Date: Thu, 13 Mar 2008 14:27:44 -0700 From: "Rao, Nikhil" <nikhil.rao@intel.com> To: "John Giacomoni" <john.giacomoni@colorado.edu>, <freebsd-hackers@freebsd.org> Subject: RE: Shared VM address range across processes Message-ID: <12A5C15467D5B94F8E0FF265D9498ADD02B6AF6A@orsmsx419.amr.corp.intel.com> In-Reply-To: <006DB5A0-3669-473B-84B6-E3C8CC3C059D@colorado.edu> References: <12A5C15467D5B94F8E0FF265D9498ADD02B35F81@orsmsx419.amr.corp.intel.com> <006DB5A0-3669-473B-84B6-E3C8CC3C059D@colorado.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
Hi John, Is the approach that you are working on based on necessarily using the kernel address space, so is this approach not feasible with user space virtual addresses ? Nikhil -----Original Message----- From: John Giacomoni [mailto:john.giacomoni@colorado.edu]=20 Sent: Thursday, March 13, 2008 9:13 AM To: freebsd-hackers@freebsd.org Cc: Rao, Nikhil Subject: Re: Shared VM address range across processes Nihkil, I'm working on something similar for a research project and the answer is that it is possible but ugly. First, are you sure you need to do this? Ensuring safety by checking pointers before dereferencing can be painful :) FreeBSD seems to have checks scattered throughout the kernel trying to ensure that the kernel address range remains unavailable to the =20 userspace address range. These checks can obviously be bypassed but they are =20 fairly invasive. Once all those checks are bipassed, you need to ensure that =20 the PTEs and PDEs are have the userspace bit set for the appropriate page ranges which then requires flushing the specific pages out of the TLB using the invlpg function, note that flushing the TLB is insufficient as kernel pages are marked global and thus won't flush with any other =20 method. files that I touched /usr/src/sys/amd64/amd64/pmap.c - pmap_enter /usr/src/sys/amd64/amd64/trap.c - trap_pfault and the allocation site needs to ensure that the user-mode bit is set on the correct PTEs and PDEs. I directly allocate memory using vm objects to help me bypass the =20 various address range checks that can be found in the higher levels of the =20 kernel. I'm planning on generalizing and cleaning my approach up in the next few months but I'll be glad to answer any specific questions you might have. For the FreeBSD kernel developers, Is there a reason to enforce the high/low mem address range as strongly as is done in FreeBSD? It seems that if the higher-levels of the kernel allow a mapping, the lower-levels should respect that. John G On Mar 12, 2008, at 12:46 PM, Rao, Nikhil wrote: > Hi, > > > > I want to map device memory into the same virtual address range in > multiple processes, this means I would have to add a vm_map_entry per > address range in every process, since the list of processes can be > potentially huge .. Is it allowed to point to the same list of > vm_map_entrys from multiple vm_spaces ? BSD3 had a field in the > vm_map_entry that could be a share map - would it be an idea that I > could reuse ? > > > > Nikhil > > _______________________________________________ > freebsd-hackers@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-hackers > To unsubscribe, send any mail to "freebsd-hackers-unsubscribe@freebsd.org=20 > " -- John.Giacomoni@colorado.edu University of Colorado at Boulder Department of Computer Science Engineering Center, ECCR 1B50 430 UCB Boulder, CO 80303-0430 USA
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?12A5C15467D5B94F8E0FF265D9498ADD02B6AF6A>