Oom Adjuster Designs

Purpose of Oom Adjuster

The Android OS runs with limited hardware resources, i.e. CPU/RAM/Power. To strive for the better performance, Oom Adjuster is introduced to tweak the following 3 major factors:

  • Process State
    • Widely used by the System Server, i.e., determine if it's foreground or not, change the GC behavior, etc.
    • Defined in ActivityManager#PROCESS_STATE_*
  • Oom Adj score
    • Used by the lmkd to determine which process should be expunged on memory pressure.
    • Defined in ProcessList#*_ADJ
  • Scheduler Group
    • Used to tweak the process group, thread priorities.
    • Top process is scheduled to be running on a dedicated big core, while foreground processes take the other big cores; background processes stay with LITTLE cores instead.

Process Capabilities

Besides the above 3 major factors, Android R introduced the Process Capabilities ActivityManager#PROCESS_CAPABILITY_*. It's a new attribute to process record, mainly designed for supporting the “while-in-use” permission model - in addition to the traditional Android permissions, whether or not a process has access to a given API, will be guarded by its current process state as well. The OomAdjuster will compute the process capabilities during updating the oom adj. Meanwhile, the flag ActivityManager#BIND_INCLUDE_CAPABILITIES enables the possibility to “transfer” the capability from a client process to the service process it binds to.

Rationale of Oom Adjuster

System server keeps a list of recent used app processes. Given the 4 types of entities that an Android processes could have: Activity, Service, Content Provider and Broadcast Receiver, the System Server has to adjust the above 3 factors to give the users the best performance according to the states of the entities. A typical case would be that: foreground app A binds into a background service B in order to serve the user, in the case of memory pressure, the background service B should be avoided from being expunged since it would result in user-perceptible interruption of service. The Oom Adjuster is to tweak the aforementioned 3 factors for those app processes.

The timing of updating the Oom Adj score is vital: assume a camera process in background gets launched into foreground, launching camera typically incurs high memory pressure, which could incur low memory kills - if the camera process isn't moved out of the background adj group, it could get killed by lmkd. Therefore the updates have to be called pretty frequently: in case there is an activity start, service binding, etc.

The update procedure basically consists of 3 parts:

  • Find out the process record to be updated
    • There are two categories of updateOomAdjLocked: one with the target process record to be updated, while the other one is to update all process records.
    • Besides that, while computing the Oom Aj score, the clients of service connections or content providers of the present process record, which forms a process dependency graph actually, will be evaluated as well.
    • Starting from Android R, when updating a specific process record, an optimization is made that only the reachable process records starting from this process record in the process dependency graph will be re-evaluated.
    • The cached Oom Adj scores are grouped in bucket, which is used in the isolated processes: they could be correlated - assume one isolated Chrome process is at Oom Adj score 920 and another one is 980; the later one could get expunged much earlier than the former one, which doesn't make sense; grouping them would be a big relief for this case.
  • Compute Oom Adj score
    • This procedure returns true if there is a score change, false if there is no.
    • The curAdj field in the process record is used as an intermediate value during the computation.
    • Initialize the Process State to PROCESS_STATE_CACHED_EMPTY, which is the lowest importance.
    • Calculate the scores based on various factors:

      • If it‘s not allowed to be lower than ProcessList#FOREGROUND_APP_ADJ, meaning it’s probably a persistent process, there is no too much to do here.

      • Exame if the process is the top app, running remote animation, running instrumentation, receiving broadcast, executing services, running on top but sleeping (screen off), update the intermediate values.

      • Ask Window Manager (yes, ActivityTaskManager is with WindowManager now) to tell each activity's visibility information.

      • Check if the process has recent tasks, check if it's hosting a foreground service, overlay UI, toast etc. Note for the foreground service, if it was in foreground status, allow it to stay in higher rank in memory for a while: Assuming a camera capturing case, where the camera app is still processing the picture while being switched out of foreground - keep it stay in higher rank in memory would ensure the pictures are persisted correctly.

      • Check if the process is the heavyweight process, whose launching/exiting would be slow and it's better to keep it in the memory. Note there should be only one heavyweight process across the system.

      • For sure the Home process shouldn't be expunged frequently as well.

      • The next two factors are either it was the previous process with visible UI to the user, or it's a backup agent.

      • And then it goes to the massive searches against the service connections and the content providers, each of the clients will be evaluated, and the Oom Adj score could get updated according to its clients' scores. However there are a bunch of service binding flags which could impact the result:

        • Below table captures the results with given various service binding states:
        Condition #1Condition #2Condition #3Condition #4Result
        BIND_WAIVE_PRIORITY not setBIND_ALLOW_OOM_MANAGEMENT setShown UI && Not HomeUse the app's own Adj
        Inactive for a whileUse the app's own Adj
        Client has a higher importanceShown UI && Not Home && client is invisibleUse the app's own Adj
        BIND_ABOVE_CLIENT and BIND_IMPORTANT setClient is not persistentTry client's Adj
        Client is persistentTry persistent Adj
        BIND_NOT_PERCEPTIBLE setclient < perceptible && app > low perceptibleTry low perceptible Adj
        BIND_NOT_VISIBLE setclient < perceptible && app > perceptibleTry perceptible Adj
        Client >= perceptibleTry client's Adj
        Adj > visibleMax of client/Own Adj
        Use the app's own Adj
        BIND_NOT_FOREGROUND+BIND_IMPORTANT_BACKGROUND not setClient‘s sched group > app’sBIND_IMPORTANT is setUse client's sched group
        Use default sched group
        Client's process state < topBIND_FOREGROUND_SERVICE is setProcState = bound fg
        BIND_FOREGROUND_SERVICE_WHILE_AWAKE + screen ONProcState = bound fg
        ProcState = important fg
        Client's process state = topProcState = bound top
        BIND_IMPORTANT_BACKGROUND not setClient's process state < transient bgProcState = transient bg
        BIND_NOT_FOREGROUND or BIND_IMPORTANT_BACKGROUND setClient's process state < important bgProcState = important bg
        BIND_ADJUST_WITH_ACTIVITY setAdj > fg && App visibleAdj = foreground
        BIND_NOT_FOREGROUND not setBIND_IMPORTANT is setSched = top app bound
        BIND_IMPORTANT is NOT setSched = default
        • Below table captures the results with given various content provider binding states:
        Condition #1Condition #2Condition #3Result
        Client's process state >= cachedClient ProcState = empty
        Adj > Client AdjNot shown UI or is Home, or Client's Adj <= perceptibleClient's Adj <= foreground AdjTry foreground Adj
        Client's Adj > foreground AdjTry client's Adj
        Client's process state <= fg svcClient's process state is topProcState = bound top
        Client's process state is NOT topProcState = bound fg svc
        Has external dependenciesAdj > fg appadj = fg app
        Process state > important foregroundProcState = important fg
        Still within retain timeAdj > previous app Adjadj = previous app adj
        Process state > last activityProcState = last activity
        • Some additional tweaks after the above ones:
        Condition #1Condition #2Condition #3Result
        Process state >= cached emptyHas client activitiesProcState = cached activity client
        treat like activity (IME)ProcState = cached activity
        Adj is service adjcomputing all process recordsNum of new service A > 1/3 of servicesPush it to service B
        Low on RAM and app process's PSS is largePush it to service B
  • Apply the scores, which consists of: write into kernel sysfs entries to update the Oom Adj scores; call kernel API to set the thread priorities, and then tell the world the new process state

Cycles, Cycles, Cycles

Another interesting aspect of the Oom Adjuster is the cycles of the dependencies. A simple example would be like the illustration below, process A is hosting a service which is bound by process B; meanwhile process B is hosting a service which is bound by process A.

There could be very complicated cases, which could involve multiple cycles, and in the dependency graph, each of the process record nodes could have different importance.

The Oom Adjuster maintains a global sequence ID mAdjSeq to track the current Oom Adjuster calling. And each of the process records has a field to track in which sequence the process record is evaluated. If during the Oom Adj computation, a process record with sequence ID as same as the current global sequence ID, this would mean that a cycle is detected; in this case:

  • Decrement the sequence ID of each process if there is a cycle.
  • Re-evaluate each of the process records within the cycle until nothing was promoted.
  • Iterate the processes from least important to most important ones.
  • A maximum retries of 10 is enforced, while in practice, the maximum retries could reach only 2 to 3.

The Modern Implementation

As aforementioned, the OomAdjuster makes the computation in a recursive way, while this is inefficient in dealing with the cycles. The overall code complexity should be around O((1 + num(retries)) * num(procs) * num(binding connections)). In addition, depending on the ordering of the input, the algorithm may produce different results and sometimes it's wrong.

The new “Modern Implementation” is based on the rationale that, apps can't promote the service/provider it connects to, to a higher bucket than itself. We are introducing a bucket based, breadth first search algorithm, as illustrated below:

for all processes in the process list
  compute the state of each process, but, excluding its clients
  put each process to the corresponding bucket according to the state value
done

for each bucket, starting from the top most to the bottom most
  for each process in the bucket
     for each process it binds to
           if the state of the bindee process could be elevated because of the binding; then
              move the bindee process to the higher bucket
           fi
      done
  done
done

The overall code complexity should be around O(num(procs) * num(binding connections)), which saves the retry time from the existing algorithm.