[libvirt] [PATCH] conf: use proper maximum for parsing memory values
The value is stored in unsigned long long, so ULLONG_MAX is the proper upper limit to use. Signed-off-by: Martin Kletzander <mkletzan@redhat.com> --- Even though this is a build-breaker (for 32bit systems memtune-unlimited fails to parse), I'm not pushing it as one because it feels odd that such change wouldn't break anything else. src/conf/domain_conf.c | 12 +++--------- 1 file changed, 3 insertions(+), 9 deletions(-) diff --git a/src/conf/domain_conf.c b/src/conf/domain_conf.c index a351382..beb3d26 100644 --- a/src/conf/domain_conf.c +++ b/src/conf/domain_conf.c @@ -6388,16 +6388,10 @@ virDomainParseMemory(const char *xpath, xmlXPathContextPtr ctxt, unsigned long long *mem, bool required) { int ret = -1; - unsigned long long bytes, max; - - /* On 32-bit machines, our bound is 0xffffffff * KiB. On 64-bit - * machines, our bound is off_t (2^63). */ - if (sizeof(unsigned long) < sizeof(long long)) - max = 1024ull * ULONG_MAX; - else - max = LLONG_MAX; + unsigned long long bytes; - ret = virDomainParseScaledValue(xpath, ctxt, &bytes, 1024, max, required); + ret = virDomainParseScaledValue(xpath, ctxt, &bytes, 1024, + ULLONG_MAX, required); if (ret < 0) goto cleanup; -- 2.1.2
On 10/30/2014 09:49 AM, Martin Kletzander wrote:
The value is stored in unsigned long long, so ULLONG_MAX is the proper upper limit to use.
No, it's not.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com> --- Even though this is a build-breaker (for 32bit systems memtune-unlimited fails to parse), I'm not pushing it as one because it feels odd that such change wouldn't break anything else.
What's the compile failure? This patch is intentionally trying to fit the largest value that will fit in an unsigned long when scaled by kilobytes, while still parsing by bytes. Thus, on 64-bit machines, you can parse 0xffffffffffffffff bytes, then scale down by 1024 and store in unsigned long; on 32-bit machines, your maximum has to be limited to 0xffffffff*1024 bytes before scaling back down.
src/conf/domain_conf.c | 12 +++--------- 1 file changed, 3 insertions(+), 9 deletions(-)
diff --git a/src/conf/domain_conf.c b/src/conf/domain_conf.c index a351382..beb3d26 100644 --- a/src/conf/domain_conf.c +++ b/src/conf/domain_conf.c @@ -6388,16 +6388,10 @@ virDomainParseMemory(const char *xpath, xmlXPathContextPtr ctxt, unsigned long long *mem, bool required) { int ret = -1; - unsigned long long bytes, max; - - /* On 32-bit machines, our bound is 0xffffffff * KiB. On 64-bit - * machines, our bound is off_t (2^63). */ - if (sizeof(unsigned long) < sizeof(long long)) - max = 1024ull * ULONG_MAX; - else - max = LLONG_MAX; + unsigned long long bytes;
- ret = virDomainParseScaledValue(xpath, ctxt, &bytes, 1024, max, required); + ret = virDomainParseScaledValue(xpath, ctxt, &bytes, 1024, + ULLONG_MAX, required);
NACK. I need to see the compiler failure to see the proper fix, but this fix is not going to work. -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
On 10/30/2014 09:55 AM, Eric Blake wrote:
On 10/30/2014 09:49 AM, Martin Kletzander wrote:
The value is stored in unsigned long long, so ULLONG_MAX is the proper upper limit to use.
No, it's not.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com> --- Even though this is a build-breaker (for 32bit systems memtune-unlimited fails to parse), I'm not pushing it as one because it feels odd that such change wouldn't break anything else.
What's the compile failure? This patch is intentionally trying to fit the largest value that will fit in an unsigned long when scaled by kilobytes, while still parsing by bytes. Thus, on 64-bit machines, you can parse 0xffffffffffffffff bytes, then scale down by 1024 and store in unsigned long; on 32-bit machines, your maximum has to be limited to 0xffffffff*1024 bytes before scaling back down.
Sounds like the real bug is in the memtune-unlimited test. 2^53-1 is the maximum memory for a 64-bit host, but invalid on a 32-bit host. Any interface that uses 'unsigned long' as its capping point has to have different maximum limits on 32-bit hosts. -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
On 10/30/2014 10:09 AM, Eric Blake wrote:
On 10/30/2014 09:55 AM, Eric Blake wrote:
On 10/30/2014 09:49 AM, Martin Kletzander wrote:
The value is stored in unsigned long long, so ULLONG_MAX is the proper upper limit to use.
No, it's not.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com> --- Even though this is a build-breaker (for 32bit systems memtune-unlimited fails to parse), I'm not pushing it as one because it feels odd that such change wouldn't break anything else.
What's the compile failure? This patch is intentionally trying to fit the largest value that will fit in an unsigned long when scaled by kilobytes, while still parsing by bytes. Thus, on 64-bit machines, you can parse 0xffffffffffffffff bytes, then scale down by 1024 and store in unsigned long; on 32-bit machines, your maximum has to be limited to 0xffffffff*1024 bytes before scaling back down.
Sounds like the real bug is in the memtune-unlimited test. 2^53-1 is the maximum memory for a 64-bit host, but invalid on a 32-bit host. Any interface that uses 'unsigned long' as its capping point has to have different maximum limits on 32-bit hosts.
Or maybe the problem is that at some point we used unsigned long, and later moved to unsigned long long, but never updated the comment? I'm trying to investigate the history of this code... -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
On 10/30/2014 10:23 AM, Eric Blake wrote:
Or maybe the problem is that at some point we used unsigned long, and later moved to unsigned long long, but never updated the comment? I'm trying to investigate the history of this code...
Looking a bit deeper, commit 4888f0fb5 was where we changed all internal tracking to use unsigned long long; but we still have the public API virDomain{Get,Set}MaxMemory that are tied to 'unsigned long'. I guess what I was trying to do in 2e22f23 (both in 2012) was to ensure that virDomainGetMaxMemory would never have to fail due to overflow, by capping the input to something that would fit in the output. And we still got it wrong then, since commit 19e7c04 (in 2013) fixed a case where 32-bit hosts were truncating incorrectly. Meanwhile, it looks like virDomain{Get,Set}MemoryParamters have ALWAYS used ullong. But this interface doesn't report max memory - so we still have to live with the fact that there is currently no interface for the user getting at max memory other than crippled API. Maybe the solution is to enhance virDomain{Get,Set}MemoryParameters to be a superset of the older APIs of virDomain{Get,Set}MaxMemory, and teach the older APIs to fail with VIR_ERR_OVERFLOW when the value can only be passed through the newer API. Or maybe the solution is to make virDomainParseMemory add a bool parameter that says whether the caller is old-style (32-bit cap) or new-style (full 64-bit width), and set the maximum according to that information. That's probably the easiest for doing right now. I can take a stab at it, since it was my commit in 2012 that originally introduced the weird 32-bit cap even when parsing a 64-bit field. -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
On 10/30/2014 10:51 AM, Eric Blake wrote:
On 10/30/2014 10:23 AM, Eric Blake wrote:
Or maybe the problem is that at some point we used unsigned long, and later moved to unsigned long long, but never updated the comment? I'm trying to investigate the history of this code...
Or maybe the solution is to make virDomainParseMemory add a bool parameter that says whether the caller is old-style (32-bit cap) or new-style (full 64-bit width), and set the maximum according to that information. That's probably the easiest for doing right now. I can take a stab at it, since it was my commit in 2012 that originally introduced the weird 32-bit cap even when parsing a 64-bit field.
I went with this idea: https://www.redhat.com/archives/libvir-list/2014-October/msg01084.html -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
participants (2)
-
Eric Blake -
Martin Kletzander