Check out the new USENIX Web site. next up previous
Next: Signal delivery Up: Design and implementation Previous: System calls

Context switching

If a process sleeps instead of returning immediately from a system call, it calls schedule. The scheduler selects a new process to run and calls the architecture-specific context switching code to actually perform the switch. In this port, that involves the running process sending a message to the tracing thread that it is being switched out in favor of another process. Since each process in the virtual machine is also a process in the host, the tracing thread performs the switch by stopping the old process and continuing the new one. The new process returns from the context switch that it entered when it last ran and continues whatever it was doing.

Sometimes, after a process is switched back in, its address space will need some updating. If some of its address space had been paged out while it was sleeping, those physical pages, with their new contents, will still be mapped. So, in this situation, the process will need to update its address space by unmapping pages which are no longer valid in its context. These pages are listed in a circular buffer whose address is stored in the process mm_struct. When a page is swapped out from a process while it's asleep, its address is appended to this buffer. When the process wakes up, it looks at this buffer to see if there are any changes to its address space that it hasn't applied yet. It then updates its address space and sets an index in its thread structure to point to the end of the buffer. This private index is necessary because many processes might share a virtual address space and an mm_struct. In general, their host address spaces will be in different states, depending on how long it's been since they've been updated. So, each process keeps track of how many address changes in this buffer it has seen.

There is a possibility that the buffer might wrap around while a process is asleep and some of the address space changes it needs to make have been lost. To avoid this, the index into the buffer that each process maintains is really an absolute count. The index is obtained by dividing the count by the buffer size and taking the remainder. If the buffer has wrapped, then the process count of address space changes will differ from the actual count by more than the number of slots in the buffer. In this case, the entire address space will be scanned and compared to the process page tables. Any differences will be fixed by remapping or unmapping the page concerned.


next up previous
Next: Signal delivery Up: Design and implementation Previous: System calls
Jeff Dike 2000-08-23