On Wed, Mar 17, 2021 at 04:38:57PM +0100, Greg Kroah-Hartman wrote: > On Wed, Mar 17, 2021 at 04:20:52PM +0100, Michal Hocko wrote: > > On Wed 17-03-21 15:56:44, Greg KH wrote: > > > On Wed, Mar 17, 2021 at 03:44:16PM +0100, Michal Hocko wrote: > > > > On Wed 17-03-21 14:34:27, Greg KH wrote: > > > > > On Wed, Mar 17, 2021 at 01:08:21PM +0100, Michal Hocko wrote: > > > > > > Btw. I still have problems with the approach. seq_file is intended to > > > > > > provide safe way to dump values to the userspace. Sacrificing > > > > > > performance just because of some abuser seems like a wrong way to go as > > > > > > Al pointed out earlier. Can we simply stop the abuse and disallow to > > > > > > manipulate the buffer directly? I do realize this might be more tricky > > > > > > for reasons mentioned in other emails but this is definitely worth > > > > > > doing. > > > > > > > > > > We have to provide a buffer to "write into" somehow, so what is the best > > > > > way to stop "abuse" like this? > > > > > > > > What is wrong about using seq_* interface directly? > > > > > > Right now every show() callback of sysfs would have to be changed :( > > > > Is this really the case? Would it be too ugly to have an intermediate > > buffer and then seq_puts it into the seq file inside sysfs_kf_seq_show. > > Oh, good idea. > > > Sure one copy more than necessary but it this shouldn't be a hot path or > > even visible on small strings. So that might be worth destroying an > > inherently dangerous seq API (seq_get_buf). > > I'm all for that, let me see if I can carve out some time tomorrow to > try this out. The trouble has been that C string APIs are just so impossibly fragile. We just get too many bugs with it, so we really do need to rewrite the callbacks to use seq_file, since it has a safe API. I've been trying to write coccinelle scripts to do some of this refactoring, but I have not found a silver bullet. (This is why I've suggested adding the temporary "seq_show" and "seq_store" functions, so we can transition all the callbacks without a flag day.) > But, you don't get rid of the "ability" to have a driver write more than > a PAGE_SIZE into the buffer passed to it. I guess I could be paranoid > and do some internal checks (allocate a bunch of memory and check for > overflow by hand), if this is something to really be concerned about... Besides the CFI prototype enforcement changes (which I can build into the new seq_show/seq_store callbacks), the buffer management is the primary issue: we just can't hand drivers a string (even with a length) because the C functions are terrible. e.g. just look at the snprintf vs scnprintf -- we constantly have to just build completely new API when what we need is a safe way (i.e. obfuscated away from the caller) to build a string. Luckily seq_file does this already, so leaning into that is good here. -- Kees Cook
