On 12/18/18 10:47 AM, Daniel P. Berrangé wrote: > On Tue, Dec 18, 2018 at 10:39:20AM -0500, John Ferlan wrote: >> Commit 912c6b22 modified the virCommandAddArg and virCommandAddArgPair >> to perform NULL argument checking; however, no corresponding change >> to the prototypes was made to remove the ATTRIBUTE_NONNULL, so the >> Coverity build failed. Adjust the prototypes accordingly. > > This sounds wrong or at least different from the past. > I didn't push yet... The last time this happened is/was commit 425aac3abf. Still, before this patch if you build with STATIC_ANALYSIS set you'd see the same phenomena. If I modify config.h to have STATIC_ANALYSIS be 1 instead of 0, I get: util/vircommand.c: In function 'virCommandAddArg': util/vircommand.c:1501:8: error: nonnull argument 'val' compared to NULL [-Werror=nonnull-compare] if (val == NULL) { ^ util/vircommand.c: In function 'virCommandAddArgPair': util/vircommand.c:1604:14: error: nonnull argument 'name' compared to NULL [-Werror=nonnull-compare] if (name == NULL || val == NULL) { ^ util/vircommand.c:1604:29: error: nonnull argument 'val' compared to NULL [-Werror=nonnull-compare] if (name == NULL || val == NULL) { ^ Is there a different way you'd prefer this resolved? I could leave it as a my coverity environment only, but I would assume clang would be impacted too, although I don't use clang. > We have historically still added nonnull annotations even when > having checks for NULL in the impl. We had to explicitly disabble > the nonnull annotation when building under GCC to prevent it from > optimizing out the NULL checks. We left it enabled for STATIC_ANALYSIS > so that coverity would still report if any callers passed a NULL into > the methods. > That wasn't my recollection. Adding ATTRIBUTE_NONNULL has been "less important" because we found that it really only helps if someone calls some API with NULL and it does not help if the argument itself is NULL. My recollection is always towards removing it once the attribute itself was checked in the API. I also have a faint recollection of a discussion we've had before on this with the thought towards removing all the ATTRIBUTE_NONNULL's and replacing with real argument == NULL checks, but that got to be too many patches and checks in too many places.

On 12/18/18 11:16 AM, Daniel P. Berrangé wrote: > On Tue, Dec 18, 2018 at 11:10:34AM -0500, John Ferlan wrote: >> >> >> On 12/18/18 10:47 AM, Daniel P. Berrangé wrote: >>> On Tue, Dec 18, 2018 at 10:39:20AM -0500, John Ferlan wrote: >>>> Commit 912c6b22 modified the virCommandAddArg and virCommandAddArgPair >>>> to perform NULL argument checking; however, no corresponding change >>>> to the prototypes was made to remove the ATTRIBUTE_NONNULL, so the >>>> Coverity build failed. Adjust the prototypes accordingly. >>> >>> This sounds wrong or at least different from the past. >>> >> >> I didn't push yet... >> >> The last time this happened is/was commit 425aac3abf. >> >> Still, before this patch if you build with STATIC_ANALYSIS set you'd see >> the same phenomena. If I modify config.h to have STATIC_ANALYSIS be 1 >> instead of 0, I get: >> >> util/vircommand.c: In function 'virCommandAddArg': >> util/vircommand.c:1501:8: error: nonnull argument 'val' compared to NULL >> [-Werror=nonnull-compare] >> if (val == NULL) { >> ^ >> util/vircommand.c: In function 'virCommandAddArgPair': >> util/vircommand.c:1604:14: error: nonnull argument 'name' compared to >> NULL [-Werror=nonnull-compare] >> if (name == NULL || val == NULL) { >> ^ >> util/vircommand.c:1604:29: error: nonnull argument 'val' compared to >> NULL [-Werror=nonnull-compare] >> if (name == NULL || val == NULL) { >> ^ >> Is there a different way you'd prefer this resolved? I could leave it as >> a my coverity environment only, but I would assume clang would be >> impacted too, although I don't use clang. >> >> >>> We have historically still added nonnull annotations even when >>> having checks for NULL in the impl. We had to explicitly disabble >>> the nonnull annotation when building under GCC to prevent it from >>> optimizing out the NULL checks. We left it enabled for STATIC_ANALYSIS >>> so that coverity would still report if any callers passed a NULL into >>> the methods. >>> >> >> That wasn't my recollection. Adding ATTRIBUTE_NONNULL has been "less >> important" because we found that it really only helps if someone calls >> some API with NULL and it does not help if the argument itself is NULL. >> My recollection is always towards removing it once the attribute itself >> was checked in the API. I also have a faint recollection of a discussion >> we've had before on this with the thought towards removing all the >> ATTRIBUTE_NONNULL's and replacing with real argument == NULL checks, but >> that got to be too many patches and checks in too many places. > > This goes way back to this commit: > > commit eefb881d4683d50882b43e5b28b0e94657cd0c9c > Author: Laine Stump <laine(a)laine.org&gt; > Date: Wed Apr 25 16:10:01 2012 -0400 > > build: make ATTRIBUTE_NONNULL() a NOP unless STATIC_ANALYSIS is on > > The ATTRIBUTE_NONNULL(m) macro normally resolves to the gcc builtin > __attribute__((__nonnull__(m))). The effect of this in gcc is > unfortunately only to make gcc believe that "m" can never possibly be > NULL, *not* to add in any checks to guarantee that it isn't ever NULL > (i.e. it is an optimization aid, *not* something to verify code > correctness.) - see the following gcc bug report for more details: > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17308 And Eric's comment from Apr 2012 eventually answered in Sept 2015 would seem to indicate to me the exact thing that happened in this code and why removing the ATTRIBUTE_NONNULL from the header file would be necessary. Maybe I'm reading it wrong, but it appears to say that if ATTRIBUTE_NONNULL was provided in the prototype and the code did a comparison of the variable to NULL, then some sort of warning would be issued. And that's what I see in the code provided. I really only skimmed and didn't spend too many cycles thinking about it though. > > Static source analyzers such as clang and coverity apparently can use > ATTRIBUTE_NONNULL(), though, to detect dead code (in the case that the > arg really is guaranteed non-NULL), as well as situations where an > obviously NULL arg is given to the function. > > https://bugzilla.redhat.com/show_bug.cgi?id=815270 is a good example > of a bug caused by erroneous application of ATTRIBUTE_NONNULL(). > Several people spent a long time staring at this code and not finding > the problem, because the problem wasn't in the function itself, but in > the prototype that specified ATTRIBUTE_NONNULL() for an arg that > actually *wasn't* always non-NULL, and caused a segv when dereferenced > (even though the code that dereferenced the pointer was inside an if() > that checked for a NULL pointer, that code was optimized out by gcc). > > There may be some very small gain to be had from the optimizations > that can be inferred from ATTRIBUTE_NONNULL(), but it seems safer to > err on the side of generating code that behaves as expected, while > turning on the attribute for static analyzers. > > > We had methods doing "if (foo == NULL) return 0" and had code that > was (mistakenly) passing in NULL which would have been safe, except > gcc had deleted the "if (foo == NULL)" check. Rather than removing > all attribute(nonnull) checks, we merely disabling it under gcc. We > left it under STATIC_ANALYSIS so that coverity could still report on > places which passed an explicit NULL into a method annotated nonnull. > > Um, OK. Well, I've been removing ATTRIBUTE_NONNULL when encountering this same error in the past. It just seemed to be the path of least resistance. After all now regardless of whether its "bar(NULL)" or "bar(foo)" where foo==NULL, if the code in bar does "if (arg1 == NULL) xxx;" then we're covered. In any case, I'll just leave this change in my local Coverity tree and drop this patch.

John -- libvir-list mailing list libvir-list(a)redhat.com https://www.redhat.com/mailman/listinfo/libvir-list

> We have historically still added nonnull annotations even when > having checks for NULL in the impl. We had to explicitly disabble > the nonnull annotation when building under GCC to prevent it from > optimizing out the NULL checks. We left it enabled for STATIC_ANALYSIS > so that coverity would still report if any callers passed a NULL into > the methods. Looking at the gcc bug and following some of the links mentioned therein, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=17308 https://gcc.gnu.org/ml/gcc-patches/2016-02/msg01103.html It looks like modern gcc has added -Wnonnull-compare that noisily informs us any time we have an ATTRIBUTE_NONNULL mismatch with the body of a function comparing that parameter against NULL after all. And gnulib's manywarnings module turns that on automatically (at least for gcc 8.2.1 on Fedora 29). That is sufficient to fix the bug that led us to historically disable ATTRIBUTE_NONNULL under gcc (commit eefb881), so maybe it's time to just blindly enable ATTRIBUTE_NONNULL everywhere on the grounds that we have enough developers and CI coverage to now immediately catch inconsistent attributes rather than risking silently mis-compiled code that omits the branch after a NULL comparison as dead code.

>> >> We have historically still added nonnull annotations even when >> having checks for NULL in the impl. We had to explicitly disabble >> the nonnull annotation when building under GCC to prevent it from >> optimizing out the NULL checks. We left it enabled for STATIC_ANALYSIS >> so that coverity would still report if any callers passed a NULL into >> the methods. > > Looking at the gcc bug and following some of the links mentioned therein, > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=17308 > https://gcc.gnu.org/ml/gcc-patches/2016-02/msg01103.html > > It looks like modern gcc has added -Wnonnull-compare that noisily > informs us any time we have an ATTRIBUTE_NONNULL mismatch with the body > of a function comparing that parameter against NULL after all. And > gnulib's manywarnings module turns that on automatically (at least for > gcc 8.2.1 on Fedora 29). > > That is sufficient to fix the bug that led us to historically disable > ATTRIBUTE_NONNULL under gcc (commit eefb881), so maybe it's time to just > blindly enable ATTRIBUTE_NONNULL everywhere on the grounds that we have > enough developers and CI coverage to now immediately catch inconsistent > attributes rather than risking silently mis-compiled code that omits the > branch after a NULL comparison as dead code. Yes & no. Originally the desire was that we would have functions which were marked NONNULL and *also* have "if (foo == NULL) ... return..' as a second safety net.

IOW, we were aiming to get *both* compile time and runtime checking for NULL, since compile time checks were unsufficiently safe on their own. The compiler was "clever" and optimized out the runtime checks, leaving only the incomplete compile time checks.

We compromised and only set __attribute__(nonnull) under coverity. The idea was that coverity would get us the static checks as a substiute for the compile time checks, while we still have the runtime checks.

Using -Wnonnull-compare would detect when gcc did this dangerous optimization, but it would still force us to either remove either the compile time check, or the runtime check.

I do wonder if we can enable __attribute__(nonnull) under GCC and then set '-fdelete-null-pointer-checks' to stop it deleting our runtime NULL checks, so we get both compile time & runtime protection.

For any pointer parameter there's then two possibilities. Either NULL is a valid value and results in some useful behaviour, or NULL is not valid and libvirt will catch that and report an error.

The latter are essentially all programmer bugs, but we wanted to be robust and catch them at runtime & report graceful error instead of referencing NULL or assert()ing. We added ATTRIBUTE_NONNULL widely to try to catch the latter at compile time too. This also served as documentation in the header files to indicate that you shouldn't be passing NULL for that parameter. We later discovered our usage of ATTRIBUTE_NONNULL was based on a mis-understanding of the functional impact of that annotation, and so #defined it to a no-op, so it merely served as documentation that you should not pass NULL to a method. It is still desirable for us to find a way to do all three tasks 1. document that NULL should not be passed for a param 2. detect mistakes passing NULL via static analysis 3. retain runtime graceful NULL param checking. AFAICT we have a choice either a. Replace ATTRIBUTE_NONNULL to just a plain comment /* nonnull */ against each parameter. This covers items 1 & 2, mostly as we do today, while stopping coverity getting upset b. Re-enable ATTRIBUTE_NONNULL to expand to attribute(nonnull) but only under modern GCC, and add -fdelete-null-pointer-checks This covers items 1, 2 & 3, under GCC, while stopping coverity getting upset