Merge "Try to avoid overlapping job execution." into tm-dev

This commit is contained in:
Kweku Adams
2022-05-04 19:05:14 +00:00
committed by Android (Google) Code Review
5 changed files with 93 additions and 11 deletions

View File

@@ -1777,7 +1777,10 @@ public class JobInfo implements Parcelable {
* {@link Build.VERSION_CODES#S}, but starting from Android version
* {@link Build.VERSION_CODES#TIRAMISU}, expedited jobs for the foreground app are
* guaranteed to be started before {@link JobScheduler#schedule(JobInfo)} returns (assuming
* all requested constraints are satisfied), similar to foreground services.
* all requested constraints are satisfied), similar to foreground services. However, this
* start guarantee means there is a higher chance of overlapping executions, as noted in
* {@link JobService}, so be sure to handle that properly if you intend to reschedule the
* job while it's actively running.
*
* @see JobInfo#isExpedited()
*/

View File

@@ -107,7 +107,9 @@ public abstract class JobScheduler {
/**
* Schedule a job to be executed. Will replace any currently scheduled job with the same
* ID with the new information in the {@link JobInfo}. If a job with the given ID is currently
* running, it will be stopped.
* running, it will be stopped. Note that in some cases, the newly scheduled job may be started
* before the previously running job has been fully stopped. See {@link JobService} for
* additional details.
*
* <p class="caution"><strong>Note:</strong> Scheduling a job can have a high cost, even if it's
* rescheduling the same job and the job didn't execute, especially on platform versions before
@@ -131,7 +133,7 @@ public abstract class JobScheduler {
* job. If a job with the same ID is already scheduled, it will be replaced with the
* new {@link JobInfo}, but any previously enqueued work will remain and be dispatched the
* next time it runs. If a job with the same ID is already running, the new work will be
* enqueued for it.
* enqueued for it without stopping the job.
*
* <p>The work you enqueue is later retrieved through
* {@link JobParameters#dequeueWork() JobParameters.dequeueWork}. Be sure to see there

View File

@@ -31,6 +31,17 @@ import android.os.IBinder;
* in blocking any future callbacks from the JobManager - specifically
* {@link #onStopJob(android.app.job.JobParameters)}, which is meant to inform you that the
* scheduling requirements are no longer being met.</p>
*
* As a subclass of {@link Service}, there will only be one active instance of any JobService
* subclasses, regardless of job ID. This means that if you schedule multiple jobs with different
* job IDs but using the same JobService class, that JobService may receive multiple calls to
* {@link #onStartJob(JobParameters)} and {@link #onStopJob(JobParameters)}, with each call being
* for the separate jobs.
*
* <p class="note">Note that if you cancel and reschedule an already executing job,
* there may be a small period of time where {@link #onStartJob(JobParameters)} has been called for
* the newly scheduled job instance before {@link #onStopJob(JobParameters)} has been called or
* fully processed for the old job.</p>
*/
public abstract class JobService extends Service {
private static final String TAG = "JobService";

View File

@@ -542,6 +542,22 @@ class JobConcurrencyManager {
return mRunningJobs.contains(job);
}
/**
* Returns true if a job that is "similar" to the provided job is currently running.
* "Similar" in this context means any job that the {@link JobStore} would consider equivalent
* and replace one with the other.
*/
@GuardedBy("mLock")
private boolean isSimilarJobRunningLocked(JobStatus job) {
for (int i = mRunningJobs.size() - 1; i >= 0; --i) {
JobStatus js = mRunningJobs.valueAt(i);
if (job.getUid() == js.getUid() && job.getJobId() == js.getJobId()) {
return true;
}
}
return false;
}
/** Return {@code true} if the state was updated. */
@GuardedBy("mLock")
private boolean refreshSystemStateLocked() {
@@ -699,6 +715,21 @@ class JobConcurrencyManager {
continue;
}
final boolean isTopEj = nextPending.shouldTreatAsExpeditedJob()
&& nextPending.lastEvaluatedBias == JobInfo.BIAS_TOP_APP;
// Avoid overlapping job execution as much as possible.
if (!isTopEj && isSimilarJobRunningLocked(nextPending)) {
if (DEBUG) {
Slog.w(TAG, "Delaying execution of job because of similarly running one: "
+ nextPending);
}
// It would be nice to let the JobService running the other similar job know about
// this new job so that it doesn't unbind from the JobService and we can call
// onStartJob as soon as the older job finishes.
// TODO: optimize the job reschedule flow to reduce service binding churn
continue;
}
// Find an available slot for nextPending. The context should be one of the following:
// 1. Unused
// 2. Its job should have used up its minimum execution guarantee so it
@@ -707,8 +738,6 @@ class JobConcurrencyManager {
ContextAssignment selectedContext = null;
final int allWorkTypes = getJobWorkTypes(nextPending);
final boolean pkgConcurrencyOkay = !isPkgConcurrencyLimitedLocked(nextPending);
final boolean isTopEj = nextPending.shouldTreatAsExpeditedJob()
&& nextPending.lastEvaluatedBias == JobInfo.BIAS_TOP_APP;
final boolean isInOverage = projectedRunningCount > STANDARD_CONCURRENCY_LIMIT;
boolean startingJob = false;
if (idle.size() > 0) {
@@ -1177,6 +1206,15 @@ class JobConcurrencyManager {
continue;
}
// Avoid overlapping job execution as much as possible.
if (isSimilarJobRunningLocked(nextPending)) {
if (DEBUG) {
Slog.w(TAG, "Avoiding execution of job because of similarly running one: "
+ nextPending);
}
continue;
}
if (worker.getPreferredUid() != nextPending.getUid()) {
if (backupJob == null && !isPkgConcurrencyLimitedLocked(nextPending)) {
int allWorkTypes = getJobWorkTypes(nextPending);
@@ -1260,6 +1298,15 @@ class JobConcurrencyManager {
continue;
}
// Avoid overlapping job execution as much as possible.
if (isSimilarJobRunningLocked(nextPending)) {
if (DEBUG) {
Slog.w(TAG, "Avoiding execution of job because of similarly running one: "
+ nextPending);
}
continue;
}
if (isPkgConcurrencyLimitedLocked(nextPending)) {
continue;
}

View File

@@ -1208,12 +1208,22 @@ public class JobSchedulerService extends com.android.server.SystemService
// This may throw a SecurityException.
jobStatus.prepareLocked();
final boolean canExecuteImmediately;
if (toCancel != null) {
// Implicitly replaces the existing job record with the new instance
cancelJobImplLocked(toCancel, jobStatus, JobParameters.STOP_REASON_CANCELLED_BY_APP,
JobParameters.INTERNAL_STOP_REASON_CANCELED, "job rescheduled by app");
final boolean wasJobExecuting = cancelJobImplLocked(toCancel, jobStatus,
JobParameters.STOP_REASON_CANCELLED_BY_APP,
JobParameters.INTERNAL_STOP_REASON_CANCELED,
"job rescheduled by app");
// Avoid overlapping job executions. Don't push for immediate execution if an old
// job with the same ID was running, but let TOP EJs start immediately.
canExecuteImmediately = !wasJobExecuting
|| (jobStatus.isRequestedExpeditedJob()
&& mUidBiasOverride.get(jobStatus.getSourceUid(), JobInfo.BIAS_DEFAULT)
== JobInfo.BIAS_TOP_APP);
} else {
startTrackingJobLocked(jobStatus, null);
canExecuteImmediately = true;
}
if (work != null) {
@@ -1256,7 +1266,12 @@ public class JobSchedulerService extends com.android.server.SystemService
// list and try to run it.
mJobPackageTracker.notePending(jobStatus);
mPendingJobQueue.add(jobStatus);
maybeRunPendingJobsLocked();
if (canExecuteImmediately) {
// Don't ask the JobConcurrencyManager to try to run the job immediately. The
// JobServiceContext will ask the JobConcurrencyManager for another job once
// it finishes cleaning up the old job.
maybeRunPendingJobsLocked();
}
} else {
evaluateControllerStatesLocked(jobStatus);
}
@@ -1377,8 +1392,10 @@ public class JobSchedulerService extends com.android.server.SystemService
* is null, the cancelled job is removed outright from the system. If
* {@code incomingJob} is non-null, it replaces {@code cancelled} in the store of
* currently scheduled jobs.
*
* @return true if the cancelled job was running
*/
private void cancelJobImplLocked(JobStatus cancelled, JobStatus incomingJob,
private boolean cancelJobImplLocked(JobStatus cancelled, JobStatus incomingJob,
@JobParameters.StopReason int reason, int internalReasonCode, String debugReason) {
if (DEBUG) Slog.d(TAG, "CANCEL: " + cancelled.toShortString());
cancelled.unprepareLocked();
@@ -1389,7 +1406,7 @@ public class JobSchedulerService extends com.android.server.SystemService
}
mChangedJobList.remove(cancelled);
// Cancel if running.
mConcurrencyManager.stopJobOnServiceContextLocked(
boolean wasRunning = mConcurrencyManager.stopJobOnServiceContextLocked(
cancelled, reason, internalReasonCode, debugReason);
// If this is a replacement, bring in the new version of the job
if (incomingJob != null) {
@@ -1397,6 +1414,7 @@ public class JobSchedulerService extends com.android.server.SystemService
startTrackingJobLocked(incomingJob, cancelled);
}
reportActiveLocked();
return wasRunning;
}
void updateUidState(int uid, int procState) {
@@ -1755,7 +1773,7 @@ public class JobSchedulerService extends com.android.server.SystemService
// same job ID), we remove it from the JobStore and tell the JobServiceContext to stop
// running the job. Once the job stops running, we then call this method again.
// TODO: rework code so we don't intentionally call this method twice for the same job
Slog.w(TAG, "Job didn't exist in JobStore");
Slog.w(TAG, "Job didn't exist in JobStore: " + jobStatus.toShortString());
}
if (mReadyToRock) {
for (int i = 0; i < mControllers.size(); i++) {
@@ -3813,6 +3831,7 @@ public class JobSchedulerService extends com.android.server.SystemService
// Double indent for readability
pw.increaseIndent();
pw.increaseIndent();
pw.println(job.toShortString());
job.dump(pw, true, nowElapsed);
pw.decreaseIndent();
pw.decreaseIndent();