Merge "Add header that defines strerror()"
diff --git a/adb/client/adb_client.cpp b/adb/client/adb_client.cpp
index 0f621eb..d91ae35 100644
--- a/adb/client/adb_client.cpp
+++ b/adb/client/adb_client.cpp
@@ -400,9 +400,15 @@
}
bool adb_get_feature_set(FeatureSet* feature_set, std::string* error) {
- std::string result;
- if (adb_query(format_host_command("features"), &result, error)) {
- *feature_set = StringToFeatureSet(result);
+ static FeatureSet* features = nullptr;
+ if (!features) {
+ std::string result;
+ if (adb_query(format_host_command("features"), &result, error)) {
+ features = new FeatureSet(StringToFeatureSet(result));
+ }
+ }
+ if (features) {
+ *feature_set = *features;
return true;
}
feature_set->clear();
diff --git a/adb/client/adb_install.cpp b/adb/client/adb_install.cpp
index f1f080a..a6e8998 100644
--- a/adb/client/adb_install.cpp
+++ b/adb/client/adb_install.cpp
@@ -16,6 +16,7 @@
#include "adb_install.h"
+#include <fcntl.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
@@ -40,18 +41,31 @@
static constexpr int kFastDeployMinApi = 24;
#endif
+namespace {
+
+enum InstallMode {
+ INSTALL_DEFAULT,
+ INSTALL_PUSH,
+ INSTALL_STREAM,
+};
+
+}
+
static bool can_use_feature(const char* feature) {
FeatureSet features;
std::string error;
if (!adb_get_feature_set(&features, &error)) {
fprintf(stderr, "error: %s\n", error.c_str());
- return true;
+ return false;
}
return CanUseFeature(features, feature);
}
-static bool use_legacy_install() {
- return !can_use_feature(kFeatureCmd);
+static InstallMode best_install_mode() {
+ if (can_use_feature(kFeatureCmd)) {
+ return INSTALL_STREAM;
+ }
+ return INSTALL_PUSH;
}
static bool is_apex_supported() {
@@ -112,7 +126,7 @@
}
int uninstall_app(int argc, const char** argv) {
- if (use_legacy_install()) {
+ if (best_install_mode() == INSTALL_PUSH) {
return uninstall_app_legacy(argc, argv);
}
return uninstall_app_streamed(argc, argv);
@@ -200,32 +214,49 @@
return 1;
}
+#ifdef __linux__
+ posix_fadvise(local_fd.get(), 0, 0, POSIX_FADV_SEQUENTIAL | POSIX_FADV_NOREUSE);
+#endif
+
+ const bool use_abb = can_use_feature(kFeatureAbb);
std::string error;
- std::string cmd = "exec:cmd package";
+ std::vector<std::string> cmd_args = {use_abb ? "package" : "exec:cmd package"};
+ cmd_args.reserve(argc + 3);
// don't copy the APK name, but, copy the rest of the arguments as-is
while (argc-- > 1) {
- cmd += " " + escape_arg(std::string(*argv++));
+ if (use_abb) {
+ cmd_args.push_back(*argv++);
+ } else {
+ cmd_args.push_back(escape_arg(*argv++));
+ }
}
// add size parameter [required for streaming installs]
// do last to override any user specified value
- cmd += " " + android::base::StringPrintf("-S %" PRIu64, static_cast<uint64_t>(sb.st_size));
+ cmd_args.push_back("-S");
+ cmd_args.push_back(
+ android::base::StringPrintf("%" PRIu64, static_cast<uint64_t>(sb.st_size)));
if (is_apex) {
- cmd += " --apex";
+ cmd_args.push_back("--apex");
}
- unique_fd remote_fd(adb_connect(cmd, &error));
+ unique_fd remote_fd;
+ if (use_abb) {
+ remote_fd = send_abb_exec_command(cmd_args, &error);
+ } else {
+ remote_fd.reset(adb_connect(android::base::Join(cmd_args, " "), &error));
+ }
if (remote_fd < 0) {
fprintf(stderr, "adb: connect error for write: %s\n", error.c_str());
return 1;
}
- char buf[BUFSIZ];
copy_to_file(local_fd.get(), remote_fd.get());
- read_status_line(remote_fd.get(), buf, sizeof(buf));
+ char buf[BUFSIZ];
+ read_status_line(remote_fd.get(), buf, sizeof(buf));
if (!strncmp("Success", buf, 7)) {
fputs(buf, stdout);
return 0;
@@ -256,8 +287,7 @@
int result = -1;
std::vector<const char*> apk_file = {argv[last_apk]};
- std::string apk_dest =
- "/data/local/tmp/" + android::base::Basename(argv[last_apk]);
+ std::string apk_dest = "/data/local/tmp/" + android::base::Basename(argv[last_apk]);
if (use_fastdeploy == true) {
#if defined(ENABLE_FASTDEPLOY)
@@ -292,11 +322,7 @@
int install_app(int argc, const char** argv) {
std::vector<int> processedArgIndicies;
- enum installMode {
- INSTALL_DEFAULT,
- INSTALL_PUSH,
- INSTALL_STREAM
- } installMode = INSTALL_DEFAULT;
+ InstallMode installMode = INSTALL_DEFAULT;
bool use_fastdeploy = false;
bool is_reinstall = false;
bool use_localagent = false;
@@ -337,14 +363,10 @@
}
if (installMode == INSTALL_DEFAULT) {
- if (use_legacy_install()) {
- installMode = INSTALL_PUSH;
- } else {
- installMode = INSTALL_STREAM;
- }
+ installMode = best_install_mode();
}
- if (installMode == INSTALL_STREAM && use_legacy_install() == true) {
+ if (installMode == INSTALL_STREAM && best_install_mode() == INSTALL_PUSH) {
error_exit("Attempting to use streaming install on unsupported device");
}
@@ -420,7 +442,7 @@
if (first_apk == -1) error_exit("need APK file on command line");
std::string install_cmd;
- if (use_legacy_install()) {
+ if (best_install_mode() == INSTALL_PUSH) {
install_cmd = "exec:pm";
} else {
install_cmd = "exec:cmd package";
@@ -545,7 +567,7 @@
if (first_package == -1) error_exit("need APK or APEX files on command line");
- if (use_legacy_install()) {
+ if (best_install_mode() == INSTALL_PUSH) {
fprintf(stderr, "adb: multi-package install is not supported on this device\n");
return EXIT_FAILURE;
}
diff --git a/adb/client/commandline.cpp b/adb/client/commandline.cpp
index 7d0eb40..599e0e6 100644
--- a/adb/client/commandline.cpp
+++ b/adb/client/commandline.cpp
@@ -357,7 +357,7 @@
}
void copy_to_file(int inFd, int outFd) {
- std::vector<char> buf(32 * 1024);
+ std::vector<char> buf(64 * 1024);
int len;
long total = 0;
int old_stdin_mode = -1;
diff --git a/adb/client/commandline.h b/adb/client/commandline.h
index 6cfd4f7..cd5933a 100644
--- a/adb/client/commandline.h
+++ b/adb/client/commandline.h
@@ -17,7 +17,11 @@
#ifndef COMMANDLINE_H
#define COMMANDLINE_H
+#include <android-base/strings.h>
+
#include "adb.h"
+#include "adb_client.h"
+#include "adb_unique_fd.h"
// Callback used to handle the standard streams (stdout and stderr) sent by the
// device's upon receiving a command.
@@ -105,4 +109,17 @@
const std::string& command, bool disable_shell_protocol = false,
StandardStreamsCallbackInterface* callback = &DEFAULT_STANDARD_STREAMS_CALLBACK);
+// Connects to the device "abb" service with |command| and returns the fd.
+template <typename ContainerT>
+unique_fd send_abb_exec_command(const ContainerT& command_args, std::string* error) {
+ std::string service_string = "abb_exec:" + android::base::Join(command_args, ABB_ARG_DELIMETER);
+
+ unique_fd fd(adb_connect(service_string, error));
+ if (fd < 0) {
+ fprintf(stderr, "adb: failed to run abb_exec. Error: %s\n", error->c_str());
+ return unique_fd{};
+ }
+ return fd;
+}
+
#endif // COMMANDLINE_H
diff --git a/adb/daemon/abb.cpp b/adb/daemon/abb.cpp
index 425438e..17c25e8 100644
--- a/adb/daemon/abb.cpp
+++ b/adb/daemon/abb.cpp
@@ -24,6 +24,7 @@
#include "adb_io.h"
#include "adb_utils.h"
#include "shell_service.h"
+#include "sysdeps.h"
namespace {
@@ -69,6 +70,11 @@
} // namespace
static int execCmd(std::string_view args, borrowed_fd in, borrowed_fd out, borrowed_fd err) {
+ int max_buf = LINUX_MAX_SOCKET_SIZE;
+ adb_setsockopt(in, SOL_SOCKET, SO_SNDBUF, &max_buf, sizeof(max_buf));
+ adb_setsockopt(out, SOL_SOCKET, SO_SNDBUF, &max_buf, sizeof(max_buf));
+ adb_setsockopt(err, SOL_SOCKET, SO_SNDBUF, &max_buf, sizeof(max_buf));
+
AdbFdTextOutput oin(out);
AdbFdTextOutput oerr(err);
return cmdMain(parseCmdArgs(args), oin, oerr, in.get(), out.get(), err.get(),
@@ -98,6 +104,8 @@
}
unique_fd result = StartCommandInProcess(std::string(name), &execCmd, protocol);
+ int max_buf = LINUX_MAX_SOCKET_SIZE;
+ adb_setsockopt(result, SOL_SOCKET, SO_SNDBUF, &max_buf, sizeof(max_buf));
if (android::base::SendFileDescriptors(fd, "", 1, result.get()) != 1) {
PLOG(ERROR) << "Failed to send an inprocess fd for command: " << data;
break;
diff --git a/adb/daemon/abb_service.cpp b/adb/daemon/abb_service.cpp
index a435279..e1df4a5 100644
--- a/adb/daemon/abb_service.cpp
+++ b/adb/daemon/abb_service.cpp
@@ -53,14 +53,13 @@
return error_fd;
}
- if (!SendProtocolString(socket_fd_, std::string(command))) {
+ if (!SendProtocolString(socket_fd_, command)) {
PLOG(ERROR) << "failed to send command to abb";
socket_fd_.reset();
continue;
}
unique_fd fd;
- std::string error;
char buf;
if (android::base::ReceiveFileDescriptors(socket_fd_, &buf, 1, &fd) != 1) {
PLOG(ERROR) << "failed to receive FD from abb";
diff --git a/adb/fdevent/fdevent.h b/adb/fdevent/fdevent.h
index ccb0c92..2424252 100644
--- a/adb/fdevent/fdevent.h
+++ b/adb/fdevent/fdevent.h
@@ -20,6 +20,7 @@
#include <stddef.h>
#include <stdint.h>
+#include <atomic>
#include <chrono>
#include <deque>
#include <functional>
diff --git a/fs_mgr/fs_mgr_overlayfs.cpp b/fs_mgr/fs_mgr_overlayfs.cpp
index ac15ce4..4ee7db9 100644
--- a/fs_mgr/fs_mgr_overlayfs.cpp
+++ b/fs_mgr/fs_mgr_overlayfs.cpp
@@ -150,6 +150,31 @@
return (vst.f_bfree >= (vst.f_blocks * kPercentThreshold / 100));
}
+const auto kPhysicalDevice = "/dev/block/by-name/"s;
+
+bool fs_mgr_update_blk_device(FstabEntry* entry) {
+ if (entry->fs_mgr_flags.logical) {
+ fs_mgr_update_logical_partition(entry);
+ }
+ if (fs_mgr_access(entry->blk_device)) {
+ return true;
+ }
+ if (entry->blk_device != "/dev/root") {
+ return false;
+ }
+
+ // special case for system-as-root (taimen and others)
+ auto blk_device = kPhysicalDevice + "system";
+ if (!fs_mgr_access(blk_device)) {
+ blk_device += fs_mgr_get_slot_suffix();
+ if (!fs_mgr_access(blk_device)) {
+ return false;
+ }
+ }
+ entry->blk_device = blk_device;
+ return true;
+}
+
bool fs_mgr_overlayfs_enabled(FstabEntry* entry) {
// readonly filesystem, can not be mount -o remount,rw
// for squashfs, erofs or if free space is (near) zero making such a remount
@@ -157,19 +182,19 @@
if (!fs_mgr_filesystem_has_space(entry->mount_point)) {
return true;
}
- if (entry->fs_mgr_flags.logical) {
- fs_mgr_update_logical_partition(entry);
+
+ // blk_device needs to be setup so we can check superblock.
+ // If we fail here, because during init first stage and have doubts.
+ if (!fs_mgr_update_blk_device(entry)) {
+ return true;
}
+
+ // check if ext4 de-dupe
auto save_errno = errno;
- errno = 0;
auto has_shared_blocks = fs_mgr_has_shared_blocks(entry->mount_point, entry->blk_device);
if (!has_shared_blocks && (entry->mount_point == "/system")) {
has_shared_blocks = fs_mgr_has_shared_blocks("/", entry->blk_device);
}
- // special case for first stage init for system as root (taimen)
- if (!has_shared_blocks && (errno == ENOENT) && (entry->blk_device == "/dev/root")) {
- has_shared_blocks = true;
- }
errno = save_errno;
return has_shared_blocks;
}
@@ -388,8 +413,6 @@
return SlotNumberForSlotSuffix(fs_mgr_get_slot_suffix());
}
-const auto kPhysicalDevice = "/dev/block/by-name/"s;
-
std::string fs_mgr_overlayfs_super_device(uint32_t slot_number) {
return kPhysicalDevice + fs_mgr_get_super_partition_name(slot_number);
}
diff --git a/fs_mgr/libdm/include/libdm/dm.h b/fs_mgr/libdm/include/libdm/dm.h
index f5783cb..cf306f3 100644
--- a/fs_mgr/libdm/include/libdm/dm.h
+++ b/fs_mgr/libdm/include/libdm/dm.h
@@ -197,6 +197,7 @@
struct TargetInfo {
struct dm_target_spec spec;
std::string data;
+ TargetInfo() {}
TargetInfo(const struct dm_target_spec& spec, const std::string& data)
: spec(spec), data(data) {}
};
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
index 062e00b..f7608dc 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
@@ -22,7 +22,8 @@
#include <vector>
#include <android-base/unique_fd.h>
-#include <libdm/dm_target.h>
+#include <libdm/dm.h>
+#include <libfiemap/image_manager.h>
#ifndef FRIEND_TEST
#define FRIEND_TEST(test_set_name, individual_test) \
@@ -38,7 +39,7 @@
namespace snapshot {
-enum class UpdateState {
+enum class UpdateState : unsigned int {
// No update or merge is in progress.
None,
@@ -51,6 +52,10 @@
// The kernel is merging in the background.
Merging,
+ // Post-merge cleanup steps could not be completed due to a transient
+ // error, but the next reboot will finish any pending operations.
+ MergeNeedsReboot,
+
// Merging is complete, and needs to be acknowledged.
MergeCompleted,
@@ -94,8 +99,23 @@
// Wait for the current merge to finish, then perform cleanup when it
// completes. It is necessary to call this after InitiateMerge(), or when
- // a merge is detected for the first time after boot.
- bool WaitForMerge();
+ // a merge state is detected during boot.
+ //
+ // Note that after calling WaitForMerge(), GetUpdateState() may still return
+ // that a merge is in progress:
+ // MergeFailed indicates that a fatal error occurred. WaitForMerge() may
+ // called any number of times again to attempt to make more progress, but
+ // we do not expect it to succeed if a catastrophic error occurred.
+ //
+ // MergeNeedsReboot indicates that the merge has completed, but cleanup
+ // failed. This can happen if for some reason resources were not closed
+ // properly. In this case another reboot is needed before we can take
+ // another OTA. However, WaitForMerge() can be called again without
+ // rebooting, to attempt to finish cleanup anyway.
+ //
+ // MergeCompleted indicates that the update has fully completed.
+ // GetUpdateState will return None, and a new update can begin.
+ UpdateState WaitForMerge();
// Find the status of the current update, if any.
//
@@ -109,9 +129,14 @@
FRIEND_TEST(SnapshotTest, CreateSnapshot);
FRIEND_TEST(SnapshotTest, MapSnapshot);
FRIEND_TEST(SnapshotTest, MapPartialSnapshot);
+ FRIEND_TEST(SnapshotTest, NoMergeBeforeReboot);
+ FRIEND_TEST(SnapshotTest, Merge);
+ FRIEND_TEST(SnapshotTest, MergeCannotRemoveCow);
friend class SnapshotTest;
+ using DmTargetSnapshot = android::dm::DmTargetSnapshot;
using IImageManager = android::fiemap::IImageManager;
+ using TargetInfo = android::dm::DeviceMapper::TargetInfo;
explicit SnapshotManager(IDeviceInfo* info);
@@ -126,16 +151,18 @@
// this. It also serves as a proof-of-lock for some functions.
class LockedFile final {
public:
- LockedFile(const std::string& path, android::base::unique_fd&& fd)
- : path_(path), fd_(std::move(fd)) {}
+ LockedFile(const std::string& path, android::base::unique_fd&& fd, int lock_mode)
+ : path_(path), fd_(std::move(fd)), lock_mode_(lock_mode) {}
~LockedFile();
const std::string& path() const { return path_; }
int fd() const { return fd_; }
+ int lock_mode() const { return lock_mode_; }
private:
std::string path_;
android::base::unique_fd fd_;
+ int lock_mode_;
};
std::unique_ptr<LockedFile> OpenFile(const std::string& file, int open_flags, int lock_flags);
bool Truncate(LockedFile* file);
@@ -189,6 +216,7 @@
std::unique_ptr<LockedFile> LockExclusive();
UpdateState ReadUpdateState(LockedFile* file);
bool WriteUpdateState(LockedFile* file, UpdateState state);
+ std::string GetStateFilePath() const;
// This state is persisted per-snapshot in /metadata/ota/snapshots/.
struct SnapshotStatus {
@@ -200,11 +228,38 @@
uint64_t metadata_sectors = 0;
};
+ // Helpers for merging.
+ bool SwitchSnapshotToMerge(LockedFile* lock, const std::string& name);
+ bool RewriteSnapshotDeviceTable(const std::string& dm_name);
+ bool MarkSnapshotMergeCompleted(LockedFile* snapshot_lock, const std::string& snapshot_name);
+ void AcknowledgeMergeSuccess(LockedFile* lock);
+ void AcknowledgeMergeFailure();
+
+ // Note that these require the name of the device containing the snapshot,
+ // which may be the "inner" device. Use GetsnapshotDeviecName().
+ bool QuerySnapshotStatus(const std::string& dm_name, std::string* target_type,
+ DmTargetSnapshot::Status* status);
+ bool IsSnapshotDevice(const std::string& dm_name, TargetInfo* target = nullptr);
+
+ // Internal callback for when merging is complete.
+ bool OnSnapshotMergeComplete(LockedFile* lock, const std::string& name,
+ const SnapshotStatus& status);
+ bool CollapseSnapshotDevice(const std::string& name, const SnapshotStatus& status);
+
+ // Only the following UpdateStates are used here:
+ // UpdateState::Merging
+ // UpdateState::MergeCompleted
+ // UpdateState::MergeFailed
+ // UpdateState::MergeNeedsReboot
+ UpdateState CheckMergeState();
+ UpdateState CheckMergeState(LockedFile* lock);
+ UpdateState CheckTargetMergeState(LockedFile* lock, const std::string& name);
+
// Interact with status files under /metadata/ota/snapshots.
- std::unique_ptr<LockedFile> OpenSnapshotStatusFile(const std::string& name, int open_flags,
- int lock_flags);
- bool WriteSnapshotStatus(LockedFile* file, const SnapshotStatus& status);
- bool ReadSnapshotStatus(LockedFile* file, SnapshotStatus* status);
+ bool WriteSnapshotStatus(LockedFile* lock, const std::string& name,
+ const SnapshotStatus& status);
+ bool ReadSnapshotStatus(LockedFile* lock, const std::string& name, SnapshotStatus* status);
+ std::string GetSnapshotStatusFilePath(const std::string& name);
// Return the name of the device holding the "snapshot" or "snapshot-merge"
// target. This may not be the final device presented via MapSnapshot(), if
diff --git a/fs_mgr/libsnapshot/snapshot.cpp b/fs_mgr/libsnapshot/snapshot.cpp
index ef56179..63a01f3 100644
--- a/fs_mgr/libsnapshot/snapshot.cpp
+++ b/fs_mgr/libsnapshot/snapshot.cpp
@@ -19,6 +19,8 @@
#include <sys/types.h>
#include <sys/unistd.h>
+#include <thread>
+
#include <android-base/file.h>
#include <android-base/logging.h>
#include <android-base/parseint.h>
@@ -140,9 +142,6 @@
LOG(INFO) << "Snapshot " << name << " will have COW size " << cow_size;
- auto status_file = OpenSnapshotStatusFile(name, O_RDWR | O_CREAT, LOCK_EX);
- if (!status_file) return false;
-
// Note, we leave the status file hanging around if we fail to create the
// actual backing image. This is harmless, since it'll get removed when
// CancelUpdate is called.
@@ -151,7 +150,7 @@
.device_size = device_size,
.snapshot_size = snapshot_size,
};
- if (!WriteSnapshotStatus(status_file.get(), status)) {
+ if (!WriteSnapshotStatus(lock, name, status)) {
PLOG(ERROR) << "Could not write snapshot status: " << name;
return false;
}
@@ -168,11 +167,13 @@
CHECK(lock);
if (!EnsureImageManager()) return false;
- auto status_file = OpenSnapshotStatusFile(name, O_RDWR, LOCK_EX);
- if (!status_file) return false;
-
SnapshotStatus status;
- if (!ReadSnapshotStatus(status_file.get(), &status)) {
+ if (!ReadSnapshotStatus(lock, name, &status)) {
+ return false;
+ }
+ if (status.state == "merge-completed") {
+ LOG(ERROR) << "Should not create a snapshot device for " << name
+ << " after merging has completed.";
return false;
}
@@ -210,19 +211,31 @@
std::string cow_dev;
if (!images_->MapImageDevice(cow_name, timeout_ms, &cow_dev)) {
+ LOG(ERROR) << "Could not map image device: " << cow_name;
return false;
}
auto& dm = DeviceMapper::Instance();
- // Merging is a global state, not per-snapshot. We do however track the
- // progress of individual snapshots' merges.
+ // Note that merging is a global state. We do track whether individual devices
+ // have completed merging, but the start of the merge process is considered
+ // atomic.
SnapshotStorageMode mode;
- UpdateState update_state = ReadUpdateState(lock);
- if (update_state == UpdateState::Merging || update_state == UpdateState::MergeCompleted) {
- mode = SnapshotStorageMode::Merge;
- } else {
- mode = SnapshotStorageMode::Persistent;
+ switch (ReadUpdateState(lock)) {
+ case UpdateState::MergeCompleted:
+ case UpdateState::MergeNeedsReboot:
+ LOG(ERROR) << "Should not create a snapshot device for " << name
+ << " after global merging has completed.";
+ return false;
+ case UpdateState::Merging:
+ case UpdateState::MergeFailed:
+ // Note: MergeFailed indicates that a merge is in progress, but
+ // is possibly stalled. We still have to honor the merge.
+ mode = SnapshotStorageMode::Merge;
+ break;
+ default:
+ mode = SnapshotStorageMode::Persistent;
+ break;
}
// The kernel (tested on 4.19) crashes horribly if a device has both a snapshot
@@ -267,11 +280,8 @@
CHECK(lock);
if (!EnsureImageManager()) return false;
- auto status_file = OpenSnapshotStatusFile(name, O_RDWR, LOCK_EX);
- if (!status_file) return false;
-
SnapshotStatus status;
- if (!ReadSnapshotStatus(status_file.get(), &status)) {
+ if (!ReadSnapshotStatus(lock, name, &status)) {
return false;
}
@@ -300,31 +310,525 @@
CHECK(lock);
if (!EnsureImageManager()) return false;
- // Take the snapshot's lock after Unmap, since it will also try to lock.
- auto status_file = OpenSnapshotStatusFile(name, O_RDONLY, LOCK_EX);
- if (!status_file) return false;
-
auto cow_name = GetCowName(name);
if (!images_->BackingImageExists(cow_name)) {
return true;
}
+ if (images_->IsImageMapped(cow_name) && !images_->UnmapImageDevice(cow_name)) {
+ return false;
+ }
if (!images_->DeleteBackingImage(cow_name)) {
return false;
}
- if (!android::base::RemoveFileIfExists(status_file->path())) {
- LOG(ERROR) << "Failed to remove status file: " << status_file->path();
+ std::string error;
+ auto file_path = GetSnapshotStatusFilePath(name);
+ if (!android::base::RemoveFileIfExists(file_path, &error)) {
+ LOG(ERROR) << "Failed to remove status file " << file_path << ": " << error;
return false;
}
return true;
}
bool SnapshotManager::InitiateMerge() {
- return false;
+ auto lock = LockExclusive();
+ if (!lock) return false;
+
+ UpdateState state = ReadUpdateState(lock.get());
+ if (state != UpdateState::Unverified) {
+ LOG(ERROR) << "Cannot begin a merge if an update has not been verified";
+ return false;
+ }
+ if (!device_->IsRunningSnapshot()) {
+ LOG(ERROR) << "Cannot begin a merge if the device is not booted off a snapshot";
+ return false;
+ }
+
+ std::vector<std::string> snapshots;
+ if (!ListSnapshots(lock.get(), &snapshots)) {
+ LOG(ERROR) << "Could not list snapshots";
+ return false;
+ }
+
+ auto& dm = DeviceMapper::Instance();
+ for (const auto& snapshot : snapshots) {
+ // The device has to be mapped, since everything should be merged at
+ // the same time. This is a fairly serious error. We could forcefully
+ // map everything here, but it should have been mapped during first-
+ // stage init.
+ if (dm.GetState(snapshot) == DmDeviceState::INVALID) {
+ LOG(ERROR) << "Cannot begin merge; device " << snapshot << " is not mapped.";
+ return false;
+ }
+ }
+
+ // Point of no return - mark that we're starting a merge. From now on every
+ // snapshot must be a merge target.
+ if (!WriteUpdateState(lock.get(), UpdateState::Merging)) {
+ return false;
+ }
+
+ bool rewrote_all = true;
+ for (const auto& snapshot : snapshots) {
+ // If this fails, we have no choice but to continue. Everything must
+ // be merged. This is not an ideal state to be in, but it is safe,
+ // because we the next boot will try again.
+ if (!SwitchSnapshotToMerge(lock.get(), snapshot)) {
+ LOG(ERROR) << "Failed to switch snapshot to a merge target: " << snapshot;
+ rewrote_all = false;
+ }
+ }
+
+ // If we couldn't switch everything to a merge target, pre-emptively mark
+ // this merge as failed. It will get acknowledged when WaitForMerge() is
+ // called.
+ if (!rewrote_all) {
+ WriteUpdateState(lock.get(), UpdateState::MergeFailed);
+ }
+
+ // Return true no matter what, because a merge was initiated.
+ return true;
}
-bool SnapshotManager::WaitForMerge() {
- return false;
+bool SnapshotManager::SwitchSnapshotToMerge(LockedFile* lock, const std::string& name) {
+ SnapshotStatus status;
+ if (!ReadSnapshotStatus(lock, name, &status)) {
+ return false;
+ }
+ if (status.state != "created") {
+ LOG(WARNING) << "Snapshot " << name << " has unexpected state: " << status.state;
+ }
+
+ // After this, we return true because we technically did switch to a merge
+ // target. Everything else we do here is just informational.
+ auto dm_name = GetSnapshotDeviceName(name, status);
+ if (!RewriteSnapshotDeviceTable(dm_name)) {
+ return false;
+ }
+
+ status.state = "merging";
+
+ DmTargetSnapshot::Status dm_status;
+ if (!QuerySnapshotStatus(dm_name, nullptr, &dm_status)) {
+ LOG(ERROR) << "Could not query merge status for snapshot: " << dm_name;
+ }
+ status.sectors_allocated = dm_status.sectors_allocated;
+ status.metadata_sectors = dm_status.metadata_sectors;
+ if (!WriteSnapshotStatus(lock, name, status)) {
+ LOG(ERROR) << "Could not update status file for snapshot: " << name;
+ }
+ return true;
+}
+
+bool SnapshotManager::RewriteSnapshotDeviceTable(const std::string& dm_name) {
+ auto& dm = DeviceMapper::Instance();
+
+ std::vector<DeviceMapper::TargetInfo> old_targets;
+ if (!dm.GetTableInfo(dm_name, &old_targets)) {
+ LOG(ERROR) << "Could not read snapshot device table: " << dm_name;
+ return false;
+ }
+ if (old_targets.size() != 1 || DeviceMapper::GetTargetType(old_targets[0].spec) != "snapshot") {
+ LOG(ERROR) << "Unexpected device-mapper table for snapshot: " << dm_name;
+ return false;
+ }
+
+ std::string base_device, cow_device;
+ if (!DmTargetSnapshot::GetDevicesFromParams(old_targets[0].data, &base_device, &cow_device)) {
+ LOG(ERROR) << "Could not derive underlying devices for snapshot: " << dm_name;
+ return false;
+ }
+
+ DmTable table;
+ table.Emplace<DmTargetSnapshot>(0, old_targets[0].spec.length, base_device, cow_device,
+ SnapshotStorageMode::Merge, kSnapshotChunkSize);
+ if (!dm.LoadTableAndActivate(dm_name, table)) {
+ LOG(ERROR) << "Could not swap device-mapper tables on snapshot device " << dm_name;
+ return false;
+ }
+ LOG(INFO) << "Successfully switched snapshot device to a merge target: " << dm_name;
+ return true;
+}
+
+enum class TableQuery {
+ Table,
+ Status,
+};
+
+static bool GetSingleTarget(const std::string& dm_name, TableQuery query,
+ DeviceMapper::TargetInfo* target) {
+ auto& dm = DeviceMapper::Instance();
+ if (dm.GetState(dm_name) == DmDeviceState::INVALID) {
+ return false;
+ }
+
+ std::vector<DeviceMapper::TargetInfo> targets;
+ bool result;
+ if (query == TableQuery::Status) {
+ result = dm.GetTableStatus(dm_name, &targets);
+ } else {
+ result = dm.GetTableInfo(dm_name, &targets);
+ }
+ if (!result) {
+ LOG(ERROR) << "Could not query device: " << dm_name;
+ return false;
+ }
+ if (targets.size() != 1) {
+ return false;
+ }
+
+ *target = std::move(targets[0]);
+ return true;
+}
+
+bool SnapshotManager::IsSnapshotDevice(const std::string& dm_name, TargetInfo* target) {
+ DeviceMapper::TargetInfo snap_target;
+ if (!GetSingleTarget(dm_name, TableQuery::Status, &snap_target)) {
+ return false;
+ }
+ auto type = DeviceMapper::GetTargetType(snap_target.spec);
+ if (type != "snapshot" && type != "snapshot-merge") {
+ return false;
+ }
+ if (target) {
+ *target = std::move(snap_target);
+ }
+ return true;
+}
+
+bool SnapshotManager::QuerySnapshotStatus(const std::string& dm_name, std::string* target_type,
+ DmTargetSnapshot::Status* status) {
+ DeviceMapper::TargetInfo target;
+ if (!IsSnapshotDevice(dm_name, &target)) {
+ LOG(ERROR) << "Device " << dm_name << " is not a snapshot or snapshot-merge device";
+ return false;
+ }
+ if (!DmTargetSnapshot::ParseStatusText(target.data, status)) {
+ LOG(ERROR) << "Could not parse snapshot status text: " << dm_name;
+ return false;
+ }
+ if (target_type) {
+ *target_type = DeviceMapper::GetTargetType(target.spec);
+ }
+ return true;
+}
+
+// Note that when a merge fails, we will *always* try again to complete the
+// merge each time the device boots. There is no harm in doing so, and if
+// the problem was transient, we might manage to get a new outcome.
+UpdateState SnapshotManager::WaitForMerge() {
+ while (true) {
+ UpdateState state = CheckMergeState();
+ if (state != UpdateState::Merging) {
+ // Either there is no merge, or the merge was finished, so no need
+ // to keep waiting.
+ return state;
+ }
+
+ // This wait is not super time sensitive, so we have a relatively
+ // low polling frequency.
+ std::this_thread::sleep_for(2s);
+ }
+}
+
+UpdateState SnapshotManager::CheckMergeState() {
+ auto lock = LockExclusive();
+ if (!lock) {
+ AcknowledgeMergeFailure();
+ return UpdateState::MergeFailed;
+ }
+
+ auto state = CheckMergeState(lock.get());
+ if (state == UpdateState::MergeCompleted) {
+ AcknowledgeMergeSuccess(lock.get());
+ } else if (state == UpdateState::MergeFailed) {
+ AcknowledgeMergeFailure();
+ }
+ return state;
+}
+
+UpdateState SnapshotManager::CheckMergeState(LockedFile* lock) {
+ UpdateState state = ReadUpdateState(lock);
+ switch (state) {
+ case UpdateState::None:
+ case UpdateState::MergeCompleted:
+ // Harmless races are allowed between two callers of WaitForMerge,
+ // so in both of these cases we just propagate the state.
+ return state;
+
+ case UpdateState::Merging:
+ case UpdateState::MergeNeedsReboot:
+ case UpdateState::MergeFailed:
+ // We'll poll each snapshot below. Note that for the NeedsReboot
+ // case, we always poll once to give cleanup another opportunity to
+ // run.
+ break;
+
+ default:
+ LOG(ERROR) << "No merge exists, cannot wait. Update state: "
+ << static_cast<uint32_t>(state);
+ return UpdateState::None;
+ }
+
+ std::vector<std::string> snapshots;
+ if (!ListSnapshots(lock, &snapshots)) {
+ return UpdateState::MergeFailed;
+ }
+
+ bool failed = false;
+ bool merging = false;
+ bool needs_reboot = false;
+ for (const auto& snapshot : snapshots) {
+ UpdateState snapshot_state = CheckTargetMergeState(lock, snapshot);
+ switch (snapshot_state) {
+ case UpdateState::MergeFailed:
+ failed = true;
+ break;
+ case UpdateState::Merging:
+ merging = true;
+ break;
+ case UpdateState::MergeNeedsReboot:
+ needs_reboot = true;
+ break;
+ case UpdateState::MergeCompleted:
+ break;
+ default:
+ LOG(ERROR) << "Unknown merge status: " << static_cast<uint32_t>(snapshot_state);
+ failed = true;
+ break;
+ }
+ }
+
+ if (merging) {
+ // Note that we handle "Merging" before we handle anything else. We
+ // want to poll until *nothing* is merging if we can, so everything has
+ // a chance to get marked as completed or failed.
+ return UpdateState::Merging;
+ }
+ if (failed) {
+ // Note: since there are many drop-out cases for failure, we acknowledge
+ // it in WaitForMerge rather than here and elsewhere.
+ return UpdateState::MergeFailed;
+ }
+ if (needs_reboot) {
+ WriteUpdateState(lock, UpdateState::MergeNeedsReboot);
+ return UpdateState::MergeNeedsReboot;
+ }
+ return UpdateState::MergeCompleted;
+}
+
+UpdateState SnapshotManager::CheckTargetMergeState(LockedFile* lock, const std::string& name) {
+ SnapshotStatus snapshot_status;
+ if (!ReadSnapshotStatus(lock, name, &snapshot_status)) {
+ return UpdateState::MergeFailed;
+ }
+
+ std::string dm_name = GetSnapshotDeviceName(name, snapshot_status);
+
+ // During a check, we decided the merge was complete, but we were unable to
+ // collapse the device-mapper stack and perform COW cleanup. If we haven't
+ // rebooted after this check, the device will still be a snapshot-merge
+ // target. If the have rebooted, the device will now be a linear target,
+ // and we can try cleanup again.
+ if (snapshot_status.state == "merge-complete" && !IsSnapshotDevice(dm_name)) {
+ // NB: It's okay if this fails now, we gave cleanup our best effort.
+ OnSnapshotMergeComplete(lock, name, snapshot_status);
+ return UpdateState::MergeCompleted;
+ }
+
+ std::string target_type;
+ DmTargetSnapshot::Status status;
+ if (!QuerySnapshotStatus(dm_name, &target_type, &status)) {
+ return UpdateState::MergeFailed;
+ }
+ if (target_type != "snapshot-merge") {
+ // We can get here if we failed to rewrite the target type in
+ // InitiateMerge(). If we failed to create the target in first-stage
+ // init, boot would not succeed.
+ LOG(ERROR) << "Snapshot " << name << " has incorrect target type: " << target_type;
+ return UpdateState::MergeFailed;
+ }
+
+ // These two values are equal when merging is complete.
+ if (status.sectors_allocated != status.metadata_sectors) {
+ if (snapshot_status.state == "merge-complete") {
+ LOG(ERROR) << "Snapshot " << name << " is merging after being marked merge-complete.";
+ return UpdateState::MergeFailed;
+ }
+ return UpdateState::Merging;
+ }
+
+ // Merging is done. First, update the status file to indicate the merge
+ // is complete. We do this before calling OnSnapshotMergeComplete, even
+ // though this means the write is potentially wasted work (since in the
+ // ideal case we'll immediately delete the file).
+ //
+ // This makes it simpler to reason about the next reboot: no matter what
+ // part of cleanup failed, first-stage init won't try to create another
+ // snapshot device for this partition.
+ snapshot_status.state = "merge-complete";
+ if (!WriteSnapshotStatus(lock, name, snapshot_status)) {
+ return UpdateState::MergeFailed;
+ }
+ if (!OnSnapshotMergeComplete(lock, name, snapshot_status)) {
+ return UpdateState::MergeNeedsReboot;
+ }
+ return UpdateState::MergeCompleted;
+}
+
+void SnapshotManager::AcknowledgeMergeSuccess(LockedFile* lock) {
+ if (!WriteUpdateState(lock, UpdateState::None)) {
+ // We'll try again next reboot, ad infinitum.
+ return;
+ }
+}
+
+void SnapshotManager::AcknowledgeMergeFailure() {
+ // Log first, so worst case, we always have a record of why the calls below
+ // were being made.
+ LOG(ERROR) << "Merge could not be completed and will be marked as failed.";
+
+ auto lock = LockExclusive();
+ if (!lock) return;
+
+ // Since we released the lock in between WaitForMerge and here, it's
+ // possible (1) the merge successfully completed or (2) was already
+ // marked as a failure. So make sure to check the state again, and
+ // only mark as a failure if appropriate.
+ UpdateState state = ReadUpdateState(lock.get());
+ if (state != UpdateState::Merging && state != UpdateState::MergeNeedsReboot) {
+ return;
+ }
+
+ WriteUpdateState(lock.get(), UpdateState::MergeFailed);
+}
+
+bool SnapshotManager::OnSnapshotMergeComplete(LockedFile* lock, const std::string& name,
+ const SnapshotStatus& status) {
+ auto dm_name = GetSnapshotDeviceName(name, status);
+ if (IsSnapshotDevice(dm_name)) {
+ // We are extra-cautious here, to avoid deleting the wrong table.
+ std::string target_type;
+ DmTargetSnapshot::Status dm_status;
+ if (!QuerySnapshotStatus(dm_name, &target_type, &dm_status)) {
+ return false;
+ }
+ if (target_type != "snapshot-merge") {
+ LOG(ERROR) << "Unexpected target type " << target_type
+ << " for snapshot device: " << dm_name;
+ return false;
+ }
+ if (dm_status.sectors_allocated != dm_status.metadata_sectors) {
+ LOG(ERROR) << "Merge is unexpectedly incomplete for device " << dm_name;
+ return false;
+ }
+ if (!CollapseSnapshotDevice(name, status)) {
+ LOG(ERROR) << "Unable to collapse snapshot: " << name;
+ return false;
+ }
+ // Note that collapsing is implicitly an Unmap, so we don't need to
+ // unmap the snapshot.
+ }
+
+ if (!DeleteSnapshot(lock, name)) {
+ LOG(ERROR) << "Could not delete snapshot: " << name;
+ return false;
+ }
+ return true;
+}
+
+bool SnapshotManager::CollapseSnapshotDevice(const std::string& name,
+ const SnapshotStatus& status) {
+ // Ideally, we would complete the following steps to collapse the device:
+ // (1) Rewrite the snapshot table to be identical to the base device table.
+ // (2) Rewrite the verity table to use the "snapshot" (now linear) device.
+ // (3) Delete the base device.
+ //
+ // This should be possible once libsnapshot understands LpMetadata. In the
+ // meantime, we implement a simpler solution: rewriting the snapshot table
+ // to be a single dm-linear segment against the base device. While not as
+ // ideal, it still lets us remove the COW device. We can remove this
+ // implementation once the new method has been tested.
+ auto& dm = DeviceMapper::Instance();
+ auto dm_name = GetSnapshotDeviceName(name, status);
+
+ DeviceMapper::TargetInfo target;
+ if (!GetSingleTarget(dm_name, TableQuery::Table, &target)) {
+ return false;
+ }
+ if (DeviceMapper::GetTargetType(target.spec) != "snapshot-merge") {
+ // This should be impossible, it was checked above.
+ LOG(ERROR) << "Snapshot device has invalid target type: " << dm_name;
+ return false;
+ }
+
+ std::string base_device, cow_device;
+ if (!DmTargetSnapshot::GetDevicesFromParams(target.data, &base_device, &cow_device)) {
+ LOG(ERROR) << "Could not parse snapshot device " << dm_name
+ << " parameters: " << target.data;
+ return false;
+ }
+
+ uint64_t num_sectors = status.snapshot_size / kSectorSize;
+ if (num_sectors * kSectorSize != status.snapshot_size) {
+ LOG(ERROR) << "Snapshot " << name
+ << " size is not sector aligned: " << status.snapshot_size;
+ return false;
+ }
+
+ if (dm_name != name) {
+ // We've derived the base device, but we actually need to replace the
+ // table of the outermost device. Do a quick verification that this
+ // device looks like we expect it to.
+ std::vector<DeviceMapper::TargetInfo> outer_table;
+ if (!dm.GetTableInfo(name, &outer_table)) {
+ LOG(ERROR) << "Could not validate outer snapshot table: " << name;
+ return false;
+ }
+ if (outer_table.size() != 2) {
+ LOG(ERROR) << "Expected 2 dm-linear targets for tabble " << name
+ << ", got: " << outer_table.size();
+ return false;
+ }
+ for (const auto& target : outer_table) {
+ auto target_type = DeviceMapper::GetTargetType(target.spec);
+ if (target_type != "linear") {
+ LOG(ERROR) << "Outer snapshot table may only contain linear targets, but " << name
+ << " has target: " << target_type;
+ return false;
+ }
+ }
+ uint64_t sectors = outer_table[0].spec.length + outer_table[1].spec.length;
+ if (sectors != num_sectors) {
+ LOG(ERROR) << "Outer snapshot " << name << " should have " << num_sectors
+ << ", got: " << sectors;
+ return false;
+ }
+ }
+
+ // Note: we are replacing the OUTER table here, so we do not use dm_name.
+ DmTargetLinear new_target(0, num_sectors, base_device, 0);
+ LOG(INFO) << "Replacing snapshot device " << name
+ << " table with: " << new_target.GetParameterString();
+
+ DmTable table;
+ table.Emplace<DmTargetLinear>(new_target);
+ if (!dm.LoadTableAndActivate(name, table)) {
+ return false;
+ }
+
+ if (dm_name != name) {
+ // Attempt to delete the snapshot device. Nothing should be depending on
+ // the device, and device-mapper should have flushed remaining I/O. We
+ // could in theory replace with dm-zero (or re-use the table above), but
+ // for now it's better to know why this would fail.
+ if (!dm.DeleteDevice(dm_name)) {
+ LOG(ERROR) << "Unable to delete snapshot device " << dm_name << ", COW cannot be "
+ << "reclaimed until after reboot.";
+ return false;
+ }
+ }
+ return true;
}
bool SnapshotManager::RemoveAllSnapshots(LockedFile* lock) {
@@ -342,6 +846,12 @@
}
UpdateState SnapshotManager::GetUpdateState(double* progress) {
+ // If we've never started an update, the state file won't exist.
+ auto state_file = GetStateFilePath();
+ if (access(state_file.c_str(), F_OK) != 0 && errno == ENOENT) {
+ return UpdateState::None;
+ }
+
auto file = LockShared();
if (!file) {
return UpdateState::None;
@@ -388,7 +898,10 @@
PLOG(ERROR) << "Acquire flock failed: " << file;
return nullptr;
}
- return std::make_unique<LockedFile>(file, std::move(fd));
+ // For simplicity, we want to CHECK that lock_mode == LOCK_EX, in some
+ // calls, so strip extra flags.
+ int lock_mode = lock_flags & (LOCK_EX | LOCK_SH);
+ return std::make_unique<LockedFile>(file, std::move(fd), lock_mode);
}
SnapshotManager::LockedFile::~LockedFile() {
@@ -397,9 +910,13 @@
}
}
+std::string SnapshotManager::GetStateFilePath() const {
+ return metadata_dir_ + "/state"s;
+}
+
std::unique_ptr<SnapshotManager::LockedFile> SnapshotManager::OpenStateFile(int open_flags,
int lock_flags) {
- auto state_file = metadata_dir_ + "/state"s;
+ auto state_file = GetStateFilePath();
return OpenFile(state_file, open_flags, lock_flags);
}
@@ -434,6 +951,10 @@
return UpdateState::Merging;
} else if (contents == "merge-completed") {
return UpdateState::MergeCompleted;
+ } else if (contents == "merge-needs-reboot") {
+ return UpdateState::MergeNeedsReboot;
+ } else if (contents == "merge-failed") {
+ return UpdateState::MergeFailed;
} else {
LOG(ERROR) << "Unknown merge state in update state file";
return UpdateState::None;
@@ -458,6 +979,12 @@
case UpdateState::MergeCompleted:
contents = "merge-completed";
break;
+ case UpdateState::MergeNeedsReboot:
+ contents = "merge-needs-reboot";
+ break;
+ case UpdateState::MergeFailed:
+ contents = "merge-failed";
+ break;
default:
LOG(ERROR) << "Unknown update state";
return false;
@@ -471,51 +998,66 @@
return true;
}
-auto SnapshotManager::OpenSnapshotStatusFile(const std::string& name, int open_flags,
- int lock_flags) -> std::unique_ptr<LockedFile> {
+std::string SnapshotManager::GetSnapshotStatusFilePath(const std::string& name) {
auto file = metadata_dir_ + "/snapshots/"s + name;
- return OpenFile(file, open_flags, lock_flags);
+ return file;
}
-bool SnapshotManager::ReadSnapshotStatus(LockedFile* file, SnapshotStatus* status) {
- // Reset position since some calls read+write.
- if (lseek(file->fd(), 0, SEEK_SET) < 0) {
- PLOG(ERROR) << "lseek status file failed";
+bool SnapshotManager::ReadSnapshotStatus(LockedFile* lock, const std::string& name,
+ SnapshotStatus* status) {
+ CHECK(lock);
+ auto path = GetSnapshotStatusFilePath(name);
+
+ unique_fd fd(open(path.c_str(), O_RDONLY | O_CLOEXEC | O_NOFOLLOW));
+ if (fd < 0) {
+ PLOG(ERROR) << "Open failed: " << path;
return false;
}
std::string contents;
- if (!android::base::ReadFdToString(file->fd(), &contents)) {
- PLOG(ERROR) << "read status file failed";
+ if (!android::base::ReadFdToString(fd, &contents)) {
+ PLOG(ERROR) << "read failed: " << path;
return false;
}
auto pieces = android::base::Split(contents, " ");
if (pieces.size() != 5) {
- LOG(ERROR) << "Invalid status line for snapshot: " << file->path();
+ LOG(ERROR) << "Invalid status line for snapshot: " << path;
return false;
}
status->state = pieces[0];
if (!android::base::ParseUint(pieces[1], &status->device_size)) {
- LOG(ERROR) << "Invalid device size in status line for: " << file->path();
+ LOG(ERROR) << "Invalid device size in status line for: " << path;
return false;
}
if (!android::base::ParseUint(pieces[2], &status->snapshot_size)) {
- LOG(ERROR) << "Invalid snapshot size in status line for: " << file->path();
+ LOG(ERROR) << "Invalid snapshot size in status line for: " << path;
return false;
}
if (!android::base::ParseUint(pieces[3], &status->sectors_allocated)) {
- LOG(ERROR) << "Invalid snapshot size in status line for: " << file->path();
+ LOG(ERROR) << "Invalid snapshot size in status line for: " << path;
return false;
}
if (!android::base::ParseUint(pieces[4], &status->metadata_sectors)) {
- LOG(ERROR) << "Invalid snapshot size in status line for: " << file->path();
+ LOG(ERROR) << "Invalid snapshot size in status line for: " << path;
return false;
}
return true;
}
-bool SnapshotManager::WriteSnapshotStatus(LockedFile* file, const SnapshotStatus& status) {
+bool SnapshotManager::WriteSnapshotStatus(LockedFile* lock, const std::string& name,
+ const SnapshotStatus& status) {
+ // The caller must take an exclusive lock to modify snapshots.
+ CHECK(lock);
+ CHECK(lock->lock_mode() == LOCK_EX);
+
+ auto path = GetSnapshotStatusFilePath(name);
+ unique_fd fd(open(path.c_str(), O_RDWR | O_CLOEXEC | O_NOFOLLOW | O_CREAT | O_SYNC, 0660));
+ if (fd < 0) {
+ PLOG(ERROR) << "Open failed: " << path;
+ return false;
+ }
+
std::vector<std::string> pieces = {
status.state,
std::to_string(status.device_size),
@@ -525,9 +1067,8 @@
};
auto contents = android::base::Join(pieces, " ");
- if (!Truncate(file)) return false;
- if (!android::base::WriteStringToFd(contents, file->fd())) {
- PLOG(ERROR) << "write to status file failed: " << file->path();
+ if (!android::base::WriteStringToFd(contents, fd)) {
+ PLOG(ERROR) << "write failed: " << path;
return false;
}
return true;
diff --git a/fs_mgr/libsnapshot/snapshot_test.cpp b/fs_mgr/libsnapshot/snapshot_test.cpp
index 9cc9bd7..4903224 100644
--- a/fs_mgr/libsnapshot/snapshot_test.cpp
+++ b/fs_mgr/libsnapshot/snapshot_test.cpp
@@ -22,13 +22,20 @@
#include <chrono>
#include <iostream>
+#include <android-base/file.h>
+#include <android-base/properties.h>
#include <android-base/strings.h>
+#include <android-base/unique_fd.h>
#include <gtest/gtest.h>
+#include <libdm/dm.h>
#include <libfiemap/image_manager.h>
namespace android {
namespace snapshot {
+using android::base::unique_fd;
+using android::dm::DeviceMapper;
+using android::dm::DmDeviceState;
using namespace std::chrono_literals;
using namespace std::string_literals;
@@ -48,12 +55,15 @@
TestDeviceInfo* test_device = nullptr;
class SnapshotTest : public ::testing::Test {
+ public:
+ SnapshotTest() : dm_(DeviceMapper::Instance()) {}
+
protected:
void SetUp() override {
test_device->set_is_running_snapshot(false);
if (sm->GetUpdateState() != UpdateState::None) {
- ASSERT_TRUE(sm->CancelUpdate());
+ CleanupTestArtifacts();
}
ASSERT_TRUE(sm->BeginUpdate());
ASSERT_TRUE(sm->EnsureImageManager());
@@ -65,13 +75,32 @@
void TearDown() override {
lock_ = nullptr;
- if (sm->GetUpdateState() != UpdateState::None) {
- ASSERT_TRUE(sm->CancelUpdate());
+ CleanupTestArtifacts();
+ }
+
+ void CleanupTestArtifacts() {
+ // Normally cancelling inside a merge is not allowed. Since these
+ // are tests, we don't care, destroy everything that might exist.
+ std::vector<std::string> snapshots = {"test-snapshot"};
+ for (const auto& snapshot : snapshots) {
+ DeleteSnapshotDevice(snapshot);
+ temp_images_.emplace_back(snapshot + "-cow");
+
+ auto status_file = sm->GetSnapshotStatusFilePath(snapshot);
+ android::base::RemoveFileIfExists(status_file);
}
+
+ // Remove all images.
+ temp_images_.emplace_back("test-snapshot-cow");
for (const auto& temp_image : temp_images_) {
image_manager_->UnmapImageDevice(temp_image);
image_manager_->DeleteBackingImage(temp_image);
}
+
+ if (sm->GetUpdateState() != UpdateState::None) {
+ auto state_file = sm->GetStateFilePath();
+ unlink(state_file.c_str());
+ }
}
bool AcquireLock() {
@@ -87,6 +116,17 @@
return image_manager_->MapImageDevice(name, 10s, path);
}
+ bool DeleteSnapshotDevice(const std::string& snapshot) {
+ if (dm_.GetState(snapshot) != DmDeviceState::INVALID) {
+ if (!dm_.DeleteDevice(snapshot)) return false;
+ }
+ if (dm_.GetState(snapshot + "-inner") != DmDeviceState::INVALID) {
+ if (!dm_.DeleteDevice(snapshot + "-inner")) return false;
+ }
+ return true;
+ }
+
+ DeviceMapper& dm_;
std::unique_ptr<SnapshotManager::LockedFile> lock_;
std::vector<std::string> temp_images_;
android::fiemap::IImageManager* image_manager_ = nullptr;
@@ -106,11 +146,8 @@
// Scope so delete can re-acquire the snapshot file lock.
{
- auto file = sm->OpenSnapshotStatusFile("test-snapshot", O_RDONLY, LOCK_SH);
- ASSERT_NE(file, nullptr);
-
SnapshotManager::SnapshotStatus status;
- ASSERT_TRUE(sm->ReadSnapshotStatus(file.get(), &status));
+ ASSERT_TRUE(sm->ReadSnapshotStatus(lock_.get(), "test-snapshot", &status));
ASSERT_EQ(status.state, "created");
ASSERT_EQ(status.device_size, kDeviceSize);
ASSERT_EQ(status.snapshot_size, kDeviceSize);
@@ -151,6 +188,117 @@
ASSERT_TRUE(android::base::StartsWith(snap_device, "/dev/block/dm-"));
}
+TEST_F(SnapshotTest, NoMergeBeforeReboot) {
+ ASSERT_TRUE(AcquireLock());
+
+ // Set the state to Unverified, as if we finished an update.
+ ASSERT_TRUE(sm->WriteUpdateState(lock_.get(), UpdateState::Unverified));
+
+ // Release the lock.
+ lock_ = nullptr;
+
+ // Merge should fail, since we didn't mark the device as rebooted.
+ ASSERT_FALSE(sm->InitiateMerge());
+}
+
+TEST_F(SnapshotTest, Merge) {
+ ASSERT_TRUE(AcquireLock());
+
+ static const uint64_t kDeviceSize = 1024 * 1024;
+ ASSERT_TRUE(sm->CreateSnapshot(lock_.get(), "test-snapshot", kDeviceSize, kDeviceSize,
+ kDeviceSize));
+
+ std::string base_device, snap_device;
+ ASSERT_TRUE(CreateTempDevice("base-device", kDeviceSize, &base_device));
+ ASSERT_TRUE(sm->MapSnapshot(lock_.get(), "test-snapshot", base_device, 10s, &snap_device));
+
+ std::string test_string = "This is a test string.";
+ {
+ unique_fd fd(open(snap_device.c_str(), O_RDWR | O_CLOEXEC | O_SYNC));
+ ASSERT_GE(fd, 0);
+ ASSERT_TRUE(android::base::WriteFully(fd, test_string.data(), test_string.size()));
+ }
+
+ // Note: we know the name of the device is test-snapshot because we didn't
+ // request a linear segment.
+ DeviceMapper::TargetInfo target;
+ ASSERT_TRUE(sm->IsSnapshotDevice("test-snapshot", &target));
+ ASSERT_EQ(DeviceMapper::GetTargetType(target.spec), "snapshot");
+
+ // Set the state to Unverified, as if we finished an update.
+ ASSERT_TRUE(sm->WriteUpdateState(lock_.get(), UpdateState::Unverified));
+
+ // Release the lock.
+ lock_ = nullptr;
+
+ test_device->set_is_running_snapshot(true);
+ ASSERT_TRUE(sm->InitiateMerge());
+
+ // The device should have been switched to a snapshot-merge target.
+ ASSERT_TRUE(sm->IsSnapshotDevice("test-snapshot", &target));
+ ASSERT_EQ(DeviceMapper::GetTargetType(target.spec), "snapshot-merge");
+
+ // We should not be able to cancel an update now.
+ ASSERT_FALSE(sm->CancelUpdate());
+
+ ASSERT_EQ(sm->WaitForMerge(), UpdateState::MergeCompleted);
+ ASSERT_EQ(sm->GetUpdateState(), UpdateState::None);
+
+ // The device should no longer be a snapshot or snapshot-merge.
+ ASSERT_FALSE(sm->IsSnapshotDevice("test-snapshot"));
+
+ // Test that we can read back the string we wrote to the snapshot.
+ unique_fd fd(open(base_device.c_str(), O_RDONLY | O_CLOEXEC));
+ ASSERT_GE(fd, 0);
+
+ std::string buffer(test_string.size(), '\0');
+ ASSERT_TRUE(android::base::ReadFully(fd, buffer.data(), buffer.size()));
+ ASSERT_EQ(test_string, buffer);
+}
+
+TEST_F(SnapshotTest, MergeCannotRemoveCow) {
+ ASSERT_TRUE(AcquireLock());
+
+ static const uint64_t kDeviceSize = 1024 * 1024;
+ ASSERT_TRUE(sm->CreateSnapshot(lock_.get(), "test-snapshot", kDeviceSize, kDeviceSize,
+ kDeviceSize));
+
+ std::string base_device, snap_device;
+ ASSERT_TRUE(CreateTempDevice("base-device", kDeviceSize, &base_device));
+ ASSERT_TRUE(sm->MapSnapshot(lock_.get(), "test-snapshot", base_device, 10s, &snap_device));
+
+ // Keep an open handle to the cow device. This should cause the merge to
+ // be incomplete.
+ auto cow_path = android::base::GetProperty("gsid.mapped_image.test-snapshot-cow", "");
+ unique_fd fd(open(cow_path.c_str(), O_RDONLY | O_CLOEXEC));
+ ASSERT_GE(fd, 0);
+
+ // Set the state to Unverified, as if we finished an update.
+ ASSERT_TRUE(sm->WriteUpdateState(lock_.get(), UpdateState::Unverified));
+
+ // Release the lock.
+ lock_ = nullptr;
+
+ test_device->set_is_running_snapshot(true);
+ ASSERT_TRUE(sm->InitiateMerge());
+
+ // COW cannot be removed due to open fd, so expect a soft failure.
+ ASSERT_EQ(sm->WaitForMerge(), UpdateState::MergeNeedsReboot);
+
+ // Forcefully delete the snapshot device, so it looks like we just rebooted.
+ ASSERT_TRUE(DeleteSnapshotDevice("test-snapshot"));
+
+ // Map snapshot should fail now, because we're in a merge-complete state.
+ ASSERT_TRUE(AcquireLock());
+ ASSERT_FALSE(sm->MapSnapshot(lock_.get(), "test-snapshot", base_device, 10s, &snap_device));
+
+ // Release everything and now the merge should complete.
+ fd = {};
+ lock_ = nullptr;
+
+ ASSERT_EQ(sm->WaitForMerge(), UpdateState::MergeCompleted);
+}
+
} // namespace snapshot
} // namespace android