Re: [Qemu-devel] [PATCH v3 4/9] {monitor, hw/pvrdma}: Expose device internals via monitor interface

[Markus Armbruster]

Yuval Shaia <address@hidden> writes:

> On Fri, Mar 01, 2019 at 08:17:02AM +0100, Markus Armbruster wrote:
>> Marcel Apfelbaum <address@hidden> writes:
>> 
>> > Hi Yuval,
>> >
>> > On 2/27/19 4:06 PM, Yuval Shaia wrote:
>> >> Allow interrogating device internals through HMP interface.
>> >> The exposed indicators can be used for troubleshooting by developers or
>> >> sysadmin.
>> >> There is no need to expose these attributes to a management system (e.x.
>> >> libvirt) because (1) most of them are not "device-management' related
>> >> info and (2) there is no guarantee the interface is stable.
>> >>
>> >> Signed-off-by: Yuval Shaia <address@hidden>
>> >> ---
>> >>   hmp-commands-info.hx      | 16 ++++++++
>> >>   hw/rdma/rdma_backend.c    | 70 ++++++++++++++++++++++++++---------
>> >>   hw/rdma/rdma_rm.c         |  7 ++++
>> >>   hw/rdma/rdma_rm_defs.h    | 27 +++++++++++++-
>> >>   hw/rdma/vmw/pvrdma.h      |  5 +++
>> >>   hw/rdma/vmw/pvrdma_hmp.h  | 21 +++++++++++
>> >>   hw/rdma/vmw/pvrdma_main.c | 77 +++++++++++++++++++++++++++++++++++++++
>> >>   monitor.c                 | 10 +++++
>> >>   8 files changed, 215 insertions(+), 18 deletions(-)
>> >>   create mode 100644 hw/rdma/vmw/pvrdma_hmp.h
>> >>
>> >> diff --git a/hmp-commands-info.hx b/hmp-commands-info.hx
>> >> index cbee8b944d..9153c33974 100644
>> >> --- a/hmp-commands-info.hx
>> >> +++ b/hmp-commands-info.hx
>> >> @@ -524,6 +524,22 @@ STEXI
>> >>   Show CPU statistics.
>> >>   ETEXI
>> >>   +#if defined(CONFIG_PVRDMA)
>> >> +    {
>> >> +        .name       = "pvrdmacounters",
>> >> +        .args_type  = "",
>> >> +        .params     = "",
>> >> +        .help       = "show pvrdma device counters",
>> >> +        .cmd        = hmp_info_pvrdmacounters,
>> >> +    },
>> >> +
>> >> +STEXI
>> >> address@hidden info pvrdmacounters
>> >> address@hidden info pvrdmacounters
>> >> +Show pvrdma device counters.
>> >> +ETEXI
>> >> +#endif
>> >> +
>> >>   #if defined(CONFIG_SLIRP)
>> >>       {
>> >>           .name       = "usernet",
>> [...]
>> >> diff --git a/hw/rdma/vmw/pvrdma_main.c b/hw/rdma/vmw/pvrdma_main.c
>> >> index b6061f4b6e..85101368c5 100644
>> >> --- a/hw/rdma/vmw/pvrdma_main.c
>> >> +++ b/hw/rdma/vmw/pvrdma_main.c
>> >> @@ -14,6 +14,7 @@
>> >>    */
>> >>     #include "qemu/osdep.h"
>> >> +#include "qemu/units.h"
>> >>   #include "qapi/error.h"
>> >>   #include "hw/hw.h"
>> >>   #include "hw/pci/pci.h"
>> >> @@ -25,6 +26,7 @@
>> >>   #include "cpu.h"
>> >>   #include "trace.h"
>> >>   #include "sysemu/sysemu.h"
>> >> +#include "monitor/monitor.h"
>> >>     #include "../rdma_rm.h"
>> >>   #include "../rdma_backend.h"
>> >> @@ -32,10 +34,13 @@
>> >>     #include <infiniband/verbs.h>
>> >>   #include "pvrdma.h"
>> >> +#include "pvrdma_hmp.h"
>> >>   #include "standard-headers/rdma/vmw_pvrdma-abi.h"
>> >>   #include 
>> >> "standard-headers/drivers/infiniband/hw/vmw_pvrdma/pvrdma_dev_api.h"
>> >>   #include "pvrdma_qp_ops.h"
>> >>   +GSList *devices;
>> 
>> "devices" is far too generic for an external identifier.  Are you
>> missing a 'static' here?  Even if static, I'd recommend "rdma_devices".
>
> Yep, thanks.
>
>> 
>> >> +
>> >>   static Property pvrdma_dev_properties[] = {
>> >>       DEFINE_PROP_STRING("netdev", PVRDMADev, backend_eth_device_name),
>> >>       DEFINE_PROP_STRING("ibdev", PVRDMADev, backend_device_name),
>> >> @@ -55,6 +60,71 @@ static Property pvrdma_dev_properties[] = {
>> [...]
>> >> +}
>> >> +
>> >> +void pvrdma_dump_counters(Monitor *mon)
>> >> +{
>> >> +    g_slist_foreach(devices, pvrdma_dump_device_counters, mon);
>> >> +}
>> 
>> Note for later: does nothing when @devices is empty.
>
> But that is fine, isn't it?

Yes.  It's just an observation for use in a later comment.

>> 
>> >> +
>> >>   static void free_dev_ring(PCIDevice *pci_dev, PvrdmaRing *ring,
>> >>                             void *ring_state)
>> >>   {
>> >> @@ -304,6 +374,8 @@ static void pvrdma_fini(PCIDevice *pdev)
>> >>         rdma_info_report("Device %s %x.%x is down", pdev->name,
>> >>                        PCI_SLOT(pdev->devfn), PCI_FUNC(pdev->devfn));
>> >> +
>> >> +    devices = g_slist_remove(devices, pdev);
>> >>   }
>> >>     static void pvrdma_stop(PVRDMADev *dev)
>> >> @@ -394,6 +466,7 @@ static void pvrdma_regs_write(void *opaque, hwaddr 
>> >> addr, uint64_t val,
>> >>           if (val == 0) {
>> >>               trace_pvrdma_regs_write(addr, val, "REQUEST", "");
>> >>               pvrdma_exec_cmd(dev);
>> >> +            dev->stats.commands++;
>> >>           }
>> >>           break;
>> >>       default:
>> >> @@ -612,9 +685,13 @@ static void pvrdma_realize(PCIDevice *pdev, Error 
>> >> **errp)
>> >>           goto out;
>> >>       }
>> >>   +    memset(&dev->stats, 0, sizeof(dev->stats));
>> >> +
>> >>       dev->shutdown_notifier.notify = pvrdma_shutdown_notifier;
>> >>       qemu_register_shutdown_notifier(&dev->shutdown_notifier);
>> >>   +    devices = g_slist_append(devices, pdev);
>> >> +
>> >>   out:
>> >>       if (rc) {
>> >>           pvrdma_fini(pdev);
>> 
>> Note for later: @devices contains the realized "pvrdma" devices.
>
> This happens only when rc indicates an error and we jumped here before
> adding the device to the list.

I'm referring to the update of @devices, not the out: label.

>> You could find these devices with object_child_foreach_recursive()
>> instead of maintaining a separate list.
>
> Hmmm...interesting.
> I will check if it fits my needs and will change accordingly if yes.

Examples include hmp_info_irq() and hmp_info_pic().

>> >> diff --git a/monitor.c b/monitor.c
>> >> index defa129319..53ecb5bc70 100644
>> >> --- a/monitor.c
>> >> +++ b/monitor.c
>> >> @@ -85,6 +85,9 @@
>> >>   #include "sysemu/iothread.h"
>> >>   #include "qemu/cutils.h"
>> >>   #include "tcg/tcg.h"
>> >> +#ifdef CONFIG_PVRDMA
>> >> +#include "hw/rdma/vmw/pvrdma_hmp.h"
>> >> +#endif
>> >>     #if defined(TARGET_S390X)
>> >>   #include "hw/s390x/storage-keys.h"
>> >> @@ -1362,6 +1365,13 @@ static void hmp_info_cpustats(Monitor *mon, const 
>> >> QDict *qdict)
>> >>       cpu_dump_statistics(cs, (FILE *)mon, &monitor_fprintf, 0);
>> >>   }
>> >>   +#ifdef CONFIG_PVRDMA
>> >> +static void hmp_info_pvrdmacounters(Monitor *mon, const QDict *qdict)
>> >> +{
>> >> +    pvrdma_dump_counters(mon);
>> >
>> > Compilation fails on archs with no PCI support:
>> >
>> >     /usr/bin/ld: monitor.o: in function `hmp_info_pvrdmacounters':
>> >     /home/marcel/git/qemu/monitor.c:1371: undefined reference to
>> > `pvrdma_dump_counters'
>> >    collect2: error: ld returned 1 exit status
>> >    make[1]: *** [Makefile:210: qemu-system-m68k] Error 1
>> >
>> >
>> > The below patch solves it by adding a pci stub:
>> >
>> > diff --git a/hw/pci/pci-stub.c b/hw/pci/pci-stub.c
>> > index b941a0e842..cab364f93d 100644
>> > --- a/hw/pci/pci-stub.c
>> > +++ b/hw/pci/pci-stub.c
>> > @@ -26,6 +26,7 @@
>> >  #include "qapi/qmp/qerror.h"
>> >  #include "hw/pci/pci.h"
>> >  #include "hw/pci/msi.h"
>> > +#include "hw/rdma/vmw/pvrdma_hmp.h"
>> >
>> >  bool msi_nonbroken;
>> >  bool pci_available;
>> > @@ -53,3 +54,9 @@ uint16_t pci_requester_id(PCIDevice *dev)
>> >      g_assert(false);
>> >      return 0;
>> >  }
>> > +
>> > +void pvrdma_dump_counters(Monitor *mon)
>> > +{
>> > +    monitor_printf(mon, "PVRDMA requires PCI support\n");
>> > +}
>> > +
>> 
>> When CONFIG_PCI is enabled, "info pvrdmacounters" does nothing when
>> there are no "pvrdma" devices.
>> 
>> When CONFIG_PCI is disabled, there are no "pvrdma" devices.  Therefore,
>> "info pvrdmacounters" should also do nothing then, shouldn't it?
>
> Yeah, problem was in case that pvrdma was selected in ./configure phase but
> platform does not have PCI -> CONFIG_PCI is diabeled -> pvrdma code is not
> compiled -> pvrdma_dump_counters is missing in link phase.
>
> If you have a better alternative then i'm fine, meanwhile i'm taking
> Marcel's proposal.

Marcel's idea to fix compilation with a stub is spot on.  But should it
print a message?  I don't think so.

>> 
>> > However you should include a generic rdma header as hw/rdma/rdma_hmp.h
>> > and not a vmw specific one.
>> >
>> >
>> > Thanks,
>> > Marcel
>> >
>> >
>> >> +}
>> >> +#endif
>> >> +
>> >>   static void hmp_info_trace_events(Monitor *mon, const QDict *qdict)
>> >>   {
>> >>       const char *name = qdict_get_try_str(qdict, "name");
>>