Alternate signal stack is used on Linux to support stack banging.

Stack banging needs to touch page below current ESP; however, this will
generate SIGSEGV on Linux. To work around the problem, we install alternate
signal stack for SIGSEGV at the lower end of thread stack. When we receive
a SEGV due to stack bang, we will touch the same page again from SEGV
handler. Because we are now running on alternate signal stack with a smaller
ESP, this time we can succeed.

There are other ways to get around the problem without using alt signal
stack:

1. mmap() the page directly. This may have some performance issue, as mmap()
   on old systems will create a new memory region. Using mmap() for stack
   bang may create too many memory regions. With Tiger, we'll drop some old
   systems, so we might be able to use mmap(). Need to test.
2. Temporarily adjust ESP to lower address within SEGV handler. However, when
   stack space is already low, and if there's another signal delivered after
   SEGV, changing ESP might leave too little space to handle the new signal.
   We can get around this by temporarily masking all signals.

Here are the benefits from not using alternate signal stack:

1. Allow smaller thread stack, and more threads. Currently, every thread
   needs 40K alternate signal stack, minimum stack size is 96K. Removing alt
   signal stack can reduce minimum stack size by about 40%.
2. Only need one set of guard pages. Right now, we have HotSpot guard pages
   between alt signal stack and normal thread stack, and regular glibc guard
   page below alt signal stack. We don't need glibc guard pages when there
   is not alternate signal stack. HotSpot guard pages are enough. Creating
   guard page on Linux is very expensive. Using a micro benchmark, not
   creating glibc guard (or HotSpot guard) can reduce thread creation time
   by about 40%.
3. Easier to maintain. Same model on all platforms.

###@###.### 2003-04-23