On Sat, 21 Aug 2010 18:30:24 +1000 Nick Piggin <npiggin@xxxxxxxxx> wrote: > Thanks, I had both of the same concerns as Christoph with API > change and exposing symlink fds last time I looked at the patces, > actually. > > But they can probably be worked around or avoided. I think the more > important thing is whether it is worth supporting. This is > all restricted to root (or CAP_DAC_READ_SEARCH) only, right, and > what exact semantics they want. I would like to see more discussion > of what this enables and some results. They allow a credible user-space implementation of the server for some network filesystem protocols such as NFS and apparently 9P. > > For the case of avoiding expensive network revalidations in path name > lookup, do we even need to open symlinks? Could the security issues be > avoided by always having handle attached to an open fd? I don't see what you are getting at here... which particular security isses, and what fd would you use? As I understand it there are only two security issues that have been noted. 1/ lookup-by-filehandle can bypass any 'search' permission tests on ancestor directories. I cannot see any way to avoid this except require CAP_DAC_READ_SEARCH 2/ Creating a hardlink to an 'fd' allows a process that was given an 'fd' that it could not have opened itself to prevent that file from being removed (and space reclaimed) by creating a private hardlink. This could be avoided by requiring CAP_DAC_READ_SEARCH for that particular operation (and probably requiring i_nlink > 0 anyway) but that feels like a very special-case restriction. Was it one of these that you were referring to? Thanks, NeilBrown > > On Sat, Aug 21, 2010 at 10:09:00AM +1000, Neil Brown wrote: > > [[email address for Nick Piggin changed to npiggin@xxxxxxxxx]] > > > > On Fri, 20 Aug 2010 12:51:35 +0100 > > Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > > > > > On Fri, Aug 20, 2010 at 07:53:03PM +1000, Neil Brown wrote: > > > > On Fri, 20 Aug 2010 04:30:57 -0400 > > > > Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: > > > > > > > > > Suddenly getting an file pointer for a symlink which could never happen > > > > > before is a really bad idea. Just add a proper readlink_by_handle > > > > > system call, similar to what's done in the XFS interface. > > > > > > > > Why is that? > > > > With futexes we suddenly get a file descriptor for something we could never > > > > get a file descriptor on before and that doesn't seem to be a problem. > > > > > > > > Why should symlinks be special as the only thing that you cannot have a file > > > > descriptor for? Uniformity of interface is a very valuable property. > > > > > > You are welcome to review the codepaths around pathname resolution for > > > assumptions of presense of ->follow_link() and friends; there _are_ > > > subtle cases and dumping your "opened symlinks" in there is far from > > > a trivial change. Note that it affects more than just the starting > > > points of lookups; /proc/*/fd/* stuff is also involved. > > > > So as I understand it you aren't rejecting the concept in principle, but you > > believe non-trivial code review is required before it can be considered an > > acceptable change? > > That's quite reasonable. I hope to find time to have a look at the code. > > > > > > > > BTW, speaking of NULL pathname, linkat() variant that allows creating a link > > > to an opened file is also a very dubious thing; at the very least, you get > > > non-trivial security implications, since now a process that got an opened > > > descriptor passed to it by somebody else may create hardlinks to the sucker. > > > > Fair comment, and while one could imagine ways around this (such as requiring > > some Capability to link an fd) they wouldn't be very elegant. > > But then nor is inventing a pile of new syscalls for doing different things > > with handles such as the list Aneesh posted. > > > > Maybe a different approach is needed. > > > > How about a new AT flag: AT_FILE_HANDLE > > > > Meaning is that the 'dirfd' is used only to identify a filesystem (vfsmnt) and > > the 'name' pointer actually points to a filehandle fragment interpreted in > > that filesystem. > > > > One problem is that there is no way to pass the length... > > Options: > > fragment is at most 64 bytes nul padded at the end > > fragment is hex encoded and nul terminated > > ?? > > > > I think I prefer the hex encoding, but I'm hoping someone else has a better > > idea. > > > > NeilBrown > -- > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html > Please read the FAQ at http://www.tux.org/lkml/ -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
