Hi 2022. november 23., szerda 20:06 keltezéssel, Greg Kroah-Hartman írta: > On Wed, Nov 23, 2022 at 02:25:59PM -0400, Jason Gunthorpe wrote: > > > On Wed, Nov 23, 2022 at 07:10:49PM +0100, Greg Kroah-Hartman wrote: > > > > > On Wed, Nov 23, 2022 at 05:49:36PM +0000, Matthew Wilcox wrote: > > > > > > > On Wed, Nov 23, 2022 at 01:29:56PM -0400, Jason Gunthorpe wrote: > > > > > > > > > #define generic_container_of(in_type, in, out_type, out_member) \ > > > > > _Generic(in, \ > > > > > const in_type *: ((const out_type *)container_of(in, out_type, out_member)), \ > > > > > in_type *: ((out_type *)container_of(in, out_type, out_member)) \ > > > > > ) > > > > > > > > There's a neat trick I found in seqlock.h: > > > > > > > > #define generic_container_of(in_t, in, out_t, m) \ > > > > _Generic(*(in), \ > > > > const in_t: ((const out_t *)container_of(in, out_t, m)), \ > > > > in_t: ((out_t *)container_of(in, out_type, m)) \ > > > > ) > > > > > > > > and now it fits in 80 columns ;-) > > > > > > Nice trick! Dropping the inline functions is a bit different, let me > > > see if that still gives a sane error if we pass an incorrect type or > > > mess with the const * the wrong way. I'll run some tests tomorrow > > > afternoon... > > > > The errors in some cases are very verbose, but it is somewhat > > understandable - the worst is when _Generic fails to match anything, > > but also at least clang partially expanded container_of and it throws > > other assertions too. > > > > I also wonder if this could just be rolled into the normal > > container_of. > > > I think we might be able to now, my previous attempts with inline > functions prevented that. I'll beat on that tomorrow... > > > in_type would have to be derived like: > > > > in_type = typeof((out_type *)NULL)->out_member) > > > > But I don't know if you can use typeof in a generic type matching expression.. > > > Maybe that is what threw me before, I can't remember. I do know we > tried a number of different attempts, can't recall the failed ones... > I am sorry I haven't followed the previous discussion, but has something like the following been considered? #define container_of2(ptr, type, member) \ _Generic((ptr), \ typeof_member(const type, member) *: container_of((ptr), const type, member), \ const void *: container_of((ptr), const type, member), \ default: container_of((ptr), type, member) \ ) This is not perfect by any means, but I believe it is a reasonable improvement. It only really works reliably if the member type and the specified `type` in the arguments are unqualified, but I believe that should cover the majority of cases. I have tried to build an x86-64 defconfig with this and there are numerous const violations, but I haven't seen any false positives. > [...] Regards, Barnabás Pőcze
