----- Original Message ----- > On 08/30/2013 11:07 AM, Robert O'Callahan wrote: > > Being able to sample uniformly whether or not the thread is in a syscall is > > important sometimes, so let's not lose that. > > I am not proposing to lose that facility. Am only proposing (really, > merely parroting Taras' suggestion) that it would be nice to have the option > of some kind of adaptive backoff for threads blocked in syscalls.
One way of handling this is the following: When you take a sample, replace the return address for the current stack frame with the address of a known function inside your profile (call this a "trampoline function"). Now the call stack looks something like: A --> B --> C --> D --Trampoline--> E where the sample was taken in E. The trampoline function enables us to keep track of what some portion of the stack looks like without having to unwind it. Unwinding on subsequent samples proceeds until we encounter a frame whose return address is the trampoline function. At that point, we know everything else above that frame is the same as whatever it was the last time we took a sample. There are now three interesting scenarios of what can happen before we take the next sample. 1) E does not return. It may call other functions, it may do some computation, or it may just be blocked on a syscall. Whatever the case, we take a sample, find that E's return address is our trampoline function, and stop. 2) E calls F calls G, and then a sample is taken. Just prior to the sample, we have: A --> B --> C --> D --Trampoline--> E --> F --> G The profiler fixes things up to do: A --> B --> C --> D --> E --> F --Trampoline--> G and we go about our merry way. 3) E returns. Instead of returning to D, it returns to our trampoline function. Said function then replaces the return address of D's stack frame with the address of the trampoline function, and returns to the appropriate PC in D: A --> B --> C --Trampoline--> D This can happen several times (D returns to C, etc.). Depending on how deep the call stack is and how frequently we push and pop stack frames, being able to not unwind some number of frames can be a significant savings. -Nathan _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
