| Message ID | 20230620175712.1331625-1-peter.maydell@linaro.org |
|---|---|
| State | Accepted |
| Commit | c5ffd16ba4c8fd3601742cc9d2b3cff03995dd5d |
| Headers | show |
| Series | Revert "cputlb: Restrict SavedIOTLB to system emulation" | expand |
On 6/20/23 19:57, Peter Maydell wrote: > This reverts commit d7ee93e24359703debf4137f4cc632563aa4e8d1. > > That commit tries to make a field in the CPUState struct not be > present when CONFIG_USER_ONLY is set. Unfortunately, you can't > conditionally omit fields in structs like this based on ifdefs that > are set per-target. If you try it, then code in files compiled > per-target (where CONFIG_USER_ONLY is or can be set) will disagree > about the struct layout with files that are compiled once-only (where > this kind of ifdef is never set). > > This manifests specifically in 'make check-tcg' failing, because code > in cpus-common.c that sets up the CPUState::cpu_index field puts it > at a different offset from the code in plugins/core.c in > qemu_plugin_vcpu_init_hook() which reads the cpu_index field. The > latter then hits an assert because from its point of view every > thread has a 0 cpu_index. There might be other weird behaviour too. > > Mostly we catch this kind of bug because the CONFIG_whatever is > listed in include/exec/poison.h and so the reference to it in > build-once source files will then cause a compiler error. > Unfortunately CONFIG_USER_ONLY is an exception to that: we have some > places where we use it in "safe" ways in headers that will be seen by > once-only source files (e.g. ifdeffing out function prototypes) and > it would be a lot of refactoring to be able to get to a position > where we could poison it. This leaves us in a "you have to be > careful to walk around the bear trap" situation... > > Fixes: d7ee93e243597 ("cputlb: Restrict SavedIOTLB to system emulation") > Signed-off-by: Peter Maydell <peter.maydell@linaro.org> > --- > include/hw/core/cpu.h | 6 ++---- > 1 file changed, 2 insertions(+), 4 deletions(-) Ho hum, thanks. I'll apply this directly. r~
On 21/6/23 07:19, Richard Henderson wrote: > On 6/20/23 19:57, Peter Maydell wrote: >> This reverts commit d7ee93e24359703debf4137f4cc632563aa4e8d1. >> >> That commit tries to make a field in the CPUState struct not be >> present when CONFIG_USER_ONLY is set. Unfortunately, you can't >> conditionally omit fields in structs like this based on ifdefs that >> are set per-target. If you try it, then code in files compiled >> per-target (where CONFIG_USER_ONLY is or can be set) will disagree >> about the struct layout with files that are compiled once-only (where >> this kind of ifdef is never set). Oops, sorry. >> This manifests specifically in 'make check-tcg' failing, because code >> in cpus-common.c that sets up the CPUState::cpu_index field puts it >> at a different offset from the code in plugins/core.c in >> qemu_plugin_vcpu_init_hook() which reads the cpu_index field. The >> latter then hits an assert because from its point of view every >> thread has a 0 cpu_index. There might be other weird behaviour too. Why isn't this covered by CI, and where could we add a such check? >> Mostly we catch this kind of bug because the CONFIG_whatever is >> listed in include/exec/poison.h and so the reference to it in >> build-once source files will then cause a compiler error. >> Unfortunately CONFIG_USER_ONLY is an exception to that: we have some >> places where we use it in "safe" ways in headers that will be seen by >> once-only source files (e.g. ifdeffing out function prototypes) and >> it would be a lot of refactoring to be able to get to a position >> where we could poison it. This leaves us in a "you have to be >> careful to walk around the bear trap" situation... >> >> Fixes: d7ee93e243597 ("cputlb: Restrict SavedIOTLB to system emulation") >> Signed-off-by: Peter Maydell <peter.maydell@linaro.org> >> --- >> include/hw/core/cpu.h | 6 ++---- >> 1 file changed, 2 insertions(+), 4 deletions(-) > > Ho hum, thanks. I'll apply this directly. Thanks both.
On 21/6/23 11:39, Philippe Mathieu-Daudé wrote: > On 21/6/23 07:19, Richard Henderson wrote: >> On 6/20/23 19:57, Peter Maydell wrote: >>> This reverts commit d7ee93e24359703debf4137f4cc632563aa4e8d1. >>> >>> That commit tries to make a field in the CPUState struct not be >>> present when CONFIG_USER_ONLY is set. Unfortunately, you can't >>> conditionally omit fields in structs like this based on ifdefs that >>> are set per-target. If you try it, then code in files compiled >>> per-target (where CONFIG_USER_ONLY is or can be set) will disagree >>> about the struct layout with files that are compiled once-only (where >>> this kind of ifdef is never set). > > Oops, sorry. > >>> This manifests specifically in 'make check-tcg' failing, because code >>> in cpus-common.c that sets up the CPUState::cpu_index field puts it >>> at a different offset from the code in plugins/core.c in >>> qemu_plugin_vcpu_init_hook() which reads the cpu_index field. The >>> latter then hits an assert because from its point of view every >>> thread has a 0 cpu_index. There might be other weird behaviour too. > > Why isn't this covered by CI, and where could we add a such check? Actually it is covered and failed on staging: https://gitlab.com/qemu-project/qemu/-/jobs/4503766933 Anyhow, sorry for the mess. >>> Mostly we catch this kind of bug because the CONFIG_whatever is >>> listed in include/exec/poison.h and so the reference to it in >>> build-once source files will then cause a compiler error. >>> Unfortunately CONFIG_USER_ONLY is an exception to that: we have some >>> places where we use it in "safe" ways in headers that will be seen by >>> once-only source files (e.g. ifdeffing out function prototypes) and >>> it would be a lot of refactoring to be able to get to a position >>> where we could poison it. This leaves us in a "you have to be >>> careful to walk around the bear trap" situation... >>> >>> Fixes: d7ee93e243597 ("cputlb: Restrict SavedIOTLB to system emulation") >>> Signed-off-by: Peter Maydell <peter.maydell@linaro.org> >>> --- >>> include/hw/core/cpu.h | 6 ++---- >>> 1 file changed, 2 insertions(+), 4 deletions(-) >> >> Ho hum, thanks. I'll apply this directly. > > Thanks both. >
Hello, Philippe Mathieu-Daudé <philmd@linaro.org> writes: > On 21/6/23 11:39, Philippe Mathieu-Daudé wrote: >> On 21/6/23 07:19, Richard Henderson wrote: >>> On 6/20/23 19:57, Peter Maydell wrote: >>>> This manifests specifically in 'make check-tcg' failing, because code >>>> in cpus-common.c that sets up the CPUState::cpu_index field puts it >>>> at a different offset from the code in plugins/core.c in >>>> qemu_plugin_vcpu_init_hook() which reads the cpu_index field. The >>>> latter then hits an assert because from its point of view every >>>> thread has a 0 cpu_index. There might be other weird behaviour too. >> Why isn't this covered by CI, and where could we add a such check? > > Actually it is covered and failed on staging: > https://gitlab.com/qemu-project/qemu/-/jobs/4503766933 Just for the record, it was also caught yesterday by the TCWG CI: https://ci.linaro.org/job/tcwg_gnu_cross_check_gcc--master-arm-bisect/87/
diff --git a/include/hw/core/cpu.h b/include/hw/core/cpu.h index ee8d6b40b3b..4871ad85f07 100644 --- a/include/hw/core/cpu.h +++ b/include/hw/core/cpu.h @@ -226,7 +226,7 @@ struct CPUWatchpoint { QTAILQ_ENTRY(CPUWatchpoint) entry; }; -#if defined(CONFIG_PLUGIN) && !defined(CONFIG_USER_ONLY) +#ifdef CONFIG_PLUGIN /* * For plugins we sometime need to save the resolved iotlb data before * the memory regions get moved around by io_writex. @@ -410,11 +410,9 @@ struct CPUState { #ifdef CONFIG_PLUGIN GArray *plugin_mem_cbs; -#if !defined(CONFIG_USER_ONLY) /* saved iotlb data from io_writex */ SavedIOTLB saved_iotlb; -#endif /* !CONFIG_USER_ONLY */ -#endif /* CONFIG_PLUGIN */ +#endif /* TODO Move common fields from CPUArchState here. */ int cpu_index;
This reverts commit d7ee93e24359703debf4137f4cc632563aa4e8d1. That commit tries to make a field in the CPUState struct not be present when CONFIG_USER_ONLY is set. Unfortunately, you can't conditionally omit fields in structs like this based on ifdefs that are set per-target. If you try it, then code in files compiled per-target (where CONFIG_USER_ONLY is or can be set) will disagree about the struct layout with files that are compiled once-only (where this kind of ifdef is never set). This manifests specifically in 'make check-tcg' failing, because code in cpus-common.c that sets up the CPUState::cpu_index field puts it at a different offset from the code in plugins/core.c in qemu_plugin_vcpu_init_hook() which reads the cpu_index field. The latter then hits an assert because from its point of view every thread has a 0 cpu_index. There might be other weird behaviour too. Mostly we catch this kind of bug because the CONFIG_whatever is listed in include/exec/poison.h and so the reference to it in build-once source files will then cause a compiler error. Unfortunately CONFIG_USER_ONLY is an exception to that: we have some places where we use it in "safe" ways in headers that will be seen by once-only source files (e.g. ifdeffing out function prototypes) and it would be a lot of refactoring to be able to get to a position where we could poison it. This leaves us in a "you have to be careful to walk around the bear trap" situation... Fixes: d7ee93e243597 ("cputlb: Restrict SavedIOTLB to system emulation") Signed-off-by: Peter Maydell <peter.maydell@linaro.org> --- include/hw/core/cpu.h | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-)