[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] [PATCH 1/4] latency: Introduce new members for _virDomainBlockStats

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:
  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:

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 ... };

    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

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]