mm/filemap.c: atomic file read(2)/write(2) ?

hi all

judging from mm/filemap.c it seems like
ordinary reads/writes should be atomic to each other (read sees write completely or not at all,
not only where it "can be proven to be after the write"), if
- the fs uses the generic code from filemap.c, like ext2 and most do
- the file region affected fits within one cache page, like nice little B-Tree blocks do,
so there is only one copy_from/to_user per call
- the respective userspace memory regions fit within one page, so the copy will not be interrupted
- we're not interfering with another processor (but depending on mm hardware it could even work on SMP ?)

correct?

Would be a nice property to avoid read locks on a L-B-Tree

thx + cheers
Paul

Nope. There are file descriptors that have atomicity guarantees (pipes(, but regular files do not. In particular, even on UP you'll see nonatomic reads with PREEMPT enabled.

Even without preempt you could see interesting partial updates if you take a page fault. That's true even if your user space only ever reads or writes within a whole page, since what you can get on the read() path is:

- read one word from the page cache
- try to write it to user space
- take a page fault - sleep

- writer now comes in and updates the region

- the reader comes back after the page fault
- the reader completes the copy_to_user(), but the _first_ word was read from the old page cache contents and not re-read.

(This will depend on which particular version of copy_to_user() you use: a "rep movs" will re-read the page cache and give an "atomic" read, while all other forms of memory copies will only re-try the destination write instruction, with the old value read from the source).

Linus