On Wed, Oct 24 2018, Theodore Y. Ts'o wrote: > On Wed, Oct 24, 2018 at 12:47:57PM +1100, NeilBrown wrote: >> >> I doubt it was copied - more likely independent evolution. >> But on reflection, I see that it is probably reasonable that it >> shouldn't be used this way - or at all in this context. >> I'll try to understand what the issues really are and see if I can >> find a solution that doesn't depend on this interface. >> Thanks for your help. > > At least for ext4, the primary problem is that we want to use a 64-bit > telldir/seekdir cookie if all 64-bits will make it to user space, and > a 32-bit telldir cookie if only 32 bits will make it userspace. This > impacts NFS as well because if there are people who are still using > NFSv2, which has 32-bit directory offsets, we need to constrain the > telldir/seekdir cookies we give to NFS to be a 32 has as opposed to a > 64-bit hash.
NFSd uses FMODE_32BITHASH or FMODE64BITHASH to allow ext4 to do the right thing. FMODE_32BITHASH is set for NFSv2 only. Maybe sys_getdents needs to set FMODE_32BITHASH, and sys_getdent64 needs to set FMODE_64BITHASH - or something like that. For lustre it is a bit more complex. The internal "inode number" is 128 bits and we (sort of) hash it to 32 or 64 bits. cp_compat_stat() just says -EOVERFLOW if we give a 64 bit number when 32 are expected, and there is no flag to say "this is a 32-bit 'stat' request". But I need to dig into exactly what that "sort-of" means - maybe there is an answer in there. NeilBrown
signature.asc
Description: PGP signature
