Fix pty test in vscode integrated terminal

CHANGELOG.md

@@ -4,6 +4,30 @@ All notable changes to virtkit will be documented in this file.
 
 ## [Unreleased]
 
+### Added
+
+- **DAX maps only the files worth mapping.** A share's window now serves regular files of
+  1M and more by default (`dax=inode`); smaller files read through the guest page cache as
+  without DAX, so a source tree no longer spends the window's overhead on files too small to
+  benefit. `vk run --dax`, `x-virtkit.dax` and `[executor.vm] dax` take `<size>:always` for
+  the previous behaviour and `<size>:inode=<min>` for another floor.
+
+### Changed
+
+- **Reading a large tree from a read-only share is faster.** A guest that reads many small
+  files — from an `:overlay` volume or the executor's host checkout — no longer pays a
+  per-file host cost that could dominate file-heavy jobs.
+
+### Fixed
+
+- **A stalled network transfer inside a guest now fails fast instead of hanging for the rest
+  of the job.** When the host was briefly overloaded a guest download could wedge and never
+  recover — a `cargo` fetch stuck at 0 bytes was the usual symptom. It now errors out
+  promptly and the application reconnects.
+- **`host_checkout` jobs no longer stall on the first git command inside the guest.** The
+  job's first `git status` or `git checkout` now runs at full speed instead of stalling once
+  per job.
+
 ## [0.68.0] - 2026-09-11
 
 ### Changed

Cargo.toml

@@ -16,7 +16,7 @@ members = [
 #
 # third_party/imago is a single vendored leaf crate (see third_party/imago/VENDOR.md), a
 # dependency of third_party/libkrun's block device — excluded for the same reason.
-exclude = ["third_party/libkrun", "third_party/imago"]
+exclude = ["third_party/libkrun", "third_party/imago", "third_party/ipstack"]
 
 [workspace.package]
 version = "0.68.0"
@@ -161,3 +161,9 @@ indicatif = "0.18"
 unicode-width = "0.2"
 # Use indicatif's transitive ANSI classifier in dashboard layout tests.
 console = { version = "0.16", default-features = false, features = ["ansi-parsing"] }
+
+# Vendored ipstack (third_party/ipstack/VENDOR.md): 1.0.1 plus one local patch — a TCP
+# connection whose retransmissions are exhausted is reset instead of left Established with
+# a hole in its stream.
+[patch.crates-io]
+ipstack = { path = "third_party/ipstack" }

README.md

@@ -283,9 +283,13 @@ in per-fault latency.
 
 The window reserves address space, not memory, and costs nothing until mapped. It defaults
 to 8G per share; `vk run --dax`, a service's `x-virtkit.dax` and the executor's `[executor.vm] dax`
-resize it or turn it `off`. Each guest supports 64G of windows — eight at the default size,
-with further shares served without DAX. Guests with more than 63.25G of RAM have no room
-for windows and receive none. DAX requires the built-in VMM. Under
+resize it or turn it `off`. Each mapping costs the host an mmap and the guest an EPT
+invalidation per 2 MiB range whatever the file's size, so by default only regular files of
+1M and more go through the window (`dax=inode`; the host marks them) and smaller files read
+through the guest's page cache as without DAX; `<size>:always` maps every file, and
+`<size>:inode=<min>` moves the floor. Each guest supports 64G of windows — eight at the
+default size, with further shares served without DAX. Guests with more than 63.25G of RAM
+have no room for windows and receive none. DAX requires the built-in VMM. Under
 `VIRTKIT_VMM=cloud-hypervisor`, and for single-file binds and `vk build` stage guests,
 shares are served the ordinary way.
 

docs/gitlab-ci.md

@@ -629,14 +629,20 @@ copy: ten concurrent job VMs reading a tools tree held eleven copies on the host
 
 ```toml
 [executor.vm]
-dax = "8G"            # per share; the default
+dax = "8G"            # per share; the default: files of 1M and more through the window
+# dax = "8G:always"   # every file through the window
+# dax = "8G:inode=64K" # another size floor
 ```
 
 The window reserves guest address space, not memory, and costs nothing until mapped.
-`"off"` disables it. Each guest supports 64G of windows — eight at the default size, with
-further shares served without DAX. Guests with more than 63.25G of RAM have no room for
-windows and receive none. Mappings are 4 KiB-granular, so the benefit is avoiding a tools
-tree's memory cost per VM, not per-fault latency. Compose services inherit the job VM's
+`"off"` disables it. Each mapping costs the host an mmap and the guest an EPT invalidation
+per 2 MiB range whatever the file's size, which a source tree's small files never repay:
+by default only regular files of 1M and more are mapped (`dax=inode`; the host marks them),
+the rest read through the guest page cache as without DAX. Each guest supports 64G of
+windows — eight at the default size, with further shares served without DAX. Guests with
+more than 63.25G of RAM have no room for windows and receive none.
+Mappings are 4 KiB-granular, so the benefit is avoiding a tools tree's memory cost per
+VM, not per-fault latency. Compose services inherit the job VM's
 setting unless they declare `x-virtkit.dax`. DAX requires the built-in VMM;
 cloud-hypervisor has no DAX path and serves shares the ordinary way whatever this says.