This fixes 3 problems in syslogd related to sizing receive buffers...
- A call was misplaced at the wrong level of nested if blocks, so that
the buffers for unix domain sockets (/dev/log, /dev/klog) were never
increased at all; they remained at a way-too-small default size of 4096.
- The function that was supposed to double the size of the buffer
sometimes did nothing, and sometimes installed a wildly-wrong buffer
size (either too large or too small) due to an unitialized 'slen'
variable passed to getsockopt(). Most often it doubled the UDP buffers
from 40k to 80k because accidentally there would be harmless stack
garbage in the unitialized variables.
- The whole concept of blindly doubling a socket's buffer size without
knowing what size it started at is a design flaw that has to be called a
bug. If the double_rbuf() function had worked at all (I.E., if the
other two bugs didn't exist) this would lead to UDP sockets having an
80k buffer while unix dgram sockets get an 8k buffer. There's nothing
about the problem being solved that requires larger buffers for UDP than
for unix dgram sockets -- the buffering requirements are the same
regardless of socket type.
This change renames the double_rbuf() function to increase_rbuf() and
increases the buffer size on all types of sockets to 80k. 80k was
chosen only because it appears to be the size the original change was
shooting for, and it certainly seems to be reasonably large (I might
have picked 64k in the absence of any historical guidance).
Add ENETUNREACH and EADDRNOTAVAIL to the list of errors that are potentially
transient and shouldn't result in closing the socket and giving up forever.
3360970b01