This is the mail archive of the cygwin mailing list for the Cygwin project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Cygwin Python/PIL TCL/TK fork rebase solution


--On 16 January 2007 04:58 -0800 Brian Dessent wrote:

Because in order to emulate fork(), Cygwin has to be able to re-execute
the binary and have it load with the same memory layout.  If there are
DLLs that overlap and need remapping by the OS then the memory layout
becomes non-deterministic.  If Cygwin cannot create a child process with
the same memory layout as the parent, then it cannot fork and you get
the "unable to remap" error.

Also note that this explains why binaries that dynamically load modules
at runtime with dlopen() (such as Perl, Python, Apache, ...) are
particular likely to be affected by this whereas if you are just running
ordinary binaries that do no dynamic loading you almost never have to
mess with rebasing.  Specifically, if a process loads modules at runtime
and those modules need to be relocated by the OS then the memory layout
depends now on countless details of the execution of the script or the
logic flow of the binary.  Or put differently, the Windows process image
loader (i.e. the thing that loads a .exe and all DLLs that it was linked
against at link time) is relatively deterministic and repeatable, so you
can get away with unbased DLLs if there is no runtime loading since the
memory layout is still somewhat repeatably reproducable.

Brian,


Thanks for the explanations. So, if I've understood things correctly, the difficulty boils down to cloning a parent process's address space layout within that of a child, which includes ensuring that DLLs appear at the same base within both processes.

For this to be the problem it appears to be, I'm guessing that there must be some shortcoming in the Windows APIs in this area when compared with facilities available within other Posix-compliant OSs.

How does Linux deal with the same issues of having libraries (or whatever are logically equivalent to DLLs) potentially linked at different bases in the two address spaces?

--
Robin Walker (Junior Bursar), Queens' College, Cambridge CB3 9ET, UK
rdhw@cam.ac.uk  http://www.queens.cam.ac.uk/  Tel:+44 1223 335528

Attachment: p7s00000.p7s
Description: S/MIME cryptographic signature


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]