fix address space size issue
v2 cover fix address space size issue
A bug reported by Luiz Capitulino let us to find several bugs in memory address space setup.
One issue is that gdb stub can give us arbitrary addresses and we'll try to access them.
Since our lookup ignored high bits in the address, we hit a wrong section and got a crash.
In fact, PCI devices can access arbitrary addresses too, so we should just make lookup robust against this case.
Another issue has to do with size of regions. memory API uses UINT64_MAX so say "all 64 bit" but some devices mistakenly used INT64_MAX.
It should not affect most systems in practice as everything should be limited by address space size, but it's an API misuse that we should not keep around, and it will become a problem if a system with 64 bit target address hits this path.
Patch 1 introduces TARGET_PHYS_ADDR_SPACE_MAX that is
the max size for memory regions rendered by exec.
Patches 2-3 limits the size of memory regions used by exec.c. Patch 4 fixes an actual bug.
The rest of patches make code cleaner and more robust.
Marcel Apfelbaum (3):
exec: declare TARGET_PHYS_ADDR_SPACE_MAX to limit memory regions
rendered by exec
hw/alpha: limit iommu-typhoon memory size
hw/ppc: limit iommu-spapr memory size
Michael S. Tsirkin (4):
exec: don't ignore high address bits on lookup
pci: fix address space size for bridge
exec: don't ignore high address bits on set
spapr_pci: s/INT64_MAX/UINT64_MAX/
Paolo Bonzini (1):
pc: s/INT64_MAX/UINT64_MAX/
exec.c | 18 ++++++++++++++++++ hw/alpha/typhoon.c | 2 +- hw/i386/pc_piix.c | 2 +- hw/i386/pc_q35.c | 2 +- hw/pci/pci_bridge.c | 2 +- hw/ppc/spapr_iommu.c | 2 +- hw/ppc/spapr_pci.c | 2 +- include/exec/address-spaces.h | 4 ++++ 8 files changed, 28 insertions(+), 6 deletions(-)
Signature · 1 line
1.8.3.1
PATCH v2 1/8 exec: declare TARGET_PHYS_ADDR_SPACE_MAX to limit memory regions rendered by exec
The page table logic in exec.c assumes
that memory addresses are at most TARGET_PHYS_ADDR_SPACE_BITS. Use TARGET_PHYS_ADDR_SPACE_MAX as max size for memory regions rendered by exec.
--- include/exec/address-spaces.h | 4 ++++ 1 file changed, 4 insertions(+)
include/exec/address-spaces.h +4 -0
@@ -23,6 +23,10 @@#ifndef CONFIG_USER_ONLY +#define TARGET_PHYS_ADDR_SPACE_MAX \ + (TARGET_PHYS_ADDR_SPACE_BITS == 64 ? \ + UINT64_MAX : (0x1ULL << TARGET_PHYS_ADDR_SPACE_BITS)) +/* Get the root memory region. This interface should only be used temporarily * until a proper bus interface is available. */
Signature · 1 line
1.8.3.1
Re: [Qemu-devel] [PATCH for-1.7 v2 1/8] exec: declare TARGET_PHYS_ADDR_SPACE_MAX to limit memory regions rendered by exec
Quoted 15 lines · Marcel Apfelbaum
The page table logic in
exec.cassumes
that memory addresses are at mostTARGET_PHYS_ADDR_SPACE_BITS. UseTARGET_PHYS_ADDR_SPACE_MAXas max size for memory regions rendered by exec.Signed-off-by Marcel Apfelbaum <[email protected]>--- include/exec/address-spaces.h | 4 ++++ 1 file changed, 4 insertions(+)include/exec/address-spaces.h +0
-0
include/exec/address-spaces.h line 29
Quoted 7 lines
include/exec/address-spaces.h +4
-0@@ -23,6 +23,10 @@#ifndef CONFIG_USER_ONLY +#define TARGET_PHYS_ADDR_SPACE_MAX \ + (TARGET_PHYS_ADDR_SPACE_BITS == 64 ? \ + UINT64_MAX : (0x1ULL << TARGET_PHYS_ADDR_SPACE_BITS)) +
I think it's worth adding a comment that this is a size intended for use in memory_region_init() calls and so follows the odd convention used by that API that it is a size in bytes with the exception that UINT64_MAX represents 2^64.
(it follows from this that using the #define anywhere except in a memory_region_init() call is probably a bug)
-- PMM
Re: [Qemu-devel] [PATCH for-1.7 v2 1/8] exec: declare TARGET_PHYS_ADDR_SPACE_MAX to limit memory regions rendered by exec
Quoted 19 lines · Peter Maydell
Quoted 18 lines · Marcel Apfelbaum
The page table logic in
exec.cassumes
that memory addresses are at mostTARGET_PHYS_ADDR_SPACE_BITS. UseTARGET_PHYS_ADDR_SPACE_MAXas max size for memory regions rendered by exec.Signed-off-by Marcel Apfelbaum <[email protected]>--- include/exec/address-spaces.h | 4 ++++ 1 file changed, 4 insertions(+)include/exec/address-spaces.h +0
-0@@ -23,6 +23,10 @@#ifndef CONFIG_USER_ONLY
Quoted 10 lines
include/exec/address-spaces.h line 29
Quoted 4 lines
include/exec/address-spaces.h +4
-0+#define TARGET_PHYS_ADDR_SPACE_MAX \ + (TARGET_PHYS_ADDR_SPACE_BITS == 64 ? \ + UINT64_MAX : (0x1ULL << TARGET_PHYS_ADDR_SPACE_BITS)) +I think it's worth adding a comment that this is a size intended for use in
memory_region_init()calls and so follows the odd convention used by that API that it is a size in bytes with the exception thatUINT64_MAXrepresents 2^64.
I will add the comment in v3.
Thanks,
Marcel
Quoted 5 lines
(it follows from this that using the #define anywhere except in a
memory_region_init()call is probably a bug)-- PMM
Re: [Qemu-devel] [PATCH for-1.7 v2 1/8] exec: declare TARGET_PHYS_ADDR_SPACE_MAX to limit memory regions rendered by exec
In reply to Peter MaydellQuoted 30 lines · Peter Maydell
Quoted 15 lines · Marcel Apfelbaum
The page table logic in
exec.cassumes
that memory addresses are at mostTARGET_PHYS_ADDR_SPACE_BITS. UseTARGET_PHYS_ADDR_SPACE_MAXas max size for memory regions rendered by exec.Signed-off-by Marcel Apfelbaum <[email protected]>--- include/exec/address-spaces.h | 4 ++++ 1 file changed, 4 insertions(+)include/exec/address-spaces.h +0
-0include/exec/address-spaces.h line 29
Quoted 7 lines
include/exec/address-spaces.h +4
-0@@ -23,6 +23,10 @@#ifndef CONFIG_USER_ONLY +#define TARGET_PHYS_ADDR_SPACE_MAX \ + (TARGET_PHYS_ADDR_SPACE_BITS == 64 ? \ + UINT64_MAX : (0x1ULL << TARGET_PHYS_ADDR_SPACE_BITS)) +I think it's worth adding a comment that this is a size intended for use in
memory_region_init()calls and so follows the odd convention used by that API that it is a size in bytes with the exception thatUINT64_MAXrepresents 2^64.
Quoted 4 lines
(it follows from this that using the #define anywhere except in a
memory_region_init()call is probably a bug)-- PMM
BTW how about we change the API to pass in int128? Not for 1.7 of course.
This will help make sure it's only used for MRs.
Signature · 1 line
MST
Re: [Qemu-devel] [PATCH for-1.7 v2 1/8] exec: declare TARGET_PHYS_ADDR_SPACE_MAX to limit memory regions rendered by exec
Il 07/11/2013 13:04, Michael S. Tsirkin ha scritto:
Quoted 6 lines
Quoted 4 lines
Quoted 4 lines
(it follows from this that using the #define anywhere except in a
memory_region_init()call is probably a bug)-- PMM
BTW how about we change the API to pass in int128?
The vast majority of memory regions have small sizes, it would add boilerplate to wrap the size with an int128_make64 everywhere.
Paolo
Quoted 3 lines
Not for 1.7 of course.
This will help make sure it's only used for MRs.
Re: [Qemu-devel] [PATCH for-1.7 v2 1/8] exec: declare TARGET_PHYS_ADDR_SPACE_MAX to limit memory regions rendered by exec
Quoted 4 lines · Paolo Bonzini
Il 07/11/2013 13:04, Michael S. Tsirkin ha scritto:
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
(it follows from this that using the #define anywhere except in a
memory_region_init()call is probably a bug)
Quoted 8 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
-- PMM
BTW how about we change the API to pass in int128?
The vast majority of memory regions have small sizes, it would add boilerplate to wrap the size with an
int128_make64everywhere.Paolo
Well let's have two APIs for 64 bit and for 128 bit.
Quoted 3 lines
Quoted 3 lines
Not for 1.7 of course.
This will help make sure it's only used for MRs.
PATCH v2 2/8 hw/alpha: limit iommu-typhoon memory size
In reply to Marcel ApfelbaumThe page table logic in exec.c assumes
that memory addresses are at most TARGET_PHYS_ADDR_SPACE_BITS. Limit iommu-typhoon accordingly.
--- hw/alpha/typhoon.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
hw/alpha/typhoon.c +1 -1
@@ -888,7 +888,7 @@ PCIBus *typhoon_init(ram_addr_t ram_size, ISABus **isa_bus,/* Host memory as seen from the PCI side, via the IOMMU. */ memory_region_init_iommu (&s->pchip.iommu,OBJECT (s), &typhoon_iommu_ops, -"iommu-typhoon" ,UINT64_MAX ); +"iommu-typhoon" ,TARGET_PHYS_ADDR_SPACE_MAX );address_space_init (&s->pchip.iommu_as, &s->pchip.iommu,"pchip0-pci" );pci_setup_iommu (b, typhoon_pci_dma_iommu, s);
Signature · 1 line
1.8.3.1
PATCH v2 3/8 hw/ppc: limit iommu-spapr memory size
In reply to Marcel ApfelbaumThe page table logic in exec.c assumes
that memory addresses are at most TARGET_PHYS_ADDR_SPACE_BITS. Limit iommu-spapr accordingly.
--- hw/ppc/spapr_iommu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
hw/ppc/spapr_iommu.c +1 -1
@@ -136,7 +136,7 @@ static int spapr_tce_table_realize(DeviceState *dev)trace_spapr_iommu_new_table (tcet->liobn, tcet, tcet->table, tcet->fd);memory_region_init_iommu (&tcet->iommu,OBJECT (dev), &spapr_iommu_ops, -"iommu-spapr" ,UINT64_MAX ); +"iommu-spapr" ,TARGET_PHYS_ADDR_SPACE_MAX );QLIST_INSERT_HEAD (&spapr_tce_tables, tcet, list);
Signature · 1 line
1.8.3.1
PATCH v2 4/8 exec: don't ignore high address bits on lookup
In reply to Marcel ApfelbaumFrom: "Michael S. Tsirkin" <[email protected]>
Lookup of address > target address space should return an unassigned section, instead of silently ignoring high bits.
--- exec.c | 7 +++++++ 1 file changed, 7 insertions(+)
exec.c +7 -0
@@ -197,6 +197,9 @@ static void phys_page_set_level(PhysPageEntry *lp, hwaddr *index, } } +#define TARGET_PHYS_MAX_ADDR \ + (0x1ULL << (TARGET_PHYS_ADDR_SPACE_BITS - TARGET_PAGE_BITS)) +static void phys_page_set (AddressSpaceDispatch *d,hwaddr index,hwaddr nb,uint16_t leaf)
@@ -213,6 +216,10 @@ static MemoryRegionSection *phys_page_find(PhysPageEntry lp, hwaddr index,PhysPageEntry *p;int i; +if (index >=TARGET_PHYS_MAX_ADDR ) { +return §ions[PHYS_SECTION_UNASSIGNED ]; + } +for (i =P_L2_LEVELS -1 ; i >=0 && !lp.is_leaf; i--) {if (lp.ptr ==PHYS_MAP_NODE_NIL ) {return §ions[PHYS_SECTION_UNASSIGNED ];
Signature · 1 line
1.8.3.1
PATCH v2 5/8 pci: fix address space size for bridge
In reply to Marcel ApfelbaumFrom: "Michael S. Tsirkin" <[email protected]>
Address space size for bridge should be full 64 bit, so we should use UINT64_MAX not INT64_MAX as it's size.
--- hw/pci/pci_bridge.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
hw/pci/pci_bridge.c +1 -1
@@ -372,7 +372,7 @@ int pci_bridge_initfn(PCIDevice *dev, const char *typename) sec_bus->parent_dev = dev; sec_bus->map_irq = br->map_irq ? br->map_irq : pci_swizzle_map_irq_fn; sec_bus->address_space_mem = &br->address_space_mem; -memory_region_init (&br->address_space_mem,OBJECT (br),"pci_bridge_pci" ,INT64_MAX ); +memory_region_init (&br->address_space_mem,OBJECT (br),"pci_bridge_pci" ,UINT64_MAX ); sec_bus->address_space_io = &br->address_space_io;memory_region_init (&br->address_space_io,OBJECT (br),"pci_bridge_io" ,65536 ); br->windows =pci_bridge_region_init (br);
Signature · 1 line
1.8.3.1
PATCH v2 6/8 exec: don't ignore high address bits on set
In reply to Marcel ApfelbaumFrom: "Michael S. Tsirkin" <[email protected]>
We should never add pages outside target address space, we had such bugs in the past so add debug message to catch that quickly if we do, otherwise at least don't corrupt the page tables.
--- exec.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
exec.c +11 -0
@@ -52,6 +52,7 @@#include "exec/memory-internal.h" //#define DEBUG_SUBPAGE +/* #define DEBUG_PHYS_PAGE_SET */ #if !defined(CONFIG_USER_ONLY )static int in_migration;
@@ -204,6 +205,16 @@ static void phys_page_set(AddressSpaceDispatch *d, hwaddr index, hwaddr nb, uint16_t leaf) { + if (nb && (index + nb - 1) >= TARGET_PHYS_MAX_ADDR) { +#ifdef DEBUG_PHYS_PAGE_SET +printf ("%s: page table corruption: page address " TARGET_FMT_plx +" > max supported address %llx\n" , __func__, (index + nb -1 ), +TARGET_PHYS_MAX_ADDR ); +#endif +/* at least don't corrupt the page tables */ +return ; + } +/* Wildly overreserve - it doesn't matter much. */ phys_map_node_reserve (3 *P_L2_LEVELS );
Signature · 1 line
1.8.3.1
PATCH v2 7/8 pc: s/INT64_MAX/UINT64_MAX/
In reply to Marcel ApfelbaumFrom: Paolo Bonzini <[email protected]>
It doesn't make sense for a region to be INT64_MAX in size:
memory core uses UINT64_MAX as a special value meaning "all 64 bit" this is what was meant here.
While this should never affect the PC system which at the moment always has < 63 bit size, this makes us hit all kind of corner case bugs with sub-pages, so users are probably better off if we just use UINT64_MAX instead.
--- hw/i386/pc_piix.c | 2 +- hw/i386/pc_q35.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
hw/i386/pc_piix.c +1 -1
@@ -115,7 +115,7 @@ static void pc_init1(QEMUMachineInitArgs *args,if (pci_enabled) { pci_memory =g_new (MemoryRegion,1 ); -memory_region_init (pci_memory,NULL ,"pci" ,INT64_MAX ); +memory_region_init (pci_memory,NULL ,"pci" ,UINT64_MAX ); rom_memory = pci_memory; }else { pci_memory =NULL ;
hw/i386/pc_q35.c +1 -1
@@ -102,7 +102,7 @@ static void pc_q35_init(QEMUMachineInitArgs *args)/* pci enabled */ if (pci_enabled) { pci_memory =g_new (MemoryRegion,1 ); -memory_region_init (pci_memory,NULL ,"pci" ,INT64_MAX ); +memory_region_init (pci_memory,NULL ,"pci" ,UINT64_MAX ); rom_memory = pci_memory; }else { pci_memory =NULL ;
Signature · 1 line
1.8.3.1
Re: [Qemu-devel] [PATCH for-1.7 v2 7/8] pc: s/INT64_MAX/UINT64_MAX/
Il 07/11/2013 11:41, Marcel Apfelbaum ha scritto:
Quoted 11 lines
From: Paolo Bonzini <[email protected]>
It doesn't make sense for a region to be
INT64_MAXin size:
memory core usesUINT64_MAXas a special value meaning "all 64 bit" this is what was meant here.While this should never affect the PC system which at the moment always has < 63 bit size, this makes us hit all kind of corner case bugs with sub-pages, so users are probably better off if we just use
UINT64_MAXinstead.
Quoted 4 lines
Reported-by Luiz Capitulino <[email protected]>Tested-by Luiz Capitulino <[email protected]>Reviewed-by Michael S. Tsirkin <[email protected]>Signed-off-by Michael S. Tsirkin <[email protected]>
Michael, is this patch (and 8/8: spapr_pci: s/INT64_MAX/UINT64_MAX/) queued for 1.7 and/or 1.8?
Paolo
Quoted 32 lines
--- hw/i386/pc_piix.c | 2 +- hw/i386/pc_q35.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)hw/i386/pc_piix.c +1
-1@@ -115,7 +115,7 @@ static void pc_init1(QEMUMachineInitArgs *args,if (pci_enabled) { pci_memory =g_new (MemoryRegion,1 ); -memory_region_init (pci_memory,NULL ,"pci" ,INT64_MAX ); +memory_region_init (pci_memory,NULL ,"pci" ,UINT64_MAX ); rom_memory = pci_memory; }else { pci_memory =NULL ;hw/i386/pc_q35.c +1
-1@@ -102,7 +102,7 @@ static void pc_q35_init(QEMUMachineInitArgs *args)/* pci enabled */ if (pci_enabled) { pci_memory =g_new (MemoryRegion,1 ); -memory_region_init (pci_memory,NULL ,"pci" ,INT64_MAX ); +memory_region_init (pci_memory,NULL ,"pci" ,UINT64_MAX ); rom_memory = pci_memory; }else { pci_memory =NULL ;
Re: [Qemu-devel] [PATCH for-1.7 v2 7/8] pc: s/INT64_MAX/UINT64_MAX/
Quoted 17 lines · Paolo Bonzini
Il 07/11/2013 11:41, Marcel Apfelbaum ha scritto:
Quoted 11 lines
From: Paolo Bonzini <[email protected]>
It doesn't make sense for a region to be
INT64_MAXin size:
memory core usesUINT64_MAXas a special value meaning "all 64 bit" this is what was meant here.While this should never affect the PC system which at the moment always has < 63 bit size, this makes us hit all kind of corner case bugs with sub-pages, so users are probably better off if we just use
UINT64_MAXinstead.Quoted 4 lines
Reported-by Luiz Capitulino <[email protected]>Tested-by Luiz Capitulino <[email protected]>Reviewed-by Michael S. Tsirkin <[email protected]>Signed-off-by Michael S. Tsirkin <[email protected]>
Quoted 4 lines
Michael, is this patch (and 8/8:
spapr_pci: s/INT64_MAX/UINT64_MAX/) queued for 1.7 and/or 1.8?Paolo
I queued this for 1.8 - I don't think it fixes any known specific bugs. You can see my queue atgit://git.kernel.org/pub/scm/virt/kvm/mst/qemu.git pci
Quoted 32 lines
Quoted 32 lines
--- hw/i386/pc_piix.c | 2 +- hw/i386/pc_q35.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)hw/i386/pc_piix.c +1
-1@@ -115,7 +115,7 @@ static void pc_init1(QEMUMachineInitArgs *args,if (pci_enabled) { pci_memory =g_new (MemoryRegion,1 ); -memory_region_init (pci_memory,NULL ,"pci" ,INT64_MAX ); +memory_region_init (pci_memory,NULL ,"pci" ,UINT64_MAX ); rom_memory = pci_memory; }else { pci_memory =NULL ;hw/i386/pc_q35.c +1
-1@@ -102,7 +102,7 @@ static void pc_q35_init(QEMUMachineInitArgs *args)/* pci enabled */ if (pci_enabled) { pci_memory =g_new (MemoryRegion,1 ); -memory_region_init (pci_memory,NULL ,"pci" ,INT64_MAX ); +memory_region_init (pci_memory,NULL ,"pci" ,UINT64_MAX ); rom_memory = pci_memory; }else { pci_memory =NULL ;
PATCH v2 8/8 spapr_pci: s/INT64_MAX/UINT64_MAX/
In reply to Marcel ApfelbaumFrom: "Michael S. Tsirkin" <[email protected]>
It doesn't make sense for a region to be INT64_MAX in size:
memory core uses UINT64_MAX as a special value meaning "all 64 bit" this is what was meant here.
While this should never affect the spapr system which at the moment always has < 63 bit size, this makes us hit all kind of corner case bugs with sub-pages, so users are probably better off if we just use UINT64_MAX instead.
--- hw/ppc/spapr_pci.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
hw/ppc/spapr_pci.c +1 -1
@@ -555,7 +555,7 @@ static int spapr_phb_init(SysBusDevice *s)/* Initialize memory regions */ sprintf (namebuf,"%s.mmio" , sphb->dtbusname); -memory_region_init (&sphb->memspace,OBJECT (sphb), namebuf,INT64_MAX ); +memory_region_init (&sphb->memspace,OBJECT (sphb), namebuf,UINT64_MAX );sprintf (namebuf,"%s.mmio-alias" , sphb->dtbusname); memory_region_init_alias(&sphb->memwindow, OBJECT(sphb),
Signature · 1 line
1.8.3.1
Quoted 20 lines · Marcel Apfelbaum
A bug reported by Luiz Capitulino let us to find several bugs in memory address space setup.
One issue is that gdb stub can give us arbitrary addresses and we'll try to access them.
Since our lookup ignored high bits in the address, we hit a wrong section and got a crash.
In fact, PCI devices can access arbitrary addresses too, so we should just make lookup robust against this case.Another issue has to do with size of regions. memory API uses
UINT64_MAXso say "all 64 bit" but some devices mistakenly usedINT64_MAX.It should not affect most systems in practice as everything should be limited by address space size, but it's an API misuse that we should not keep around, and it will become a problem if a system with 64 bit target address hits this path.
Quoted 5 lines
Patch 1 introduces
TARGET_PHYS_ADDR_SPACE_MAXthat is
the max size for memory regions rendered by exec.
Patches 2-3 limits the size of memory regions used byexec.c. Patch 4 fixes an actual bug.
The rest of patches make code cleaner and more robust.
ppc bits are:
Alex
Il 07/11/2013 11:41, Marcel Apfelbaum ha scritto:
Quoted 20 lines
A bug reported by Luiz Capitulino let us to find several bugs in memory address space setup.
One issue is that gdb stub can give us arbitrary addresses and we'll try to access them.
Since our lookup ignored high bits in the address, we hit a wrong section and got a crash.
In fact, PCI devices can access arbitrary addresses too, so we should just make lookup robust against this case.Another issue has to do with size of regions. memory API uses
UINT64_MAXso say "all 64 bit" but some devices mistakenly usedINT64_MAX.It should not affect most systems in practice as everything should be limited by address space size, but it's an API misuse that we should not keep around, and it will become a problem if a system with 64 bit target address hits this path.
Quoted 5 lines
Patch 1 introduces
TARGET_PHYS_ADDR_SPACE_MAXthat is
the max size for memory regions rendered by exec.
Patches 2-3 limits the size of memory regions used byexec.c. Patch 4 fixes an actual bug.
The rest of patches make code cleaner and more robust.
I think this is wrong.
There is no reason why boards should be using TARGET_PHYS_ADDR_SPACE_MAX if the PCI address space is 64-bit. Many files in hw/ do not even have access to TARGET_PHYS_ADDR_SPACE_MAX, which is only available to files that are compiled per-target.
The fact that you have to reduce the IOMMU address spaces from 64 bits (UINT64_MAX, not the buggy INT64_MAX) to something else is a red flag. If the IOMMU supports 64-bit addressing, the IOMMU regions should have 2^64 bits as their length. Yes, it works by chance now, but it _does_ work for 2^64-bit size regions: you can do DMA to a high address (say 0xf << 60), the IOMMU tables will convert to a low address that fits within TARGET_PHYS_ADDR_SPACE_BITS, and the everything will be fine. Patches 4 and 6 change that.
So, ack for patch 5-7-8, which should also be enough to fix the problem that Luiz reported. For everything else the solution is to make memory dispatch use the whole 64 bits, as in
http://permalink.gmane.org/gmane.comp.emulators.qemu/240229. If you want to delay that to 1.8 that's fine.
Paolo
Quoted 26 lines
Marcel Apfelbaum (3):
exec: declareTARGET_PHYS_ADDR_SPACE_MAXto limit memory regions
rendered by exec
hw/alpha: limit iommu-typhoon memory size
hw/ppc: limit iommu-spapr memory sizeMichael S. Tsirkin (4):
exec: don't ignore high address bits on lookup
pci: fix address space size for bridge
exec: don't ignore high address bits on set
spapr_pci: s/INT64_MAX/UINT64_MAX/Paolo Bonzini (1):
pc: s/INT64_MAX/UINT64_MAX/exec.c | 18 ++++++++++++++++++ hw/alpha/typhoon.c | 2 +- hw/i386/pc_piix.c | 2 +- hw/i386/pc_q35.c | 2 +- hw/pci/pci_bridge.c | 2 +- hw/ppc/spapr_iommu.c | 2 +- hw/ppc/spapr_pci.c | 2 +- include/exec/address-spaces.h | 4 ++++ 8 files changed, 28 insertions(+), 6 deletions(-)
Quoted 43 lines · Paolo Bonzini
Il 07/11/2013 11:41, Marcel Apfelbaum ha scritto:
Quoted 20 lines
A bug reported by Luiz Capitulino let us to find several bugs in memory address space setup.
One issue is that gdb stub can give us arbitrary addresses and we'll try to access them.
Since our lookup ignored high bits in the address, we hit a wrong section and got a crash.
In fact, PCI devices can access arbitrary addresses too, so we should just make lookup robust against this case.Another issue has to do with size of regions. memory API uses
UINT64_MAXso say "all 64 bit" but some devices mistakenly usedINT64_MAX.It should not affect most systems in practice as everything should be limited by address space size, but it's an API misuse that we should not keep around, and it will become a problem if a system with 64 bit target address hits this path.
Quoted 5 lines
Patch 1 introduces
TARGET_PHYS_ADDR_SPACE_MAXthat is
the max size for memory regions rendered by exec.
Patches 2-3 limits the size of memory regions used byexec.c. Patch 4 fixes an actual bug.
The rest of patches make code cleaner and more robust.I think this is wrong.
There is no reason why boards should be using
TARGET_PHYS_ADDR_SPACE_MAXif the PCI address space is 64-bit. Many files in hw/ do not even have access toTARGET_PHYS_ADDR_SPACE_MAX, which is only available to files that are compiled per-target.The fact that you have to reduce the IOMMU address spaces from 64 bits (
UINT64_MAX, not the buggyINT64_MAX) to something else is a red flag. If the IOMMU supports 64-bit addressing, the IOMMU regions should have 2^64 bits as their length. Yes, it works by chance now, but it _does_ work for 2^64-bit size regions: you can do DMA to a high address (say0xf<< 60), the IOMMU tables will convert to a low address that fits withinTARGET_PHYS_ADDR_SPACE_BITS, and the everything will be fine. Patches 4 and 6 change that.
Quoted 2 lines
So, ack for patch 5-7-8, which should also be enough to fix the problem that Luiz reported.
Not at all. As long as exec.c ignores high bits, any access there will end up in the wrong region.
We might not get a crash but we'll get guest memory corruption. Not good.
Quoted 33 lines
For everything else the solution is to make memory
dispatch use the whole 64 bits, as in
http://permalink.gmane.org/gmane.comp.emulators.qemu/240229. If you want to delay that to 1.8 that's fine.Paolo
Quoted 26 lines
Marcel Apfelbaum (3):
exec: declareTARGET_PHYS_ADDR_SPACE_MAXto limit memory regions
rendered by exec
hw/alpha: limit iommu-typhoon memory size
hw/ppc: limit iommu-spapr memory sizeMichael S. Tsirkin (4):
exec: don't ignore high address bits on lookup
pci: fix address space size for bridge
exec: don't ignore high address bits on set
spapr_pci: s/INT64_MAX/UINT64_MAX/Paolo Bonzini (1):
pc: s/INT64_MAX/UINT64_MAX/exec.c | 18 ++++++++++++++++++ hw/alpha/typhoon.c | 2 +- hw/i386/pc_piix.c | 2 +- hw/i386/pc_q35.c | 2 +- hw/pci/pci_bridge.c | 2 +- hw/ppc/spapr_iommu.c | 2 +- hw/ppc/spapr_pci.c | 2 +- include/exec/address-spaces.h | 4 ++++ 8 files changed, 28 insertions(+), 6 deletions(-)
Il 07/11/2013 16:05, Michael S. Tsirkin ha scritto:
Quoted 4 lines
Quoted 2 lines
Quoted 2 lines
So, ack for patch 5-7-8, which should also be enough to fix the problem that Luiz reported.
Not at all. As long as
exec.cignores high bits, any access there will end up in the wrong region.
... unless it happens to be the correct region because the whole 2^64-bit range is covered by a single region. This is what happens for IOMMUs.
Paolo
Quoted 1 line
We might not get a crash but we'll get guest memory corruption.
Quoted 11 lines · Paolo Bonzini
Il 07/11/2013 16:05, Michael S. Tsirkin ha scritto:
Quoted 4 lines
Quoted 2 lines
Quoted 2 lines
So, ack for patch 5-7-8, which should also be enough to fix the problem that Luiz reported.
Not at all. As long as
exec.cignores high bits, any access there will end up in the wrong region.... unless it happens to be the correct region because the whole 2^64-bit range is covered by a single region. This is what happens for IOMMUs.
Paolo
Only if you are lucky.
What makes sure ignoring high bits does not hit the e1000 MMIO? Nothing I think.
Quoted 1 line
Quoted 1 line
We might not get a crash but we'll get guest memory corruption.