Serious problem with multithreaded I/O on Linux

Thu Jul 15 07:45:00 GMT 1999

Bryce McKinlay wrote:
> Why are blocking IO calls not gc-safe, anyway?

Blocking I/O isn't the problem. The GC plays tricks with mprotect() for
its incremental mode... by write-protecting a heap address, it can trap
page faults and detect when a heap block is changed. That works fine in
user mode. But if the kernel tries to modify a read-only block, such as
during a read() system call, the call will fail without generating a
page fault.
If we don't use incremental mode, I think the wrapped system calls can
safely be removed.
-- 
Jeff Sturm
jsturm@sigma6.com