Re: [PATCH 0/2] Better handling of slow remote transfers

[Gary Benson <gbenson@redhat.com>]

Pedro Alves wrote:
> On 08/12/2015 01:32 PM, Gary Benson wrote:
> > Pedro Alves wrote:
> > > IIUC, it still auto fetches the executable and then the solibs
> > > from the target by default (e.g., after "attach"), so still
> > > subject to lack of interruptibility?
> > 
> > Yes and no.  It will fetch the executable from the remote iff one
> > has not been otherwise specified (i.e. by "file", or on the
> > command line).  It will *only* fetch libraries from the remote if
> > the parent executable has a target prefix.  So:
> > 
> >   (gdb) file a.out
> >   (gdb) target remote :9999
> > 
> >    - exec_filename is "a.out"
> >    - exec_filename has no "target:" prefix
> >    - "target:" prefix is NOT applied to shared libraries
> >    - solib paths end up as "/path/to/libsolib.so.1"
> >    - solibs are NOT fetched over RSP
> > 
> 
> But to me it looks like GDB _should_ retrieve the libraries out of
> the target in this case.  You'll usually have a local copy of the
> executable, because you just compiled it, but not of the shared
> libraries.  It seems to me we're only considering this option
> because we didn't make transfers interruptible?

Right, but users who have been doing that will be using "set sysroot
remote:":

  (gdb) file a.out
  (gdb) set sysroot remote:
  (gdb) target remote :9999

  - gdb_sysroot is "target:"
  - solib paths end up as "target:/path/to/libsolib.so.1"
  - solibs ARE fetched over RSP

This series basically means that:

  a) users get to type whatever they were typing before and have the
     same thing happen, with the exception of the 0.001% of users who
     have been typing "target remote :9999" with no "file" or "set sysroot"
     commands and debugging with no symbols whatsoever.

  b) those users can add "set auto-target-prefix off" in their .gdbinit

  c) users who want GDB to connect to a remote target and Just Work
     get to type "target remote :9999" without messing about with
     "file" and "set sysroot" commands.

> > > I was only OK with trying to make transfers interruptible in the
> > > branch assuming it was something non-invasive, like a missing
> > > QUIT here and there.
> > 
> > No, gdbserver sends the data in PBUFSIZ chunks, but GDB reads the
> > data a character at a time.
> 
> Can you expand on this?  What code is it that reads the data a
> character at a time?  What data is gdb getting at when it does that?

I was looking in getpkt_or_notif_sane_1, but I think maybe I misread
it.  I'll get back to you on this...

Cheers,
Gary

-- 
http://gbenson.net/