #1198

Are you sure you want to change the base?

Properly CoW page tables to the scratch region #1198

Uh oh!

Conversation

syntactically commented Jan 28, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

andreiltd left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dblnz left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

simongdavies left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

syntactically commented Jan 29, 2026

Uh oh!

ludfjig left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment