Re: [Qemu-devel] [PATCH 4/4] target-arm: add minimal dump-guest-memory support

[Rabin Vincent <rabin@rab.in>]

On Thu, Jun 28, 2012 at 05:46:02PM +0100, Peter Maydell wrote:
> On 20 June 2012 18:28, Rabin Vincent <rabin@rab.in> wrote:
> > Add a minimal dump-guest-memory support for ARM.  The -p option is not
> > supported and we don't add any QEMU-specific notes.
> 
> So what does this patch give us? This commit message is pretty
> short and I couldn't find a cover message for the patchset...

It makes the dump-guest-memory command work for arm-softmmu.  The
resulting core dump can be analysed with a tool such as the crash
utility.

> 
> > Signed-off-by: Rabin Vincent <rabin@rab.in>
> > ---
> >  configure                        |    4 +--
> >  target-arm/Makefile.objs         |    2 +-
> >  target-arm/arch_dump.c           |   59 ++++++++++++++++++++++++++++++++++++++
> >  target-arm/arch_memory_mapping.c |   13 +++++++++
> >  4 files changed, 75 insertions(+), 3 deletions(-)
> >  create mode 100644 target-arm/arch_dump.c
> >  create mode 100644 target-arm/arch_memory_mapping.c
> >
> > diff --git a/configure b/configure
> > index b68c0ca..a20ad19 100755
> > --- a/configure
> > +++ b/configure
> > @@ -3727,7 +3727,7 @@ case "$target_arch2" in
> >     fi
> >  esac
> >  case "$target_arch2" in
> > -  i386|x86_64)
> > +  arm|i386|x86_64)
> >     echo "CONFIG_HAVE_GET_MEMORY_MAPPING=y" >> $config_target_mak
> >  esac
> >  if test "$target_arch2" = "ppc64" -a "$fdt" = "yes"; then
> > @@ -3746,7 +3746,7 @@ if test "$target_softmmu" = "yes" ; then
> >     echo "subdir-$target: subdir-libcacard" >> $config_host_mak
> >   fi
> >   case "$target_arch2" in
> > -    i386|x86_64)
> > +    arm|i386|x86_64)
> >       echo "CONFIG_HAVE_CORE_DUMP=y" >> $config_target_mak
> >   esac
> >  fi
> > diff --git a/target-arm/Makefile.objs b/target-arm/Makefile.objs
> > index f447c4f..837b374 100644
> > --- a/target-arm/Makefile.objs
> > +++ b/target-arm/Makefile.objs
> > @@ -1,5 +1,5 @@
> >  obj-y += arm-semi.o
> > -obj-$(CONFIG_SOFTMMU) += machine.o
> > +obj-$(CONFIG_SOFTMMU) += machine.o arch_memory_mapping.o arch_dump.o
> >  obj-y += translate.o op_helper.o helper.o cpu.o
> >  obj-y += neon_helper.o iwmmxt_helper.o
> >
> > diff --git a/target-arm/arch_dump.c b/target-arm/arch_dump.c
> > new file mode 100644
> > index 0000000..47a7e40
> > --- /dev/null
> > +++ b/target-arm/arch_dump.c
> > @@ -0,0 +1,59 @@
> > +#include "cpu.h"
> > +#include "cpu-all.h"
> > +#include "dump.h"
> > +#include "elf.h"
> > +
> > +typedef struct {
> > +    char pad1[24];
> > +    uint32_t pid;
> > +    char pad2[44];
> > +    uint32_t regs[18];
> > +    char pad3[4];
> > +} arm_elf_prstatus;
> 
> I'm guessing this is following some specification's structure layout;
> what specification?

struct elf_prstatus from the Linux kernel's include/linux/elfcore.h.

> 
> > +
> > +int cpu_write_elf64_note(write_core_dump_function f, CPUArchState *env,
> > +                         int cpuid, void *opaque)
> 
> Should these APIs really be taking a CPUArchState* rather rather than
> an ARMCPU* ? (Andreas?)

No idea.  Cc'ing Wen, who added the APIs.

> 
> > +{
> > +    return -1;
> > +}
> > +
> > +int cpu_write_elf32_note(write_core_dump_function f, CPUArchState *env,
> > +                         int cpuid, void *opaque)
> > +{
> > +    arm_elf_prstatus prstatus;
> > +
> > +    memset(&prstatus, 0, sizeof(prstatus));
> > +    memcpy(&(prstatus.regs), env->regs, sizeof(env->regs));
> 
> This looks a bit odd -- env->regs[] is a 16 word array but
> prstatus.regs is 18 words. What are the last two words for?

CPSR and orig_r0.  orig_r0 is not useful, but I think we can save the
CPSR in there.

> 
> > +    prstatus.pid = cpuid;
> > +
> > +    return dump_write_elf_note(ELFCLASS32, "CORE", NT_PRSTATUS,
> > +                               &prstatus, sizeof(prstatus),
> > +                               f, opaque);
> > +}
> > +
> > +int cpu_write_elf64_qemunote(write_core_dump_function f, CPUArchState *env,
> > +                             void *opaque)
> > +{
> > +    return -1;
> > +}
> > +
> > +int cpu_write_elf32_qemunote(write_core_dump_function f, CPUArchState *env,
> > +                             void *opaque)
> > +{
> > +    return 0;
> > +}
> > +
> > +int cpu_get_dump_info(ArchDumpInfo *info)
> > +{
> > +    info->d_machine = EM_ARM;
> > +    info->d_endian = ELFDATA2LSB;
> 
> ...even for big endian ARM?

I'll use TARGET_WORDS_BIGENDIAN to check.

Though it appears we don't have a armbe-softmmu?

> 
> > +    info->d_class = ELFCLASS32;
> > +
> > +    return 0;
> > +}
> > +
> > +ssize_t cpu_get_note_size(int class, int machine, int nr_cpus)
> > +{
> > +    return nr_cpus * dump_get_note_size(ELFCLASS32, "CORE",
> > +                                        sizeof(arm_elf_prstatus));
> > +}
> > diff --git a/target-arm/arch_memory_mapping.c b/target-arm/arch_memory_mapping.c
> > new file mode 100644
> > index 0000000..eeaaf09
> > --- /dev/null
> > +++ b/target-arm/arch_memory_mapping.c
> > @@ -0,0 +1,13 @@
> > +#include "cpu.h"
> > +#include "cpu-all.h"
> > +#include "memory_mapping.h"
> > +
> > +bool cpu_paging_enabled(CPUArchState *env)
> > +{
> > +    return 0;
> > +}
> > +
> > +int cpu_get_memory_mapping(MemoryMappingList *list, CPUArchState *env)
> > +{
> > +    return -1;
> > +}
> 
> Why do we need these null implementations and why do they
> work better than the default ones in memory_mapping-stub.c ?

The implementations are to make the dump-guest-memory command build.  A
full implementation would add support for the "-p" option which afaics
is supposed to walk the page tables and dump only the pages which are
mapped instead of the complete RAM.  I personally have no need for this
option, so they are only null implementations which result in an error
if this option is used.

The current config code keeps memory-mapping.c and memory-mapping-stub.c
exclusive.  I think we should be able to make some changes there to
allow us to use memory-mapping-stub.c instead of this
arch_memory_mapping.c.