| Message ID | 1488171164-28319-1-git-send-email-xyjxie@linux.vnet.ibm.com |
|---|---|
| State | New |
| Headers | show |
On Mon, Feb 27, 2017 at 12:52:44PM +0800, Yongji Xie wrote: > At the moment ram device's memory regions are DEVICE_NATIVE_ENDIAN. It's > incorrect. This memory region is backed by a MMIO area in host, so the > uint64_t data that MemoryRegionOps read from/write to this area should be > host-endian rather than target-endian. Hence, current code does not work > when target and host endianness are different which is the most common case > on PPC64. To fix it, this introduces DEVICE_HOST_ENDIAN for the ram device. > > This has been tested on PPC64 BE/LE host/guest in all possible combinations > including TCG. > > Suggested-by: Paolo Bonzini <pbonzini@redhat.com> > Signed-off-by: Yongji Xie <xyjxie@linux.vnet.ibm.com> Reviewed-by: David Gibson <david@gibson.dropbear.id.au> The effect of the patch is certainly correct. I remain a little concerned that the name "host endian" might cause more confusion than it resolves, but a better term isn't immediately obvious to me. > --- > include/exec/cpu-common.h | 6 ++++++ > memory.c | 2 +- > 2 files changed, 7 insertions(+), 1 deletions(-) > > diff --git a/include/exec/cpu-common.h b/include/exec/cpu-common.h > index bd15853..eef74df 100644 > --- a/include/exec/cpu-common.h > +++ b/include/exec/cpu-common.h > @@ -36,6 +36,12 @@ enum device_endian { > DEVICE_LITTLE_ENDIAN, > }; > > +#if defined(HOST_WORDS_BIGENDIAN) > +#define DEVICE_HOST_ENDIAN DEVICE_BIG_ENDIAN > +#else > +#define DEVICE_HOST_ENDIAN DEVICE_LITTLE_ENDIAN > +#endif > + > /* address in the RAM (different from a physical address) */ > #if defined(CONFIG_XEN_BACKEND) > typedef uint64_t ram_addr_t; > diff --git a/memory.c b/memory.c > index ed8b5aa..17cfada 100644 > --- a/memory.c > +++ b/memory.c > @@ -1180,7 +1180,7 @@ static void memory_region_ram_device_write(void *opaque, hwaddr addr, > static const MemoryRegionOps ram_device_mem_ops = { > .read = memory_region_ram_device_read, > .write = memory_region_ram_device_write, > - .endianness = DEVICE_NATIVE_ENDIAN, > + .endianness = DEVICE_HOST_ENDIAN, > .valid = { > .min_access_size = 1, > .max_access_size = 8,
On 28/02/17 11:41, David Gibson wrote: > On Mon, Feb 27, 2017 at 12:52:44PM +0800, Yongji Xie wrote: >> At the moment ram device's memory regions are DEVICE_NATIVE_ENDIAN. It's >> incorrect. This memory region is backed by a MMIO area in host, so the >> uint64_t data that MemoryRegionOps read from/write to this area should be >> host-endian rather than target-endian. Hence, current code does not work >> when target and host endianness are different which is the most common case >> on PPC64. To fix it, this introduces DEVICE_HOST_ENDIAN for the ram device. >> >> This has been tested on PPC64 BE/LE host/guest in all possible combinations >> including TCG. >> >> Suggested-by: Paolo Bonzini <pbonzini@redhat.com> >> Signed-off-by: Yongji Xie <xyjxie@linux.vnet.ibm.com> > > Reviewed-by: David Gibson <david@gibson.dropbear.id.au> > > The effect of the patch is certainly correct. I remain a little > concerned that the name "host endian" might cause more confusion than > it resolves, but a better term isn't immediately obvious to me. In order to match memory_region_wrong_endianness(), it could be DEVICE_CORRECT_ENDIAN :) Just joking :) > >> --- >> include/exec/cpu-common.h | 6 ++++++ >> memory.c | 2 +- >> 2 files changed, 7 insertions(+), 1 deletions(-) >> >> diff --git a/include/exec/cpu-common.h b/include/exec/cpu-common.h >> index bd15853..eef74df 100644 >> --- a/include/exec/cpu-common.h >> +++ b/include/exec/cpu-common.h >> @@ -36,6 +36,12 @@ enum device_endian { >> DEVICE_LITTLE_ENDIAN, >> }; >> >> +#if defined(HOST_WORDS_BIGENDIAN) >> +#define DEVICE_HOST_ENDIAN DEVICE_BIG_ENDIAN >> +#else >> +#define DEVICE_HOST_ENDIAN DEVICE_LITTLE_ENDIAN >> +#endif >> + >> /* address in the RAM (different from a physical address) */ >> #if defined(CONFIG_XEN_BACKEND) >> typedef uint64_t ram_addr_t; >> diff --git a/memory.c b/memory.c >> index ed8b5aa..17cfada 100644 >> --- a/memory.c >> +++ b/memory.c >> @@ -1180,7 +1180,7 @@ static void memory_region_ram_device_write(void *opaque, hwaddr addr, >> static const MemoryRegionOps ram_device_mem_ops = { >> .read = memory_region_ram_device_read, >> .write = memory_region_ram_device_write, >> - .endianness = DEVICE_NATIVE_ENDIAN, >> + .endianness = DEVICE_HOST_ENDIAN, >> .valid = { >> .min_access_size = 1, >> .max_access_size = 8, >
on 2017/2/28 8:41, David Gibson wrote: > On Mon, Feb 27, 2017 at 12:52:44PM +0800, Yongji Xie wrote: >> At the moment ram device's memory regions are DEVICE_NATIVE_ENDIAN. It's >> incorrect. This memory region is backed by a MMIO area in host, so the >> uint64_t data that MemoryRegionOps read from/write to this area should be >> host-endian rather than target-endian. Hence, current code does not work >> when target and host endianness are different which is the most common case >> on PPC64. To fix it, this introduces DEVICE_HOST_ENDIAN for the ram device. >> >> This has been tested on PPC64 BE/LE host/guest in all possible combinations >> including TCG. >> >> Suggested-by: Paolo Bonzini <pbonzini@redhat.com> >> Signed-off-by: Yongji Xie <xyjxie@linux.vnet.ibm.com> > Reviewed-by: David Gibson <david@gibson.dropbear.id.au> > > The effect of the patch is certainly correct. I remain a little > concerned that the name "host endian" might cause more confusion than > it resolves, but a better term isn't immediately obvious to me. If the memory region's endianness indicates the endianness of multi-byte value that MemoryRegionOps read from/write to this memory region, should "host endian" be reasonable? For a mmio store, QEMU just get a bunch of bytes in the memory at the beginning. Then we use ldX_p to load a target-endian multi-byte value from the memory. Then adjust_endianness() change the endianness of the multi-byte value from target-endian to memory region's endianness. For the mmap MMIO area, we should use host-endian multi-byte value to access it. *(uint32_t *)(mr->ram_block->host + addr) = (uint32_t)data; Here it is the same as stl_he_p(). The "host-endian" means we load a bunch of bytes as a host-endian value, and write the value to the mmap MMIO area. That's my understanding. Not sure if it's correct. Thanks, Yongji
On Tue, Feb 28, 2017 at 06:12:56PM +0800, Yongji Xie wrote: > on 2017/2/28 8:41, David Gibson wrote: > > > On Mon, Feb 27, 2017 at 12:52:44PM +0800, Yongji Xie wrote: > > > At the moment ram device's memory regions are DEVICE_NATIVE_ENDIAN. It's > > > incorrect. This memory region is backed by a MMIO area in host, so the > > > uint64_t data that MemoryRegionOps read from/write to this area should be > > > host-endian rather than target-endian. Hence, current code does not work > > > when target and host endianness are different which is the most common case > > > on PPC64. To fix it, this introduces DEVICE_HOST_ENDIAN for the ram device. > > > > > > This has been tested on PPC64 BE/LE host/guest in all possible combinations > > > including TCG. > > > > > > Suggested-by: Paolo Bonzini <pbonzini@redhat.com> > > > Signed-off-by: Yongji Xie <xyjxie@linux.vnet.ibm.com> > > Reviewed-by: David Gibson <david@gibson.dropbear.id.au> > > > > The effect of the patch is certainly correct. I remain a little > > concerned that the name "host endian" might cause more confusion than > > it resolves, but a better term isn't immediately obvious to me. > > If the memory region's endianness indicates the endianness of multi-byte > value that > MemoryRegionOps read from/write to this memory region, should "host endian" > be reasonable? > > For a mmio store, QEMU just get a bunch of bytes in the memory at the > beginning. > Then we use ldX_p to load a target-endian multi-byte value from the memory. > Then > adjust_endianness() change the endianness of the multi-byte value from > target-endian > to memory region's endianness. > > For the mmap MMIO area, we should use host-endian multi-byte value to access > it. > > *(uint32_t *)(mr->ram_block->host + addr) = (uint32_t)data; > > Here it is the same as stl_he_p(). > > The "host-endian" means we load a bunch of bytes as a host-endian value, and > write the > value to the mmap MMIO area. That's my understanding. Not sure if it's > correct. That's correct. The difficulty is that generally the endian flag describes the device's endianness as it appears to the guest. The guest doesn't (and shouldn't) know the host's endianness, so describing something as "host endian" is pretty weird from that point of view. Basically the only way this can work is if the qemu device is treating all data from the guest as pieces of a bytestream and never interpreting things as multibyte values.
on 2017/3/1 8:35, David Gibson wrote: > On Tue, Feb 28, 2017 at 06:12:56PM +0800, Yongji Xie wrote: >> on 2017/2/28 8:41, David Gibson wrote: >> >>> On Mon, Feb 27, 2017 at 12:52:44PM +0800, Yongji Xie wrote: >>>> At the moment ram device's memory regions are DEVICE_NATIVE_ENDIAN. It's >>>> incorrect. This memory region is backed by a MMIO area in host, so the >>>> uint64_t data that MemoryRegionOps read from/write to this area should be >>>> host-endian rather than target-endian. Hence, current code does not work >>>> when target and host endianness are different which is the most common case >>>> on PPC64. To fix it, this introduces DEVICE_HOST_ENDIAN for the ram device. >>>> >>>> This has been tested on PPC64 BE/LE host/guest in all possible combinations >>>> including TCG. >>>> >>>> Suggested-by: Paolo Bonzini <pbonzini@redhat.com> >>>> Signed-off-by: Yongji Xie <xyjxie@linux.vnet.ibm.com> >>> Reviewed-by: David Gibson <david@gibson.dropbear.id.au> >>> >>> The effect of the patch is certainly correct. I remain a little >>> concerned that the name "host endian" might cause more confusion than >>> it resolves, but a better term isn't immediately obvious to me. >> If the memory region's endianness indicates the endianness of multi-byte >> value that >> MemoryRegionOps read from/write to this memory region, should "host endian" >> be reasonable? >> >> For a mmio store, QEMU just get a bunch of bytes in the memory at the >> beginning. >> Then we use ldX_p to load a target-endian multi-byte value from the memory. >> Then >> adjust_endianness() change the endianness of the multi-byte value from >> target-endian >> to memory region's endianness. >> >> For the mmap MMIO area, we should use host-endian multi-byte value to access >> it. >> >> *(uint32_t *)(mr->ram_block->host + addr) = (uint32_t)data; >> >> Here it is the same as stl_he_p(). >> >> The "host-endian" means we load a bunch of bytes as a host-endian value, and >> write the >> value to the mmap MMIO area. That's my understanding. Not sure if it's >> correct. > That's correct. The difficulty is that generally the endian flag > describes the device's endianness as it appears to the guest. The > guest doesn't (and shouldn't) know the host's endianness, so > describing something as "host endian" is pretty weird from that point > of view. Basically the only way this can work is if the qemu device > is treating all data from the guest as pieces of a bytestream and > never interpreting things as multibyte values. > OK, I think I know what you mean. Indeed, it's hard to describe the ram device's endianness from this point of view. Just transfer the bytestream without considering any endianness seems to be good. Thanks, Yongji
diff --git a/include/exec/cpu-common.h b/include/exec/cpu-common.h index bd15853..eef74df 100644 --- a/include/exec/cpu-common.h +++ b/include/exec/cpu-common.h @@ -36,6 +36,12 @@ enum device_endian { DEVICE_LITTLE_ENDIAN, }; +#if defined(HOST_WORDS_BIGENDIAN) +#define DEVICE_HOST_ENDIAN DEVICE_BIG_ENDIAN +#else +#define DEVICE_HOST_ENDIAN DEVICE_LITTLE_ENDIAN +#endif + /* address in the RAM (different from a physical address) */ #if defined(CONFIG_XEN_BACKEND) typedef uint64_t ram_addr_t; diff --git a/memory.c b/memory.c index ed8b5aa..17cfada 100644 --- a/memory.c +++ b/memory.c @@ -1180,7 +1180,7 @@ static void memory_region_ram_device_write(void *opaque, hwaddr addr, static const MemoryRegionOps ram_device_mem_ops = { .read = memory_region_ram_device_read, .write = memory_region_ram_device_write, - .endianness = DEVICE_NATIVE_ENDIAN, + .endianness = DEVICE_HOST_ENDIAN, .valid = { .min_access_size = 1, .max_access_size = 8,
At the moment ram device's memory regions are DEVICE_NATIVE_ENDIAN. It's incorrect. This memory region is backed by a MMIO area in host, so the uint64_t data that MemoryRegionOps read from/write to this area should be host-endian rather than target-endian. Hence, current code does not work when target and host endianness are different which is the most common case on PPC64. To fix it, this introduces DEVICE_HOST_ENDIAN for the ram device. This has been tested on PPC64 BE/LE host/guest in all possible combinations including TCG. Suggested-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Yongji Xie <xyjxie@linux.vnet.ibm.com> --- include/exec/cpu-common.h | 6 ++++++ memory.c | 2 +- 2 files changed, 7 insertions(+), 1 deletions(-)