Skip to contentSkip to conversation

exec/rom_reset: Free rom data during inmigrate skip

PATCH Superseded Dr. David Alan Gilbert (git) 2 participants 8 messages Mar 13, 2020

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 exec/rom_reset: Free rom data during inmigrate skip

From: "Dr. David Alan Gilbert" <[email protected]>

Commit 355477f8c73e9 skips rom reset when we're an incoming migration so as not to overwrite shared ram in the ignore-shared migration optimisation.
However, it's got an unexpected side effect that because it skips freeing the ROM data, when rom_reset gets called later on, after migration (e.g. during a reboot), the ROM does get reset to the original file contents. Because of seabios/x86's weird reboot process this confuses a reboot into hanging after a migration.

Fixes 355477f8c73e9 ("migration: do not rom_reset() during incoming migration")

https://bugzilla.redhat.com/show_bug.cgi?id=1809380

---
 hw/core/loader.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)
hw/core/loader.c +14 -9
index d1b78f60cd..4e583eb3bd 100644
@@ -1119,19 +1119,24 @@ static void rom_reset(void *unused)
 {
     Rom *rom;
 
-    /*
-     * We don't need to fill in the RAM with ROM data because we'll fill
-     * the data in during the next incoming migration in all cases.  Note
-     * that some of those RAMs can actually be modified by the guest on ARM
-     * so this is probably the only right thing to do here.
-     */
-    if (runstate_check(RUN_STATE_INMIGRATE))
-        return;
-
     QTAILQ_FOREACH(rom, &roms, next) {
         if (rom->fw_file) {
             continue;
         }
+        /*
+         * We don't need to fill in the RAM with ROM data because we'll fill
+         * the data in during the next incoming migration in all cases.  Note
+         * that some of those RAMs can actually be modified by the guest on ARM
+         * so this is probably the only right thing to do here.
+         */
+        if (runstate_check(RUN_STATE_INMIGRATE) && rom->data) {
+            /*
+             * Free it so that a rom_reset after migration doesn't overwrite a
+             * potentially modified 'rom'.
+             */
+            rom_free_data(rom);
+        }
+
         if (rom->data == NULL) {
             continue;
         }
Signature · 1 line
2.24.1
#1 lore raw
Quoted 27 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Commit 355477f8c73e9 skips rom reset when we're an incoming migration so as not to overwrite shared ram in the ignore-shared migration optimisation.
However, it's got an unexpected side effect that because it skips freeing the ROM data, when rom_reset gets called later on, after migration (e.g. during a reboot), the ROM does get reset to the original file contents. Because of seabios/x86's weird reboot process this confuses a reboot into hanging after a migration.

Fixes 355477f8c73e9 ("migration: do not rom_reset() during incoming migration")

https://bugzilla.redhat.com/show_bug.cgi?id=1809380

---
 hw/core/loader.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)
     QTAILQ_FOREACH(rom, &roms, next) {
         if (rom->fw_file) {
             continue;
         }
hw/core/loader.c +2 -0
+        /*
+         * We don't need to fill in the RAM with ROM data because we'll fill

hw/core/loader.c static void rom_reset(void *unused) · line 1137

Quoted 10 lines
hw/core/loader.c +10 -0
+         * the data in during the next incoming migration in all cases.  Note
+         * that some of those RAMs can actually be modified by the guest on ARM
+         * so this is probably the only right thing to do here.
+         */
+        if (runstate_check(RUN_STATE_INMIGRATE) && rom->data) {
+            /*
+             * Free it so that a rom_reset after migration doesn't overwrite a
+             * potentially modified 'rom'.
+             */
+            rom_free_data(rom);

Shouldn't this condition match the condition in rom_reset() for when we call rom_free_data()? You want the behaviour on a subsequent reset to match the behaviour you'd get if you did a reset on the source end without the migration.

hw/core/loader.c static void rom_reset(void *unused) · line 1142

Quoted 5 lines
hw/core/loader.c +2 -0
+        }
+
         if (rom->data == NULL) {
             continue;
         }

If you put this check above your new one you wouldn't need to check rom->data in it. Also it would make the loop structure better match rom_reset(), which checks rom->fw_file first, then rom->data, then has the condition for "do we need to call rom_free_data()".

thanks
-- PMM

#2 lore raw
Quoted 41 lines · Peter Maydell
Quoted 27 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Commit 355477f8c73e9 skips rom reset when we're an incoming migration so as not to overwrite shared ram in the ignore-shared migration optimisation.
However, it's got an unexpected side effect that because it skips freeing the ROM data, when rom_reset gets called later on, after migration (e.g. during a reboot), the ROM does get reset to the original file contents. Because of seabios/x86's weird reboot process this confuses a reboot into hanging after a migration.

Fixes 355477f8c73e9 ("migration: do not rom_reset() during incoming migration")

https://bugzilla.redhat.com/show_bug.cgi?id=1809380

---
 hw/core/loader.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)
     QTAILQ_FOREACH(rom, &roms, next) {
         if (rom->fw_file) {
             continue;
         }
hw/core/loader.c +2 -0
+        /*
+         * We don't need to fill in the RAM with ROM data because we'll fill

hw/core/loader.c static void rom_reset(void *unused) · line 1137

Quoted 10 lines
hw/core/loader.c +10 -0
+         * the data in during the next incoming migration in all cases.  Note
+         * that some of those RAMs can actually be modified by the guest on ARM
+         * so this is probably the only right thing to do here.
+         */
+        if (runstate_check(RUN_STATE_INMIGRATE) && rom->data) {
+            /*
+             * Free it so that a rom_reset after migration doesn't overwrite a
+             * potentially modified 'rom'.
+             */
+            rom_free_data(rom);
Quoted 4 lines

Shouldn't this condition match the condition in rom_reset() for when we call rom_free_data()? You want the behaviour on a subsequent reset to match the behaviour you'd get if you did a reset on the source end without the migration.

Wait, this *is* rom_reset(). Now I'm really confused.

thanks
-- PMM

#3 lore raw
Quoted 42 lines · * Peter Maydell ([email protected]) wrote:
Quoted 41 lines · Peter Maydell
Quoted 37 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Commit 355477f8c73e9 skips rom reset when we're an incoming migration so as not to overwrite shared ram in the ignore-shared migration optimisation.
However, it's got an unexpected side effect that because it skips freeing the ROM data, when rom_reset gets called later on, after migration (e.g. during a reboot), the ROM does get reset to the original file contents. Because of seabios/x86's weird reboot process this confuses a reboot into hanging after a migration.

Fixes 355477f8c73e9 ("migration: do not rom_reset() during incoming migration")

https://bugzilla.redhat.com/show_bug.cgi?id=1809380

---
 hw/core/loader.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)
     QTAILQ_FOREACH(rom, &roms, next) {
         if (rom->fw_file) {
             continue;
         }
hw/core/loader.c +12 -0
+        /*
+         * We don't need to fill in the RAM with ROM data because we'll fill
+         * the data in during the next incoming migration in all cases.  Note
+         * that some of those RAMs can actually be modified by the guest on ARM
+         * so this is probably the only right thing to do here.
+         */
+        if (runstate_check(RUN_STATE_INMIGRATE) && rom->data) {
+            /*
+             * Free it so that a rom_reset after migration doesn't overwrite a
+             * potentially modified 'rom'.
+             */
+            rom_free_data(rom);
Quoted 6 lines
Quoted 4 lines

Shouldn't this condition match the condition in rom_reset() for when we call rom_free_data()? You want the behaviour on a subsequent reset to match the behaviour you'd get if you did a reset on the source end without the migration.

Wait, this *is* rom_reset(). Now I'm really confused.

The exsiting rom_reset gets called multiple times:
a) During init
This actually copies the ROMs and then calls rom_free_data

b) During a subsequent reboot
This is mostly skipped because rom->data is now free because of the prior call to rom_free_data during (a)

During an inbound migrate, (a) happens before the migration, and
(b) happens during a reboot after the migration.

The problem is that 355477f8c73e9 caused (a) to be skipped then when (b) happens it actually overwrites the ROM because the rom_free_data had been skipped. What I'm doing here is doing the rom_free_data(..) which causes it to then skip this iteration during (a) AND causes it to skip it during (b).

Dave

Quoted 3 lines

thanks
-- PMM

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#4 lore raw
Quoted 59 lines · Dr. David Alan Gilbert
Quoted 42 lines · * Peter Maydell ([email protected]) wrote:
Quoted 41 lines · Peter Maydell
Quoted 37 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Commit 355477f8c73e9 skips rom reset when we're an incoming migration so as not to overwrite shared ram in the ignore-shared migration optimisation.
However, it's got an unexpected side effect that because it skips freeing the ROM data, when rom_reset gets called later on, after migration (e.g. during a reboot), the ROM does get reset to the original file contents. Because of seabios/x86's weird reboot process this confuses a reboot into hanging after a migration.

Fixes 355477f8c73e9 ("migration: do not rom_reset() during incoming migration")

https://bugzilla.redhat.com/show_bug.cgi?id=1809380

---
 hw/core/loader.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)
     QTAILQ_FOREACH(rom, &roms, next) {
         if (rom->fw_file) {
             continue;
         }
hw/core/loader.c +12 -0
+        /*
+         * We don't need to fill in the RAM with ROM data because we'll fill
+         * the data in during the next incoming migration in all cases.  Note
+         * that some of those RAMs can actually be modified by the guest on ARM
+         * so this is probably the only right thing to do here.
+         */
+        if (runstate_check(RUN_STATE_INMIGRATE) && rom->data) {
+            /*
+             * Free it so that a rom_reset after migration doesn't overwrite a
+             * potentially modified 'rom'.
+             */
+            rom_free_data(rom);
Quoted 6 lines
Quoted 4 lines

Shouldn't this condition match the condition in rom_reset() for when we call rom_free_data()? You want the behaviour on a subsequent reset to match the behaviour you'd get if you did a reset on the source end without the migration.

Wait, this *is* rom_reset(). Now I'm really confused.

The exsiting rom_reset gets called multiple times:
a) During init
This actually copies the ROMs and then calls rom_free_data

b) During a subsequent reboot
This is mostly skipped because rom->data is now free because of the prior call to rom_free_data during (a)

Quoted 8 lines

During an inbound migrate, (a) happens before the migration, and
(b) happens during a reboot after the migration.

The problem is that 355477f8c73e9 caused (a) to be skipped then when (b) happens it actually overwrites the ROM because the rom_free_data had been skipped. What I'm doing here is doing the rom_free_data(..) which causes it to then skip this iteration during (a) AND causes it to skip it during (b).

OK, but why is your condition for when to call rom_free_data() in this special case not the same as the condition that we use in the normal no-migration-involved case? I would expect those to match up.

thanks
-- PMM

#5 lore raw
Quoted 70 lines · * Peter Maydell ([email protected]) wrote:
Quoted 59 lines · Dr. David Alan Gilbert
Quoted 42 lines · * Peter Maydell ([email protected]) wrote:
Quoted 41 lines · Peter Maydell
Quoted 37 lines · Dr. David Alan Gilbert (git)

From: "Dr. David Alan Gilbert" <[email protected]>

Commit 355477f8c73e9 skips rom reset when we're an incoming migration so as not to overwrite shared ram in the ignore-shared migration optimisation.
However, it's got an unexpected side effect that because it skips freeing the ROM data, when rom_reset gets called later on, after migration (e.g. during a reboot), the ROM does get reset to the original file contents. Because of seabios/x86's weird reboot process this confuses a reboot into hanging after a migration.

Fixes 355477f8c73e9 ("migration: do not rom_reset() during incoming migration")

https://bugzilla.redhat.com/show_bug.cgi?id=1809380

---
 hw/core/loader.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)
     QTAILQ_FOREACH(rom, &roms, next) {
         if (rom->fw_file) {
             continue;
         }
hw/core/loader.c +12 -0
+        /*
+         * We don't need to fill in the RAM with ROM data because we'll fill
+         * the data in during the next incoming migration in all cases.  Note
+         * that some of those RAMs can actually be modified by the guest on ARM
+         * so this is probably the only right thing to do here.
+         */
+        if (runstate_check(RUN_STATE_INMIGRATE) && rom->data) {
+            /*
+             * Free it so that a rom_reset after migration doesn't overwrite a
+             * potentially modified 'rom'.
+             */
+            rom_free_data(rom);
Quoted 6 lines
Quoted 4 lines

Shouldn't this condition match the condition in rom_reset() for when we call rom_free_data()? You want the behaviour on a subsequent reset to match the behaviour you'd get if you did a reset on the source end without the migration.

Wait, this *is* rom_reset(). Now I'm really confused.

The exsiting rom_reset gets called multiple times:
a) During init
This actually copies the ROMs and then calls rom_free_data

b) During a subsequent reboot
This is mostly skipped because rom->data is now free because of the prior call to rom_free_data during (a)

Quoted 8 lines

During an inbound migrate, (a) happens before the migration, and
(b) happens during a reboot after the migration.

The problem is that 355477f8c73e9 caused (a) to be skipped then when (b) happens it actually overwrites the ROM because the rom_free_data had been skipped. What I'm doing here is doing the rom_free_data(..) which causes it to then skip this iteration during (a) AND causes it to skip it during (b).

Quoted 4 lines

OK, but why is your condition for when to call rom_free_data() in this special case not the same as the condition that we use in the normal no-migration-involved case? I would expect those to match up.

Ah yes, I think you're right, so something like:

  if (runstate_check(RUN_STATE_INMIGRATE) && rom->data && rom->isrom) {

I'll try that after lunch.

Dave

Quoted 3 lines

thanks
-- PMM

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#6 lore raw
Quoted 3 lines · Dr. David Alan Gilbert

Ah yes, I think you're right, so something like:

  if (runstate_check(RUN_STATE_INMIGRATE) && rom->data && rom->isrom) {

I think you would see the difference here for images loaded into RAM, rather than ROM -- they need to be reinstated on reset, because the guest can scribble on them. So we retain the data and don't free it.

thanks
-- PMM

#7 lore raw
Quoted 10 lines · * Peter Maydell ([email protected]) wrote:
Quoted 3 lines · Dr. David Alan Gilbert

Ah yes, I think you're right, so something like:

  if (runstate_check(RUN_STATE_INMIGRATE) && rom->data && rom->isrom) {

I think you would see the difference here for images loaded into RAM, rather than ROM -- they need to be reinstated on reset, because the guest can scribble on them. So we retain the data and don't free it.

Hmm, that's true; so I'm failing to skip a copy in the !isrom case, whch the original patch needed.

So what I think we'll need is:

   if (runstate_check(RUN_STATE_INMIGRATE))
   {
       if (rom->data && rom->isrom) {
           rom_free_data(rom);
       }
       continue;
   }
Quoted 3 lines

thanks
-- PMM

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK