Skip to contentSkip to conversation

dump: Set correct vaddr for ELF dump

PATCH Superseded Jon Doron 3 participants 5 messages Dec 25, 2018 – Jan 8, 2019

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Jon Doron

author
#0 lore raw

PATCH 1/1 dump: Set correct vaddr for ELF dump

vaddr needs to be equal to the paddr since the dump file represents the physical memory image.

Without setting vaddr correctly, GDB would load all the different memory regions on top of each other to vaddr 0, thus making GDB showing the wrong memory data for a given address.

---
 dump.c                       | 4 ++--
 scripts/dump-guest-memory.py | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)
dump.c +2 -2
index 4ec94c5e25..bf77a119ea 100644
@@ -192,7 +192,7 @@ static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump64(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump64(s, filesz);
     phdr.p_memsz = cpu_to_dump64(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump64(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;
 
     assert(memory_mapping->length >= filesz);
 
@@ -216,7 +216,7 @@ static void write_elf32_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump32(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump32(s, filesz);
     phdr.p_memsz = cpu_to_dump32(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump32(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;
 
     assert(memory_mapping->length >= filesz);
 
scripts/dump-guest-memory.py +1 -0
index 198cd0fe40..2c587cbefc 100644
@@ -163,6 +163,7 @@ class ELF(object):
         phdr = get_arch_phdr(self.endianness, self.elfclass)
         phdr.p_type = p_type
         phdr.p_paddr = p_paddr
+        phdr.p_vaddr = p_paddr
         phdr.p_filesz = p_size
         phdr.p_memsz = p_size
         self.segments.append(phdr)
Signature · 1 line
2.19.2
#1 lore raw

Hi

Quoted 9 lines · Jon Doron

vaddr needs to be equal to the paddr since the dump file represents the physical memory image.

Without setting vaddr correctly, GDB would load all the different memory regions on top of each other to vaddr 0, thus making GDB showing the wrong memory data for a given address.

This is a non-trivial patch! (qemu-trivial, please ignore).

Quoted 10 lines
---
 dump.c                       | 4 ++--
 scripts/dump-guest-memory.py | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)
dump.c +0 -0
index 4ec94c5e25..bf77a119ea 100644

dump.c static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping, · line 195

Quoted 5 lines
dump.c +1 -1
@@ -192,7 +192,7 @@ static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump64(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump64(s, filesz);
     phdr.p_memsz = cpu_to_dump64(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump64(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;

This is likely breaking paging=true somehow, which sets memory_mapping->virt_addr to non-0.

According to doc "If you want to use gdb to process the core, please set @paging to true."

Although I am not able to (gdb) x/10bx 0xa0000 for example on a core produced with paging. Not sure why, anybody could help?

Quoted 16 lines
dump.c +1 -1
     assert(memory_mapping->length >= filesz);
 
@@ -216,7 +216,7 @@ static void write_elf32_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump32(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump32(s, filesz);
     phdr.p_memsz = cpu_to_dump32(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump32(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;
 
     assert(memory_mapping->length >= filesz);
 
scripts/dump-guest-memory.py +0 -0
index 198cd0fe40..2c587cbefc 100644

scripts/dump-guest-memory.py class ELF(object): · line 166

Quoted 4 lines
scripts/dump-guest-memory.py +1 -0
@@ -163,6 +163,7 @@ class ELF(object):
         phdr = get_arch_phdr(self.endianness, self.elfclass)
         phdr.p_type = p_type
         phdr.p_paddr = p_paddr
+        phdr.p_vaddr = p_paddr

With your proposed change though, I can dump memory with gdb...

Quoted 7 lines
scripts/dump-guest-memory.py +0 -1
         phdr.p_filesz = p_size
         phdr.p_memsz = p_size
         self.segments.append(phdr)
--

2.19.2

Signature · 1 line
Marc-André Lureau
#2 lore raw
Quoted 51 lines · Marc-André Lureau

Hi

Quoted 9 lines · Jon Doron

vaddr needs to be equal to the paddr since the dump file represents the physical memory image.

Without setting vaddr correctly, GDB would load all the different memory regions on top of each other to vaddr 0, thus making GDB showing the wrong memory data for a given address.

This is a non-trivial patch! (qemu-trivial, please ignore).

Quoted 10 lines
---
 dump.c                       | 4 ++--
 scripts/dump-guest-memory.py | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)
dump.c +0 -0
index 4ec94c5e25..bf77a119ea 100644

dump.c static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping, · line 195

Quoted 5 lines
dump.c +1 -1
@@ -192,7 +192,7 @@ static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump64(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump64(s, filesz);
     phdr.p_memsz = cpu_to_dump64(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump64(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;

This is likely breaking paging=true somehow, which sets memory_mapping->virt_addr to non-0.

According to doc "If you want to use gdb to process the core, please set @paging to true."

Although I am not able to (gdb) x/10bx 0xa0000 for example on a core produced with paging. Not sure why, anybody could help?

Quoted 11 lines
dump.c +1 -1
     assert(memory_mapping->length >= filesz);
 
@@ -216,7 +216,7 @@ static void write_elf32_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump32(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump32(s, filesz);
     phdr.p_memsz = cpu_to_dump32(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump32(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;
 
     assert(memory_mapping->length >= filesz);
 
Quoted 21 lines

scripts/dump-guest-memory.py class ELF(object): · line 166

Quoted 9 lines
scripts/dump-guest-memory.py +1 -0
index 198cd0fe40..2c587cbefc 100644
@@ -163,6 +163,7 @@ class ELF(object):
         phdr = get_arch_phdr(self.endianness, self.elfclass)
         phdr.p_type = p_type
         phdr.p_paddr = p_paddr
+        phdr.p_vaddr = p_paddr

With your proposed change though, I can dump memory with gdb...

Quoted 7 lines
scripts/dump-guest-memory.py +0 -1
         phdr.p_filesz = p_size
         phdr.p_memsz = p_size
         self.segments.append(phdr)
--

2.19.2

I've never used paging-enabled dumps. First, because doing so requires QEMU to trust guest memory contents (see original commit 783e9b4826b9; or more recently/generally, the @dump-guest-memory docs in "qapi/misc.json"). Second, because whenever I had to deal with guest memory dumps, I always used "crash" (which needs no paging), and the subject guests were all Linux.

I can't comment on paging-enabled patches for dump, except that they shouldn't regress the paging-disabled functionality. :) If the patches satisfy that, I'm fine.

(I *am* surprised that GDB insists on p_vaddr equaling p_paddr; after all, in the guest, the virtual address is "memory_mapping->virt_addr". But, I would never claim to understand most of the ELF intricacies, and/or what GDB requires on top of those.)

Thanks
Laszlo

author
#3 lore raw

Thank you for looking into this, perhaps I could change the patch (at least in the C part not the python script) to something like:
- phdr.p_vaddr = cpu_to_dumpXX(s, memory_mapping->virt_addr);
+ phdr.p_vaddr = cpu_to_dumpXX(s, memory_mapping->virt_addr) ?
cpu_to_dumpXX(s, memory_mapping->virt_addr) : phdr.p_paddr;

So in the case of paging where virt_addr is available we will use it

Thanks,
-- Jon.

Quoted 93 lines · Laszlo Ersek
Quoted 51 lines · Marc-André Lureau

Hi

Quoted 9 lines · Jon Doron

vaddr needs to be equal to the paddr since the dump file represents the physical memory image.

Without setting vaddr correctly, GDB would load all the different memory regions on top of each other to vaddr 0, thus making GDB showing the wrong memory data for a given address.

This is a non-trivial patch! (qemu-trivial, please ignore).

Quoted 10 lines
---
 dump.c                       | 4 ++--
 scripts/dump-guest-memory.py | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)
dump.c +0 -0
index 4ec94c5e25..bf77a119ea 100644

dump.c static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping, · line 195

Quoted 5 lines
dump.c +1 -1
@@ -192,7 +192,7 @@ static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump64(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump64(s, filesz);
     phdr.p_memsz = cpu_to_dump64(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump64(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;

This is likely breaking paging=true somehow, which sets memory_mapping->virt_addr to non-0.

According to doc "If you want to use gdb to process the core, please set @paging to true."

Although I am not able to (gdb) x/10bx 0xa0000 for example on a core produced with paging. Not sure why, anybody could help?

Quoted 11 lines
dump.c +1 -1
     assert(memory_mapping->length >= filesz);
 
@@ -216,7 +216,7 @@ static void write_elf32_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump32(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump32(s, filesz);
     phdr.p_memsz = cpu_to_dump32(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump32(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;
 
     assert(memory_mapping->length >= filesz);
 
Quoted 21 lines

scripts/dump-guest-memory.py class ELF(object): · line 166

Quoted 9 lines
scripts/dump-guest-memory.py +1 -0
index 198cd0fe40..2c587cbefc 100644
@@ -163,6 +163,7 @@ class ELF(object):
         phdr = get_arch_phdr(self.endianness, self.elfclass)
         phdr.p_type = p_type
         phdr.p_paddr = p_paddr
+        phdr.p_vaddr = p_paddr

With your proposed change though, I can dump memory with gdb...

Quoted 7 lines
scripts/dump-guest-memory.py +0 -1
         phdr.p_filesz = p_size
         phdr.p_memsz = p_size
         self.segments.append(phdr)
--

2.19.2

I've never used paging-enabled dumps. First, because doing so requires QEMU to trust guest memory contents (see original commit 783e9b4826b9; or more recently/generally, the @dump-guest-memory docs in "qapi/misc.json"). Second, because whenever I had to deal with guest memory dumps, I always used "crash" (which needs no paging), and the subject guests were all Linux.

I can't comment on paging-enabled patches for dump, except that they shouldn't regress the paging-disabled functionality. :) If the patches satisfy that, I'm fine.

(I *am* surprised that GDB insists on p_vaddr equaling p_paddr; after all, in the guest, the virtual address is "memory_mapping->virt_addr". But, I would never claim to understand most of the ELF intricacies, and/or what GDB requires on top of those.)

Thanks
Laszlo

#4 lore raw
Quoted 7 lines · Jon Doron

Thank you for looking into this, perhaps I could change the patch (at least in the C part not the python script) to something like:

(quoted diff) +1 -1
-    phdr.p_vaddr = cpu_to_dumpXX(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = cpu_to_dumpXX(s, memory_mapping->virt_addr) ?

cpu_to_dumpXX(s, memory_mapping->virt_addr) : phdr.p_paddr;

So in the case of paging where virt_addr is available we will use it

I guess that would be an improvement, yes.

Regarding style: although I'm personally opposed to most GNUisms, QEMU liberally uses, for example, the GNU extension:

https://gcc.gnu.org/onlinedocs/gcc/Conditionals.html

And, the "?:" operator would apply here.

Thanks,
Laszlo

Quoted 94 lines · Laszlo Ersek
Quoted 93 lines · Marc-André Lureau
Quoted 51 lines

Hi

Quoted 9 lines · Jon Doron

vaddr needs to be equal to the paddr since the dump file represents the physical memory image.

Without setting vaddr correctly, GDB would load all the different memory regions on top of each other to vaddr 0, thus making GDB showing the wrong memory data for a given address.

This is a non-trivial patch! (qemu-trivial, please ignore).

Quoted 10 lines
---
 dump.c                       | 4 ++--
 scripts/dump-guest-memory.py | 1 +
 2 files changed, 3 insertions(+), 2 deletions(-)
dump.c +0 -0
index 4ec94c5e25..bf77a119ea 100644

dump.c static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping, · line 195

Quoted 5 lines
dump.c +1 -1
@@ -192,7 +192,7 @@ static void write_elf64_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump64(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump64(s, filesz);
     phdr.p_memsz = cpu_to_dump64(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump64(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;

This is likely breaking paging=true somehow, which sets memory_mapping->virt_addr to non-0.

According to doc "If you want to use gdb to process the core, please set @paging to true."

Although I am not able to (gdb) x/10bx 0xa0000 for example on a core produced with paging. Not sure why, anybody could help?

Quoted 11 lines
dump.c +1 -1
     assert(memory_mapping->length >= filesz);
 
@@ -216,7 +216,7 @@ static void write_elf32_load(DumpState *s, MemoryMapping *memory_mapping,
     phdr.p_paddr = cpu_to_dump32(s, memory_mapping->phys_addr);
     phdr.p_filesz = cpu_to_dump32(s, filesz);
     phdr.p_memsz = cpu_to_dump32(s, memory_mapping->length);
-    phdr.p_vaddr = cpu_to_dump32(s, memory_mapping->virt_addr);
+    phdr.p_vaddr = phdr.p_paddr;
 
     assert(memory_mapping->length >= filesz);
 
Quoted 21 lines

scripts/dump-guest-memory.py class ELF(object): · line 166

Quoted 9 lines
scripts/dump-guest-memory.py +1 -0
index 198cd0fe40..2c587cbefc 100644
@@ -163,6 +163,7 @@ class ELF(object):
         phdr = get_arch_phdr(self.endianness, self.elfclass)
         phdr.p_type = p_type
         phdr.p_paddr = p_paddr
+        phdr.p_vaddr = p_paddr

With your proposed change though, I can dump memory with gdb...

Quoted 7 lines
scripts/dump-guest-memory.py +0 -1
         phdr.p_filesz = p_size
         phdr.p_memsz = p_size
         self.segments.append(phdr)
--

2.19.2

I've never used paging-enabled dumps. First, because doing so requires QEMU to trust guest memory contents (see original commit 783e9b4826b9; or more recently/generally, the @dump-guest-memory docs in "qapi/misc.json"). Second, because whenever I had to deal with guest memory dumps, I always used "crash" (which needs no paging), and the subject guests were all Linux.

I can't comment on paging-enabled patches for dump, except that they shouldn't regress the paging-disabled functionality. :) If the patches satisfy that, I'm fine.

(I *am* surprised that GDB insists on p_vaddr equaling p_paddr; after all, in the guest, the virtual address is "memory_mapping->virt_addr". But, I would never claim to understand most of the ELF intricacies, and/or what GDB requires on top of those.)

Thanks
Laszlo