On Fri, Nov 19, 2021 at 04:57:46PM +0100, Arnd Bergmann wrote: > On Fri, Nov 19, 2021 at 4:06 PM Alejandro Colomar (man-pages) > <alx.manpages@xxxxxxxxx> wrote: > > On 11/19/21 15:47, Arnd Bergmann wrote: > > > On Fri, Nov 19, 2021 at 12:36 PM Alejandro Colomar > > > > Yes, I would like to untangle the dependencies. > > > > The main reason I started doing this splitting > > is because I wouldn't be able to include > > <linux/stddef.h> in some headers, > > because it pulled too much stuff that broke unrelated things. > > > > So that's why I started from there. > > > > I for example would like to get NULL in memberof() > > without puling anything else, > > so <linux/NULL.h> makes sense for that. > > > > It's clear that every .c wants NULL, > > but it's not so clear that every .c wants > > everything that <linux/stddef.h> pulls indirectly. > > From what I can tell, linux/stddef.h is tiny, I don't think it's really > worth optimizing this part. I have spent some time last year > trying to untangle some of the more interesting headers, but ended > up not completing this as there are some really hard problems > once you start getting to the interesting bits. > > The approach I tried was roughly: > > - For each header in the kernel, create a preprocessed version > that includes all the indirect includes, from that start a set > of lookup tables that record which header is eventually included > by which ones, and the size of each preprocessed header in > bytes > > - For a given kernel configuration (e.g. defconfig or allmodconfig) > that I'm most interested in, look at which files are built, and what > the direct includes are in the source files. > > - Sort the headers by the product of the number of direct includes > and the preprocessed size: the largest ones are those that are > worth looking at first. > > - use graphviz to visualize the directed graph showing the includes > between the top 100 headers in that list. You get something like > I had in [1], or the version afterwards at [2]. > > - split out unneeded indirect includes from the headers in the center > of that graph, typically by splitting out struct definitions. > > - repeat. > > The main problem with this approach is that as soon as you start > actually reducing the unneeded indirect includes, you end up with > countless .c files that no longer build because they are missing a > direct include for something that was always included somewhere > deep underneath, so I needed a second set of scripts to add > direct includes to every .c file. Can't it be done with cocci support? > On the plus side, I did see something on the order of a 30% > compile speed improvement with clang, which is insane > given that this only removed dead definitions. Thumb up! > > But I'll note that linux/fs.h, linux/sched.h, linux/mm.h are > > interesting headers for further splitting. > > > > > > BTW, I also have a longstanding doubt about > > how header files are organized in the kernel, > > and which headers can and cannot be included > > from which other files. > > > > For example I see that files in samples or scripts or tools, > > that redefine many things such as offsetof() or ARRAY_SIZE(), > > and I don't know if there's a good reason for that, > > or if I should simply remove all that stuff and > > include <linux/offsetof.h> everywhere I see offsetof() being used. > > The main issue here is that user space code should not > include anything outside of include/uapi/ and arch/*/include/uapi/ > > offsetof() is defined in include/linux/stddef.h, so this is by > definition not accessible here. It appears that there is also > an include/uapi/linux/stddef.h that is really strange because > it includes linux/compiler_types.h, which in turn is outside > of uapi/. This should probably be fixed. > > Arnd > > [1] https://drive.google.com/file/d/14IKifYDadg2W5fMsefxr4373jizo9bLl/view?usp=sharing > [2] https://drive.google.com/file/d/1pWQcv3_ZXGqZB8ogV-JOfoV-WJN2UNnd/view?usp=sharing -- With Best Regards, Andy Shevchenko
