Patchwork memory: add -dont-dump-guest option to reduce core dump size

login
register
mail settings
Submitter Jason Baron
Date May 4, 2012, 9:23 p.m.
Message ID <201205042123.q44LNp9x029061@int-mx01.intmail.prod.int.phx2.redhat.com>
Download mbox | patch
Permalink /patch/156995/
State New
Headers show

Comments

Jason Baron - May 4, 2012, 9:23 p.m.
Add a command line parameter to not dump guest memory in the core dump, the
command line is: -dont-dump-guest. This brought the core dump down from
383MB to 13 MB on a 1GB guest.

Signed-off-by: Jason Baron <jbaron@redhat.com>
---
 exec.c          |   13 +++++++++++++
 osdep.h         |    7 +++++++
 qemu-options.hx |    5 +++++
 sysemu.h        |    1 +
 vl.c            |    4 ++++
 5 files changed, 30 insertions(+), 0 deletions(-)
Jason Baron - May 21, 2012, 5:53 p.m.
On Fri, May 04, 2012 at 05:23:51PM -0400, Jason Baron wrote:
> Add a command line parameter to not dump guest memory in the core dump, the
> command line is: -dont-dump-guest. This brought the core dump down from
> 383MB to 13 MB on a 1GB guest.
> 
> Signed-off-by: Jason Baron <jbaron@redhat.com>
> ---
>  exec.c          |   13 +++++++++++++
>  osdep.h         |    7 +++++++
>  qemu-options.hx |    5 +++++
>  sysemu.h        |    1 +
>  vl.c            |    4 ++++
>  5 files changed, 30 insertions(+), 0 deletions(-)
> 

Any thoughts on this? Is the guest memory often helpful in debugging
when the qemu process segfaults?

This feature also seems useful if somebody is running a sensitive
workload, such that the non-sensitive qemu state can be dumped
independently of the guest state. In that case, perhaps there should
also be an option that allows the guest to be dumped on a segfault, now
separately?

Thanks,

-Jason
Marcelo Tosatti - July 6, 2012, 12:04 a.m.
On Mon, May 21, 2012 at 01:53:36PM -0400, Jason Baron wrote:
> On Fri, May 04, 2012 at 05:23:51PM -0400, Jason Baron wrote:
> > Add a command line parameter to not dump guest memory in the core dump, the
> > command line is: -dont-dump-guest. This brought the core dump down from
> > 383MB to 13 MB on a 1GB guest.
> > 
> > Signed-off-by: Jason Baron <jbaron@redhat.com>
> > ---
> >  exec.c          |   13 +++++++++++++
> >  osdep.h         |    7 +++++++
> >  qemu-options.hx |    5 +++++
> >  sysemu.h        |    1 +
> >  vl.c            |    4 ++++
> >  5 files changed, 30 insertions(+), 0 deletions(-)
> > 
> 
> Any thoughts on this? Is the guest memory often helpful in debugging
> when the qemu process segfaults?

For most development related usage, no (with large guests, its 
troublesome).

Please add documentation for the command (patch looks fine otherwise).

> This feature also seems useful if somebody is running a sensitive
> workload, such that the non-sensitive qemu state can be dumped
> independently of the guest state. In that case, perhaps there should
> also be an option that allows the guest to be dumped on a segfault, now
> separately?

Isnt what this option does? (well, disabling it).
Markus Armbruster - July 6, 2012, 5:11 a.m.
Marcelo Tosatti <mtosatti@redhat.com> writes:

> On Mon, May 21, 2012 at 01:53:36PM -0400, Jason Baron wrote:
>> On Fri, May 04, 2012 at 05:23:51PM -0400, Jason Baron wrote:
>> > Add a command line parameter to not dump guest memory in the core dump, the
>> > command line is: -dont-dump-guest. This brought the core dump down from
>> > 383MB to 13 MB on a 1GB guest.
>> > 
>> > Signed-off-by: Jason Baron <jbaron@redhat.com>
>> > ---
>> >  exec.c          |   13 +++++++++++++
>> >  osdep.h         |    7 +++++++
>> >  qemu-options.hx |    5 +++++
>> >  sysemu.h        |    1 +
>> >  vl.c            |    4 ++++
>> >  5 files changed, 30 insertions(+), 0 deletions(-)
>> > 
>> 
>> Any thoughts on this? Is the guest memory often helpful in debugging
>> when the qemu process segfaults?
>
> For most development related usage, no (with large guests, its 
> troublesome).
>
> Please add documentation for the command (patch looks fine otherwise).
>
>> This feature also seems useful if somebody is running a sensitive
>> workload, such that the non-sensitive qemu state can be dumped
>> independently of the guest state. In that case, perhaps there should

I wouldn't make claims related to guest data confidentiality, because
sensitive stuff could conceivably leak from guest RAM (not dumped) into
device model state (dumped).

>> also be an option that allows the guest to be dumped on a segfault, now
>> separately?
>
> Isnt what this option does? (well, disabling it).

It seems there could be more knobs to control than just "dump guest
state yes/no".  Therefore, extensible command line syntax like
"--core-dump guest-ram=off" seems to be advisable.  We have too many
-dont-do-FOO options already.
Jason Baron - July 6, 2012, 8:26 p.m.
On Thu, Jul 05, 2012 at 09:04:18PM -0300, Marcelo Tosatti wrote:
> On Mon, May 21, 2012 at 01:53:36PM -0400, Jason Baron wrote:
> > On Fri, May 04, 2012 at 05:23:51PM -0400, Jason Baron wrote:
> > > Add a command line parameter to not dump guest memory in the core dump, the
> > > command line is: -dont-dump-guest. This brought the core dump down from
> > > 383MB to 13 MB on a 1GB guest.
> > > 
> > > Signed-off-by: Jason Baron <jbaron@redhat.com>
> > > ---
> > >  exec.c          |   13 +++++++++++++
> > >  osdep.h         |    7 +++++++
> > >  qemu-options.hx |    5 +++++
> > >  sysemu.h        |    1 +
> > >  vl.c            |    4 ++++
> > >  5 files changed, 30 insertions(+), 0 deletions(-)
> > > 
> > 
> > Any thoughts on this? Is the guest memory often helpful in debugging
> > when the qemu process segfaults?
> 
> For most development related usage, no (with large guests, its 
> troublesome).
> 
> Please add documentation for the command (patch looks fine otherwise).
> 
> > This feature also seems useful if somebody is running a sensitive
> > workload, such that the non-sensitive qemu state can be dumped
> > independently of the guest state. In that case, perhaps there should
> > also be an option that allows the guest to be dumped on a segfault, now
> > separately?
> 
> Isnt what this option does? (well, disabling it).
> 

Right, disabling it gives us both qemu and guest bits. So I was just
suggesting a mode that dumps guest bits but not qemu bits. (Probably
that is setup in guest, though). So I'm not sure if that makes sense.
But as Markus suggested, if we make the interface more general, we are
more open to something like that later (at least in terms of the
interface).

Thanks,

-Jason
Avi Kivity - July 11, 2012, 2:14 p.m.
On 07/06/2012 08:11 AM, Markus Armbruster wrote:
> 
> It seems there could be more knobs to control than just "dump guest
> state yes/no".  Therefore, extensible command line syntax like
> "--core-dump guest-ram=off" seems to be advisable.  We have too many
> -dont-do-FOO options already.

-m 2T,dump=no

?

Patch

diff --git a/exec.c b/exec.c
index 0607c9b..f9d2b93 100644
--- a/exec.c
+++ b/exec.c
@@ -35,6 +35,7 @@ 
 #include "qemu-timer.h"
 #include "memory.h"
 #include "exec-memory.h"
+#include "sysemu.h"
 #if defined(CONFIG_USER_ONLY)
 #include <qemu.h>
 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
@@ -2605,6 +2606,7 @@  ram_addr_t qemu_ram_alloc_from_ptr(ram_addr_t size, void *host,
                                    MemoryRegion *mr)
 {
     RAMBlock *new_block;
+    int ret;
 
     size = TARGET_PAGE_ALIGN(size);
     new_block = g_malloc0(sizeof(*new_block));
@@ -2659,6 +2661,17 @@  ram_addr_t qemu_ram_alloc_from_ptr(ram_addr_t size, void *host,
     memset(ram_list.phys_dirty + (new_block->offset >> TARGET_PAGE_BITS),
            0xff, size >> TARGET_PAGE_BITS);
 
+
+    /* Use MADV_DONTDUMP, if user doesn't want the guest memory in the core */
+    if (dont_dump_guest) {
+        ret = qemu_madvise(new_block->host, size, QEMU_MADV_DONTDUMP);
+        if (ret) {
+            perror("qemu_madvise");
+            fprintf(stderr, "madvise doesn't support MADV_DONTDUMP, "
+                            "but -dont-dump-guest-core specified\n");
+        }
+    }
+
     if (kvm_enabled())
         kvm_setup_guest_memory(new_block->host, size);
 
diff --git a/osdep.h b/osdep.h
index 9db8766..db275f2 100644
--- a/osdep.h
+++ b/osdep.h
@@ -100,6 +100,11 @@  void qemu_vfree(void *ptr);
 #else
 #define QEMU_MADV_MERGEABLE QEMU_MADV_INVALID
 #endif
+#ifdef MADV_DONTDUMP
+#define QEMU_MADV_DONTDUMP MADV_DONTDUMP
+#else
+#define QEMU_MADV_DONTDUMP QEMU_MADV_INVALID
+#endif
 
 #elif defined(CONFIG_POSIX_MADVISE)
 
@@ -107,6 +112,7 @@  void qemu_vfree(void *ptr);
 #define QEMU_MADV_DONTNEED  POSIX_MADV_DONTNEED
 #define QEMU_MADV_DONTFORK  QEMU_MADV_INVALID
 #define QEMU_MADV_MERGEABLE QEMU_MADV_INVALID
+#define QEMU_MADV_DONTDUMP QEMU_MADV_INVALID
 
 #else /* no-op */
 
@@ -114,6 +120,7 @@  void qemu_vfree(void *ptr);
 #define QEMU_MADV_DONTNEED  QEMU_MADV_INVALID
 #define QEMU_MADV_DONTFORK  QEMU_MADV_INVALID
 #define QEMU_MADV_MERGEABLE QEMU_MADV_INVALID
+#define QEMU_MADV_DONTDUMP QEMU_MADV_INVALID
 
 #endif
 
diff --git a/qemu-options.hx b/qemu-options.hx
index a169792..ae12cc6 100644
--- a/qemu-options.hx
+++ b/qemu-options.hx
@@ -2724,6 +2724,11 @@  DEF("qtest-log", HAS_ARG, QEMU_OPTION_qtest_log,
     "-qtest-log LOG  specify tracing options\n",
     QEMU_ARCH_ALL)
 
+DEF("dont-dump-guest", 0, QEMU_OPTION_dont_dump_guest,
+    "-dont-dump-guest\n"
+    "                do not dump guest core\n",
+    QEMU_ARCH_ALL)
+
 HXCOMM This is the last statement. Insert new options before this line!
 STEXI
 @end table
diff --git a/sysemu.h b/sysemu.h
index bc2c788..f738f34 100644
--- a/sysemu.h
+++ b/sysemu.h
@@ -131,6 +131,7 @@  extern uint8_t *boot_splash_filedata;
 extern int boot_splash_filedata_size;
 extern uint8_t qemu_extra_params_fw[2];
 extern QEMUClock *rtc_clock;
+extern int dont_dump_guest;
 
 #define MAX_NODES 64
 extern int nb_numa_nodes;
diff --git a/vl.c b/vl.c
index ae91a8a..6259c21 100644
--- a/vl.c
+++ b/vl.c
@@ -232,6 +232,7 @@  int boot_menu;
 uint8_t *boot_splash_filedata;
 int boot_splash_filedata_size;
 uint8_t qemu_extra_params_fw[2];
+int dont_dump_guest;
 
 typedef struct FWBootEntry FWBootEntry;
 
@@ -3191,6 +3192,9 @@  int main(int argc, char **argv, char **envp)
             case QEMU_OPTION_qtest_log:
                 qtest_log = optarg;
                 break;
+            case QEMU_OPTION_dont_dump_guest:
+                dont_dump_guest = 1;
+                break;
             default:
                 os_parse_cmd_args(popt->index, optarg);
             }