The timeline is at the center of the redaction system. It provides an efficient method to find which package a thread/process belongs to.
The timeline allows queries to be connected to time. Without this, there‘s a significant privacy conern because a pid can be recycled. Just because the pid is excluded from redaction before time T, doesn’t mean it should be redacted after time T.
The timeline uses an event-based pattern using two events:
An event-based structure (compared to a span-based structure) is used as it is better suited to handle errors/issues in the underlying data. For example, if a pid doesn't explictly ends before being reused (e.g. two back-to-back open events), the event-based structure “just works”.
Open events contain the thread‘s full state. The close event only contains the information needed to reference the thread’s previous event.
struct Open { uint64_t ts; int32_t pid; int32_t ppid; uint64_t uid; }; struct Close { uint64_t ts; int32_t pid; };
The vast majory of threads will have one event, an open event provided by the ProcessTree
. For some threads, they will have multiple open (ProcessTree
, NewTask
) and close events (ProcFree
) in alternating order.
struct Slice { int32_t pid; uint64_t uid; }; class Timeline { Slice Query(uint64_t ts, int32_t pid) const; };
Events, regardless of type, are stored in contiguous memory and are ordered first by pid and second by time. This is done to allow events to be found via a binary search.
The vast majory of threads will have one event, the open event. Some threads may have close and re-open events.
To handle a query,
pid
(the first instance of pid
)ts
(for pid
)If an event was found:
if (e.type == kOpen && uid != 0) return Slice(pid, e.uid); // The pid is active, check the parent for a uid. if (e.type == kOpen && uid == 0) return Query(ts, e.ppid); return Slice(pid, kNoPackage);
If pid
does not have an immediate package (uid
), the parent must be searched. The parent-child tree is short, so the recursive search will be relatively short. To minimize this even more, a union-find operation is applied because any queries can be made.
Simple runtime overview:
Initialization:
$sort + union\ find$
$nlogn + mlogn$
Query:
$P(p) = m_p * (logn + e_p)$
Because of the union-find in initialization, $m_p \to 0$
To further reduce the runtime, the search domain is reduces by remove all open events for $pids$ that don't connect to a target $uid$. By removing open events, and close events, there are two advantages:
Removing open events are safe and simple. By removing open events, those pids can never be marked by active. Keeping the close events effectively reminds the system that the pid is not active.
The number of open events exceeds the number of close events. Removing open events will have a greater effect on the number of events.
Example:
Name | Value | Notes |
---|---|---|
tids | 3666 | Total number of threads. |
freed threads | 5 | Number of threads that were freed. |
reused threads | 0 | No threads were used more than one time. |
process tids | 64 | Total number of threads connected to the target process. |
After initialization, there would only be 64 open events and 5 close events. This means that every uid lookup would be $logn\ |\ n=64 = 6$. Finding the uid given a pid is one of the most common operations during redaction because uid determines if something needs to be redacted.
task_rename
are generated when a thread renames itself. This often happens after (but not limited to) a task_newtask
event. The task_rename
event exposes the threads old name and the threads new name.
New task event:
event { timestamp: 6702094133317685 pid: 6167 task_newtask { pid: 7972 comm: "adbd" clone_flags: 4001536 oom_score_adj: -1000 } }
Task rename event:
event { timestamp: 6702094133665498 pid: 7972 task_rename { pid: 7972 oldcomm: "adbd" newcomm: "shell svc 7971" oom_score_adj: -1000 } }
A task_rename
should be redacted when event.pid
does not belong to that target package (context.package_uid
). Since the pid's naming information will be removed everywhere, and naming information is effectively metadata, the whole event can be dropped without effecting the integrity of the trace.