On Thu, 2020-08-27 at 21:42 +0200, Julia Lawall wrote:
>
> On Thu, 27 Aug 2020, Joe Perches wrote:
>
> > On Thu, 2020-08-27 at 15:48 +0100, Alex Dewar wrote:
> > > On Thu, Aug 27, 2020 at 03:41:06PM +0200, Rasmus Villemoes wrote:
> > > > On 27/08/2020 15.18, Alex Dewar wrote:
> > > > > On Thu, Aug 27, 2020 at 09:15:37AM +0200, Greg Kroah-Hartman wrote:
> > > > > > On Thu, Aug 27, 2020 at 08:42:06AM +0200, Rasmus Villemoes wrote:
> > > > > > > On 25/08/2020 00.23, Alex Dewar wrote:
> > > > > > > > kernel/cpu.c: don't use snprintf() for sysfs attrs
> > > > > > > >
> > > > > > > > As per the documentation (Documentation/filesystems/sysfs.rst),
> > > > > > > > snprintf() should not be used for formatting values returned by sysfs.
> > > > > > > >
> > > > > > >
> > > > > > > Can we have a sysfs_sprintf() (could just be a macro that does sprintf)
> > > > > > > to make it clear to the next reader that we know we're in a sysfs show
> > > > > > > method? It would make auditing uses of sprintf() much easier.
> > > > > >
> > > > > > Code churn to keep code checkers quiet for pointless reasons? What
> > > > > > could go wrong with that...
> > > >
> > > > I did not (mean to) suggest replacing existing sprintf() calls in sysfs
> > > > show methods. But when changes _are_ being made, such as when replacing
> > > > snprintf() calls for whatever reasons, can we please not make it harder
> > > > for people doing manual audits (those are "code checkers" as well, I
> > > > suppose, but they do tend to only make noise when finding something).
> > > >
> > > > > > It should be pretty obvious to any reader that you are in a sysfs show
> > > > > > method, as almost all of them are trivially tiny and obvious.
> > > >
> > > > git grep doesn't immediately show that, not even with a suitable -C
> > > > argument, as you can't really know the potential callers unless you open
> > > > the file and see that the function is only assigned as a .show method.
> > > > And even that can be a pain because it's all hidden behind five levels
> > > > of magic macros that build identifiers with ##.
> > > >
> > > > > Perhaps I should have mentioned this in the commit message, but the problem
> > > > > is that snprintf() doesn't return the number of bytes written to the
> > > > > destination buffer,
> > > >
> > > > I'm perfectly well aware of that, TYVM (you may want to 'git log
> > > > --author Villemoes lib/vsprintf.c').
> > > >
> > > > but the number of bytes that *would have been written if
> > > > > they fitted*, which may be more than the bounds specified [1]. So "return
> > > > > snprintf(...)" for sysfs attributes is an antipattern. If you need bounded
> > > > > string ops, scnprintf() is the way to go. Using snprintf() can give a
> > > > > false sense of security, because it isn't necessarily safe.
> > > >
> > > > Huh? This all seems utterly irrelevant WRT a change that replaces
> > > > PAGE_SIZE by INT_MAX (because that's what sprintf() is going to pretend
> > > > you passed). You get the same return value.
> > > >
> > > > But I'm not at all concerned about whether one passes the proper buffer
> > > > size or not in sysfs show methods; with my embedded hat on, I'm all for
> > > > saving a few bytes of .text here and there. The problem, as far as I'm
> > > > concerned, is merely that adding sprintf() callers makes it harder to
> > > > find the problematic sprintf() instances.
> > > >
> > >
> > > Apologies, I think I might have expressed myself poorly, being a kernel noob
> > > ;-). I know that this is a stylistic change rather than a functional
> > > one -- I meant that I was hoping that it would be helpful to get rid of bad
> > > uses of snprintf().
> > >
> > > I really like your idea of helper methods though :-). If in show()
> > > methods we could have something like:
> > > return sysfs_itoa(buf, i);
> > > in place of:
> > > return sprintf(buf, "%d\n", i);
> > >
> > > ... then we wouldn't be introducing any new calls to sprintf() as you
> > > say, but we'd still be removing a call to snprintf() (which also may be
> > > problematic). Plus we'd have type checking on the argument.
> > >
> > > For returning strings, we could have a bounded and unbounded variant of
> > > the function. As it seems like only single values should be returned via
> > > sysfs, if we did things this way then it would only be these
> > > string-returning functions which could cause buffer overflow problems
> > > and kernel devs could focus their attention accordingly...
> > >
> > > What do people think? I'm happy to have a crack, provided this is
> > > actually a sensible thing to do! I'm looking for a newbie-level project
> > > to get started with.
> >
> > Not a bad idea.
> >
> > Coccinelle should be able to transform the various .show
> > methods to something sysfs_ prefixed in a fairly automated
> > way.
>
> Something like
>
> identifier f;
> fresh identifier = "sysfs" ## f;
>
> may be useful. Let me know if further help is needed.
Perhaps it's a bit more complicated.
Perhaps what's necessary is to find any
appropriate .show function and change
any use of strcpy/sprintf within those
function to some other name.
For instance:
drivers/isdn/mISDN/core.c-static ssize_t name_show(struct device *dev,
drivers/isdn/mISDN/core.c- struct device_attribute *attr, char *buf)
drivers/isdn/mISDN/core.c-{
drivers/isdn/mISDN/core.c: strcpy(buf, dev_name(dev));
drivers/isdn/mISDN/core.c- return strlen(buf);
drivers/isdn/mISDN/core.c-}
drivers/isdn/mISDN/core.c-static DEVICE_ATTR_RO(name);
and macroized uses like:
drivers/base/node.c-#define CACHE_ATTR(name, fmt) \
drivers/base/node.c-static ssize_t name##_show(struct device *dev, \
drivers/base/node.c- struct device_attribute *attr, \
drivers/base/node.c- char *buf) \
drivers/base/node.c-{ \
drivers/base/node.c- return sprintf(buf, fmt "\n", to_cache_info(dev)->cache_attrs.name);\
drivers/base/node.c-} \
drivers/base/node.c:DEVICE_ATTR_RO(name);
drivers/base/node.c-
drivers/base/node.c-CACHE_ATTR(size, "%llu")
drivers/base/node.c-CACHE_ATTR(line_size, "%u")
drivers/base/node.c-CACHE_ATTR(indexing, "%u")
drivers/base/node.c-CACHE_ATTR(write_policy, "%u")
