> -----Original Message----- > From: Marc Sune [mailto:marc.sune at bisdn.de] > Sent: Tuesday, March 03, 2015 2:39 PM > To: Ananyev, Konstantin; Richardson, Bruce > Cc: dev at dpdk.org > Subject: Re: [dpdk-dev] [PATCH] mk: add support for gdb debug info generation > > > On 03/03/15 14:31, Ananyev, Konstantin wrote: > > > >> -----Original Message----- > >> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Bruce Richardson > >> Sent: Tuesday, March 03, 2015 1:03 PM > >> To: Marc Sune > >> Cc: dev at dpdk.org > >> Subject: Re: [dpdk-dev] [PATCH] mk: add support for gdb debug info > >> generation > >> > >> On Tue, Mar 03, 2015 at 01:56:19PM +0100, Marc Sune wrote: > >>> On 03/03/15 13:40, Panu Matilainen wrote: > >>>> On 03/03/2015 02:19 PM, Marc Sune wrote: > >>>>> On 03/03/15 10:33, Bruce Richardson wrote: > >>>>>> On Mon, Mar 02, 2015 at 06:32:13PM +0100, Marc Sune wrote: > >>>>>>> On 22/02/15 12:51, Marc Sune wrote: > >>>>>>>> I don't like the proposed patch, but I am recovering this old thread > >>>>>>>> because I agree on the problem statement. > >>>>>>>> > >>>>>>>> On 04/04/14 11:57, Ananyev, Konstantin wrote: > >>>>>>>>> Hi Cyril, > >>>>>>>>> We already do have 'EXTRA_CFLAGS' and 'EXTRA_LDFLAGS' that you > >>>>>>>>> can use > >>>>>>>>> to enable debug, or any other compiler/linker options you need. > >>>>>>>>> Wonder, why that is not enough? > >>>>>>>> EXTRA_FLAGS var affects all the DPDK libraries. I was wondering why > >>>>>>>> setting individually: > >>>>>>>> > >>>>>>>> diff --git a/config/common_linuxapp b/config/common_linuxapp > >>>>>>>> index 2f9643b..04adc0d 100644 > >>>>>>>> --- a/config/common_linuxapp > >>>>>>>> +++ b/config/common_linuxapp > >>>>>>>> @@ -127,7 +127,7 @@ CONFIG_RTE_LIBRTE_KVARGS=y > >>>>>>>> # Compile generic ethernet library > >>>>>>>> # > >>>>>>>> CONFIG_RTE_LIBRTE_ETHER=y > >>>>>>>> -CONFIG_RTE_LIBRTE_ETHDEV_DEBUG=n > >>>>>>>> +CONFIG_RTE_LIBRTE_ETHDEV_DEBUG=y > >>>>>>>> > >>>>>>>> > >>>>>>>> to put an example, does not set -g and -O0 in that particular module > >>>>>>>> only. > >>>>>>>> No one would ever use something compiled in DEBUG in production > >>>>>>>> anyway. > >>>>>>>> > >>>>>>>> I always end up modifying manually Makefiles in the lib library that > >>>>>>>> I am > >>>>>>>> interested in having insides, overriding CFLAGS=-O3, which is not > >>>>>>>> that > >>>>>>>> nice. > >>>>>>> I would like some feedback on this idea. If the community sees > >>>>>>> benefit, I > >>>>>>> will work on a patch for this. > >>>>>>> > >>>>>>> Marc > >>>>>>> > >>>>>> So, your proposal is to patch things so that any time one sets DEBUG=y > >>>>>> in the > >>>>>> build-time config for a library, we change the '-O3' to '-O0' and set > >>>>>> -g also. > >>>>>> Correct? > >>>>> I am not sure what you mean by 'patch things'. I would simply enable the > >>>>> build system to override the default compilation flags (now DPDK-wide, > >>>>> or specifically librte_ wide) when _DEBUG=y for a library, changing > >>>>> compilation flags from -O3 to -O0 -g and possibly also -fno-inline. I > >>>>> have to check if -O0 already implicitly means -fno-inline (even for > >>>>> __attribute__((always_inline)) ). > >>>>> > >>>>> I did a quick test. I chose KNI because it didn't have a DEBUG flag for > >>>>> the user-space library. For other libraries, the existing _DEBUG setting > >>>>> would be enough: > >>>>> > >>>>> marc at dpdk:~/dpdk$ git diff HEAD > >>>>> diff --git a/config/common_linuxapp b/config/common_linuxapp > >>>>> index 97f1c9e..8a3cef8 100644 > >>>>> --- a/config/common_linuxapp > >>>>> +++ b/config/common_linuxapp > >>>>> @@ -405,6 +405,7 @@ CONFIG_RTE_LIBRTE_PIPELINE=y > >>>>> # > >>>>> CONFIG_RTE_LIBRTE_KNI=y > >>>>> CONFIG_RTE_KNI_PREEMPT_DEFAULT=y > >>>>> +CONFIG_RTE_LIBRTE_KNI_DEBUG=y > >>>>> CONFIG_RTE_KNI_KO_DEBUG=n > >>>>> CONFIG_RTE_KNI_VHOST=n > >>>>> CONFIG_RTE_KNI_VHOST_MAX_CACHE_SIZE=1024 > >>>>> diff --git a/lib/librte_kni/Makefile b/lib/librte_kni/Makefile > >>>>> index 7107832..895f64e 100644 > >>>>> --- a/lib/librte_kni/Makefile > >>>>> +++ b/lib/librte_kni/Makefile > >>>>> @@ -34,7 +34,7 @@ include $(RTE_SDK)/mk/rte.vars.mk > >>>>> # library name > >>>>> LIB = librte_kni.a > >>>>> > >>>>> -CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -O3 -fno-strict-aliasing > >>>>> +CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) $(CONFIG_RTE_LIBRTE_KNI_CFLAGS) > >>>>> > >>>>> EXPORT_MAP := rte_kni_version.map > >>>>> > >>>>> diff --git a/mk/rte.app.mk b/mk/rte.app.mk > >>>>> index 63a41e2..eee477d 100644 > >>>>> --- a/mk/rte.app.mk > >>>>> +++ b/mk/rte.app.mk > >>>>> @@ -78,6 +78,13 @@ endif > >>>>> ifeq ($(CONFIG_RTE_LIBRTE_KNI),y) > >>>>> ifeq ($(CONFIG_RTE_EXEC_ENV_LINUXAPP),y) > >>>>> LDLIBS += -lrte_kni > >>>>> + > >>>>> +ifeq ($(CONFIG_RTE_LIBRTE_KNI_DEBUG),y) > >>>>> +CONFIG_RTE_LIBRTE_KNI_CFLAGS = -O0 -g -fno-strict-aliasing -fno-inline > >>>>> +else > >>>>> +CONFIG_RTE_LIBRTE_KNI_CFLAGS = -O3 -fno-strict-aliasing > >>>>> +endif > >>>>> + > >>>>> endif > >>>>> endif > >>>>> > >>>>> > >>>>> Thoughts? > >>>> My 5c is that if anything, DPDK needs *less* places that muck around with > >>>> compiler flags, not more. If you something like this for all the > >>>> libraries > >>>> in DPDK the number doesn't just increase a bit, it explodes. > >>> If you check the part below this one in my original email, that you > >>> stripped > >>> out (without notice), the suggestion was also to add a global _DEBUG > >>> parameter for the entire DPDK set of libraries, to change all the CFLAGS > >>> at > >>> once (not in the attached PATCH). > >>> > >>>> I dont see that much point in this thing, but I'd approach it by defining > >>>> the debug flags someplace central, say DEBUG_FLAGS, and append that to > >>>> the > >>>> common cflags when *_DEBUG config is enabled. At least with gcc the last > >>>> option wins so if you just append -O0 when debugging then that's what > >>>> wins, the earlier -O3 does not matter. > >>> The original problem is the one you expose; libraries hardcode the CFLAGS, > >>> ignoring user-flags. There is no way to change this unless you change the > >>> Makefiles directly. > >>> > >>> But right now, each library does hardcode its *own* flags (check Makefiles > >>> for the libraries), so there is already not a unified approach here. I see > >>> for instance KNI having -fno-strict-aliasing while other libraries don't. > >>> > >>> Having said that, there are moments, specially with -O3, in which to be > >>> able > >>> to reproduce a bug, you need to compile certain parts of code with -O3 and > >>> the rest with -O0 -g (the ones to be debugged). The approach proposed > >>> (both > >>> a global *and* a lib specific) allows that. > >>> > >>> Marc > >>> > >> I believe that the global option of overriding the CFLAGS is already > >> sufficiently > >> covered - including being documented in programmers guide - by > >> EXTRA_CFLAGS. The > >> ability to turn off optimization support for a single library is not > >> covered > >> anywhere, and that suggestion seems reasonable to me. For each library, we > >> can > >> just append '-O0 -g' to the CFLAGS in that libraries makefile if the debug > >> option > >> is set. I don't see that as significantly complicating things [though I > >> wouldn't > >> make any changes to the rte.app.mk to allow this, just have it per lib in > >> the > >> lib's makefile] > >> > > Well, but then people would say 'why only -g -O0' for _DEBUG config option? > > And would ask for ability to add/change other compile options to add/change > > for particular library. > > If we need that, probably we can introduce > > RTE_LIBRTE_<LIBNAME>_EXTRA_CFLAGS config macros for each library? > > So people can set them in their config files to whatever they like? > > It is not a bad idea either. However, there are already those _DEBUG > config elements (the KNI is a special case).
> > Maybe both can be combined (i.e. RTE_LIBRTE_<LIBNAME>_EXTRA_CFLAGS > should always override _DEBUG). It would probably get a bit > over-engineered then, though. I would probably leave _DEBUG config options with what they are doing right now: perform extra checks and generate logs. After all, people might want their program to be more verbose, but they don't want compile options to be changed. Konstantin > > marc > > > > Konstantin
