On Fri, Sep 29, 2006 at 07:39:26PM +0200, Petter Reinholdtsen wrote: > [Mario Holbe] > > I personally think that something around 100k should suffice, but if > > you like to have it somehow dynamic, I attached a small sample patch > > which limits the size to one tenth of available memory. > > Thank you for the patch. If we want dosemu and user-mode-linux to > stop using /dev/shm/ to store their mmaped files, and instead use > /lib/init/rw/ or similar, the size will have to be a lot higher than > 100k.
These two filesystems serve fundamentally different purposes, and namespace collisions between those two uses should be avoided at all costs -- by keeping them completely separate. Is there any good reason to combine the two? The overhead of two [sensibly configured] tmpfs filesystems rather than one is minimal. > One idea I am considering is to make /dev/shm/ a symlink or bind-mount > to a subdirectory /lib/init/rw/shm/, and thus only have one tmpfs file > system by default. Please don't do this. Sensible defaults are all that is required in both cases. For /lib/init/rw, this could most likely be set to a tiny amount, like the 100 KiB suggested. For /dev/shm, requirements could be a lot higher, and vary from system to system, but again a sensible default would fix this. The current practice of using the kernel default of 0.5*coresize is wrong. I'm currently safe, having a good 6 GiB of swap, but for high memory systems with less swap than core, you're heading into potential DoS territory with the current approach. On a system with 8 GiB of core, a 4 GiB /lib/init/rw is a waste and a huge liability. Suggestion: choose fixed limits, and allow the user to configure both. /lib/init/rw could be fixed to a specific size, and /dev/shm could be e.g. 0.5*core up to an upper limit of 512 MiB (by default). The current SHM_SIZE in /etc/default/tmpfs is no longer sufficient. Please could you add an INIT_RW_SIZE in addition, and set it by default? (As in the patch). Also, given the widely differing sizes of the various tmpfs filesystems, TMPFS_SIZE is not really all that useful any more. Could this be deprecated or removed? Regards, Roger -- .''`. Roger Leigh : :' : Debian GNU/Linux http://people.debian.org/~rleigh/ `. `' Printing on GNU/Linux? http://gutenprint.sourceforge.net/ `- GPG Public Key: 0x25BFB848 Please GPG sign your mail.
signature.asc
Description: Digital signature