| Message ID | 20230808042001.411094-1-npiggin@gmail.com |
|---|---|
| Headers | show |
| Series | ppc: record-replay enablement and fixes | expand |
Hello, On 8/8/23 06:19, Nicholas Piggin wrote: > The patches in this series has been seen a few times in various > iterations. > > There are two main pieces, some assorted small fixes and tests for > record-replay, plus a large set of decrementer fixes. I merged these > into one series rather than send decrementer fixes alone first, because > record-replay has been very good at uncovering timer problems, so it's > good to have those test cases in at the same time IMO. > > Some of the fixes we might take to stable, but unclear which. > Decrementer fixes were a bit of a tangle so maybe we just leave those > alone since they work okay. > > The decrementer is not emulated perfectly still. Underflow from -ve > to +ve is not implemented, for one. I started doing that but it's not > trivial so better stop here for now. > > For record-replay, pseries is now quite solid with rr. Surely some > issues to iron out but it is becoming usable. > > powernv record-replay has some known problems migrating edge-triggered > decrementer, and edge triggered msgsnd. Also it seems to get stuck in > xive init somewhere when replaying from checkpoint, so there is probably > some state in xive not being reset. But at least it runs the avocado > test and seems close to working, so I've added that test case so we > don't go backwards (ha!). > > Other machine types might not be too far off if there is interest. I > found it quite difficult to find these problems though, reverse > debugging will sometimes just lock up, stop at wrong location, or abort > with wrong event. Difficult understand what went wrong. Worst case I had > to basically bisect the replay of the trace, and find the minimum length > of replay that hit the problem -- that sometimes would land near a > mtDEC or timer interrupt or similar. > > Thanks, > Nick > > Nicholas Piggin (19): > ppc/vhyp: reset exception state when handling vhyp hcall > ppc/vof: Fix missed fields in VOF cleanup > hw/ppc/ppc.c: Tidy over-long lines > hw/ppc: Introduce functions for conversion between timebase and > nanoseconds > host-utils: Add muldiv64_round_up > hw/ppc: Round up the decrementer interval when converting to ns > hw/ppc: Avoid decrementer rounding errors > target/ppc: Sign-extend large decrementer to 64-bits > hw/ppc: Always store the decrementer value > target/ppc: Migrate DECR SPR > hw/ppc: Reset timebase facilities on machine reset > hw/ppc: Read time only once to perform decrementer write > target/ppc: Fix CPU reservation migration for record-replay > target/ppc: Fix timebase reset with record-replay > spapr: Fix machine reset deadlock from replay-record > spapr: Fix record-replay machine reset consuming too many events > tests/avocado: boot ppc64 pseries replay-record test to Linux VFS > mount > tests/avocado: reverse-debugging cope with re-executing breakpoints > tests/avocado: ppc64 reverse debugging tests for pseries and powernv > > hw/ppc/mac_oldworld.c | 1 + > hw/ppc/pegasos2.c | 1 + > hw/ppc/pnv_core.c | 2 + > hw/ppc/ppc.c | 236 +++++++++++++++++++---------- > hw/ppc/prep.c | 1 + > hw/ppc/spapr.c | 32 +++- > hw/ppc/spapr_cpu_core.c | 2 + > hw/ppc/vof.c | 2 + > include/hw/ppc/ppc.h | 3 +- > include/hw/ppc/spapr.h | 2 + > include/qemu/host-utils.h | 21 ++- > target/ppc/compat.c | 19 +++ > target/ppc/cpu.h | 3 + > target/ppc/excp_helper.c | 3 + > target/ppc/machine.c | 40 ++++- > target/ppc/translate.c | 4 + > tests/avocado/replay_kernel.py | 3 +- > tests/avocado/reverse_debugging.py | 54 ++++++- > 18 files changed, 330 insertions(+), 99 deletions(-) > I am preparing a PR with this series. It is time to take a look at it if you haven't already ! Thanks, C.
The patches in this series has been seen a few times in various iterations. There are two main pieces, some assorted small fixes and tests for record-replay, plus a large set of decrementer fixes. I merged these into one series rather than send decrementer fixes alone first, because record-replay has been very good at uncovering timer problems, so it's good to have those test cases in at the same time IMO. Some of the fixes we might take to stable, but unclear which. Decrementer fixes were a bit of a tangle so maybe we just leave those alone since they work okay. The decrementer is not emulated perfectly still. Underflow from -ve to +ve is not implemented, for one. I started doing that but it's not trivial so better stop here for now. For record-replay, pseries is now quite solid with rr. Surely some issues to iron out but it is becoming usable. powernv record-replay has some known problems migrating edge-triggered decrementer, and edge triggered msgsnd. Also it seems to get stuck in xive init somewhere when replaying from checkpoint, so there is probably some state in xive not being reset. But at least it runs the avocado test and seems close to working, so I've added that test case so we don't go backwards (ha!). Other machine types might not be too far off if there is interest. I found it quite difficult to find these problems though, reverse debugging will sometimes just lock up, stop at wrong location, or abort with wrong event. Difficult understand what went wrong. Worst case I had to basically bisect the replay of the trace, and find the minimum length of replay that hit the problem -- that sometimes would land near a mtDEC or timer interrupt or similar. Thanks, Nick Nicholas Piggin (19): ppc/vhyp: reset exception state when handling vhyp hcall ppc/vof: Fix missed fields in VOF cleanup hw/ppc/ppc.c: Tidy over-long lines hw/ppc: Introduce functions for conversion between timebase and nanoseconds host-utils: Add muldiv64_round_up hw/ppc: Round up the decrementer interval when converting to ns hw/ppc: Avoid decrementer rounding errors target/ppc: Sign-extend large decrementer to 64-bits hw/ppc: Always store the decrementer value target/ppc: Migrate DECR SPR hw/ppc: Reset timebase facilities on machine reset hw/ppc: Read time only once to perform decrementer write target/ppc: Fix CPU reservation migration for record-replay target/ppc: Fix timebase reset with record-replay spapr: Fix machine reset deadlock from replay-record spapr: Fix record-replay machine reset consuming too many events tests/avocado: boot ppc64 pseries replay-record test to Linux VFS mount tests/avocado: reverse-debugging cope with re-executing breakpoints tests/avocado: ppc64 reverse debugging tests for pseries and powernv hw/ppc/mac_oldworld.c | 1 + hw/ppc/pegasos2.c | 1 + hw/ppc/pnv_core.c | 2 + hw/ppc/ppc.c | 236 +++++++++++++++++++---------- hw/ppc/prep.c | 1 + hw/ppc/spapr.c | 32 +++- hw/ppc/spapr_cpu_core.c | 2 + hw/ppc/vof.c | 2 + include/hw/ppc/ppc.h | 3 +- include/hw/ppc/spapr.h | 2 + include/qemu/host-utils.h | 21 ++- target/ppc/compat.c | 19 +++ target/ppc/cpu.h | 3 + target/ppc/excp_helper.c | 3 + target/ppc/machine.c | 40 ++++- target/ppc/translate.c | 4 + tests/avocado/replay_kernel.py | 3 +- tests/avocado/reverse_debugging.py | 54 ++++++- 18 files changed, 330 insertions(+), 99 deletions(-)