util/oslib-posix.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
When using a memory-backend object with prealloc turned on, QEMU
will memset() the first byte in every memory page to zero. While
this might have been acceptable for memory backends associated
with RAM, this corrupts application data for NVDIMMs.
Instead of setting every page to zero, read the current byte
value and then just write that same value back, so we are not
corrupting the original data.
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
---
I'm unclear if this is actually still safe in practice ? Is the
compiler permitted to optimize away the read+write since it doesn't
change the memory value. I'd hope not, but I've been surprised
before...
IMHO this is another factor in favour of requesting an API from
the kernel to provide the prealloc behaviour we want.
util/oslib-posix.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/util/oslib-posix.c b/util/oslib-posix.c
index 35012b9..8f5b656 100644
--- a/util/oslib-posix.c
+++ b/util/oslib-posix.c
@@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t memory, Error **errp)
/* MAP_POPULATE silently ignores failures */
for (i = 0; i < numpages; i++) {
- memset(area + (hpagesize * i), 0, 1);
+ char val = *(area + (hpagesize * i));
+ memset(area + (hpagesize * i), 0, val);
}
}
--
2.9.3
On 02/23/2017 11:59 AM, Daniel P. Berrange wrote: > When using a memory-backend object with prealloc turned on, QEMU > will memset() the first byte in every memory page to zero. While > this might have been acceptable for memory backends associated > with RAM, this corrupts application data for NVDIMMs. > > Instead of setting every page to zero, read the current byte > value and then just write that same value back, so we are not > corrupting the original data. > > Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > --- > > I'm unclear if this is actually still safe in practice ? Is the > compiler permitted to optimize away the read+write since it doesn't > change the memory value. I'd hope not, but I've been surprised > before... > > IMHO this is another factor in favour of requesting an API from > the kernel to provide the prealloc behaviour we want. > > util/oslib-posix.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/util/oslib-posix.c b/util/oslib-posix.c > index 35012b9..8f5b656 100644 > --- a/util/oslib-posix.c > +++ b/util/oslib-posix.c > @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t memory, Error **errp) > > /* MAP_POPULATE silently ignores failures */ > for (i = 0; i < numpages; i++) { > - memset(area + (hpagesize * i), 0, 1); > + char val = *(area + (hpagesize * i)); > + memset(area + (hpagesize * i), 0, val); I think you wanted: memset(area + (hpagesize * i), val, 1); because what you are suggesting will overwrite even more than the first byte with zeroes. Michal
On Thu, Feb 23, 2017 at 01:05:33PM +0100, Michal Privoznik wrote: > On 02/23/2017 11:59 AM, Daniel P. Berrange wrote: > > When using a memory-backend object with prealloc turned on, QEMU > > will memset() the first byte in every memory page to zero. While > > this might have been acceptable for memory backends associated > > with RAM, this corrupts application data for NVDIMMs. > > > > Instead of setting every page to zero, read the current byte > > value and then just write that same value back, so we are not > > corrupting the original data. > > > > Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > > --- > > > > I'm unclear if this is actually still safe in practice ? Is the > > compiler permitted to optimize away the read+write since it doesn't > > change the memory value. I'd hope not, but I've been surprised > > before... > > > > IMHO this is another factor in favour of requesting an API from > > the kernel to provide the prealloc behaviour we want. > > > > util/oslib-posix.c | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/util/oslib-posix.c b/util/oslib-posix.c > > index 35012b9..8f5b656 100644 > > --- a/util/oslib-posix.c > > +++ b/util/oslib-posix.c > > @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t memory, Error **errp) > > > > /* MAP_POPULATE silently ignores failures */ > > for (i = 0; i < numpages; i++) { > > - memset(area + (hpagesize * i), 0, 1); > > + char val = *(area + (hpagesize * i)); > > + memset(area + (hpagesize * i), 0, val); > > I think you wanted: > > memset(area + (hpagesize * i), val, 1); > > because what you are suggesting will overwrite even more than the first > byte with zeroes. Lol, yes, I'm stupid. Anyway, rather than repost this yet, I'm interested if this is actually the right way to fix the problem or if we should do something totally different.... Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://entangle-photo.org -o- http://search.cpan.org/~danberr/ :|
On 02/23/2017 01:07 PM, Daniel P. Berrange wrote: > On Thu, Feb 23, 2017 at 01:05:33PM +0100, Michal Privoznik wrote: >> On 02/23/2017 11:59 AM, Daniel P. Berrange wrote: >>> When using a memory-backend object with prealloc turned on, QEMU >>> will memset() the first byte in every memory page to zero. While >>> this might have been acceptable for memory backends associated >>> with RAM, this corrupts application data for NVDIMMs. >>> >>> Instead of setting every page to zero, read the current byte >>> value and then just write that same value back, so we are not >>> corrupting the original data. >>> >>> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> >>> --- >>> >>> I'm unclear if this is actually still safe in practice ? Is the >>> compiler permitted to optimize away the read+write since it doesn't >>> change the memory value. I'd hope not, but I've been surprised >>> before... >>> >>> IMHO this is another factor in favour of requesting an API from >>> the kernel to provide the prealloc behaviour we want. >>> >>> util/oslib-posix.c | 3 ++- >>> 1 file changed, 2 insertions(+), 1 deletion(-) >>> >>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c >>> index 35012b9..8f5b656 100644 >>> --- a/util/oslib-posix.c >>> +++ b/util/oslib-posix.c >>> @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t memory, Error **errp) >>> >>> /* MAP_POPULATE silently ignores failures */ >>> for (i = 0; i < numpages; i++) { >>> - memset(area + (hpagesize * i), 0, 1); >>> + char val = *(area + (hpagesize * i)); >>> + memset(area + (hpagesize * i), 0, val); >> >> I think you wanted: >> >> memset(area + (hpagesize * i), val, 1); >> >> because what you are suggesting will overwrite even more than the first >> byte with zeroes. > > Lol, yes, I'm stupid. > > Anyway, rather than repost this yet, I'm interested if this is actually > the right way to fix the problem or if we should do something totally > different.... So, I've done some analysis and if the optimizations are enabled, this whole body is optimized away. Not the loop though. Here's what I've tested: #include <stdio.h> #include <stdlib.h> #include <string.h> int main(int argc, char *argv[]) { int ret = EXIT_FAILURE; unsigned char *ptr; size_t i, j; if (!(ptr = malloc(1024 * 4))) { perror("malloc"); goto cleanup; } for (i = 0; i < 4; i++) { unsigned char val = ptr[i*1024]; memset(ptr + i * 1024, val, 1); } ret = EXIT_SUCCESS; cleanup: free(ptr); return ret; } But if I make @val volatile, I can enforce the compiler to include the body of the loop and actually read and write some bytes. BTW: if I replace memset with *(ptr + i * 1024) = val; and don't make @val volatile even the loop is optimized away. I was compiling with: gcc -S -O3 -o test.S test.c Michal
On Fri, Feb 24, 2017 at 10:05:17AM +0100, Michal Privoznik wrote: > On 02/23/2017 01:07 PM, Daniel P. Berrange wrote: > > On Thu, Feb 23, 2017 at 01:05:33PM +0100, Michal Privoznik wrote: > >> On 02/23/2017 11:59 AM, Daniel P. Berrange wrote: > >>> When using a memory-backend object with prealloc turned on, QEMU > >>> will memset() the first byte in every memory page to zero. While > >>> this might have been acceptable for memory backends associated > >>> with RAM, this corrupts application data for NVDIMMs. > >>> > >>> Instead of setting every page to zero, read the current byte > >>> value and then just write that same value back, so we are not > >>> corrupting the original data. > >>> > >>> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > >>> --- > >>> > >>> I'm unclear if this is actually still safe in practice ? Is the > >>> compiler permitted to optimize away the read+write since it doesn't > >>> change the memory value. I'd hope not, but I've been surprised > >>> before... > >>> > >>> IMHO this is another factor in favour of requesting an API from > >>> the kernel to provide the prealloc behaviour we want. > >>> > >>> util/oslib-posix.c | 3 ++- > >>> 1 file changed, 2 insertions(+), 1 deletion(-) > >>> > >>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c > >>> index 35012b9..8f5b656 100644 > >>> --- a/util/oslib-posix.c > >>> +++ b/util/oslib-posix.c > >>> @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t memory, Error **errp) > >>> > >>> /* MAP_POPULATE silently ignores failures */ > >>> for (i = 0; i < numpages; i++) { > >>> - memset(area + (hpagesize * i), 0, 1); > >>> + char val = *(area + (hpagesize * i)); > >>> + memset(area + (hpagesize * i), 0, val); > >> > >> I think you wanted: > >> > >> memset(area + (hpagesize * i), val, 1); > >> > >> because what you are suggesting will overwrite even more than the first > >> byte with zeroes. > > > > Lol, yes, I'm stupid. > > > > Anyway, rather than repost this yet, I'm interested if this is actually > > the right way to fix the problem or if we should do something totally > > different.... > > So, I've done some analysis and if the optimizations are enabled, this > whole body is optimized away. Not the loop though. Here's what I've tested: > > #include <stdio.h> > #include <stdlib.h> > #include <string.h> > > int main(int argc, char *argv[]) > { > int ret = EXIT_FAILURE; > unsigned char *ptr; > size_t i, j; > > if (!(ptr = malloc(1024 * 4))) { > perror("malloc"); > goto cleanup; > } > > for (i = 0; i < 4; i++) { > unsigned char val = ptr[i*1024]; > memset(ptr + i * 1024, val, 1); > } > > ret = EXIT_SUCCESS; > cleanup: > free(ptr); > return ret; > } > > > But if I make @val volatile, I can enforce the compiler to include the > body of the loop and actually read and write some bytes. BTW: if I > replace memset with *(ptr + i * 1024) = val; and don't make @val > volatile even the loop is optimized away. Ok, yeah, it makes sense that the compiler can optimize that away without volatile. I wonder if adding volatile has much of a performance impact on this loop... Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://entangle-photo.org -o- http://search.cpan.org/~danberr/ :|
* Daniel P. Berrange (berrange@redhat.com) wrote: > On Fri, Feb 24, 2017 at 10:05:17AM +0100, Michal Privoznik wrote: > > On 02/23/2017 01:07 PM, Daniel P. Berrange wrote: > > > On Thu, Feb 23, 2017 at 01:05:33PM +0100, Michal Privoznik wrote: > > >> On 02/23/2017 11:59 AM, Daniel P. Berrange wrote: > > >>> When using a memory-backend object with prealloc turned on, QEMU > > >>> will memset() the first byte in every memory page to zero. While > > >>> this might have been acceptable for memory backends associated > > >>> with RAM, this corrupts application data for NVDIMMs. > > >>> > > >>> Instead of setting every page to zero, read the current byte > > >>> value and then just write that same value back, so we are not > > >>> corrupting the original data. > > >>> > > >>> Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > > >>> --- > > >>> > > >>> I'm unclear if this is actually still safe in practice ? Is the > > >>> compiler permitted to optimize away the read+write since it doesn't > > >>> change the memory value. I'd hope not, but I've been surprised > > >>> before... > > >>> > > >>> IMHO this is another factor in favour of requesting an API from > > >>> the kernel to provide the prealloc behaviour we want. > > >>> > > >>> util/oslib-posix.c | 3 ++- > > >>> 1 file changed, 2 insertions(+), 1 deletion(-) > > >>> > > >>> diff --git a/util/oslib-posix.c b/util/oslib-posix.c > > >>> index 35012b9..8f5b656 100644 > > >>> --- a/util/oslib-posix.c > > >>> +++ b/util/oslib-posix.c > > >>> @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t memory, Error **errp) > > >>> > > >>> /* MAP_POPULATE silently ignores failures */ > > >>> for (i = 0; i < numpages; i++) { > > >>> - memset(area + (hpagesize * i), 0, 1); > > >>> + char val = *(area + (hpagesize * i)); > > >>> + memset(area + (hpagesize * i), 0, val); > > >> > > >> I think you wanted: > > >> > > >> memset(area + (hpagesize * i), val, 1); > > >> > > >> because what you are suggesting will overwrite even more than the first > > >> byte with zeroes. > > > > > > Lol, yes, I'm stupid. > > > > > > Anyway, rather than repost this yet, I'm interested if this is actually > > > the right way to fix the problem or if we should do something totally > > > different.... > > > > So, I've done some analysis and if the optimizations are enabled, this > > whole body is optimized away. Not the loop though. Here's what I've tested: > > > > #include <stdio.h> > > #include <stdlib.h> > > #include <string.h> > > > > int main(int argc, char *argv[]) > > { > > int ret = EXIT_FAILURE; > > unsigned char *ptr; > > size_t i, j; > > > > if (!(ptr = malloc(1024 * 4))) { > > perror("malloc"); > > goto cleanup; > > } > > > > for (i = 0; i < 4; i++) { > > unsigned char val = ptr[i*1024]; > > memset(ptr + i * 1024, val, 1); > > } > > > > ret = EXIT_SUCCESS; > > cleanup: > > free(ptr); > > return ret; > > } > > > > > > But if I make @val volatile, I can enforce the compiler to include the > > body of the loop and actually read and write some bytes. BTW: if I > > replace memset with *(ptr + i * 1024) = val; and don't make @val > > volatile even the loop is optimized away. > > Ok, yeah, it makes sense that the compiler can optimize that away without > volatile. I wonder if adding volatile has much of a performance impact on > this loop... I don't think we have anything else in QEMU to do it (other than atomic's but we don't need this to be atomic). I don't think the use of memset() helps, because the compiler is free to optimise that out as well; so I think 'volatile' is a reasonable use (although I seem to have a soft-spot for volatile and I know everyone else tells me I'm mad). However, note the performance of this loop is important as shown by Jitendra's recent patchset to parallelise it. I wonder if MADV_WILLNEED makes any difference. Dave > Regards, > Daniel > -- > |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| > |: http://libvirt.org -o- http://virt-manager.org :| > |: http://entangle-photo.org -o- http://search.cpan.org/~danberr/ :| > -- Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
On 24/02/2017 13:12, Dr. David Alan Gilbert wrote: >> Ok, yeah, it makes sense that the compiler can optimize that away without >> volatile. I wonder if adding volatile has much of a performance impact on >> this loop... > I don't think we have anything else in QEMU to do it (other than atomic's > but we don't need this to be atomic). I don't think the use of memset() > helps, because the compiler is free to optimise that out as well; so > I think 'volatile' is a reasonable use (although I seem to have a soft-spot > for volatile and I know everyone else tells me I'm mad). Yes, I think it is fine to use volatile here. Paolo
On Thu, Feb 23, 2017 at 10:59:22AM +0000, Daniel P. Berrange wrote: > When using a memory-backend object with prealloc turned on, QEMU > will memset() the first byte in every memory page to zero. While > this might have been acceptable for memory backends associated > with RAM, this corrupts application data for NVDIMMs. > > Instead of setting every page to zero, read the current byte > value and then just write that same value back, so we are not > corrupting the original data. > > Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > --- > > I'm unclear if this is actually still safe in practice ? Is the > compiler permitted to optimize away the read+write since it doesn't > change the memory value. I'd hope not, but I've been surprised > before... > > IMHO this is another factor in favour of requesting an API from > the kernel to provide the prealloc behaviour we want. > > util/oslib-posix.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/util/oslib-posix.c b/util/oslib-posix.c > index 35012b9..8f5b656 100644 > --- a/util/oslib-posix.c > +++ b/util/oslib-posix.c > @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t memory, Error **errp) > > /* MAP_POPULATE silently ignores failures */ > for (i = 0; i < numpages; i++) { > - memset(area + (hpagesize * i), 0, 1); > + char val = *(area + (hpagesize * i)); > + memset(area + (hpagesize * i), 0, val); Please include a comment in the final patch explaining why we want to preserve memory contents. In the case of NVDIMM I'm not sure if the memset is needed at all. The memory already exists - no new pages need to be allocated by the kernel. We just want the page table entries to be populated for the NVDIMM when -mem-prealloc is used. Perhaps Andrea or Rik have ideas on improving the kernel interface and whether mmap(MAP_POPULATE) should be used with NVDIMM instead of this userspace "touch every page" workaround? Stefan
On Mon, 2017-02-27 at 11:10 +0000, Stefan Hajnoczi wrote: > On Thu, Feb 23, 2017 at 10:59:22AM +0000, Daniel P. Berrange wrote: > > When using a memory-backend object with prealloc turned on, QEMU > > will memset() the first byte in every memory page to zero. While > > this might have been acceptable for memory backends associated > > with RAM, this corrupts application data for NVDIMMs. > > > > Instead of setting every page to zero, read the current byte > > value and then just write that same value back, so we are not > > corrupting the original data. > > > > Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > > --- > > > > I'm unclear if this is actually still safe in practice ? Is the > > compiler permitted to optimize away the read+write since it doesn't > > change the memory value. I'd hope not, but I've been surprised > > before... > > > > IMHO this is another factor in favour of requesting an API from > > the kernel to provide the prealloc behaviour we want. > > > > util/oslib-posix.c | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/util/oslib-posix.c b/util/oslib-posix.c > > index 35012b9..8f5b656 100644 > > --- a/util/oslib-posix.c > > +++ b/util/oslib-posix.c > > @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t > > memory, Error **errp) > > > > /* MAP_POPULATE silently ignores failures */ > > for (i = 0; i < numpages; i++) { > > - memset(area + (hpagesize * i), 0, 1); > > + char val = *(area + (hpagesize * i)); > > + memset(area + (hpagesize * i), 0, val); > > Please include a comment in the final patch explaining why we want to > preserve memory contents. > > In the case of NVDIMM I'm not sure if the memset is needed at > all. The > memory already exists - no new pages need to be allocated by the > kernel. > We just want the page table entries to be populated for the NVDIMM > when > -mem-prealloc is used. > > Perhaps Andrea or Rik have ideas on improving the kernel interface > and > whether mmap(MAP_POPULATE) should be used with NVDIMM instead of this > userspace "touch every page" workaround? Why do we need the page table entries to be populated in advance at all? The high cost of the page fault for regular memory is zeroing out the memory pages before we give them to userspace. Simply faulting in the NVDIMM memory as it is touched may make more sense than treating it like DRAM, especially given that with DAX, NVDIMM areas may be orders of magnitude larger than RAM, and we really do not want to set up all the page tables for every part of the guest DAX "disk".
On Mon, Feb 27, 2017 at 08:46:10AM -0500, Rik van Riel wrote: > On Mon, 2017-02-27 at 11:10 +0000, Stefan Hajnoczi wrote: > > On Thu, Feb 23, 2017 at 10:59:22AM +0000, Daniel P. Berrange wrote: > > > When using a memory-backend object with prealloc turned on, QEMU > > > will memset() the first byte in every memory page to zero. While > > > this might have been acceptable for memory backends associated > > > with RAM, this corrupts application data for NVDIMMs. > > > > > > Instead of setting every page to zero, read the current byte > > > value and then just write that same value back, so we are not > > > corrupting the original data. > > > > > > Signed-off-by: Daniel P. Berrange <berrange@redhat.com> > > > --- > > > > > > I'm unclear if this is actually still safe in practice ? Is the > > > compiler permitted to optimize away the read+write since it doesn't > > > change the memory value. I'd hope not, but I've been surprised > > > before... > > > > > > IMHO this is another factor in favour of requesting an API from > > > the kernel to provide the prealloc behaviour we want. > > > > > > util/oslib-posix.c | 3 ++- > > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > > > diff --git a/util/oslib-posix.c b/util/oslib-posix.c > > > index 35012b9..8f5b656 100644 > > > --- a/util/oslib-posix.c > > > +++ b/util/oslib-posix.c > > > @@ -355,7 +355,8 @@ void os_mem_prealloc(int fd, char *area, size_t > > > memory, Error **errp) > > > > > > /* MAP_POPULATE silently ignores failures */ > > > for (i = 0; i < numpages; i++) { > > > - memset(area + (hpagesize * i), 0, 1); > > > + char val = *(area + (hpagesize * i)); > > > + memset(area + (hpagesize * i), 0, val); > > > > Please include a comment in the final patch explaining why we want to > > preserve memory contents. > > > > In the case of NVDIMM I'm not sure if the memset is needed at > > all. The > > memory already exists - no new pages need to be allocated by the > > kernel. > > We just want the page table entries to be populated for the NVDIMM > > when > > -mem-prealloc is used. > > > > Perhaps Andrea or Rik have ideas on improving the kernel interface > > and > > whether mmap(MAP_POPULATE) should be used with NVDIMM instead of this > > userspace "touch every page" workaround? > > Why do we need the page table entries to be populated > in advance at all? It is a choice apps using QEMU have - they can tell QEMU whether they want prealloc or not - if they decide they do want it, then we should not be corrupting the data. > The high cost of the page fault for regular memory > is zeroing out the memory pages before we give them > to userspace. NVDIMM in the guest might be backed by regular memory in the host - QEMU doesn't require use of NVDIMM in the host. > Simply faulting in the NVDIMM memory as it is touched > may make more sense than treating it like DRAM, > especially given that with DAX, NVDIMM areas may be > orders of magnitude larger than RAM, and we really > do not want to set up all the page tables for every > part of the guest DAX "disk". Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://entangle-photo.org -o- http://search.cpan.org/~danberr/ :|
© 2016 - 2024 Red Hat, Inc.