Add support for memory pages compression

.github/workflows/check-commits.yml

@@ -19,7 +19,7 @@ jobs:
         # Checkout pull request HEAD commit instead of merge commit
         ref: ${{ github.event.pull_request.head.sha }}
     - name: Install dependencies
-      run: sudo contrib/apt-install libprotobuf-dev libprotobuf-c-dev protobuf-c-compiler protobuf-compiler python3-protobuf libnl-3-dev libnet-dev libcap-dev uuid-dev
+      run: sudo contrib/apt-install libprotobuf-dev libprotobuf-c-dev protobuf-c-compiler protobuf-compiler python3-protobuf libnl-3-dev libnet-dev libcap-dev uuid-dev liblz4-dev
     - name: Configure git user details
       run: |
         git config --global user.email "checkpoint-restore@users.noreply.github.com"
@@ -28,3 +28,8 @@ jobs:
       run: git fetch origin ${{ github.base_ref }}:${{ github.base_ref }}
     - name: Build each commit
       run: git rebase ${{ github.base_ref }} -x "make -C scripts/ci check-commit"
+    - name: Build without LZ4
+      run: |
+        make mrproper
+        make -j "$(nproc)" NO_LZ4=1
+        make unittest NO_LZ4=1

Documentation/crit.txt

@@ -18,6 +18,10 @@ SYNOPSIS
 
 *crit* 'show' [-h] in
 
+*crit* 'compress' [-h] [--in-place] [--acceleration N] dir
+
+*crit* 'decompress' [-h] [--in-place] dir
+
 DESCRIPTION
 -----------
 *crit* is a feature-rich replacement for existing *criu* show.
@@ -43,12 +47,46 @@ Positional Arguments
 *show*::
     convert *criu* image from binary to human-readable JSON
 
+*compress*::
+    compress memory page images in a checkpoint directory using CRIU's
+    per-page LZ4 format. The command rewrites *pages-*.img*,
+    *pagemap-*.img*, and *inventory.img*. Pages that are zero-filled or
+    do not meet CRIU's compression threshold are stored without an LZ4
+    payload. By default, the original files are preserved with a *.bak*
+    suffix; use *--in-place* to skip these backup files. The inventory
+    is updated to record compression metadata and the compressed-image
+    version. Existing *.bak* files are never overwritten.
+
+*decompress*::
+    decompress memory page images in a checkpoint directory back to
+    uncompressed page payloads. The command removes compression metadata
+    from the pagemap and inventory images. When the directory has no parent
+    reference, the command also restores the normal image version. It retains
+    the compressed-image version for incremental checkpoints because a parent
+    may still contain compressed payloads. By default, rewritten files are
+    preserved with a *.bak* suffix; use *--in-place* to skip these backup
+    files. Existing *.bak* files are never overwritten.
+
+Both transformations preserve the ownership, permissions, ACLs, security
+labels, and other extended attributes of rewritten images. All output is
+staged and synchronized before it is installed. If a write, rename, *SIGHUP*,
+*SIGINT*, or *SIGTERM* interrupts the commit, CRIT restores the complete
+original image set instead of leaving a partially transformed set.
+
 Optional Arguments
 ~~~~~~~~~~~~~~~~~~
 
 *-h*, *--help*::
     Print some help and exit
 
+*--in-place*::
+    Rewrite files without creating *.bak* backup files. This option is
+    accepted by *compress* and *decompress*.
+
+*--acceleration* 'N'::
+    Set the LZ4 acceleration level used by *compress*. Higher values
+    trade compression ratio for speed.
+
 SEE ALSO
 --------
 criu(8)

Documentation/criu.txt

@@ -378,6 +378,60 @@ mount -t cgroup -o devices,freezer none devices,freezer
     Allows to link unlinked files back, if possible (modifies filesystem
     during *restore*).
 
+*-c, --compress*::
+    Enable LZ4 per-page compression of memory pages during *dump*.
+    Each system page is compressed as an independent LZ4 block.
+    Zero-filled pages are detected and stored with no data; pages that
+    do not compress well are stored raw to avoid decompression overhead
+    on *restore*. The compression choice made at *dump* time is recorded
+    in the inventory image and applied automatically on *restore*.
+    Compatible with *--page-server* when *--tls* is not used, *--stream*,
+    *--lazy-pages*, iterative checkpointing, *--auto-dedup*, and the *dedup*
+    command.
+
+*--compress-region* 'size'::
+    Enable LZ4 region compression of memory pages during *dump*.
+    A run of consecutive pages totalling 'size' bytes is compressed as
+    a single LZ4 block, which yields a better ratio and faster dump
+    on heap-shaped data than per-page compression at the cost of
+    slightly slower restore on partial reads. 'size' accepts K/M/G
+    suffixes (e.g. *256K*, *1M*); it must be a multiple of the system
+    page size and at most 4M. Mutually exclusive with *--compress*.
+    Currently supports only the local image path (no
+    *--page-server* / *--stream*). Hole-punching deduplication is
+    skipped for compressed images since compressed pages are
+    variably-sized.
+
+*--compress-acceleration* 'N'::
+    Set the LZ4 acceleration level for page compression (1 to 65537).
+    Controls how thoroughly the compressor searches for matching byte
+    sequences. The default value of 1 gives the best compression
+    ratio. Higher values skip more match candidates, resulting in
+    faster compression but larger output. Decompression speed is not
+    affected. Implies *--compress* unless *--compress-region* is set.
+
+*--decompress-threads* 'N'::
+    Set worker concurrency for LZ4 decoding and eligible large zero fills,
+    including the calling restore thread. The default value of 1 keeps each
+    LZ4 decode serial and handles zero blocks without a worker pool;
+    independent private, shmem, and memfd requests can still run concurrently.
+    A value of 0 lets CRIU choose the concurrency for each batch. CRIU starts
+    with the CPUs allowed by its affinity mask, then limits the width by the
+    number and decoded size of the blocks in that batch and by the restore-wide
+    CPU budget. Small batches and zero runs with fewer than two blocks remain
+    serial. Values above 1 set an upper bound on aggregate worker concurrency
+    instead of using the automatic CPU count. Explicit requests above the
+    available CPU count are reduced with a warning. The accepted range is 0
+    through 1024; this validation ceiling does not determine the automatic
+    concurrency.
+    Encoded staging memory is bounded independently of the thread count.
+    Restore keeps at most two active 32 MiB encoded input buffers. When a
+    local reader can reserve a second buffer without waiting, its calling thread
+    reads the next payload while pool workers decode the current one. If
+    another restore request already owns the second buffer, reading remains
+    synchronous. The option can also be set through a configuration file
+    (*decompress-threads 0*).
+
 *--timeout* 'number'::
     Set a time limit in seconds for collecting tasks during the
     dump operation. The timeout is 10 seconds by default.

Makefile

@@ -237,6 +237,7 @@ else
 export CONFIG_COMPAT := y
 export CONFIG_GNUTLS := y
 export CONFIG_HAS_LIBBPF := y
+export CONFIG_LZ4 := y
 endif
 
 #
@@ -462,10 +463,13 @@ ruff:
 		test/others/criu-ns/run.py \
 		crit/*.py \
 		crit/crit/*.py \
+		test/others/crit/*.py \
 		scripts/uninstall_module.py \
 		coredump/ coredump/coredump \
 		scripts/github-indent-warnings.py \
-		contrib/criu-service-client/test/*.py
+		contrib/criu-service-client/test/*.py \
+		contrib/compression-benchmark/ \
+		test/others/compression/
 
 shellcheck:
 	shellcheck --version
@@ -478,6 +482,7 @@ shellcheck:
 	shellcheck -x test/others/config-file/*.sh
 	shellcheck -x test/others/action-script/*.sh
 	shellcheck -x contrib/criu-service-client/test/*.sh
+	shellcheck -x test/others/compression/*/*.sh
 
 codespell:
 	codespell

Makefile.config

@@ -46,6 +46,14 @@ else
         $(info $S      Install gnutls-devel (RPM) or gnutls-dev (DEB) to fix.)
 endif
 
+ifeq ($(NO_LZ4)x$(call pkg-config-check,liblz4),xy)
+        LIBS_FEATURES	+= -llz4
+        export CONFIG_LZ4 := y
+        FEATURE_DEFINES	+= -DCONFIG_LZ4
+else
+        $(info Note: Building without LZ4 compression support.)
+endif
+
 ifeq ($(call pkg-config-check,libnftables),y)
         LIB_NFTABLES	:= $(shell $(PKG_CONFIG) --libs libnftables)
         ifeq ($(call try-cc,$(FEATURE_TEST_NFTABLES_LIB_API_0),$(LIB_NFTABLES)),true)