Merge "libsnapshot: Restrict when snapshots can be deleted."

This commit is contained in:
David Anderson 2020-02-06 16:46:58 +00:00 committed by Gerrit Code Review
commit 24c70ea7ba
5 changed files with 104 additions and 49 deletions

View File

@ -22,6 +22,7 @@ namespace android {
namespace snapshot {
#ifdef LIBSNAPSHOT_USE_HAL
using android::hardware::boot::V1_0::BoolResult;
using android::hardware::boot::V1_0::CommandResult;
#endif

View File

@ -174,6 +174,7 @@ class SnapshotManager final {
// See InitiateMerge() and ProcessUpdateState() for details.
// Returns:
// - None if no merge to initiate
// - Unverified if called on the source slot
// - MergeCompleted if merge is completed
// - other states indicating an error has occurred
UpdateState InitiateMergeAndWait();
@ -273,6 +274,7 @@ class SnapshotManager final {
FRIEND_TEST(SnapshotTest, UpdateBootControlHal);
FRIEND_TEST(SnapshotUpdateTest, DataWipeAfterRollback);
FRIEND_TEST(SnapshotUpdateTest, DataWipeRollbackInRecovery);
FRIEND_TEST(SnapshotUpdateTest, FullUpdateFlow);
FRIEND_TEST(SnapshotUpdateTest, MergeCannotRemoveCow);
FRIEND_TEST(SnapshotUpdateTest, MergeInRecovery);
FRIEND_TEST(SnapshotUpdateTest, SnapshotStatusFileWithoutCow);
@ -374,7 +376,7 @@ class SnapshotManager final {
bool HandleCancelledUpdate(LockedFile* lock);
// Helper for HandleCancelledUpdate. Assumes booting from new slot.
bool HandleCancelledUpdateOnNewSlot(LockedFile* lock);
bool AreAllSnapshotsCancelled(LockedFile* lock);
// Remove artifacts created by the update process, such as snapshots, and
// set the update state to None.
@ -439,7 +441,7 @@ class SnapshotManager final {
std::string GetSnapshotStatusFilePath(const std::string& name);
std::string GetSnapshotBootIndicatorPath();
void RemoveSnapshotBootIndicator();
std::string GetRollbackIndicatorPath();
// Return the name of the device holding the "snapshot" or "snapshot-merge"
// target. This may not be the final device presented via MapSnapshot(), if
@ -503,6 +505,8 @@ class SnapshotManager final {
friend std::ostream& operator<<(std::ostream& os, SnapshotManager::Slot slot);
Slot GetCurrentSlot();
std::string ReadUpdateSourceSlotSuffix();
std::string gsid_dir_;
std::string metadata_dir_;
std::unique_ptr<IDeviceInfo> device_;

View File

@ -188,11 +188,19 @@ bool SnapshotManager::TryCancelUpdate(bool* needs_merge) {
return true;
}
SnapshotManager::Slot SnapshotManager::GetCurrentSlot() {
std::string SnapshotManager::ReadUpdateSourceSlotSuffix() {
auto boot_file = GetSnapshotBootIndicatorPath();
std::string contents;
if (!android::base::ReadFileToString(boot_file, &contents)) {
PLOG(WARNING) << "Cannot read " << boot_file;
return {};
}
return contents;
}
SnapshotManager::Slot SnapshotManager::GetCurrentSlot() {
auto contents = ReadUpdateSourceSlotSuffix();
if (contents.empty()) {
return Slot::Unknown;
}
if (device_->GetSlotSuffix() == contents) {
@ -201,6 +209,15 @@ SnapshotManager::Slot SnapshotManager::GetCurrentSlot() {
return Slot::Target;
}
static bool RemoveFileIfExists(const std::string& path) {
std::string message;
if (!android::base::RemoveFileIfExists(path, &message)) {
LOG(ERROR) << "Remove failed: " << path << ": " << message;
return false;
}
return true;
}
bool SnapshotManager::RemoveAllUpdateState(LockedFile* lock) {
LOG(INFO) << "Removing all update state.";
@ -223,7 +240,13 @@ bool SnapshotManager::RemoveAllUpdateState(LockedFile* lock) {
return false;
}
RemoveSnapshotBootIndicator();
// It's okay if these fail - first-stage init performs a deeper check after
// reading the indicator file, so it's not a problem if it still exists
// after the update completes.
std::vector<std::string> files = {GetSnapshotBootIndicatorPath(), GetRollbackIndicatorPath()};
for (const auto& file : files) {
RemoveFileIfExists(file);
}
// If this fails, we'll keep trying to remove the update state (as the
// device reboots or starts a new update) until it finally succeeds.
@ -250,6 +273,13 @@ bool SnapshotManager::FinishedSnapshotWrites() {
return false;
}
// This file is written on boot to detect whether a rollback occurred. It
// MUST NOT exist before rebooting, otherwise, we're at risk of deleting
// snapshots too early.
if (!RemoveFileIfExists(GetRollbackIndicatorPath())) {
return false;
}
// This file acts as both a quick indicator for init (it can use access(2)
// to decide how to do first-stage mounts), and it stores the old slot, so
// we can tell whether or not we performed a rollback.
@ -966,14 +996,8 @@ std::string SnapshotManager::GetSnapshotBootIndicatorPath() {
return metadata_dir_ + "/" + android::base::Basename(kBootIndicatorPath);
}
void SnapshotManager::RemoveSnapshotBootIndicator() {
// It's okay if this fails - first-stage init performs a deeper check after
// reading the indicator file, so it's not a problem if it still exists
// after the update completes.
auto boot_file = GetSnapshotBootIndicatorPath();
if (unlink(boot_file.c_str()) == -1 && errno != ENOENT) {
PLOG(ERROR) << "unlink " << boot_file;
}
std::string SnapshotManager::GetRollbackIndicatorPath() {
return metadata_dir_ + "/rollback-indicator";
}
void SnapshotManager::AcknowledgeMergeSuccess(LockedFile* lock) {
@ -1144,25 +1168,18 @@ bool SnapshotManager::HandleCancelledUpdate(LockedFile* lock) {
if (slot == Slot::Unknown) {
return false;
}
if (slot == Slot::Target) {
// We're booted into the target slot, which means we just rebooted
// after applying the update.
if (!HandleCancelledUpdateOnNewSlot(lock)) {
return false;
}
// If all snapshots were reflashed, then cancel the entire update.
if (AreAllSnapshotsCancelled(lock)) {
RemoveAllUpdateState(lock);
return true;
}
// The only way we can get here is if:
// (1) The device rolled back to the previous slot.
// (2) This function was called prematurely before rebooting the device.
// (3) fastboot set_active was used.
// (4) The device updates to the new slot but re-flashed *all* partitions
// in the new slot.
//
// In any case, delete the snapshots. It may be worth using the boot_control
// HAL to differentiate case (2).
RemoveAllUpdateState(lock);
return true;
// This unverified update might be rolled back, or it might not (b/147347110
// comment #77). Take no action, as update_engine is responsible for deciding
// whether to cancel.
LOG(ERROR) << "Update state is being processed before reboot, taking no action.";
return false;
}
std::unique_ptr<LpMetadata> SnapshotManager::ReadCurrentMetadata() {
@ -1187,7 +1204,7 @@ SnapshotManager::MetadataPartitionState SnapshotManager::GetMetadataPartitionSta
return MetadataPartitionState::Flashed;
}
bool SnapshotManager::HandleCancelledUpdateOnNewSlot(LockedFile* lock) {
bool SnapshotManager::AreAllSnapshotsCancelled(LockedFile* lock) {
std::vector<std::string> snapshots;
if (!ListSnapshots(lock, &snapshots)) {
LOG(WARNING) << "Failed to list snapshots to determine whether device has been flashed "
@ -1196,35 +1213,45 @@ bool SnapshotManager::HandleCancelledUpdateOnNewSlot(LockedFile* lock) {
return true;
}
auto source_slot_suffix = ReadUpdateSourceSlotSuffix();
if (source_slot_suffix.empty()) {
return false;
}
uint32_t source_slot = SlotNumberForSlotSuffix(source_slot_suffix);
uint32_t target_slot = (source_slot == 0) ? 1 : 0;
// Attempt to detect re-flashing on each partition.
// - If all partitions are re-flashed, we can proceed to cancel the whole update.
// - If only some of the partitions are re-flashed, snapshots for re-flashed partitions are
// deleted. Caller is responsible for merging the rest of the snapshots.
// - If none of the partitions are re-flashed, caller is responsible for merging the snapshots.
auto metadata = ReadCurrentMetadata();
if (!metadata) return false;
bool all_snapshot_cancelled = true;
//
// Note that we use target slot metadata, since if an OTA has been applied
// to the target slot, we can detect the UPDATED flag. Any kind of flash
// operation against dynamic partitions ensures that all copies of the
// metadata are in sync, so flashing all partitions on the source slot will
// remove the UPDATED flag on the target slot as well.
const auto& opener = device_->GetPartitionOpener();
auto super_device = device_->GetSuperDevice(target_slot);
auto metadata = android::fs_mgr::ReadMetadata(opener, super_device, target_slot);
if (!metadata) {
return false;
}
bool all_snapshots_cancelled = true;
for (const auto& snapshot_name : snapshots) {
if (GetMetadataPartitionState(*metadata, snapshot_name) ==
MetadataPartitionState::Updated) {
LOG(WARNING) << "Cannot cancel update because snapshot" << snapshot_name
<< " is in use.";
all_snapshot_cancelled = false;
all_snapshots_cancelled = false;
continue;
}
// Delete snapshots for partitions that are re-flashed after the update.
LOG(INFO) << "Detected re-flashing of partition " << snapshot_name << ".";
if (!DeleteSnapshot(lock, snapshot_name)) {
// This is an error, but it is okay to leave the snapshot in the short term.
// However, if all_snapshot_cancelled == false after exiting the loop, caller may
// initiate merge for this unused snapshot, which is likely to fail.
LOG(WARNING) << "Failed to delete snapshot for re-flashed partition " << snapshot_name;
}
LOG(WARNING) << "Detected re-flashing of partition " << snapshot_name << ".";
}
if (!all_snapshot_cancelled) return false;
LOG(INFO) << "All partitions are re-flashed after update, removing all update states.";
return true;
if (all_snapshots_cancelled) {
LOG(WARNING) << "All partitions are re-flashed after update, removing all update states.";
}
return all_snapshots_cancelled;
}
bool SnapshotManager::RemoveAllSnapshots(LockedFile* lock) {
@ -1344,7 +1371,16 @@ bool SnapshotManager::NeedSnapshotsInFirstStageMount() {
// the reason be clearer? Because the indicator file still exists, and
// if this was FATAL, reverting to the old slot would be broken.
auto slot = GetCurrentSlot();
if (slot != Slot::Target) {
if (slot == Slot::Source && !device_->IsRecovery()) {
// Device is rebooting into the original slot, so mark this as a
// rollback.
auto path = GetRollbackIndicatorPath();
if (!android::base::WriteStringToFile("1", path)) {
PLOG(ERROR) << "Unable to write rollback indicator: " << path;
}
}
LOG(INFO) << "Not booting from new slot. Will not mount snapshots.";
return false;
}
@ -2370,6 +2406,10 @@ UpdateState SnapshotManager::InitiateMergeAndWait() {
return state;
}
if (state == UpdateState::Unverified) {
if (GetCurrentSlot() != Slot::Target) {
LOG(INFO) << "Cannot merge until device reboots.";
return state;
}
if (!InitiateMerge()) {
LOG(ERROR) << "Failed to initiate merge.";
return state;

View File

@ -447,6 +447,9 @@ TEST_F(SnapshotTest, FirstStageMountAfterRollback) {
auto sm = SnapshotManager::NewForFirstStageMount(info);
ASSERT_NE(sm, nullptr);
ASSERT_FALSE(sm->NeedSnapshotsInFirstStageMount());
auto indicator = sm->GetRollbackIndicatorPath();
ASSERT_EQ(access(indicator.c_str(), R_OK), 0);
}
TEST_F(SnapshotTest, Merge) {
@ -1015,6 +1018,9 @@ TEST_F(SnapshotUpdateTest, FullUpdateFlow) {
ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
auto indicator = sm->GetRollbackIndicatorPath();
ASSERT_NE(access(indicator.c_str(), R_OK), 0);
// Check that the target partitions have the same content.
for (const auto& name : {"sys_b", "vnd_b", "prd_b"}) {
ASSERT_TRUE(IsPartitionUnchanged(name));
@ -1681,9 +1687,11 @@ TEST_P(FlashAfterUpdateTest, FlashSlotAfterUpdate) {
ASSERT_NE(nullptr, flashed_builder->FindPartition("prd" + flashed_slot_suffix));
flashed_builder->RemovePartition("prd" + flashed_slot_suffix);
// Note that fastbootd always updates the partition table of both slots.
auto flashed_metadata = flashed_builder->Export();
ASSERT_NE(nullptr, flashed_metadata);
ASSERT_TRUE(UpdatePartitionTable(*opener_, "super", *flashed_metadata, flashed_slot));
ASSERT_TRUE(UpdatePartitionTable(*opener_, "super", *flashed_metadata, 0));
ASSERT_TRUE(UpdatePartitionTable(*opener_, "super", *flashed_metadata, 1));
std::string path;
for (const auto& name : {"sys", "vnd"}) {

View File

@ -112,7 +112,9 @@ bool MergeCmdHandler(int argc, char** argv) {
auto state = SnapshotManager::New()->InitiateMergeAndWait();
if (state == UpdateState::None) {
// We could wind up in the Unverified state if the device rolled back or
// hasn't fully rebooted. Ignore this.
if (state == UpdateState::None || state == UpdateState::Unverified) {
return true;
}
if (state == UpdateState::MergeCompleted) {