> On 6 Aug 2025, at 18:43, Jessica Clarke <jrt...@freebsd.org> wrote: > > On 6 Aug 2025, at 18:38, Andrew Turner <and...@freebsd.org > <mailto:and...@freebsd.org>> wrote: >> >> The branch main has been updated by andrew: >> >> URL: >> https://cgit.FreeBSD.org/src/commit/?id=81f07332c03fd2ac6efa8e15b1659a573d250329 >> >> commit 81f07332c03fd2ac6efa8e15b1659a573d250329 >> Author: Harry Moulton <harry.moul...@arm.com> >> AuthorDate: 2025-07-31 14:10:57 +0000 >> Commit: Andrew Turner <and...@freebsd.org> >> CommitDate: 2025-07-31 14:27:06 +0000 >> >> arm64: tidy up Top-Byte-Ignore (TBI) in the kernel >> >> In preparation for TBI to be enabled for processes from 15.0 we need >> to clean up copying data between userspace and the kernel. These >> functions will check the address is within the valid userspace range, >> however as the userspace and kernel ranges may overlap when TBI is >> enabled we need to mask off the top 8 bits. >> >> Processes not using TBI are unaffected as the hardware will still >> check all bits in the address, however this will happen at the first >> load/store instruction. > > I thought Linux’s ABI was that addresses passed to the kernel must be > canonical. Is that false, or is there a reason we’re diverging from > that design choice?
On Linux there is a flag to enable passing tagged pointers to the kernel. It’s needed for HWASAN to work as any pointers may be tagged, and I expect it will also be needed for MTE. I don’t know of a strong reason to not allow tagged pointers in the kernel. Andrew
