[libvirt] [RFC 0/4] Report disk latency information
The QEMU side patches are not posted yet, but considering what these patche series need is only a fixed interface provided by QEMU, posting it for early reviewing, won't push it before QEMU patches are got pushed even it got ACK. Osier
The new disk latency related members include: rd_total_times wr_total_times flush_operations flush_total_times --- include/libvirt/libvirt.h.in | 4 ++++ 1 files changed, 4 insertions(+), 0 deletions(-) diff --git a/include/libvirt/libvirt.h.in b/include/libvirt/libvirt.h.in index b1bda31..cf6bf60 100644 --- a/include/libvirt/libvirt.h.in +++ b/include/libvirt/libvirt.h.in @@ -562,8 +562,12 @@ typedef struct _virDomainBlockStats virDomainBlockStatsStruct; struct _virDomainBlockStats { long long rd_req; /* number of read requests */ long long rd_bytes; /* number of read bytes */ + long long rd_total_times; /* total time spend on cache reads in nano-seconds */ long long wr_req; /* number of write requests */ long long wr_bytes; /* number of written bytes */ + long long wr_total_times; /* total time spend on cache writes in nano-seconds */ + long long flush_req; /* number of flushed bytes */ + long long flush_total_times; /* total time spend on cache flushes in nano-seconds */ long long errs; /* In Xen this returns the mysterious 'oo_req'. */ }; -- 1.7.6
On Fri, Aug 12, 2011 at 10:17:24PM +0800, Osier Yang wrote:
The new disk latency related members include: rd_total_times wr_total_times flush_operations flush_total_times --- include/libvirt/libvirt.h.in | 4 ++++ 1 files changed, 4 insertions(+), 0 deletions(-)
diff --git a/include/libvirt/libvirt.h.in b/include/libvirt/libvirt.h.in index b1bda31..cf6bf60 100644 --- a/include/libvirt/libvirt.h.in +++ b/include/libvirt/libvirt.h.in @@ -562,8 +562,12 @@ typedef struct _virDomainBlockStats virDomainBlockStatsStruct; struct _virDomainBlockStats { long long rd_req; /* number of read requests */ long long rd_bytes; /* number of read bytes */ + long long rd_total_times; /* total time spend on cache reads in nano-seconds */ long long wr_req; /* number of write requests */ long long wr_bytes; /* number of written bytes */ + long long wr_total_times; /* total time spend on cache writes in nano-seconds */ + long long flush_req; /* number of flushed bytes */ + long long flush_total_times; /* total time spend on cache flushes in nano-seconds */ long long errs; /* In Xen this returns the mysterious 'oo_req'. */ };
NACK, it is forbidden to modify any structs in the public header files since that changes the ABI. Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|
On 08/12/2011 07:53 AM, Daniel P. Berrange wrote:
On Fri, Aug 12, 2011 at 10:17:24PM +0800, Osier Yang wrote:
The new disk latency related members include: rd_total_times wr_total_times flush_operations flush_total_times --- include/libvirt/libvirt.h.in | 4 ++++ 1 files changed, 4 insertions(+), 0 deletions(-)
diff --git a/include/libvirt/libvirt.h.in b/include/libvirt/libvirt.h.in index b1bda31..cf6bf60 100644 --- a/include/libvirt/libvirt.h.in +++ b/include/libvirt/libvirt.h.in @@ -562,8 +562,12 @@ typedef struct _virDomainBlockStats virDomainBlockStatsStruct; struct _virDomainBlockStats { long long rd_req; /* number of read requests */ long long rd_bytes; /* number of read bytes */ + long long rd_total_times; /* total time spend on cache reads in nano-seconds */ long long wr_req; /* number of write requests */ long long wr_bytes; /* number of written bytes */ + long long wr_total_times; /* total time spend on cache writes in nano-seconds */ + long long flush_req; /* number of flushed bytes */ + long long flush_total_times; /* total time spend on cache flushes in nano-seconds */ long long errs; /* In Xen this returns the mysterious 'oo_req'. */ };
NACK, it is forbidden to modify any structs in the public header files since that changes the ABI.
However, what we CAN do, without breaking ABI, is the following: In libvirt.h.in: add a new struct virDomainBlockStatsStructV1 with layout identical to the old virDomainBlockStatsStruct. add a new typedef virDomainBlockStatsStructV2 that maps to the modified virDomainBlockStatsStruct. modify struct virDomainBlockStatsStruct to add new fields, but _only_ at the end of the struct - this struct could later grow if we need to go to a v3. Leave virDomainBlockStatsPtr as a typedef for *virDomainBlockStatsStruct. In libvirt.c: int virDomainBlockStats (virDomainPtr dom, const char *path, virDomainBlockStatsPtr stats, size_t size) { virConnectPtr conn; struct _virDomainBlockStatsStructV1 statsv1 = { -1, -1, -1, -1, -1 }; struct _virDomainBlockStatsStructV2 statsv2 = { -1, -1, -1, -1, -1, -1 ... }; VIR_DOMAIN_DEBUG ... if (!path || !stats || size > sizeof(statsv2)) { virLibDomainError(VIR_ERR_INVALID_ARG, __FUNCTION__); goto error; } conn = dom->conn; if (conn->driver->domainBlockStatsv2) { if (conn->driver->domainBlockStatsv2 (dom, path, &statsv2) == -1) goto error; memcpy (stats, &statsv2, size); return 0; } if (conn->driver->domainBlockStats) { if (conn->driver->domainBlockStats (dom, path, &statsv1) == -1) goto error; if (size > sizeof(statsv1)) { memset(((char *)stats)+sizeof(statsv1), -1, size - sizeof(statsv1)); size = sizeof(statsv1); } memcpy (stats, &statsv1, size); return 0; } ... In driver.h: Update the domainBlockStats callback to explicitly take *virDomainBlockStatsStructV1, and add a new domainBlockStats2 callback that explicitly takes *virDomainBlockStatsStructV2. In remote_protocol.x, add a new RPC for domain_block_stats_2 that passes the larger struct over the wire; leaving the existing RPC tied to the smaller v1 struct. In qemu_driver.c, implement the new driver callback as the only way to get at the additional fields. This proposal is ABI safe (no function signatures changed and no new entry points are added - all the decisions on how many bytes to populate are based on an existing argument, and all APIs used pointers which are constant size rather than copies where the larger struct would cause problems). However, it has the following effects, including one slight API ramification: 1. all existing clients that were pre-compiled against the older libvirt.h should have already been calling virDomainBlockStats(...,sizeof(virDomainBlockStatsStruct)), where the compilation was done with the struct at the smaller size. This will safely call the older RPC, and skip out on any new fields, whether talking to an old or new server. 2. recompiling an existing client against newer libvirt.h with no other changes will automatically pick up the larger struct. The sizeof() argument will change. Newer servers will recognize the larger struct and automatically call the newer RPC, getting all the new fields. However, older servers will now reject the larger sizeof as too large - this is a subtle limitation, if clients are recompiled without being aware of the difference it will cause! 3. recompiling an existing client with a slight change (s/virDomainBlockStatsPtr/virDomainBlockStatsStructV1Ptr/g) will force the smaller sizeof(), so that the client can continue to talk to older servers, but will lose out on the new fields in the v2 struct. This is the mitigation to use to avoid the breakage in point 2. 4. writing a new client that only has to target 0.9.5 or newer should explicitly use virDomainBlockStatsStructV2Ptr, so that the client will safely talk to newer servers and gracefully be rejected by older servers. And by explicitly using the v2 struct instead of the generic virDomainBlockStatsPtr, the client has protected itself from a recompilation to a future v3 of virDomainBlockStatsPtr causing the 0.9.5 support to stop working (that is, by exposing all three types: V1, V2, and generic, an application can decide whether recompilation should be tied to a specific struct level, or to the latest version available). If we decide that point 2 is too onerous (that is, if we want to guarantee that recompilation of existing clients against newer libvirt.h will not break the ability of that client to talk to older servers), then we can slightly tweak the changes made in libvirt.h.in so as to keep virDomainBlockStatsPtr tied to the v1 struct size, and only use new type names for larger structs. Next question is what additional fields to provide? We don't want to make a v3 any sooner than needed, so v2 should have more than just latency, but should also include things like cdrom tray status. Remember that the public API documents that all stats not supported for a given drive by the given hypervisor will be set to -1, so it doesn't hurt to add more fields than what can be populated, so long as at least one combination of hypervisor and underlying block type can usefully use that field. -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
On Wed, Aug 17, 2011 at 09:39:34AM -0600, Eric Blake wrote:
On 08/12/2011 07:53 AM, Daniel P. Berrange wrote:
On Fri, Aug 12, 2011 at 10:17:24PM +0800, Osier Yang wrote:
The new disk latency related members include: rd_total_times wr_total_times flush_operations flush_total_times --- include/libvirt/libvirt.h.in | 4 ++++ 1 files changed, 4 insertions(+), 0 deletions(-)
diff --git a/include/libvirt/libvirt.h.in b/include/libvirt/libvirt.h.in index b1bda31..cf6bf60 100644 --- a/include/libvirt/libvirt.h.in +++ b/include/libvirt/libvirt.h.in @@ -562,8 +562,12 @@ typedef struct _virDomainBlockStats virDomainBlockStatsStruct; struct _virDomainBlockStats { long long rd_req; /* number of read requests */ long long rd_bytes; /* number of read bytes */ + long long rd_total_times; /* total time spend on cache reads in nano-seconds */ long long wr_req; /* number of write requests */ long long wr_bytes; /* number of written bytes */ + long long wr_total_times; /* total time spend on cache writes in nano-seconds */ + long long flush_req; /* number of flushed bytes */ + long long flush_total_times; /* total time spend on cache flushes in nano-seconds */ long long errs; /* In Xen this returns the mysterious 'oo_req'. */ };
NACK, it is forbidden to modify any structs in the public header files since that changes the ABI.
However, what we CAN do, without breaking ABI, is the following:
In libvirt.h.in:
add a new struct virDomainBlockStatsStructV1 with layout identical to the old virDomainBlockStatsStruct.
add a new typedef virDomainBlockStatsStructV2 that maps to the modified virDomainBlockStatsStruct.
modify struct virDomainBlockStatsStruct to add new fields, but _only_ at the end of the struct - this struct could later grow if we need to go to a v3.
I would be tempted to keep virDomainBlockStatsStruct as-is, just add virDomainBlockStatsStructV2 as the larger one. That keeps the default the small structure, and minimize the risk of incompatibilities.
Leave virDomainBlockStatsPtr as a typedef for *virDomainBlockStatsStruct.
In libvirt.c:
int virDomainBlockStats (virDomainPtr dom, const char *path, virDomainBlockStatsPtr stats, size_t size) { virConnectPtr conn; struct _virDomainBlockStatsStructV1 statsv1 = { -1, -1, -1, -1, -1 };
just struct _virDomainBlockStatsStruct statsv1 = ...
struct _virDomainBlockStatsStructV2 statsv2 = { -1, -1, -1, -1, -1, -1 ... };
VIR_DOMAIN_DEBUG ... if (!path || !stats || size > sizeof(statsv2)) { virLibDomainError(VIR_ERR_INVALID_ARG, __FUNCTION__); goto error; } conn = dom->conn;
if (conn->driver->domainBlockStatsv2) { if (conn->driver->domainBlockStatsv2 (dom, path, &statsv2) == -1) goto error;
memcpy (stats, &statsv2, size); return 0; } if (conn->driver->domainBlockStats) { if (conn->driver->domainBlockStats (dom, path, &statsv1) == -1) goto error;
if (size > sizeof(statsv1)) { memset(((char *)stats)+sizeof(statsv1), -1, size - sizeof(statsv1)); size = sizeof(statsv1); } memcpy (stats, &statsv1, size); return 0; } ...
In driver.h:
Update the domainBlockStats callback to explicitly take *virDomainBlockStatsStructV1, and add a new domainBlockStats2 callback that explicitly takes *virDomainBlockStatsStructV2.
I would not change the old entry point, just add the new one
In remote_protocol.x, add a new RPC for domain_block_stats_2 that passes the larger struct over the wire; leaving the existing RPC tied to the smaller v1 struct.
agreed
In qemu_driver.c, implement the new driver callback as the only way to get at the additional fields.
agreed too
This proposal is ABI safe (no function signatures changed and no new entry points are added - all the decisions on how many bytes to populate are based on an existing argument, and all APIs used pointers which are constant size rather than copies where the larger struct would cause problems). However, it has the following effects, including one slight API ramification:
Well it does change the signature, in a compatible way ... but. Which is why I would drop the notion of a V1 and just add a V2 structure (and associated pointers).
1. all existing clients that were pre-compiled against the older libvirt.h should have already been calling virDomainBlockStats(...,sizeof(virDomainBlockStatsStruct)), where the compilation was done with the struct at the smaller size. This will safely call the older RPC, and skip out on any new fields, whether talking to an old or new server.
2. recompiling an existing client against newer libvirt.h with no other changes will automatically pick up the larger struct. The sizeof() argument will change. Newer servers will recognize the larger struct and automatically call the newer RPC, getting all the new fields. However, older servers will now reject the larger sizeof as too large - this is a subtle limitation, if clients are recompiled without being aware of the difference it will cause!
that I'm not sure about being the right way
3. recompiling an existing client with a slight change (s/virDomainBlockStatsPtr/virDomainBlockStatsStructV1Ptr/g) will force the smaller sizeof(), so that the client can continue to talk to older servers, but will lose out on the new fields in the v2 struct. This is the mitigation to use to avoid the breakage in point 2.
Agreed
4. writing a new client that only has to target 0.9.5 or newer should explicitly use virDomainBlockStatsStructV2Ptr, so that the client will safely talk to newer servers and gracefully be rejected by older servers. And by explicitly using the v2 struct instead of the generic virDomainBlockStatsPtr, the client has protected itself from a recompilation to a future v3 of virDomainBlockStatsPtr causing the 0.9.5 support to stop working (that is, by exposing all three types: V1, V2, and generic, an application can decide whether recompilation should be tied to a specific struct level, or to the latest version available).
Agreed
If we decide that point 2 is too onerous (that is, if we want to guarantee that recompilation of existing clients against newer libvirt.h will not break the ability of that client to talk to older servers), then we can slightly tweak the changes made in libvirt.h.in so as to keep virDomainBlockStatsPtr tied to the v1 struct size, and only use new type names for larger structs.
yeah I would keep V1 the default and avoid 2/ There is a small problem in this scheme though: the python bindings they don't take a structure as a parameter so there is no way to distinguish, I would just explicitely get a class virDomain: [...] def blockStatsV2(self, path): """Extracts extended block device statistics for a domain """ ret = libvirtmod.virDomainBlockStatsV2(self._o, path) if ret is None: raise libvirtError ('virDomainBlockStats() failed', dom=self) return ret from the generator, and in python/libvirt-override.c provide an explicit entry point libvirt_virDomainBlockStatsV2() using virDomainBlockStatsStructV2 statsv2; and using that for the call That is API clean except for python but I don't think it's a problem since it's limited to the interaction between libvirtmod.so and libvirt.py so fairly well defined. The python client wanting the extra data will need to call the new method but that's Okay I think
Next question is what additional fields to provide? We don't want to make a v3 any sooner than needed, so v2 should have more than just latency, but should also include things like cdrom tray status. Remember that the public API documents that all stats not supported for a given drive by the given hypervisor will be set to -1, so it doesn't hurt to add more fields than what can be populated, so long as at least one combination of hypervisor and underlying block type can usefully use that field.
I'm tempted to add some provision for extra data by packing an long long unused [20] at the end of the struct, that we could use later once we find the need for it without requiring protocol or structure extensions, Daniel -- Daniel Veillard | libxml Gnome XML XSLT toolkit http://xmlsoft.org/ daniel@veillard.com | Rpmfind RPM search engine http://rpmfind.net/ http://veillard.com/ | virtualization library http://libvirt.org/
On Wed, Aug 17, 2011 at 09:39:34AM -0600, Eric Blake wrote:
On 08/12/2011 07:53 AM, Daniel P. Berrange wrote:
On Fri, Aug 12, 2011 at 10:17:24PM +0800, Osier Yang wrote:
The new disk latency related members include: rd_total_times wr_total_times flush_operations flush_total_times --- include/libvirt/libvirt.h.in | 4 ++++ 1 files changed, 4 insertions(+), 0 deletions(-)
diff --git a/include/libvirt/libvirt.h.in b/include/libvirt/libvirt.h.in index b1bda31..cf6bf60 100644 --- a/include/libvirt/libvirt.h.in +++ b/include/libvirt/libvirt.h.in @@ -562,8 +562,12 @@ typedef struct _virDomainBlockStats virDomainBlockStatsStruct; struct _virDomainBlockStats { long long rd_req; /* number of read requests */ long long rd_bytes; /* number of read bytes */ + long long rd_total_times; /* total time spend on cache reads in nano-seconds */ long long wr_req; /* number of write requests */ long long wr_bytes; /* number of written bytes */ + long long wr_total_times; /* total time spend on cache writes in nano-seconds */ + long long flush_req; /* number of flushed bytes */ + long long flush_total_times; /* total time spend on cache flushes in nano-seconds */ long long errs; /* In Xen this returns the mysterious 'oo_req'. */ };
NACK, it is forbidden to modify any structs in the public header files since that changes the ABI.
However, what we CAN do, without breaking ABI, is the following:
In libvirt.h.in:
add a new struct virDomainBlockStatsStructV1 with layout identical to the old virDomainBlockStatsStruct.
add a new typedef virDomainBlockStatsStructV2 that maps to the modified virDomainBlockStatsStruct.
modify struct virDomainBlockStatsStruct to add new fields, but _only_ at the end of the struct - this struct could later grow if we need to go to a v3.
Leave virDomainBlockStatsPtr as a typedef for *virDomainBlockStatsStruct.
In libvirt.c:
int virDomainBlockStats (virDomainPtr dom, const char *path, virDomainBlockStatsPtr stats, size_t size) { virConnectPtr conn; struct _virDomainBlockStatsStructV1 statsv1 = { -1, -1, -1, -1, -1 }; struct _virDomainBlockStatsStructV2 statsv2 = { -1, -1, -1, -1, -1, -1 ... };
VIR_DOMAIN_DEBUG ... if (!path || !stats || size > sizeof(statsv2)) { virLibDomainError(VIR_ERR_INVALID_ARG, __FUNCTION__); goto error; } conn = dom->conn;
if (conn->driver->domainBlockStatsv2) { if (conn->driver->domainBlockStatsv2 (dom, path, &statsv2) == -1) goto error;
memcpy (stats, &statsv2, size); return 0; } if (conn->driver->domainBlockStats) { if (conn->driver->domainBlockStats (dom, path, &statsv1) == -1) goto error;
if (size > sizeof(statsv1)) { memset(((char *)stats)+sizeof(statsv1), -1, size - sizeof(statsv1)); size = sizeof(statsv1); } memcpy (stats, &statsv1, size); return 0; } ...
In driver.h:
Update the domainBlockStats callback to explicitly take *virDomainBlockStatsStructV1, and add a new domainBlockStats2 callback that explicitly takes *virDomainBlockStatsStructV2.
In remote_protocol.x, add a new RPC for domain_block_stats_2 that passes the larger struct over the wire; leaving the existing RPC tied to the smaller v1 struct.
In qemu_driver.c, implement the new driver callback as the only way to get at the additional fields.
This proposal is ABI safe (no function signatures changed and no new entry points are added - all the decisions on how many bytes to populate are based on an existing argument, and all APIs used pointers which are constant size rather than copies where the larger struct would cause problems). However, it has the following effects, including one slight API ramification:
1. all existing clients that were pre-compiled against the older libvirt.h should have already been calling virDomainBlockStats(...,sizeof(virDomainBlockStatsStruct)), where the compilation was done with the struct at the smaller size. This will safely call the older RPC, and skip out on any new fields, whether talking to an old or new server.
2. recompiling an existing client against newer libvirt.h with no other changes will automatically pick up the larger struct. The sizeof() argument will change. Newer servers will recognize the larger struct and automatically call the newer RPC, getting all the new fields. However, older servers will now reject the larger sizeof as too large - this is a subtle limitation, if clients are recompiled without being aware of the difference it will cause!
That last point is a significant problem. While we have technically preserved "ABI" compatibility, we have not preserved actual real world application API usage compatibility. I don't really like the idea of having this kind of special case functionality in just one of our APIs, where our current policy is always to add new APIs. While this is a no doubt a very neat trick in terms of ABI compatibility, I think we should go for clarity & simplicity. Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|
On 08/17/2011 10:10 AM, Daniel P. Berrange wrote:
2. recompiling an existing client against newer libvirt.h with no other changes will automatically pick up the larger struct. The sizeof() argument will change. Newer servers will recognize the larger struct and automatically call the newer RPC, getting all the new fields. However, older servers will now reject the larger sizeof as too large - this is a subtle limitation, if clients are recompiled without being aware of the difference it will cause!
That last point is a significant problem. While we have technically preserved "ABI" compatibility, we have not preserved actual real world application API usage compatibility.
I don't really like the idea of having this kind of special case functionality in just one of our APIs, where our current policy is always to add new APIs. While this is a no doubt a very neat trick in terms of ABI compatibility, I think we should go for clarity& simplicity.
OK, both DV and you have convinced me that API compatibility is important, and that we must not get in the situation of point 2 - that is, a solution is viable only if any application that uses the type virDomainBlockStatsPtr is always referring to the v1 struct size, and the only way to get the larger struct is to use a new struct name. But I'm unclear from your answer whether you are opposed to the overall idea of reusing virDomainBlockStats() with the size argument as the key on which struct was passed in (in which case, we need to rethink everything about my proposal - and rebasing to a newer .so is the only way to take advantage of the new fields), or just the idea of an API break by how I first described the changes (in which case I am confident that we can tweak libvirt.h.in to guarantee a naming choice that preserves API compatibility across recompilation, while most of my proposal stands as-is where it remains possible to backport the new fields to older libvirtd installations since no .so entry points are changed). DV also seemed to be in favor of reusing the existing function call with a new struct, as long as there is no API break for existing code. And personally, I see value in being able to add a new struct without requiring a rebase to a new ABI. I just audited libvirt.c, and found that the following libvirt interfaces that take a size parameter, and could be extensible by use of the size field: virDomainBlockStats virDomainInterfaceStats The following libvirt interfaces take a flags parameter, and could be extensible by setting a flag bit that says a larger struct is in use: virDomainGetControlInfo virNodeGetCPUStats virNodeGetMemoryStats virDomainMemoryStats virDomainGetBlockInfo virDomainGetBlockJobInfo Meanwhile, the following APIs which take a pointer to a struct are non-extensible (neither size nor flags as an escape hatch to indicate explicit use of a larger struct), extending any of these would require a new API function: virDomainGetInfo virNodeGetInfo virDomainGetVcpus virDomainGetSecurityLabel virDomainGetSecurityModel virStoragePoolGetInfo virStorageVolGetInfo virDomainGetJobInfo As a final question for this mail, if we decide that reusing the existing ABI with a new struct size is not desirable, then would you rather that the new virDomainBlockStats2 API (or whatever we name the new function) would also take a (non-extensible) struct, or would it be better to rewrite it in terms of 'virTypedParameterPtr params, int *nparams' so that all future extensibility is covered by adding name-type-value tuples rather than dealing with future struct sizing? I'd rather go through the pain of an ABI addition at most once, rather than once per time that we identify new fields to add. -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
On 08/17/2011 12:15 PM, Eric Blake wrote:
I just audited libvirt.c, and found that the following libvirt interfaces that take a size parameter, and could be extensible by use of the size field: virDomainBlockStats virDomainInterfaceStats
The following libvirt interfaces take a flags parameter, and could be extensible by setting a flag bit that says a larger struct is in use: virDomainGetControlInfo virNodeGetCPUStats virNodeGetMemoryStats virDomainMemoryStats
virDomainMemoryStats() was designed to be regularly extended by using the 'nr_stats' parameter. When I added this API, we had this discusson and the design emerged specifically for the kind of extension being discussed here. If a newer client talks to an older server, the effect is that the client asks for more stats than are available. The old server simply returns the stats it knows about. Wouldn't this be the best approach to use for block stats as well? Am I missing a downside to the virDomainMemoryStats() approach?
virDomainGetBlockInfo virDomainGetBlockJobInfo
Meanwhile, the following APIs which take a pointer to a struct are non-extensible (neither size nor flags as an escape hatch to indicate explicit use of a larger struct), extending any of these would require a new API function: virDomainGetInfo virNodeGetInfo virDomainGetVcpus virDomainGetSecurityLabel virDomainGetSecurityModel virStoragePoolGetInfo virStorageVolGetInfo virDomainGetJobInfo
-- Adam Litke IBM Linux Technology Center
On 08/18/2011 08:40 AM, Adam Litke wrote:
The following libvirt interfaces take a flags parameter, and could be extensible by setting a flag bit that says a larger struct is in use: virDomainGetControlInfo virNodeGetCPUStats virNodeGetMemoryStats virDomainMemoryStats
virDomainMemoryStats() was designed to be regularly extended by using the 'nr_stats' parameter.
Oh, right. In my audit, I forgot that virNodeMemoryStats is a subset of virTypedParameter (it's hard-coded to long long stats only); any call that uses the virTypedParameter paradigm is extensible by adding more name-value (or name-type-value) tuples. Still, virNodeGetMemoryStats has a flags argument, so if we ever come up with a memory stat that is better represented as a different type than long long, we could use the flag bit to state that virTypedParameter is in use instead.
When I added this API, we had this discusson and the design emerged specifically for the kind of extension being discussed here. If a newer client talks to an older server, the effect is that the client asks for more stats than are available. The old server simply returns the stats it knows about. Wouldn't this be the best approach to use for block stats as well? Am I missing a downside to the virDomainMemoryStats() approach?
No, other than the fact that it is using its own name-value pair instead of reusing the more generic name-type-value tuple of virTypedParameter, but I recall that we discussed that at the time and agreed to use of a less generic type. And your question is precisely the tradeoff we're discussing for virDomainBlockStats - do we reuse the function call as is with a larger size, or do we add a new function? And if we add a new function, do we follow the virTypedParameter/virNodeMemoryStats paradigm of arbitrarily long name-[type-]value listings, or do we follow the paradigm of passing back a struct with every known value? -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
于 2011年08月18日 01:15, Eric Blake 写道:
On 08/17/2011 10:10 AM, Daniel P. Berrange wrote:
2. recompiling an existing client against newer libvirt.h with no other changes will automatically pick up the larger struct. The sizeof() argument will change. Newer servers will recognize the larger struct and automatically call the newer RPC, getting all the new fields. However, older servers will now reject the larger sizeof as too large - this is a subtle limitation, if clients are recompiled without being aware of the difference it will cause!
That last point is a significant problem. While we have technically preserved "ABI" compatibility, we have not preserved actual real world application API usage compatibility.
I don't really like the idea of having this kind of special case functionality in just one of our APIs, where our current policy is always to add new APIs. While this is a no doubt a very neat trick in terms of ABI compatibility, I think we should go for clarity& simplicity.
OK, both DV and you have convinced me that API compatibility is important, and that we must not get in the situation of point 2 - that is, a solution is viable only if any application that uses the type virDomainBlockStatsPtr is always referring to the v1 struct size, and the only way to get the larger struct is to use a new struct name.
But I'm unclear from your answer whether you are opposed to the overall idea of reusing virDomainBlockStats() with the size argument as the key on which struct was passed in (in which case, we need to rethink everything about my proposal - and rebasing to a newer .so is the only way to take advantage of the new fields), or just the idea of an API break by how I first described the changes (in which case I am confident that we can tweak libvirt.h.in to guarantee a naming choice that preserves API compatibility across recompilation, while most of my proposal stands as-is where it remains possible to backport the new fields to older libvirtd installations since no .so entry points are changed). DV also seemed to be in favor of reusing the existing function call with a new struct, as long as there is no API break for existing code. And personally, I see value in being able to add a new struct without requiring a rebase to a new ABI.
I just audited libvirt.c, and found that the following libvirt interfaces that take a size parameter, and could be extensible by use of the size field: virDomainBlockStats virDomainInterfaceStats
The following libvirt interfaces take a flags parameter, and could be extensible by setting a flag bit that says a larger struct is in use: virDomainGetControlInfo virNodeGetCPUStats virNodeGetMemoryStats virDomainMemoryStats virDomainGetBlockInfo virDomainGetBlockJobInfo
Meanwhile, the following APIs which take a pointer to a struct are non-extensible (neither size nor flags as an escape hatch to indicate explicit use of a larger struct), extending any of these would require a new API function: virDomainGetInfo virNodeGetInfo virDomainGetVcpus virDomainGetSecurityLabel virDomainGetSecurityModel virStoragePoolGetInfo virStorageVolGetInfo virDomainGetJobInfo
As a final question for this mail, if we decide that reusing the existing ABI with a new struct size is not desirable, then would you rather that the new virDomainBlockStats2 API (or whatever we name the new function) would also take a (non-extensible) struct, or would it be better to rewrite it in terms of 'virTypedParameterPtr params, int *nparams' so that all future extensibility is covered by adding name-type-value tuples rather than dealing with future struct sizing? I'd rather go through the pain of an ABI addition at most once, rather than once per time that we identify new fields to add.
Vote for the second method (rewrite the struct in terms of virTypedParameterPtr). Also we might need to use virTypedParameterPtr when implementing v2 APIs for these functions? virDomainGetInfo virNodeGetInfo virDomainGetVcpus virDomainGetSecurityLabel virDomainGetSecurityModel virStoragePoolGetInfo virStorageVolGetInfo virDomainGetJobInfo Osier
On Wed, Aug 17, 2011 at 11:15:29AM -0600, Eric Blake wrote:
But I'm unclear from your answer whether you are opposed to the overall idea of reusing virDomainBlockStats() with the size argument as the key on which struct was passed in [...]
I will just say that (rightly or wrongly) the size field was designed to allow the struct to be increased in size in future. Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming blog: http://rwmj.wordpress.com Fedora now supports 80 OCaml packages (the OPEN alternative to F#) http://cocan.org/getting_started_with_ocaml_on_red_hat_and_fedora
--- src/qemu/qemu_monitor.c | 12 ++++++++++++ src/qemu/qemu_monitor.h | 4 ++++ src/qemu/qemu_monitor_json.c | 36 +++++++++++++++++++++++++++++++++++- src/qemu/qemu_monitor_json.h | 5 ++++- src/qemu/qemu_monitor_text.c | 39 +++++++++++++++++++++++++++++---------- src/qemu/qemu_monitor_text.h | 4 ++++ 6 files changed, 88 insertions(+), 12 deletions(-) diff --git a/src/qemu/qemu_monitor.c b/src/qemu/qemu_monitor.c index db6107c..aac4b21 100644 --- a/src/qemu/qemu_monitor.c +++ b/src/qemu/qemu_monitor.c @@ -1201,8 +1201,12 @@ int qemuMonitorGetBlockStatsInfo(qemuMonitorPtr mon, const char *devname, long long *rd_req, long long *rd_bytes, + long long *rd_total_times, long long *wr_req, long long *wr_bytes, + long long *wr_total_times, + long long *flush_req, + long long *flush_total_times, long long *errs) { int ret; @@ -1217,12 +1221,20 @@ int qemuMonitorGetBlockStatsInfo(qemuMonitorPtr mon, if (mon->json) ret = qemuMonitorJSONGetBlockStatsInfo(mon, devname, rd_req, rd_bytes, + rd_total_times, wr_req, wr_bytes, + wr_total_times, + flush_req, + flush_total_times, errs); else ret = qemuMonitorTextGetBlockStatsInfo(mon, devname, rd_req, rd_bytes, + rd_total_times, wr_req, wr_bytes, + wr_total_times, + flush_req, + flush_total_times, errs); return ret; } diff --git a/src/qemu/qemu_monitor.h b/src/qemu/qemu_monitor.h index f241c9e..d7695e9 100644 --- a/src/qemu/qemu_monitor.h +++ b/src/qemu/qemu_monitor.h @@ -212,8 +212,12 @@ int qemuMonitorGetBlockStatsInfo(qemuMonitorPtr mon, const char *devname, long long *rd_req, long long *rd_bytes, + long long *rd_total_times, long long *wr_req, long long *wr_bytes, + long long *wr_total_times, + long long *flush_req, + long long *flush_total_times, long long *errs); int qemuMonitorGetBlockExtent(qemuMonitorPtr mon, diff --git a/src/qemu/qemu_monitor_json.c b/src/qemu/qemu_monitor_json.c index 2a9a078..c917978 100644 --- a/src/qemu/qemu_monitor_json.c +++ b/src/qemu/qemu_monitor_json.c @@ -1318,8 +1318,12 @@ int qemuMonitorJSONGetBlockStatsInfo(qemuMonitorPtr mon, const char *devname, long long *rd_req, long long *rd_bytes, + long long *rd_total_times, long long *wr_req, long long *wr_bytes, + long long *wr_total_times, + long long *flush_req, + long long *flush_total_times, long long *errs) { int ret; @@ -1330,7 +1334,9 @@ int qemuMonitorJSONGetBlockStatsInfo(qemuMonitorPtr mon, virJSONValuePtr reply = NULL; virJSONValuePtr devices; - *rd_req = *rd_bytes = *wr_req = *wr_bytes = *errs = 0; + *rd_req = *rd_bytes = *rd_total_times = 0; + *wr_req = *wr_bytes = *wr_total_times = 0; + *flush_req = *flush_total_times = *errs = 0; if (!cmd) return -1; @@ -1396,6 +1402,13 @@ int qemuMonitorJSONGetBlockStatsInfo(qemuMonitorPtr mon, "rd_operations"); goto cleanup; } + if (virJSONValueObjectGetNumberLong(stats, "rd_total_times_ns", + rd_total_times) < 0) { + qemuReportError(VIR_ERR_INTERNAL_ERROR, + _("cannot read %s statistic"), + "rd_total_times_ns"); + goto cleanup; + } if (virJSONValueObjectGetNumberLong(stats, "wr_bytes", wr_bytes) < 0) { qemuReportError(VIR_ERR_INTERNAL_ERROR, _("cannot read %s statistic"), @@ -1408,6 +1421,27 @@ int qemuMonitorJSONGetBlockStatsInfo(qemuMonitorPtr mon, "wr_operations"); goto cleanup; } + if (virJSONValueObjectGetNumberLong(stats, "wr_total_times_ns", + wr_total_times) < 0) { + qemuReportError(VIR_ERR_INTERNAL_ERROR, + _("cannot read %s statistic"), + "wr_total_times_ns"); + goto cleanup; + } + if (virJSONValueObjectGetNumberLong(stats, "flush_operations", + flush_req) < 0) { + qemuReportError(VIR_ERR_INTERNAL_ERROR, + _("cannot read %s statistic"), + "flush_operations"); + goto cleanup; + } + if (virJSONValueObjectGetNumberLong(stats, "flush_total_times_ns", + flush_total_times) < 0) { + qemuReportError(VIR_ERR_INTERNAL_ERROR, + _("cannot read %s statistic"), + "flush_total_times_ns"); + goto cleanup; + } } if (!found) { diff --git a/src/qemu/qemu_monitor_json.h b/src/qemu/qemu_monitor_json.h index 9512793..a1ab39d 100644 --- a/src/qemu/qemu_monitor_json.h +++ b/src/qemu/qemu_monitor_json.h @@ -64,14 +64,17 @@ int qemuMonitorJSONGetBlockStatsInfo(qemuMonitorPtr mon, const char *devname, long long *rd_req, long long *rd_bytes, + long long *rd_total_times, long long *wr_req, long long *wr_bytes, + long long *wr_total_times, + long long *flush_req, + long long *flush_total_times, long long *errs); int qemuMonitorJSONGetBlockExtent(qemuMonitorPtr mon, const char *devname, unsigned long long *extent); - int qemuMonitorJSONSetVNCPassword(qemuMonitorPtr mon, const char *password); int qemuMonitorJSONSetPassword(qemuMonitorPtr mon, diff --git a/src/qemu/qemu_monitor_text.c b/src/qemu/qemu_monitor_text.c index 7bf733d..fbea855 100644 --- a/src/qemu/qemu_monitor_text.c +++ b/src/qemu/qemu_monitor_text.c @@ -674,8 +674,12 @@ int qemuMonitorTextGetBlockStatsInfo(qemuMonitorPtr mon, const char *devname, long long *rd_req, long long *rd_bytes, + long long *rd_total_times, long long *wr_req, long long *wr_bytes, + long long *wr_total_times, + long long *flush_req, + long long *flush_total_times, long long *errs) { char *info = NULL; @@ -702,11 +706,9 @@ int qemuMonitorTextGetBlockStatsInfo(qemuMonitorPtr mon, goto cleanup; } - *rd_req = -1; - *rd_bytes = -1; - *wr_req = -1; - *wr_bytes = -1; - *errs = -1; + *rd_req = *rd_bytes = *rd_total_times = -1; + *wr_req = *wr_bytes = *wr_total_times = -1; + *flush_req = *flush_total_times = *errs = -1; /* The output format for both qemu & KVM is: * blockdevice: rd_bytes=% wr_bytes=% rd_operations=% wr_operations=% @@ -734,23 +736,40 @@ int qemuMonitorTextGetBlockStatsInfo(qemuMonitorPtr mon, while (*p) { if (STRPREFIX (p, "rd_bytes=")) { - p += 9; + p += strlen("rd_bytes"); if (virStrToLong_ll (p, &dummy, 10, rd_bytes) == -1) VIR_DEBUG ("error reading rd_bytes: %s", p); } else if (STRPREFIX (p, "wr_bytes=")) { - p += 9; + p += strlen("wr_bytes"); if (virStrToLong_ll (p, &dummy, 10, wr_bytes) == -1) VIR_DEBUG ("error reading wr_bytes: %s", p); } else if (STRPREFIX (p, "rd_operations=")) { - p += 14; + p += strlen("rd_operations"); if (virStrToLong_ll (p, &dummy, 10, rd_req) == -1) VIR_DEBUG ("error reading rd_req: %s", p); } else if (STRPREFIX (p, "wr_operations=")) { - p += 14; + p += strlen("wr_operations="); if (virStrToLong_ll (p, &dummy, 10, wr_req) == -1) VIR_DEBUG ("error reading wr_req: %s", p); - } else + } else if (STRPREFIX (p, "rd_total_times_ns=")) { + p += strlen("rd_total_times_ns="); + if (virStrToLong_ll (p, &dummy, 10, rd_total_times) == -1) + VIR_DEBUG ("error reading rd_total_times: %s", p); + } else if (STRPREFIX (p, "wr_total_times_ns=")) { + p += strlen("wr_total_times_ns="); + if (virStrToLong_ll (p, &dummy, 10, wr_total_times) == -1) + VIR_DEBUG ("error reading wr_total_times: %s", p); + } else if (STRPREFIX (p, "flush_operations=")) { + p += strlen("flush_operations="); + if (virStrToLong_ll (p, &dummy, 10, flush_req) == -1) + VIR_DEBUG ("error reading flush_req: %s", p); + } else if (STRPREFIX (p, "flush_total_times_ns=")) { + p += strlen("flush_total_times_ns="); + if (virStrToLong_ll (p, &dummy, 10, flush_total_times) == -1) + VIR_DEBUG ("error reading flush_total_times: %s", p); + } else { VIR_DEBUG ("unknown block stat near %s", p); + } /* Skip to next label. */ p = strchr (p, ' '); diff --git a/src/qemu/qemu_monitor_text.h b/src/qemu/qemu_monitor_text.h index b250738..f7d9f5c 100644 --- a/src/qemu/qemu_monitor_text.h +++ b/src/qemu/qemu_monitor_text.h @@ -61,8 +61,12 @@ int qemuMonitorTextGetBlockStatsInfo(qemuMonitorPtr mon, const char *devname, long long *rd_req, long long *rd_bytes, + long long *rd_total_times, long long *wr_req, long long *wr_bytes, + long long *wr_total_times, + long long *flush_req, + long long *flush_total_times, long long *errs); int qemuMonitorTextGetBlockExtent(qemuMonitorPtr mon, const char *devname, -- 1.7.6
* src/qemu/qemu_driver.c: Update function qemudDomainBlockStats to support the new disk latency related members. --- src/qemu/qemu_driver.c | 4 ++++ 1 files changed, 4 insertions(+), 0 deletions(-) diff --git a/src/qemu/qemu_driver.c b/src/qemu/qemu_driver.c index ce19be7..36d0d3f 100644 --- a/src/qemu/qemu_driver.c +++ b/src/qemu/qemu_driver.c @@ -6801,8 +6801,12 @@ qemudDomainBlockStats (virDomainPtr dom, disk->info.alias, &stats->rd_req, &stats->rd_bytes, + &stats->rd_total_times, &stats->wr_req, &stats->wr_bytes, + &stats->wr_total_times, + &stats->flush_req, + &stats->flush_total_times, &stats->errs); qemuDomainObjExitMonitor(driver, vm); -- 1.7.6
--- tools/virsh.c | 13 +++++++++++++ 1 files changed, 13 insertions(+), 0 deletions(-) diff --git a/tools/virsh.c b/tools/virsh.c index b053ed0..555f278 100644 --- a/tools/virsh.c +++ b/tools/virsh.c @@ -1082,12 +1082,25 @@ cmdDomblkstat (vshControl *ctl, const vshCmd *cmd) if (stats.rd_bytes >= 0) vshPrint (ctl, "%s rd_bytes %lld\n", device, stats.rd_bytes); + if (stats.rd_total_times >= 0) + vshPrint (ctl, "%sns rd_total_times %lld\n", device, stats.rd_total_times); + if (stats.wr_req >= 0) vshPrint (ctl, "%s wr_req %lld\n", device, stats.wr_req); if (stats.wr_bytes >= 0) vshPrint (ctl, "%s wr_bytes %lld\n", device, stats.wr_bytes); + if (stats.wr_total_times >= 0) + vshPrint (ctl, "%sns wr_total_times %lld\n", device, stats.wr_total_times); + + if (stats.flush_req >= 0) + vshPrint (ctl, "%s flush_req %lld\n", device, stats.flush_req); + + if (stats.flush_total_times >= 0) + vshPrint (ctl, "%sns flush_total_times %lld\n", device, + stats.flush_total_times); + if (stats.errs >= 0) vshPrint (ctl, "%s errs %lld\n", device, stats.errs); -- 1.7.6
participants (6)
-
Adam Litke -
Daniel P. Berrange -
Daniel Veillard -
Eric Blake -
Osier Yang -
Richard W.M. Jones