----- On Jul 15, 2020, at 10:58 AM, Florian Weimer fweimer@xxxxxxxxxx wrote: > * Mathieu Desnoyers: > >> ----- On Jul 15, 2020, at 9:42 AM, Florian Weimer fweimer@xxxxxxxxxx wrote: >>> * Mathieu Desnoyers: >>> >> [...] >>>> How would this allow early-rseq-adopter libraries to interact with >>>> glibc ? >>> >>> Under all extension proposals I've seen so far, early adopters are >>> essentially incompatible with glibc rseq registration. I don't think >>> you can have it both ways. >> >> The basic question I'm not sure about is whether we are allowed to increase >> the size and alignement of __rseq_abi from e.g. glibc 2.32 to glibc 2.33. > > With the current mechanism (global TLS data symbol), we can do that > using symbol versioning. That means that we can only do this on a > release boundary, That should not be a problem. > and that it's incompatible with other libraries which > use an interposing unversioned symbol. We have the freedom to define the ABI of this shared __rseq_abi symbol right now. Maybe it's not such a good thing to let early adopters use unversioned __rseq_abi symbols. Let me wrap my head around this scenario then, please let me know if I'm misunderstanding something: 1) glibc 2.32 exposes: __rseq_abi (GLIBC_2.32) with size == 32. __rseq_abi with size == 32 is available as a private symbol within glibc - both symbols alias the same contents. 2) glibc 2.33 exposes: __rseq_abi (GLIBC_2.32) with size == 32. __rseq_abi (GLIBC_2.33) with size == 64. __rseq_abi with size == 64 is available as a private symbol within glibc - the three symbols alias the same contents. Then what happens if we have a program or preloaded library defining __rseq_abi (without version) with size == 32 loaded with a glibc 2.33 ? Or what happens if we have a program or preloaded libary defining __rseq_abi (GLIBC_2.32) with size == 32 loaded with a glibc 2.33 ? I wonder if "GLIBC_*" is the right version namespace for this. Considering that the layout of this structure is defined by the Linux kernel UAPI, maybe we'd want version named as "RSEQ_1.0", "RSEQ_2.0" or something similar. Thanks, Mathieu -- Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com