Recently, the addition of some vector register save restore logic from:
https://github.com/openssl/openssl/pull/28990
Triggered an issue in our quic_multistream_test, notably causing a
connection shutdown with error code 1, rather than successful
completion.
The problem has been root caused to an error in how the ppc aes code
saves and restores vector registers.
The aes gcm code uses VSR instructions (vsldoi, etc), to maniuplate
vector registers, which only operate on the upper half of the vector
register file (VS32-63), and operands are implicitly offset to do this.
(i.e. <instr> v31 actually operates on register vs63).
However, the SAVE/RESTORE macros which save and restore those values use
stxv/lxv instructions to do the save and restore, which are VSX
instructions that have access to the full vector register file, and so
by restoring V1, rather than V33, etc, we actuall restore values to the
wrong registers, causing various issues (most notably the above
misconfiguration of the quic SSL shutdown event, which sets the error
code erroneously to 1 instead of the expected 0 value).
Fix it by offsetting the SAVE_REGS and RESTORE_REGS macros to reference
the proper registers when doing save and restore.
Fixes #29516
Reviewed-by: Saša Nedvědický <sashan@openssl.org>
Reviewed-by: Tim Hudson <tjh@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/29528)
std 23, 184(1)
std 24, 192(1)
- stxv 20, 256(1)
- stxv 21, 256+16(1)
- stxv 22, 256+32(1)
- stxv 23, 256+48(1)
- stxv 24, 256+64(1)
- stxv 25, 256+80(1)
- stxv 26, 256+96(1)
- stxv 27, 256+112(1)
- stxv 28, 256+128(1)
- stxv 29, 256+144(1)
- stxv 30, 256+160(1)
- stxv 31, 256+176(1)
+ stxv 32+20, 256(1)
+ stxv 32+21, 256+16(1)
+ stxv 32+22, 256+32(1)
+ stxv 32+23, 256+48(1)
+ stxv 32+24, 256+64(1)
+ stxv 32+25, 256+80(1)
+ stxv 32+26, 256+96(1)
+ stxv 32+27, 256+112(1)
+ stxv 32+28, 256+128(1)
+ stxv 32+29, 256+144(1)
+ stxv 32+30, 256+160(1)
+ stxv 32+31, 256+176(1)
.endm # SAVE_REGS
.macro RESTORE_REGS
- lxv 20, 256(1)
- lxv 21, 256+16(1)
- lxv 22, 256+32(1)
- lxv 23, 256+48(1)
- lxv 24, 256+64(1)
- lxv 25, 256+80(1)
- lxv 26, 256+96(1)
- lxv 27, 256+112(1)
- lxv 28, 256+128(1)
- lxv 29, 256+144(1)
- lxv 30, 256+160(1)
- lxv 31, 256+176(1)
+ lxv 32+20, 256(1)
+ lxv 32+21, 256+16(1)
+ lxv 32+22, 256+32(1)
+ lxv 32+23, 256+48(1)
+ lxv 32+24, 256+64(1)
+ lxv 32+25, 256+80(1)
+ lxv 32+26, 256+96(1)
+ lxv 32+27, 256+112(1)
+ lxv 32+28, 256+128(1)
+ lxv 32+29, 256+144(1)
+ lxv 32+30, 256+160(1)
+ lxv 32+31, 256+176(1)
ld 14, 112(1)
ld 15, 120(1)