| Message ID | 20210520100528.322846-2-rasmus.villemoes@prevas.dk |
|---|---|
| State | Superseded |
| Delegated to: | Tom Rini |
| Headers | show |
| Series | assorted printf-related patches | expand |
Hi Rasmus, On Thu, 20 May 2021 at 04:05, Rasmus Villemoes <rasmus.villemoes@prevas.dk> wrote: > > Most callers (or callers of callers, etc.) of vsnprintf() are not > prepared for it to return a negative value. > > The only case where that can currently happen is %pD, and it's IMO > more user-friendly to produce some output that clearly shows that some > "impossible" thing happened instead of having the message completely > ignored - or mishandled as for example log.c would currently do. > > Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> > --- > lib/vsprintf.c | 10 +--------- > 1 file changed, 1 insertion(+), 9 deletions(-) I think that is debatable. If we want the calling code to be fixed, then it needs to get an error code back. Otherwise the error will be apparent to the user but (perhaps) not ever debugged. The definition of printf() allows for the possibility of a negative return value. We should add a prototype to vsprintf.h to indicate this as the only existing one (in exports.h) has no comment. > > diff --git a/lib/vsprintf.c b/lib/vsprintf.c > index 9dc96c81c6..0050110683 100644 > --- a/lib/vsprintf.c > +++ b/lib/vsprintf.c > @@ -310,7 +310,7 @@ static char *device_path_string(char *buf, char *end, void *dp, int field_width, > > str = efi_dp_str((struct efi_device_path *)dp); > if (!str) > - return ERR_PTR(-ENOMEM); > + return string(buf, end, "<%pD:ENOMEM>", field_width, precision, flags); > > buf = string16(buf, end, str, field_width, precision, flags); > efi_free_pool(str); > @@ -631,8 +631,6 @@ repeat: > str = pointer(fmt + 1, str, end, > va_arg(args, void *), > field_width, precision, flags); > - if (IS_ERR(str)) > - return PTR_ERR(str); > /* Skip all alphanumeric pointer suffixes */ > while (isalnum(fmt[1])) > fmt++; > @@ -798,9 +796,6 @@ int printf(const char *fmt, ...) > i = vscnprintf(printbuffer, sizeof(printbuffer), fmt, args); > va_end(args); > > - /* Handle error */ > - if (i <= 0) > - return i; > /* Print the string */ > puts(printbuffer); > return i; > @@ -817,9 +812,6 @@ int vprintf(const char *fmt, va_list args) > */ > i = vscnprintf(printbuffer, sizeof(printbuffer), fmt, args); > > - /* Handle error */ > - if (i <= 0) > - return i; > /* Print the string */ > puts(printbuffer); > return i; > -- > 2.29.2 >
On 20/05/2021 19.51, Simon Glass wrote: > Hi Rasmus, > > On Thu, 20 May 2021 at 04:05, Rasmus Villemoes > <rasmus.villemoes@prevas.dk> wrote: >> >> Most callers (or callers of callers, etc.) of vsnprintf() are not >> prepared for it to return a negative value. >> >> The only case where that can currently happen is %pD, and it's IMO >> more user-friendly to produce some output that clearly shows that some >> "impossible" thing happened instead of having the message completely >> ignored - or mishandled as for example log.c would currently do. >> >> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> >> --- >> lib/vsprintf.c | 10 +--------- >> 1 file changed, 1 insertion(+), 9 deletions(-) > > I think that is debatable. If we want the calling code to be fixed, > then it needs to get an error code back. Otherwise the error will be > apparent to the user but (perhaps) not ever debugged. But it is not the calling code that is at fault for the vsnprintf() implementation (1) being able to fail and (2) actually encountering an ENOMEM situation. There's _nothing_ the calling code can do about that. The calling code can be said to be responsible for not passing NULL pointers, but that case is actually handled gracefully in various places in the printf code (both for %pD, but also plain %s). > The definition of printf() allows for the possibility of a negative > return value. First, please distinguish printf() from vsnprintf(). The former (in the normal userspace version) obviously can fail for the obvious EIO, ENOSPC reasons. The latter is indeed allowed to fail per the posix spec, but from a QoI perspective, I'd say it's much better to have a guarantee _for our particular implementation_ that it does not fail (meaning: returns a negative result). There's simply too many direct and indirect users of vsnprintf() that assume the result is non-negative; if we do not provide that guarantee, the alternative is to play a whack-a-mole game and add tons of error-checking code (adding bloat to the image), with almost never any good way to handle it. Take that log_info(" ... %pD") as an example. Suppose we "fix" log.c so that it ignores the message if vsnprintf (or vscnprintf, whatever) returns a negative result, just as print() currently does [which is the other thing that log_info could end up being handled by]. That means nothing gets printed on the console, and nobody gets told about the ENOMEM. In contrast, with this patch, we get Booting <%pD:ENOMEM> printed on the console, so at least _some_ part of the message gets out, and it's apparent that something odd happened. Of course, all of that is in the entirely unlikely sitation where the (efi) allocation would actually fail. If we don't want that <%pD:ENOMEM> thing, I'd still argue that we should ensure vsnprintf returns non-negative; e.g. by changing the "return PTR_ERR()" to a "goto out", i.e. simply stop the processing of the format string at the %pD which failed, but still go through the epilogue that ensures the resulting string becomes nul-terminated (another reasonable assumption made by tons of callers), and return how much got printed till then. Rasmus
On 21.05.21 14:53, Rasmus Villemoes wrote: > On 20/05/2021 19.51, Simon Glass wrote: >> Hi Rasmus, >> >> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes >> <rasmus.villemoes@prevas.dk> wrote: >>> >>> Most callers (or callers of callers, etc.) of vsnprintf() are not >>> prepared for it to return a negative value. >>> >>> The only case where that can currently happen is %pD, and it's IMO >>> more user-friendly to produce some output that clearly shows that some >>> "impossible" thing happened instead of having the message completely >>> ignored - or mishandled as for example log.c would currently do. >>> >>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> >>> --- >>> lib/vsprintf.c | 10 +--------- >>> 1 file changed, 1 insertion(+), 9 deletions(-) >> >> I think that is debatable. If we want the calling code to be fixed, >> then it needs to get an error code back. Otherwise the error will be >> apparent to the user but (perhaps) not ever debugged. > > But it is not the calling code that is at fault for the vsnprintf() > implementation (1) being able to fail and (2) actually encountering an > ENOMEM situation. There's _nothing_ the calling code can do about that. include/vsnprintf.h states: "This function follows C99 vsnprintf, but has some extensions:". The C99 spec says: "The vsnprintf function returns the number of characters that would have been written had n been sufficiently large, not counting the terminating null character, or a negative value if an encoding error occurred." It is obvious that the calling code needs to be fixed if it cannot handle negative return values. So NAK to the patch. Best regards Heinrich > > The calling code can be said to be responsible for not passing NULL > pointers, but that case is actually handled gracefully in various places > in the printf code (both for %pD, but also plain %s). > >> The definition of printf() allows for the possibility of a negative >> return value. > > First, please distinguish printf() from vsnprintf(). The former (in the > normal userspace version) obviously can fail for the obvious EIO, ENOSPC > reasons. The latter is indeed allowed to fail per the posix spec, but > from a QoI perspective, I'd say it's much better to have a guarantee > _for our particular implementation_ that it does not fail (meaning: > returns a negative result). There's simply too many direct and indirect > users of vsnprintf() that assume the result is non-negative; if we do > not provide that guarantee, the alternative is to play a whack-a-mole > game and add tons of error-checking code (adding bloat to the image), > with almost never any good way to handle it. > > Take that log_info(" ... %pD") as an example. Suppose we "fix" log.c so > that it ignores the message if vsnprintf (or vscnprintf, whatever) > returns a negative result, just as print() currently does [which is the > other thing that log_info could end up being handled by]. That means > nothing gets printed on the console, and nobody gets told about the > ENOMEM. In contrast, with this patch, we get > > Booting <%pD:ENOMEM> > > printed on the console, so at least _some_ part of the message gets out, > and it's apparent that something odd happened. Of course, all of that is > in the entirely unlikely sitation where the (efi) allocation would > actually fail. > > If we don't want that <%pD:ENOMEM> thing, I'd still argue that we should > ensure vsnprintf returns non-negative; e.g. by changing the "return > PTR_ERR()" to a "goto out", i.e. simply stop the processing of the > format string at the %pD which failed, but still go through the epilogue > that ensures the resulting string becomes nul-terminated (another > reasonable assumption made by tons of callers), and return how much got > printed till then. > > Rasmus >
On Fri, May 21, 2021 at 04:15:39PM +0200, Heinrich Schuchardt wrote: > On 21.05.21 14:53, Rasmus Villemoes wrote: > > On 20/05/2021 19.51, Simon Glass wrote: > >> Hi Rasmus, > >> > >> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes > >> <rasmus.villemoes@prevas.dk> wrote: > >>> > >>> Most callers (or callers of callers, etc.) of vsnprintf() are not > >>> prepared for it to return a negative value. > >>> > >>> The only case where that can currently happen is %pD, and it's IMO > >>> more user-friendly to produce some output that clearly shows that some > >>> "impossible" thing happened instead of having the message completely > >>> ignored - or mishandled as for example log.c would currently do. > >>> > >>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> > >>> --- > >>> lib/vsprintf.c | 10 +--------- > >>> 1 file changed, 1 insertion(+), 9 deletions(-) > >> > >> I think that is debatable. If we want the calling code to be fixed, > >> then it needs to get an error code back. Otherwise the error will be > >> apparent to the user but (perhaps) not ever debugged. > > > > But it is not the calling code that is at fault for the vsnprintf() > > implementation (1) being able to fail and (2) actually encountering an > > ENOMEM situation. There's _nothing_ the calling code can do about that. > > include/vsnprintf.h states: > > "This function follows C99 vsnprintf, but has some extensions:". > > The C99 spec says: > > "The vsnprintf function returns the number of characters that would have > been written had n been sufficiently large, not counting the > terminating null character, or a negative value if an encoding error > occurred." > > It is obvious that the calling code needs to be fixed if it cannot > handle negative return values. > > So NAK to the patch. > > Best regards > > Heinrich > > > > > The calling code can be said to be responsible for not passing NULL > > pointers, but that case is actually handled gracefully in various places > > in the printf code (both for %pD, but also plain %s). > > > >> The definition of printf() allows for the possibility of a negative > >> return value. > > > > First, please distinguish printf() from vsnprintf(). The former (in the > > normal userspace version) obviously can fail for the obvious EIO, ENOSPC > > reasons. The latter is indeed allowed to fail per the posix spec, but > > from a QoI perspective, I'd say it's much better to have a guarantee > > _for our particular implementation_ that it does not fail (meaning: > > returns a negative result). There's simply too many direct and indirect > > users of vsnprintf() that assume the result is non-negative; if we do > > not provide that guarantee, the alternative is to play a whack-a-mole > > game and add tons of error-checking code (adding bloat to the image), > > with almost never any good way to handle it. > > > > Take that log_info(" ... %pD") as an example. Suppose we "fix" log.c so > > that it ignores the message if vsnprintf (or vscnprintf, whatever) > > returns a negative result, just as print() currently does [which is the > > other thing that log_info could end up being handled by]. That means > > nothing gets printed on the console, and nobody gets told about the > > ENOMEM. In contrast, with this patch, we get > > > > Booting <%pD:ENOMEM> > > > > printed on the console, so at least _some_ part of the message gets out, > > and it's apparent that something odd happened. Of course, all of that is > > in the entirely unlikely sitation where the (efi) allocation would > > actually fail. > > > > If we don't want that <%pD:ENOMEM> thing, I'd still argue that we should > > ensure vsnprintf returns non-negative; e.g. by changing the "return > > PTR_ERR()" to a "goto out", i.e. simply stop the processing of the > > format string at the %pD which failed, but still go through the epilogue > > that ensures the resulting string becomes nul-terminated (another > > reasonable assumption made by tons of callers), and return how much got > > printed till then. So, how can we fix the callers without the above noted problems?
On 21/05/2021 16.15, Heinrich Schuchardt wrote: > On 21.05.21 14:53, Rasmus Villemoes wrote: >> On 20/05/2021 19.51, Simon Glass wrote: >>> Hi Rasmus, >>> >>> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes >>> <rasmus.villemoes@prevas.dk> wrote: >>>> >>>> Most callers (or callers of callers, etc.) of vsnprintf() are not >>>> prepared for it to return a negative value. >>>> >>>> The only case where that can currently happen is %pD, and it's IMO >>>> more user-friendly to produce some output that clearly shows that some >>>> "impossible" thing happened instead of having the message completely >>>> ignored - or mishandled as for example log.c would currently do. >>>> >>>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> >>>> --- >>>> lib/vsprintf.c | 10 +--------- >>>> 1 file changed, 1 insertion(+), 9 deletions(-) >>> >>> I think that is debatable. If we want the calling code to be fixed, >>> then it needs to get an error code back. Otherwise the error will be >>> apparent to the user but (perhaps) not ever debugged. >> >> But it is not the calling code that is at fault for the vsnprintf() >> implementation (1) being able to fail and (2) actually encountering an >> ENOMEM situation. There's _nothing_ the calling code can do about that. > > include/vsnprintf.h states: > > "This function follows C99 vsnprintf, but has some extensions:". Happy to update that comment (which is copied from linux BTW, and in the kernel there's a very simple rule: "This is printk. We want it to work." - that extends to the workhorse vsnprintf() which is not allowed to take locks or do allocations) with an amendment "... among which is that it never returns a negative value." > It is obvious that the calling code needs to be fixed if it cannot > handle negative return values. > > So NAK to the patch. So you'd rather fix the ~200 places that use the return value assuming it's non-negative, plus the unknown number of places that assume the output buffer is a valid nul-terminated string after vsnprintf() returns? And, again taking that %pD example, do you really rather want _nothing_ printed (which is the only thing log.c could sensibly do if vsnprintf() returned something negative) in the rare case of allocation failure? I must admit I completely fail to see how this can possibly be better than improving the guarantees provided by U-Boot's vsnprintf() implementation. Rasmus
On 21.05.21 16:27, Tom Rini wrote: > On Fri, May 21, 2021 at 04:15:39PM +0200, Heinrich Schuchardt wrote: >> On 21.05.21 14:53, Rasmus Villemoes wrote: >>> On 20/05/2021 19.51, Simon Glass wrote: >>>> Hi Rasmus, >>>> >>>> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes >>>> <rasmus.villemoes@prevas.dk> wrote: >>>>> >>>>> Most callers (or callers of callers, etc.) of vsnprintf() are not >>>>> prepared for it to return a negative value. >>>>> >>>>> The only case where that can currently happen is %pD, and it's IMO >>>>> more user-friendly to produce some output that clearly shows that some >>>>> "impossible" thing happened instead of having the message completely >>>>> ignored - or mishandled as for example log.c would currently do. >>>>> >>>>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> >>>>> --- >>>>> lib/vsprintf.c | 10 +--------- >>>>> 1 file changed, 1 insertion(+), 9 deletions(-) >>>> >>>> I think that is debatable. If we want the calling code to be fixed, >>>> then it needs to get an error code back. Otherwise the error will be >>>> apparent to the user but (perhaps) not ever debugged. >>> >>> But it is not the calling code that is at fault for the vsnprintf() >>> implementation (1) being able to fail and (2) actually encountering an >>> ENOMEM situation. There's _nothing_ the calling code can do about that. >> >> include/vsnprintf.h states: >> >> "This function follows C99 vsnprintf, but has some extensions:". >> >> The C99 spec says: >> >> "The vsnprintf function returns the number of characters that would have >> been written had n been sufficiently large, not counting the >> terminating null character, or a negative value if an encoding error >> occurred." >> >> It is obvious that the calling code needs to be fixed if it cannot >> handle negative return values. >> >> So NAK to the patch. >> >> Best regards >> >> Heinrich >> >>> >>> The calling code can be said to be responsible for not passing NULL >>> pointers, but that case is actually handled gracefully in various places >>> in the printf code (both for %pD, but also plain %s). >>> >>>> The definition of printf() allows for the possibility of a negative >>>> return value. >>> >>> First, please distinguish printf() from vsnprintf(). The former (in the >>> normal userspace version) obviously can fail for the obvious EIO, ENOSPC >>> reasons. The latter is indeed allowed to fail per the posix spec, but >>> from a QoI perspective, I'd say it's much better to have a guarantee >>> _for our particular implementation_ that it does not fail (meaning: >>> returns a negative result). There's simply too many direct and indirect >>> users of vsnprintf() that assume the result is non-negative; if we do >>> not provide that guarantee, the alternative is to play a whack-a-mole >>> game and add tons of error-checking code (adding bloat to the image), >>> with almost never any good way to handle it. >>> >>> Take that log_info(" ... %pD") as an example. Suppose we "fix" log.c so >>> that it ignores the message if vsnprintf (or vscnprintf, whatever) >>> returns a negative result, just as print() currently does [which is the >>> other thing that log_info could end up being handled by]. That means >>> nothing gets printed on the console, and nobody gets told about the >>> ENOMEM. In contrast, with this patch, we get >>> >>> Booting <%pD:ENOMEM> >>> >>> printed on the console, so at least _some_ part of the message gets out, >>> and it's apparent that something odd happened. Of course, all of that is >>> in the entirely unlikely sitation where the (efi) allocation would >>> actually fail. >>> >>> If we don't want that <%pD:ENOMEM> thing, I'd still argue that we should >>> ensure vsnprintf returns non-negative; e.g. by changing the "return >>> PTR_ERR()" to a "goto out", i.e. simply stop the processing of the >>> format string at the %pD which failed, but still go through the epilogue >>> that ensures the resulting string becomes nul-terminated (another >>> reasonable assumption made by tons of callers), and return how much got >>> printed till then. > > So, how can we fix the callers without the above noted problems? > The assumption that vsnprintf() is used to print to the console and that writing some arbitrary string to the buffer is allowable is utterly wrong. vsnprintf_internal() is used to implement snprintf(). snprintf() is used in numerous places where it will not lead to console output. Trying to solve one problem this patch creates a bunch of new ones. Best regards Heinrich
On 5/21/21 10:15 AM, Heinrich Schuchardt wrote: > On 21.05.21 14:53, Rasmus Villemoes wrote: >> On 20/05/2021 19.51, Simon Glass wrote: >>> Hi Rasmus, >>> >>> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes >>> <rasmus.villemoes@prevas.dk> wrote: >>>> >>>> Most callers (or callers of callers, etc.) of vsnprintf() are not >>>> prepared for it to return a negative value. >>>> >>>> The only case where that can currently happen is %pD, and it's IMO >>>> more user-friendly to produce some output that clearly shows that some >>>> "impossible" thing happened instead of having the message completely >>>> ignored - or mishandled as for example log.c would currently do. >>>> >>>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> >>>> --- >>>> lib/vsprintf.c | 10 +--------- >>>> 1 file changed, 1 insertion(+), 9 deletions(-) >>> >>> I think that is debatable. If we want the calling code to be fixed, >>> then it needs to get an error code back. Otherwise the error will be >>> apparent to the user but (perhaps) not ever debugged. >> >> But it is not the calling code that is at fault for the vsnprintf() >> implementation (1) being able to fail and (2) actually encountering an >> ENOMEM situation. There's _nothing_ the calling code can do about that. > > include/vsnprintf.h states: > > "This function follows C99 vsnprintf, but has some extensions:". > > The C99 spec says: > > "The vsnprintf function returns the number of characters that would have > been written had n been sufficiently large, not counting the > terminating null character, or a negative value if an encoding error > occurred." But is this an encoding error? And of course, the most common use of this function which checks this return value will be n = vsnprintf(NULL, 0, fmt, ...); buf = malloc(n); if (!buf) /* out of memory */ vsnprintf(buf, n, fmt, ...); Which effectively already checks for ENOMEM. While it is standard-compliant to return negative, it is also compliant to just use a bogus value or to write nothing. There are many callers of [v]s[c]nprintf which use the same pattern as the code above. If you would like to fix all of them, go ahead. But I think this fix is more efficient a cut. --Sean > > It is obvious that the calling code needs to be fixed if it cannot > handle negative return values. > > So NAK to the patch. > > Best regards > > Heinrich > >> >> The calling code can be said to be responsible for not passing NULL >> pointers, but that case is actually handled gracefully in various places >> in the printf code (both for %pD, but also plain %s). >> >>> The definition of printf() allows for the possibility of a negative >>> return value. >> >> First, please distinguish printf() from vsnprintf(). The former (in the >> normal userspace version) obviously can fail for the obvious EIO, ENOSPC >> reasons. The latter is indeed allowed to fail per the posix spec, but >> from a QoI perspective, I'd say it's much better to have a guarantee >> _for our particular implementation_ that it does not fail (meaning: >> returns a negative result). There's simply too many direct and indirect >> users of vsnprintf() that assume the result is non-negative; if we do >> not provide that guarantee, the alternative is to play a whack-a-mole >> game and add tons of error-checking code (adding bloat to the image), >> with almost never any good way to handle it. >> >> Take that log_info(" ... %pD") as an example. Suppose we "fix" log.c so >> that it ignores the message if vsnprintf (or vscnprintf, whatever) >> returns a negative result, just as print() currently does [which is the >> other thing that log_info could end up being handled by]. That means >> nothing gets printed on the console, and nobody gets told about the >> ENOMEM. In contrast, with this patch, we get >> >> Booting <%pD:ENOMEM> >> >> printed on the console, so at least _some_ part of the message gets out, >> and it's apparent that something odd happened. Of course, all of that is >> in the entirely unlikely sitation where the (efi) allocation would >> actually fail. >> >> If we don't want that <%pD:ENOMEM> thing, I'd still argue that we should >> ensure vsnprintf returns non-negative; e.g. by changing the "return >> PTR_ERR()" to a "goto out", i.e. simply stop the processing of the >> format string at the %pD which failed, but still go through the epilogue >> that ensures the resulting string becomes nul-terminated (another >> reasonable assumption made by tons of callers), and return how much got >> printed till then. >> >> Rasmus >> >
On 21.05.21 16:40, Rasmus Villemoes wrote: > On 21/05/2021 16.15, Heinrich Schuchardt wrote: >> On 21.05.21 14:53, Rasmus Villemoes wrote: >>> On 20/05/2021 19.51, Simon Glass wrote: >>>> Hi Rasmus, >>>> >>>> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes >>>> <rasmus.villemoes@prevas.dk> wrote: >>>>> >>>>> Most callers (or callers of callers, etc.) of vsnprintf() are not >>>>> prepared for it to return a negative value. >>>>> >>>>> The only case where that can currently happen is %pD, and it's IMO >>>>> more user-friendly to produce some output that clearly shows that some >>>>> "impossible" thing happened instead of having the message completely >>>>> ignored - or mishandled as for example log.c would currently do. >>>>> >>>>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> >>>>> --- >>>>> lib/vsprintf.c | 10 +--------- >>>>> 1 file changed, 1 insertion(+), 9 deletions(-) >>>> >>>> I think that is debatable. If we want the calling code to be fixed, >>>> then it needs to get an error code back. Otherwise the error will be >>>> apparent to the user but (perhaps) not ever debugged. >>> >>> But it is not the calling code that is at fault for the vsnprintf() >>> implementation (1) being able to fail and (2) actually encountering an >>> ENOMEM situation. There's _nothing_ the calling code can do about that. >> >> include/vsnprintf.h states: >> >> "This function follows C99 vsnprintf, but has some extensions:". > > Happy to update that comment (which is copied from linux BTW, and in the > kernel there's a very simple rule: "This is printk. We want it to work." > - that extends to the workhorse vsnprintf() which is not allowed to take > locks or do allocations) with an amendment "... among which is that it > never returns a negative value." > >> It is obvious that the calling code needs to be fixed if it cannot >> handle negative return values. >> >> So NAK to the patch. > > So you'd rather fix the ~200 places that use the return value assuming > it's non-negative, As you pointed out in patch 5 any usage of (v)snprintf assuming that the return value is the number of characters copied to the buffer has to be corrected anyways. plus the unknown number of places that assume the > output buffer is a valid nul-terminated string after vsnprintf() returns? Ensuring that the buffer is always nul-terminated would not contradict the C99 specification. Best regards Heinrich > > And, again taking that %pD example, do you really rather want _nothing_ > printed (which is the only thing log.c could sensibly do if vsnprintf() > returned something negative) in the rare case of allocation failure? > > I must admit I completely fail to see how this can possibly be better > than improving the guarantees provided by U-Boot's vsnprintf() > implementation. > > Rasmus >
On 21/05/2021 16.42, Heinrich Schuchardt wrote: > On 21.05.21 16:27, Tom Rini wrote: >> On Fri, May 21, 2021 at 04:15:39PM +0200, Heinrich Schuchardt wrote: >>> On 21.05.21 14:53, Rasmus Villemoes wrote: >>>> On 20/05/2021 19.51, Simon Glass wrote: >>>>> Hi Rasmus, >>>>> >>>>> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes >>>>> <rasmus.villemoes@prevas.dk> wrote: >>>>>> >>>>>> Most callers (or callers of callers, etc.) of vsnprintf() are not >>>>>> prepared for it to return a negative value. >>>>>> >>>>>> The only case where that can currently happen is %pD, and it's IMO >>>>>> more user-friendly to produce some output that clearly shows that some >>>>>> "impossible" thing happened instead of having the message completely >>>>>> ignored - or mishandled as for example log.c would currently do. >>>>>> >>>>>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> >>>>>> --- >>>>>> lib/vsprintf.c | 10 +--------- >>>>>> 1 file changed, 1 insertion(+), 9 deletions(-) >>>>> >>>>> I think that is debatable. If we want the calling code to be fixed, >>>>> then it needs to get an error code back. Otherwise the error will be >>>>> apparent to the user but (perhaps) not ever debugged. >>>> >>>> But it is not the calling code that is at fault for the vsnprintf() >>>> implementation (1) being able to fail and (2) actually encountering an >>>> ENOMEM situation. There's _nothing_ the calling code can do about that. >>> >>> include/vsnprintf.h states: >>> >>> "This function follows C99 vsnprintf, but has some extensions:". >>> >>> The C99 spec says: >>> >>> "The vsnprintf function returns the number of characters that would have >>> been written had n been sufficiently large, not counting the >>> terminating null character, or a negative value if an encoding error >>> occurred." >>> >>> It is obvious that the calling code needs to be fixed if it cannot >>> handle negative return values. >>> >>> So NAK to the patch. >>> >>> Best regards >>> >>> Heinrich >>> >>>> >>>> The calling code can be said to be responsible for not passing NULL >>>> pointers, but that case is actually handled gracefully in various places >>>> in the printf code (both for %pD, but also plain %s). >>>> >>>>> The definition of printf() allows for the possibility of a negative >>>>> return value. >>>> >>>> First, please distinguish printf() from vsnprintf(). The former (in the >>>> normal userspace version) obviously can fail for the obvious EIO, ENOSPC >>>> reasons. The latter is indeed allowed to fail per the posix spec, but >>>> from a QoI perspective, I'd say it's much better to have a guarantee >>>> _for our particular implementation_ that it does not fail (meaning: >>>> returns a negative result). There's simply too many direct and indirect >>>> users of vsnprintf() that assume the result is non-negative; if we do >>>> not provide that guarantee, the alternative is to play a whack-a-mole >>>> game and add tons of error-checking code (adding bloat to the image), >>>> with almost never any good way to handle it. >>>> >>>> Take that log_info(" ... %pD") as an example. Suppose we "fix" log.c so >>>> that it ignores the message if vsnprintf (or vscnprintf, whatever) >>>> returns a negative result, just as print() currently does [which is the >>>> other thing that log_info could end up being handled by]. That means >>>> nothing gets printed on the console, and nobody gets told about the >>>> ENOMEM. In contrast, with this patch, we get >>>> >>>> Booting <%pD:ENOMEM> >>>> >>>> printed on the console, so at least _some_ part of the message gets out, >>>> and it's apparent that something odd happened. Of course, all of that is >>>> in the entirely unlikely sitation where the (efi) allocation would >>>> actually fail. >>>> >>>> If we don't want that <%pD:ENOMEM> thing, I'd still argue that we should >>>> ensure vsnprintf returns non-negative; e.g. by changing the "return >>>> PTR_ERR()" to a "goto out", i.e. simply stop the processing of the >>>> format string at the %pD which failed, but still go through the epilogue >>>> that ensures the resulting string becomes nul-terminated (another >>>> reasonable assumption made by tons of callers), and return how much got >>>> printed till then. >> >> So, how can we fix the callers without the above noted problems? >> > > The assumption that vsnprintf() is used to print to the console and that > writing some arbitrary string to the buffer is allowable is utterly wrong. > > vsnprintf_internal() is used to implement snprintf(). snprintf() is used > in numerous places where it will not lead to console output. > > Trying to solve one problem this patch creates a bunch of new ones. Heinrich, you do realize that the error handling you added in 256060e when you made it possible for vsnprintf() to return something negative is broken and incomplete? In multiple ways, even. First, let's look at vscnprint, which wasn't touched by 256060e. int vscnprintf(char *buf, size_t size, const char *fmt, va_list args) { int i; i = vsnprintf(buf, size, fmt, args); if (likely(i < size)) return i; Integer promotion says that, should i be -ENOMEM or some other random -Esomething, that comparison is false (for any realistic value of the size parameter), so we won't actually pass on that negative value. Instead, we'll fall through to the logic that handles "oh, vsnprintf() didn't have enough room, so the length of the generated string, which is what I'm supposed to return, is size-1". Hence printf(), which uses vscnprintf(), would in that case receive CONFIG_SYS_PBSIZE-1 as result, and then it would go on to puts() the printbuffer[] - which isn't a nul-terminated string because you did that early return, so it has stack garbage. Let us look at printf() in more detail. Assuming vscnprintf() forwarded a negative value directly, do you see the problem here: uint i; i = vscnprintf(printbuffer, sizeof(printbuffer), fmt, args); /* Handle error */ if (i <= 0) return i; So even if vscnprintf() was updated, printf() _still_ would go on to pass stack garbage to puts(). ==== Let me re-iterate what I believe any vsnprintf-implementation (and its close wrappers [v]s[c][n]printf) in a kernel or bootloader context must satisfy: - never return a negative value - given a non-zero size argument, must guarantee that the output buffer is a proper nul-terminated string. Without those guarantees, bugs like those described above will creep in, even if somebody fixes the issues I just pointed out. I'm not gonna send band-aid patches which will just propagate the problems to all other users. Not only do these guarantees make it easier to use the sprintf family, it also avoids lots of code bloat from unnecessary 'ret < 0' checks. It really has nothing to do with whether the output is destined for the console, _all_ users (direct or through several layers of helpers) of the sprintf family benefit from an implementation that provides these guarantees. Rasmus
Hi Rasmus, On Thu, 27 May 2021 at 17:01, Rasmus Villemoes <rasmus.villemoes@prevas.dk> wrote: > > On 21/05/2021 16.42, Heinrich Schuchardt wrote: > > On 21.05.21 16:27, Tom Rini wrote: > >> On Fri, May 21, 2021 at 04:15:39PM +0200, Heinrich Schuchardt wrote: > >>> On 21.05.21 14:53, Rasmus Villemoes wrote: > >>>> On 20/05/2021 19.51, Simon Glass wrote: > >>>>> Hi Rasmus, > >>>>> > >>>>> On Thu, 20 May 2021 at 04:05, Rasmus Villemoes > >>>>> <rasmus.villemoes@prevas.dk> wrote: > >>>>>> > >>>>>> Most callers (or callers of callers, etc.) of vsnprintf() are not > >>>>>> prepared for it to return a negative value. > >>>>>> > >>>>>> The only case where that can currently happen is %pD, and it's IMO > >>>>>> more user-friendly to produce some output that clearly shows that some > >>>>>> "impossible" thing happened instead of having the message completely > >>>>>> ignored - or mishandled as for example log.c would currently do. > >>>>>> > >>>>>> Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> > >>>>>> --- > >>>>>> lib/vsprintf.c | 10 +--------- > >>>>>> 1 file changed, 1 insertion(+), 9 deletions(-) > >>>>> > >>>>> I think that is debatable. If we want the calling code to be fixed, > >>>>> then it needs to get an error code back. Otherwise the error will be > >>>>> apparent to the user but (perhaps) not ever debugged. > >>>> > >>>> But it is not the calling code that is at fault for the vsnprintf() > >>>> implementation (1) being able to fail and (2) actually encountering an > >>>> ENOMEM situation. There's _nothing_ the calling code can do about that. > >>> > >>> include/vsnprintf.h states: > >>> > >>> "This function follows C99 vsnprintf, but has some extensions:". > >>> > >>> The C99 spec says: > >>> > >>> "The vsnprintf function returns the number of characters that would have > >>> been written had n been sufficiently large, not counting the > >>> terminating null character, or a negative value if an encoding error > >>> occurred." > >>> > >>> It is obvious that the calling code needs to be fixed if it cannot > >>> handle negative return values. > >>> > >>> So NAK to the patch. > >>> > >>> Best regards > >>> > >>> Heinrich > >>> > >>>> > >>>> The calling code can be said to be responsible for not passing NULL > >>>> pointers, but that case is actually handled gracefully in various places > >>>> in the printf code (both for %pD, but also plain %s). > >>>> > >>>>> The definition of printf() allows for the possibility of a negative > >>>>> return value. > >>>> > >>>> First, please distinguish printf() from vsnprintf(). The former (in the > >>>> normal userspace version) obviously can fail for the obvious EIO, ENOSPC > >>>> reasons. The latter is indeed allowed to fail per the posix spec, but > >>>> from a QoI perspective, I'd say it's much better to have a guarantee > >>>> _for our particular implementation_ that it does not fail (meaning: > >>>> returns a negative result). There's simply too many direct and indirect > >>>> users of vsnprintf() that assume the result is non-negative; if we do > >>>> not provide that guarantee, the alternative is to play a whack-a-mole > >>>> game and add tons of error-checking code (adding bloat to the image), > >>>> with almost never any good way to handle it. > >>>> > >>>> Take that log_info(" ... %pD") as an example. Suppose we "fix" log.c so > >>>> that it ignores the message if vsnprintf (or vscnprintf, whatever) > >>>> returns a negative result, just as print() currently does [which is the > >>>> other thing that log_info could end up being handled by]. That means > >>>> nothing gets printed on the console, and nobody gets told about the > >>>> ENOMEM. In contrast, with this patch, we get > >>>> > >>>> Booting <%pD:ENOMEM> > >>>> > >>>> printed on the console, so at least _some_ part of the message gets out, > >>>> and it's apparent that something odd happened. Of course, all of that is > >>>> in the entirely unlikely sitation where the (efi) allocation would > >>>> actually fail. > >>>> > >>>> If we don't want that <%pD:ENOMEM> thing, I'd still argue that we should > >>>> ensure vsnprintf returns non-negative; e.g. by changing the "return > >>>> PTR_ERR()" to a "goto out", i.e. simply stop the processing of the > >>>> format string at the %pD which failed, but still go through the epilogue > >>>> that ensures the resulting string becomes nul-terminated (another > >>>> reasonable assumption made by tons of callers), and return how much got > >>>> printed till then. > >> > >> So, how can we fix the callers without the above noted problems? > >> > > > > The assumption that vsnprintf() is used to print to the console and that > > writing some arbitrary string to the buffer is allowable is utterly wrong. > > > > vsnprintf_internal() is used to implement snprintf(). snprintf() is used > > in numerous places where it will not lead to console output. > > > > Trying to solve one problem this patch creates a bunch of new ones. > > Heinrich, you do realize that the error handling you added in 256060e > when you made it possible for vsnprintf() to return something negative > is broken and incomplete? In multiple ways, even. > > First, let's look at vscnprint, which wasn't touched by 256060e. > > int vscnprintf(char *buf, size_t size, const char *fmt, va_list args) > { > int i; > > i = vsnprintf(buf, size, fmt, args); > > if (likely(i < size)) > return i; > > Integer promotion says that, should i be -ENOMEM or some other random > -Esomething, that comparison is false (for any realistic value of the > size parameter), so we won't actually pass on that negative value. > Instead, we'll fall through to the logic that handles "oh, vsnprintf() > didn't have enough room, so the length of the generated string, which is > what I'm supposed to return, is size-1". > > Hence printf(), which uses vscnprintf(), would in that case receive > CONFIG_SYS_PBSIZE-1 as result, and then it would go on to puts() the > printbuffer[] - which isn't a nul-terminated string because you did that > early return, so it has stack garbage. > > Let us look at printf() in more detail. Assuming vscnprintf() forwarded > a negative value directly, do you see the problem here: > > uint i; > > i = vscnprintf(printbuffer, sizeof(printbuffer), fmt, args); > > /* Handle error */ > if (i <= 0) > return i; > > So even if vscnprintf() was updated, printf() _still_ would go on to > pass stack garbage to puts(). > > ==== > > Let me re-iterate what I believe any vsnprintf-implementation (and its > close wrappers [v]s[c][n]printf) in a kernel or bootloader context must > satisfy: > > - never return a negative value > - given a non-zero size argument, must guarantee that the output buffer > is a proper nul-terminated string. > > Without those guarantees, bugs like those described above will creep in, > even if somebody fixes the issues I just pointed out. I'm not gonna send > band-aid patches which will just propagate the problems to all other users. > > Not only do these guarantees make it easier to use the sprintf family, > it also avoids lots of code bloat from unnecessary 'ret < 0' checks. It > really has nothing to do with whether the output is destined for the > console, _all_ users (direct or through several layers of helpers) of > the sprintf family benefit from an implementation that provides these > guarantees. Gosh what a lot of stuff to read. Thank you both for all the analysis. I feel that s...printf() should never return an error code, so I agree with Rasmus, I suppose. EFI's oddity here is just going to have to be EFI's problem. I am not becoming more sympathetic to the EFI-ification of firmware, nor does this sort of corner case progress the cause If we need to return an error somehow, then I suggest: - create a new function, e.g. vsprintf_err() - carry around a flag in vsprintf.c to indicate the behaviour on error - put support for the flag behind a CONFIG to avoid code-size bloat Reviewed-by: Simon Glass <sjg@chromium.org> (but please update docs for pointer() to explain args and return value) Regards, Simon
diff --git a/lib/vsprintf.c b/lib/vsprintf.c index 9dc96c81c6..0050110683 100644 --- a/lib/vsprintf.c +++ b/lib/vsprintf.c @@ -310,7 +310,7 @@ static char *device_path_string(char *buf, char *end, void *dp, int field_width, str = efi_dp_str((struct efi_device_path *)dp); if (!str) - return ERR_PTR(-ENOMEM); + return string(buf, end, "<%pD:ENOMEM>", field_width, precision, flags); buf = string16(buf, end, str, field_width, precision, flags); efi_free_pool(str); @@ -631,8 +631,6 @@ repeat: str = pointer(fmt + 1, str, end, va_arg(args, void *), field_width, precision, flags); - if (IS_ERR(str)) - return PTR_ERR(str); /* Skip all alphanumeric pointer suffixes */ while (isalnum(fmt[1])) fmt++; @@ -798,9 +796,6 @@ int printf(const char *fmt, ...) i = vscnprintf(printbuffer, sizeof(printbuffer), fmt, args); va_end(args); - /* Handle error */ - if (i <= 0) - return i; /* Print the string */ puts(printbuffer); return i; @@ -817,9 +812,6 @@ int vprintf(const char *fmt, va_list args) */ i = vscnprintf(printbuffer, sizeof(printbuffer), fmt, args); - /* Handle error */ - if (i <= 0) - return i; /* Print the string */ puts(printbuffer); return i;
Most callers (or callers of callers, etc.) of vsnprintf() are not prepared for it to return a negative value. The only case where that can currently happen is %pD, and it's IMO more user-friendly to produce some output that clearly shows that some "impossible" thing happened instead of having the message completely ignored - or mishandled as for example log.c would currently do. Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk> --- lib/vsprintf.c | 10 +--------- 1 file changed, 1 insertion(+), 9 deletions(-)