Date: Mon, 10 Apr 2023 21:42:18 +0300 From: Christos Margiolis <christos@freebsd.org> To: status-updates@freebsdfoundation.org Cc: freebsd-dtrace@freebsd.org, markj@freebsd.org, jrm@freebsd.org, mhorne@freebsd.org Subject: [Development report #7] Improve the kinst DTrace provider Message-ID: <20230410184218.i6gez6eusmxz2zli@pleb>
next in thread | raw e-mail | index | archive | help
The past few days I've been working on the forthcoming RISC-V port of
kinst.
Tracing of individual instructions in kinst so far (i.e in amd64) works
by using a trampoline which contains the actual instruction, followed by
an unconditional jump back to the next instruction so that we can resume
execution. In RISC-V however this turns out to not really be possible,
because we cannot encode large displacement values in a single
instruction in order to form a far jump [1]. To compensate for this, the
RISC-V trampoline contains the instruction, followed by an EBREAK
(breakpoint) instruction. kinst_invop() is able to detect if the
breakpoint was triggered by the trampoline, where in this case, it
manually sets the program counter to the next instruction, instead of
using a jump.
So far only instructions that do not address memory can be traced
without problems. Since the displacement issue I mentioned above
re-appears here as well, instructions that need modification [2] will
most likely not use the trampoline at all, and be emulated in software
instead, like we do in amd64 kinst for call instructions.
[1] Trampoline addresses are located above the kernel base address,
which can be several MBs away from the instruction we are tracing.
[2] Some instructions need modification, that is, for instructions that
use relative addressing (e.g PC-relative ones), we have to
recalculate the displacement to be relative to the trampoline
address.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20230410184218.i6gez6eusmxz2zli>