Hi Luiz, On 04/10/14 10:24, Laszlo Ersek wrote: > This is for <https://bugzilla.redhat.com/show_bug.cgi?id=616415>. > > In general, we want to propagate non-fatal (ie. non-exit()ing, > non-abort()ing) errors to the QMP caller, rather than printing them > on-site. The series focuses on errors raised at PCI assignment time (ie. > reachable from assigned_initfn()), other errors are not converted. > > Errors are not propagated through assigned_initfn(); let's wait for > someone else to convert "kvm-pci-assign" from qdev to QOM. The series is > nonetheless an improvement, because the forwarding of initialization > errors now stops just before device_realize(). We set the > stored/persistent monitor error there. > > Informative and warning messages (that report about circumstances that > don't break the assignment operation) cannot terminate in > qerror_report_err(), because that would set the persistent monitor > error, breaking the high level (QMP) operation. > > A call graph rooted in assigned_initfn() was generated with cflow. > Leaves that could never fail were removed from the graph (and this > property similarly propagated upwards as far as possible). > > The patchset loosely follows a bottom-up algorithm on this calltree. Any > leaf that reports an error internally and returns a related failure is > converted to "throw" an Error structure instead. All direct callers of > the converted leaf are addressed at once, in the same patch, to consume > the error (and they become the new leaves gradually). > > When the leaf to be converted is also called outside of > "hw/i386/kvm/pci-assign.c", the conversion keeps a compatibility > function under the original name, so that conversion of callers > unrelated to PCI assignment can be deferred. > > Reviewers should copy the call graph to a text file, and mark, as the > series progresses, fully converted functions. (Ie. functions that now > report their terminating error messages with Error objects only.) > > assigned_initfn() > error_report() > get_real_device() > monitor_handle_fd_param() > error_report() > get_real_vendor_id() > get_real_id() > error_report() > get_real_device_id() > get_real_id() > error_report() > assigned_device_pci_cap_init() > check_irqchip_in_kernel() > error_report() > pci_add_capability() > error_report() > assigned_dev_register_msix_mmio() > error_report() > assigned_dev_register_regions() > error_report() > assign_device() > error_report() > assign_failed_examine() > get_real_vendor_id() > get_real_id() > error_report() > get_real_device_id() > get_real_id() > error_report() > error_printf() > error_report() > assign_intx() > check_irqchip_in_kernel() > error_report() > error_report() > error_printf() > > Laszlo Ersek (16): > cutils: tighten qemu_parse_fd() > monitor: add Error-propagating monitor_handle_fd_param2() > pci-assign: accept Error from monitor_handle_fd_param2() > pci-assign: make assign_failed_examine() just format the cause > pci-assign: propagate errors from get_real_id() > pci-assign: propagate Error from check_irqchip_in_kernel() > pci: add Error-propagating pci_add_capability2() > pci-assign: accept Error from pci_add_capability2() > pci-assign: assignment should fail if we can't read config space > pci-assign: propagate errors from get_real_device() > pci-assign: propagate errors from assigned_device_pci_cap_init() > pci-assign: propagate errors from assigned_dev_register_msix_mmio() > pci-assign: propagate errors from assigned_dev_register_regions() > pci-assign: propagate errors from assign_device() > pci-assign: propagate errors from assign_intx() > pci-assign: assigned_initfn(): set monitor error in common error > handler > > include/hw/pci/pci.h | 4 + > include/monitor/monitor.h | 1 + > hw/i386/kvm/pci-assign.c | 273 > ++++++++++++++++++++++++++++------------------ > hw/pci/pci.c | 32 +++++- > monitor.c | 29 ++++- > util/cutils.c | 13 ++- > 6 files changed, 232 insertions(+), 120 deletions(-) >
I forgot to CC you on this. Thanks Laszlo
