% scalable virtual memory w/ RCU

# administrivia

project proposal due this Friday 11pm

work in groups (preferred; not required)

close to your research

talk to us

# today's plan

virtual address space

read-copy-update (RCU) 

functional data structure

# "big" questions

Q1: what's the problem & for what kind of applications?

Q2: is this the "right" solution for this problem?

multithreaded, VM-intensitve apps?

compare to exokernel/library OS (e.g., Corey, Barrelfish)?

# memory mapping

app ↔ kernel: mmap/munmap

kernel ↔ hw: pagefault

`cat /proc/self/maps`

virtual memory areas (VMAs): list (traverse) & tree (lookup)

tree examples: rb (linux), avl (windows), splay (freebsd)

# concurrency quiz: output?

```c++
#include <iostream>
#include <thread>
static int flag = 0;
void signal_proc() { // t1
  std::cout << "before signal" << std::endl;
  flag = 1;
}
void wait_proc() {   // t2
  while (flag == 0) {}
  std::cout << "after wait" << std::endl;
}
int main() {
  std::thread t1(signal_proc), t2(wait_proc);
  t1.join(); t2.join();
}
```
. . .

undefined behavior: `g++-4.9 -O2`, infinite loop

# concurrency primitives

atomic reads/writes, memory barriers

mutual exclusion

rw locks: concurrent reads; reads/write block each other

RCU: wait-free reads; single pointer update; delay free

RCU-friendly balanced trees: eliminate read locks

see also:
[rcutorture](https://www.kernel.org/doc/Documentation/RCU/torture.txt),
[verification challenge](http://paulmck.livejournal.com/tag/rcu)

# functional data structures

path copying: slist, tree

optimization in bonsai tree: update non-root pointers

other apps: file snapshots (btrfs), model checking (klee)