| Message ID | 3252fd83141aa9e0e6001acee1dd98e87c676b9a.camel@perches.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [RFC] vsprintf: Allow %pe to print non PTR_ERR %pe uses as decimal | expand |
On 24/03/2021 18.20, Joe Perches wrote: > On Wed, 2021-03-24 at 09:52 -0700, Joe Perches wrote: >> On Wed, 2021-03-24 at 17:42 +0100, Arnd Bergmann wrote: >>> On Wed, Mar 24, 2021 at 3:20 PM Joe Perches <joe@perches.com> wrote: >> [] >>>>> diff --git a/drivers/gpu/drm/imx/imx-ldb.c b/drivers/gpu/drm/imx/imx-ldb.c >>>> [] >>>>> @@ -197,6 +197,12 @@ static void imx_ldb_encoder_enable(struct drm_encoder *encoder) >>>>> int dual = ldb->ldb_ctrl & LDB_SPLIT_MODE_EN; >>>>> int mux = drm_of_encoder_active_port_id(imx_ldb_ch->child, encoder); >>>>> >>>>> + if (mux < 0 || mux >= ARRAY_SIZE(ldb->clk_sel)) { >>>>> + dev_warn(ldb->dev, "%s: invalid mux %d\n", >>>>> + __func__, ERR_PTR(mux)); >>>> >>>> This does not compile without warnings. >>>> >>>> drivers/gpu/drm/imx/imx-ldb.c: In function ‘imx_ldb_encoder_enable’: >>>> drivers/gpu/drm/imx/imx-ldb.c:201:22: warning: format ‘%d’ expects argument of type ‘int’, but argument 4 has type ‘void *’ [-Wformat=] >>>> 201 | dev_warn(ldb->dev, "%s: invalid mux %d\n", >>>> | ^~~~~~~~~~~~~~~~~~~~~~ >>>> >>>> If you want to use ERR_PTR, the %d should be %pe as ERR_PTR >>>> is converting an int a void * to decode the error type and >>>> emit it as a string. >>> >>> Sorry about that. >>> >>> I decided against using ERR_PTR() in order to also check for >>> positive array overflow, but the version I tested was different from >>> the version I sent. >>> >>> v3 coming. >> >> Thanks. No worries. >> >> Up to you, vsprintf would emit the positive mux as a funky hashed >> hex value by default if you use ERR_PTR with mux > ARRAY_SIZE so >> perhaps %d without the ERR_PTR use makes the most sense. >> > > Maybe it's better to output non PTR_ERR %pe uses as decimal so this > sort of code would work. No, because that would leak the pointer value when somebody has accidentally passed a real kernel pointer to %pe. If the code wants a cute -EFOO string explaining what's wrong, what about "%pe", ERR_PTR(mux < 0 : mux : -ERANGE)? Or two separate error messages if (mux < 0) ... else if (mux >= ARRAY_SIZE()) ... Rasmus
On Wed, 2021-03-24 at 18:33 +0100, Rasmus Villemoes wrote: > On 24/03/2021 18.20, Joe Perches wrote: > > On Wed, 2021-03-24 at 09:52 -0700, Joe Perches wrote: > > > On Wed, 2021-03-24 at 17:42 +0100, Arnd Bergmann wrote: > > > > On Wed, Mar 24, 2021 at 3:20 PM Joe Perches <joe@perches.com> wrote: > > > [] > > > > > > diff --git a/drivers/gpu/drm/imx/imx-ldb.c b/drivers/gpu/drm/imx/imx-ldb.c > > > > > [] > > > > > > @@ -197,6 +197,12 @@ static void imx_ldb_encoder_enable(struct drm_encoder *encoder) > > > > > > int dual = ldb->ldb_ctrl & LDB_SPLIT_MODE_EN; > > > > > > int mux = drm_of_encoder_active_port_id(imx_ldb_ch->child, encoder); > > > > > > > > > > > > + if (mux < 0 || mux >= ARRAY_SIZE(ldb->clk_sel)) { > > > > > > + dev_warn(ldb->dev, "%s: invalid mux %d\n", > > > > > > + __func__, ERR_PTR(mux)); > > > > > > > > > > This does not compile without warnings. > > > > > > > > > > drivers/gpu/drm/imx/imx-ldb.c: In function ‘imx_ldb_encoder_enable’: > > > > > drivers/gpu/drm/imx/imx-ldb.c:201:22: warning: format ‘%d’ expects argument of type ‘int’, but argument 4 has type ‘void *’ [-Wformat=] > > > > > 201 | dev_warn(ldb->dev, "%s: invalid mux %d\n", > > > > > | ^~~~~~~~~~~~~~~~~~~~~~ > > > > > > > > > > If you want to use ERR_PTR, the %d should be %pe as ERR_PTR > > > > > is converting an int a void * to decode the error type and > > > > > emit it as a string. > > > > > > > > Sorry about that. > > > > > > > > I decided against using ERR_PTR() in order to also check for > > > > positive array overflow, but the version I tested was different from > > > > the version I sent. > > > > > > > > v3 coming. > > > > > > Thanks. No worries. > > > > > > Up to you, vsprintf would emit the positive mux as a funky hashed > > > hex value by default if you use ERR_PTR with mux > ARRAY_SIZE so > > > perhaps %d without the ERR_PTR use makes the most sense. > > > > > > > > Maybe it's better to output non PTR_ERR %pe uses as decimal so this > > sort of code would work. > > No, because that would leak the pointer value when somebody has > accidentally passed a real kernel pointer to %pe. I think it's not really an issue. _All_ code that uses %p<foo> extensions need inspection anyway. It's already possible to intentionally 'leak' the ptr value by using %pe, -ptr so I think that's not really an issue. > > If the code wants a cute -EFOO string explaining what's wrong, what > about "%pe", ERR_PTR(mux < 0 : mux : -ERANGE)? Or two separate error > messages > > if (mux < 0) > ... > else if (mux >= ARRAY_SIZE()) > ... Multiple tests, more unnecessary code, multiple format strings, etc...
On 24/03/2021 20.24, Joe Perches wrote: > On Wed, 2021-03-24 at 18:33 +0100, Rasmus Villemoes wrote: >> On 24/03/2021 18.20, Joe Perches wrote: >> >>> >>> Maybe it's better to output non PTR_ERR %pe uses as decimal so this >>> sort of code would work. >> >> No, because that would leak the pointer value when somebody has >> accidentally passed a real kernel pointer to %pe. > > I think it's not really an issue. > > _All_ code that uses %p<foo> extensions need inspection anyway. There are now a bunch of sanity checks in place that catch e.g. an ERR_PTR passed to an extension that would derefence the pointer; enforcing that only ERR_PTRs are passed to %pe (or falling back to %p) is another of those safeguards. > It's already possible to intentionally 'leak' the ptr value > by using %pe, -ptr so I think that's not really an issue. > Huh, what? I assume -ptr is shorthand for (void*)-(unsigned long)ptr. How would that leak the value if ptr is an ordinary kernel pointer? That's not an ERR_PTR unless (unsigned long)ptr is < 4095 or so. If you want to print the pointer value just do %px. No need for silly games. What I'm talking about is preventing _un_intentionally leaking a valid kernel pointer value. So no, a non-ERR_PTR passed to %pe is not going to be printed as-is, not in decimal or hexadecimal or roman numerals. >> If the code wants a cute -EFOO string explaining what's wrong, what >> about "%pe", ERR_PTR(mux < 0 : mux : -ERANGE)? Or two separate error >> messages >> >> if (mux < 0) >> ... >> else if (mux >= ARRAY_SIZE()) >> ... > > Multiple tests, more unnecessary code, multiple format strings, etc... Agreed, I'm not really advocating for the latter; the former suggestion is IMO a pretty concise way of providing useful information in dmesg. Rasmus
On Wed, 2021-03-24 at 22:27 +0100, Rasmus Villemoes wrote: > On 24/03/2021 20.24, Joe Perches wrote: > > On Wed, 2021-03-24 at 18:33 +0100, Rasmus Villemoes wrote: > > > On 24/03/2021 18.20, Joe Perches wrote: > > > > > > > > > > > Maybe it's better to output non PTR_ERR %pe uses as decimal so this > > > > sort of code would work. > > > > > > No, because that would leak the pointer value when somebody has > > > accidentally passed a real kernel pointer to %pe. > > > > I think it's not really an issue. > > > > _All_ code that uses %p<foo> extensions need inspection anyway. > > There are now a bunch of sanity checks in place that catch e.g. an > ERR_PTR passed to an extension that would derefence the pointer; > enforcing that only ERR_PTRs are passed to %pe (or falling back to %p) > is another of those safeguards. > > > It's already possible to intentionally 'leak' the ptr value > > by using %pe, -ptr so I think that's not really an issue. > > > > Huh, what? I assume -ptr is shorthand for (void*)-(unsigned long)ptr. > How would that leak the value if ptr is an ordinary kernel pointer? > That's not an ERR_PTR unless (unsigned long)ptr is < 4095 or so. You are confusing ERR_PTR with IS_ERR ERR_PTR is just include/linux/err.h:static inline void * __must_check ERR_PTR(long error) include/linux/err.h-{ include/linux/err.h- return (void *) error; include/linux/err.h-}f > If you want to print the pointer value just do %px. No need for silly > games. There's no silly game here. %pe would either print a string or a value. It already does that in 2 cases.
On 24/03/2021 23.18, Joe Perches wrote: > On Wed, 2021-03-24 at 22:27 +0100, Rasmus Villemoes wrote: >> On 24/03/2021 20.24, Joe Perches wrote: >>> On Wed, 2021-03-24 at 18:33 +0100, Rasmus Villemoes wrote: >>>> On 24/03/2021 18.20, Joe Perches wrote: >>>> >>>>> >>>>> Maybe it's better to output non PTR_ERR %pe uses as decimal so this >>>>> sort of code would work. >>>> >>>> No, because that would leak the pointer value when somebody has >>>> accidentally passed a real kernel pointer to %pe. >>> >>> I think it's not really an issue. >>> >>> _All_ code that uses %p<foo> extensions need inspection anyway. >> >> There are now a bunch of sanity checks in place that catch e.g. an >> ERR_PTR passed to an extension that would derefence the pointer; >> enforcing that only ERR_PTRs are passed to %pe (or falling back to %p) >> is another of those safeguards. >> >>> It's already possible to intentionally 'leak' the ptr value >>> by using %pe, -ptr so I think that's not really an issue. >>> >> >> Huh, what? I assume -ptr is shorthand for (void*)-(unsigned long)ptr. >> How would that leak the value if ptr is an ordinary kernel pointer? >> That's not an ERR_PTR unless (unsigned long)ptr is < 4095 or so. > > You are confusing ERR_PTR with IS_ERR No I'm not, I'm just being slightly sloppy - obviously when I say "not an ERR_PTR" I mean "not the result of ERR_PTR applied to a negative errno value", or "not the result of a valid invocation of ERR_PTR". But yes, feel free to read "not an ERR_PTR" as "something for which IS_ERR is false". Can you expand on why you think %pe, -ptr would leak the value of ptr? >> If you want to print the pointer value just do %px. No need for silly >> games. > > There's no silly game here. %pe would either print a string or a value. A hashed value, that is, never the raw value. > It already does that in 2 cases. Yes, if you pass it ERR_PTR(-1234) (where no E symbol exists) or ERR_PTR(-EINVAL) but CONFIG_SYMBOLIC_ERRNAME=n, it prints the value in decimal, because people will probably recognize "-22" and values in that range don't reveal anything about the kernel image. Anything outside [-4095,0] or so is hashed. Rasmus
On Wed, 2021-03-24 at 23:36 +0100, Rasmus Villemoes wrote: > On 24/03/2021 23.18, Joe Perches wrote: > > There's no silly game here. %pe would either print a string or a value. > > A hashed value, that is, never the raw value. There is value in printing the raw value. As discussed, it can simplify the code. The worry about exposing a ptr value is IMO overstated. It's trivial to inspect the uses and _all_ %p<FOO> uses need inspection and validation at acceptance anyway.
diff --git a/lib/vsprintf.c b/lib/vsprintf.c index 3600db686fa4..debdd1c62038 100644 --- a/lib/vsprintf.c +++ b/lib/vsprintf.c @@ -619,19 +619,23 @@ static char *string_nocheck(char *buf, char *end, const char *s, return widen_string(buf, len, end, spec); } -static char *err_ptr(char *buf, char *end, void *ptr, - struct printf_spec spec) +static noinline_for_stack +char *err_ptr(char *buf, char *end, void *ptr, struct printf_spec spec) { int err = PTR_ERR(ptr); - const char *sym = errname(err); - if (sym) - return string_nocheck(buf, end, sym, spec); + if (IS_ERR(ptr)) { + const char *sym = errname(err); + + if (sym) + return string_nocheck(buf, end, sym, spec); + } /* - * Somebody passed ERR_PTR(-1234) or some other non-existing - * Efoo - or perhaps CONFIG_SYMBOLIC_ERRNAME=n. Fall back to - * printing it as its decimal representation. + * Somebody passed ERR_PTR(-1234) or some other non-existing -E<FOO> + * or perhaps CONFIG_SYMBOLIC_ERRNAME=n + * or perhaps a positive number like an array index + * Fall back to printing it as its decimal representation. */ spec.flags |= SIGN; spec.base = 10; @@ -2407,9 +2411,7 @@ char *pointer(const char *fmt, char *buf, char *end, void *ptr, case 'x': return pointer_string(buf, end, ptr, spec); case 'e': - /* %pe with a non-ERR_PTR gets treated as plain %p */ - if (!IS_ERR(ptr)) - break; + /* %pe with a non-ERR_PTR(ptr) gets treated as %ld */ return err_ptr(buf, end, ptr, spec); case 'u': case 'k':