On Thu, Mar 18, 2021 at 08:51:45AM -0700, Kees Cook wrote: > On Thu, Mar 18, 2021 at 09:07:45AM +0100, Greg Kroah-Hartman wrote: > > On Wed, Mar 17, 2021 at 02:30:47PM -0700, Kees Cook wrote: > > > 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. > > > > But, is it really worth the churn here? > > > > Yes, strings in C is "hard", but this _should_ be a simple thing for any > > driver to handle: > > return sysfs_emit(buffer, "%d\n", my_dev->value); > > > > To change that to: > > return seq_printf(seq, "%d\n", my_dev->value); > > feels very much "don't we have other more valuable things we could be > > doing?" > > > > So far we have found 1 driver that messed up and overflowed the buffer > > that I know of. While reworking apis to make it "hard to get wrong" is > > a great goal, the work involved here vs. any "protection" feels very > > low. > > I haven't been keeping a list, but it's not the only one. The _other_ > reason we need seq_file is so we can perform checks against f_cred for > things like %p obfuscation (as was needed for modules that I hacked > around) and is needed a proper bug fix for the kernel pointer exposure > bug from the same batch. So now I'm up to 3 distinct reasons that the > sysfs API is lacking -- I think it's worth the churn and time. Ok, if you think so. But if we do this, can we not do a "raw" seqfile api? I would like to see only 1 function that works like sysfs_emit() does. Perhaps: void sysfs_printf(struct attribute *attr, const char *fmt, ...); and then from there we can "derive" things like: void device_printf(struct device_attribute *attr, const char *fmt, ...); You can "hide" the needed seq_file structure in the attribute structure for the buffer management, but I don't think we need the crazy multiple ways that seq_printf() has morphed into over the years, right? seq_path() anyone? binary attribute files are a totally different thing, and probably can just be left alone for now. > > How about moving everyone to sysfs_emit() first? That way it becomes > > much more "obvious" when drivers are doing stupid things with their > > sysfs buffer. But even then, it would not have caught the iscsi issue > > as that was printing a user-provided string so maybe I'm just feeling > > grumpy about the potential churn here... > > I need to fix the prototypes for CFI sanity too. Switching to seq_file > solves 2 problems, and if we have to change the prototype once for that, > we can include the prototype fixes for CFI at the same time to avoid > double the churn. Yes, let's not go through this twice... thanks, greg k-h
