On Tue, Apr 14, 2015 at 01:29:50PM +0800, Minfei Huang wrote:
> On 04/14/15 at 12:11P, Josh Poimboeuf wrote:
> > On Tue, Apr 14, 2015 at 01:03:48PM +0800, Minfei Huang wrote:
> > > On 04/13/15 at 11:57P, Josh Poimboeuf wrote:
> > > > On Tue, Apr 14, 2015 at 08:26:29AM +0800, Minfei Huang wrote:
> > > > > On 04/13/15 at 06:13P, Josh Poimboeuf wrote:
> > > > > > On Sun, Apr 12, 2015 at 09:15:54PM +0800, Minfei Huang wrote:
> > > > > > > For now, the kallsyms will only store the first
> > > > > > > (KSYM_NAME_LEN-1). The
> > > > > > > kallsyms name is same for the function which first
> > > > > > > (KSYM_NAME_LEN-1) is
> > > > > > > same, but the rest is not.
> > > > > > >
> > > > > > > Then function will never be patched, although function name and
> > > > > > > address
> > > > > > > are provided both. The reason caused this bug is livepatch cannt
> > > > > > > recognize the function name.
> > > > > > >
> > > > > > > Now, livepatch will verify the function name with first
> > > > > > > (KSYM_NAME_LEN-1)
> > > > > > > and address, if provided. Once they are matched, we can confirm
> > > > > > > that the
> > > > > > > patched function is found.
> > > > > >
> > > > > > From scripts/kallsyms.c:
> > > > > >
> > > > > > if (strlen(str) > KSYM_NAME_LEN) {
> > > > > > fprintf(stderr, "Symbol %s too long for kallsyms (%zu
> > > > > > vs %d).\n"
> > > > > > "Please increase KSYM_NAME_LEN both in
> > > > > > kernel and kallsyms.c\n",
> > > > > > str, strlen(str), KSYM_NAME_LEN);
> > > > > > return -1;
> > > > > > }
> > > > > >
> > > > > > So I think such a long symbol name wouldn't be added to the kallsyms
> > > > > > database in the first place.
> > > > > >
> > > > >
> > > > > Actually, kernel allows overlength function name to be used. Following
> > > > > is my testing module.
> > > > >
> > > > > We can got the address in /proc/kallsyms.
> > > > > $ cat /proc/kallsyms | grep sysfs_print
> > > > > ffffffffa0000000 t
> > > > > sys_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_pri
> > > > > [sysfs_print]
> > > > > ffffffffa0000010 t kobj_release [sysfs_print]
> > > > > ffffffffa0000020 t
> > > > > sys_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_pri
> > > > > [sysfs_print]
> > > > > ffffffffa00004e0 b root_kobj [sysfs_print]
> > > > > ffffffffa0000200 d print_ktype [sysfs_print]
> > > > > ffffffffa00004a0 b print_kobj [sysfs_print]
> > > > > ffffffffa000004c t sys_print_exit [sysfs_print]
> > > > > ffffffffa0000144 r __func__.14514 [sysfs_print]
> > > > > ffffffffa0000230 d kobj_attrs [sysfs_print]
> > > > > ffffffffa0000240 d sys_print_kobj_attr [sysfs_print]
> > > > > ffffffffa0000260 d __this_module [sysfs_print]
> > > > > ffffffffa000004c t cleanup_module [sysfs_print]
> > > > >
> > > > > Code:
> > > > >
> > > > > static ssize_t
> > > > > sys_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_store(struct
> > > > > kobject *kobj, s
> > > > > const char *buf, size_t count)
> > > > > {
> > > > > return count;
> > > > > }
> > > > >
> > > > > static ssize_t
> > > > > sys_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_show(struct
> > > > > kobject *kobj,
> > > > > struct kobj_attribute *attr, char *buf)
> > > > > {
> > > > > return snprintf(buf, PAGE_SIZE-1, "%s\n", "This is printed by
> > > > > module");
> > > > > }
> > > > >
> > > > > static struct kobj_attribute sys_print_kobj_attr =
> > > > > __ATTR_RW(sys_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_print_p
> > > > > static struct attribute *kobj_attrs[] = {
> > > > > &sys_print_kobj_attr.attr,
> > > > > NULL
> > > > > };
> > > > >
> > > >
> > > > Hm, this seems like a kallsyms bug. IMO it should either fail the build
> > > > or omit the symbol from the kallsyms db. Truncating it seems dangerous
> > > > and counterintuitive.
> > > >
> > >
> > > Kallsyms will record all of the function name, without truncating it.
> > > But the kallsyms will return the truncated function name which is max to
> > > 127.
> > >
> > > > But regardless I really don't see a good reason to encourage this kind
> > > > of insanity in the livepatch code.
> > > >
> > >
> > > Yes, the above code is terrible, but we cannt stop user composing like
> > > that.
> > >
> > > Once the function name is like above, user will never have chance to use
> > > livepatch.
> >
> > Again, this seems like a kallsyms bug. Fix the bug and the real world
> > need for this patch set goes away. The user will be forced to either
> > shorten their function name or increase KSYM_NAME_LEN.
> >
>
> kallsyms bug? I donot think increasing the KSYM_NAME_LEN is a good idea.
Well, neither is having a function name > KSYM_NAME_LEN :-)
>
> For end user, they may know litter about restriction of kallsyms and
> livepatch. How can they know the restriction that function name is
> limited to 127?
As I mentioned above, I think kallsyms.c should fail the build if it
encounters a symbol longer than KSYM_NAME_LEN.
--
Josh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
