[Xen-devel] [PATCH v8 00/19] x86/PMU: Xen PMU PV(H) support

[Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>]

Here is the eighth version of PV(H) PMU patches.

Changes in v8:

* Cleaned up a bit definitions of struct xenpf_symdata and xen_pmu_params
* Added compat checks for vpmu structures
* Converted vpmu flag manipulation macros to inline routines
* Reimplemented vpmu_unload_all() to avoid long loops
* Reworked PMU fault generation and handling (new patch #12)
* Added checks for domain->vcpu[] non-NULLness
* Added more comments, renamed some routines and macros, code style cleanup


Changes in v7:

* When reading hypervisor symbols make the caller pass buffer length
  (as opposed to having this length be part of the API). Make the
  hypervisor buffer static, make xensyms_read() return zero-length
  string on end-of-symbols. Make 'type' field of xenpf_symdata a char,
  drop compat_pf_symdata definition.
* Spread PVH support across patches as opposed to lumping it into a
  separate patch
* Rename vpmu_is_set_all() to vpmu_are_all_set()
* Split VPMU cleanup patch in two
* Use memmove when copying VMX guest and host MSRs
* Make padding of xen_arch_pmu's context union a constand that does not
  depend on arch context size.
* Set interface version to 0.1
* Check pointer validity in pvpmu_init/destroy()
* Fixed crash in core2_vpmu_dump()
* Fixed crash in vmx_add_msr()
* Break handling of Intel and AMD MSRs in traps.c into separate cases
* Pass full CS selector to guests
* Add lock in pvpmu init code to prevent potential race


Changes in v6:

* Two new patches:
  o Merge VMX MSR add/remove routines in vmcs.c (patch 5)
  o Merge VPMU read/write MSR routines in vpmu.c (patch 14)
* Check for pending NMI softirq after saving VPMU context to prevent a 
newly-scheduled
  guest from overwriting sampled_vcpu written by de-scheduled VPCU.
* Keep track of enabled counters on Intel. This was removed in earlier patches 
and
  was a mistake. As result of this change struct vpmu will have a pointer to 
private
  context data (i.e. data that is not exposed to a PV(H) guest). Use this 
private pointer
  on SVM as well for storing MSR bitmap status (it was unnecessarily exposed to 
PV guests
  earlier).
  Dropped Reviewed-by: and Tested-by: tags from patch 4 since it needs to be 
reviewed
  agan (core2_vpmu_do_wrmsr() routine, mostly)
* Replaced references to dom0 with hardware_domain (and is_control_domain with
  is_hardware_domain for consistency)
* Prevent non-privileged domains from reading PMU MSRs in VPMU_PRIV_MODE
* Reverted unnecessary changes in vpmu_initialise()'s switch statement
* Fixed comment in vpmu_do_interrupt


Changes in v5:

* Dropped patch number 2 ("Stop AMD counters when called from 
vpmu_save_force()")
  as no longer needed
* Added patch number 2 that marks context as loaded before PMU registers are
  loaded. This prevents situation where a PMU interrupt may occur while context
  is still viewed as not loaded. (This is really a bug fix for exsiting VPMU
  code)
* Renamed xenpmu.h files to pmu.h
* More careful use of is_pv_domain(), is_hvm_domain(, is_pvh_domain and
  has_hvm_container_domain(). Also explicitly disabled support for PVH until
  patch 16 to make distinction between usage of the above macros more clear.
* Added support for disabling VPMU support during runtime.
* Disable VPMUs for non-privileged domains when switching to privileged
  profiling mode
* Added ARM stub for xen_arch_pmu_t
* Separated vpmu_mode from vpmu_features
* Moved CS register query to make sure we use appropriate query mechanism for
  various guest types.
* LVTPC is now set from value in shared area, not copied from dom0
* Various code and comments cleanup as suggested by Jan.

Changes in v4:

* Added support for PVH guests:
  o changes in pvpmu_init() to accommodate both PV and PVH guests, still in 
patch 10
  o more careful use of is_hvm_domain
  o Additional patch (16)
* Moved HVM interrupt handling out of vpmu_do_interrupt() for NMI-safe handling
* Fixed dom0's VCPU selection in privileged mode
* Added a cast in register copy for 32-bit PV guests cpu_user_regs_t in 
vpmu_do_interrupt.
  (don't want to expose compat_cpu_user_regs in a public header)
* Renamed public structures by prefixing them with "xen_"
* Added an entry for xenpf_symdata in xlat.lst
* Fixed pv_cpuid check for vpmu-specific cpuid adjustments
* Varios code style fixes
* Eliminated anonymous unions
* Added more verbiage to NMI patch description


Changes in v3:

* Moved PMU MSR banks out from architectural context data structures to allow
for future expansion without protocol changes
* PMU interrupts can be either NMIs or regular vector interrupts (the latter
is the default)
* Context is now marked as PMU_CACHED by the hypervisor code to avoid certain
race conditions with the guest
* Fixed races with PV guest in MSR access handlers
* More Intel VPMU cleanup
* Moved NMI-unsafe code from NMI handler
* Dropped changes to vcpu->is_running
* Added LVTPC apic handling (cached for PV guests)
* Separated privileged profiling mode into a standalone patch
* Separated NMI handling into a standalone patch


Changes in v2:

* Xen symbols are exported as data structure (as opoosed to a set of formatted
strings in v1). Even though one symbol per hypercall is returned performance
appears to be acceptable: reading whole file from dom0 userland takes on average
about twice as long as reading /proc/kallsyms
* More cleanup of Intel VPMU code to simplify publicly exported structures
* There is an architecture-independent and x86-specific public include files 
(ARM
has a stub)
* General cleanup of public include files to make them more presentable (and
to make auto doc generation better)
* Setting of vcpu->is_running is now done on ARM in schedule_tail as well 
(making
changes to common/schedule.c architecture-independent). Note that this is not
tested since I don't have access to ARM hardware.
* PCPU ID of interrupted processor is now passed to PV guest


The following patch series adds PMU support in Xen for PV(H)
guests. There is a companion patchset for Linux kernel. In addition,
another set of changes will be provided (later) for userland perf
code.

This version has following limitations:
* For accurate profiling of dom0/Xen dom0 VCPUs should be pinned.
* Hypervisor code is only profiled on processors that have running dom0 VCPUs
on them.
* No backtrace support.
* Will fail to load under XSM: we ran out of bits in permissions vector and
this needs to be fixed separately

A few notes that may help reviewing:

* A shared data structure (xenpmu_data_t) between each PV VPCU and hypervisor
CPU is used for passing registers' values as well as PMU state at the time of
PMU interrupt.
* PMU interrupts are taken by hypervisor either as NMIs or regular vector
interrupts for both HVM and PV(H). The interrupts are sent as NMIs to HVM guests
and as virtual interrupts to PV(H) guests
* PV guest's interrupt handler does not read/write PMU MSRs directly. Instead, 
it
accesses xenpmu_data_t and flushes it to HW it before returning.
* PMU mode is controlled at runtime via 
/sys/hypervisor/pmu/pmu/{pmu_mode,pmu_flags}
in addition to 'vpmu' boot option (which is preserved for back compatibility).
The following modes are provided:
  * disable: VPMU is off
  * enable: VPMU is on. Guests can profile themselves, dom0 profiles itself and 
Xen
  * priv_enable: dom0 only profiling. dom0 collects samples for everyone. 
Sampling
    in guests is suspended.
* /proc/xen/xensyms file exports hypervisor's symbols to dom0 (similar to
/proc/kallsyms)
* VPMU infrastructure is now used for HVM, PV and PVH and therefore has been 
moved
up from hvm subtree


Boris Ostrovsky (19):
  common/symbols: Export hypervisor symbols to privileged guest
  x86/VPMU: Set MSR bitmaps only for HVM/PVH guests
  x86/VPMU: Make vpmu macros a bit more efficient
  intel/VPMU: Clean up Intel VPMU code
  vmx: Merge MSR management routines
  x86/VPMU: Handle APIC_LVTPC accesses
  intel/VPMU: MSR_CORE_PERF_GLOBAL_CTRL should be initialized to zero
  x86/VPMU: Add public xenpmu.h
  x86/VPMU: Make vpmu not HVM-specific
  x86/VPMU: Interface for setting PMU mode and flags
  x86/VPMU: Initialize PMU for PV(H) guests
  x86/VPMU: When handling MSR accesses, leave fault injection to callers
  x86/VPMU: Add support for PMU register handling on PV guests
  x86/VPMU: Handle PMU interrupts for PV guests
  x86/VPMU: Merge vpmu_rdmsr and vpmu_wrmsr
  x86/VPMU: Add privileged PMU mode
  x86/VPMU: Save VPMU state for PV guests during context switch
  x86/VPMU: NMI-based VPMU support
  x86/VPMU: Move VPMU files up from hvm/ directory

 xen/arch/x86/Makefile                              |   1 +
 xen/arch/x86/domain.c                              |  17 +-
 xen/arch/x86/hvm/Makefile                          |   1 -
 xen/arch/x86/hvm/hvm.c                             |   3 +-
 xen/arch/x86/hvm/svm/Makefile                      |   1 -
 xen/arch/x86/hvm/svm/svm.c                         |  19 +-
 xen/arch/x86/hvm/vlapic.c                          |   5 +-
 xen/arch/x86/hvm/vmx/Makefile                      |   1 -
 xen/arch/x86/hvm/vmx/vmcs.c                        | 113 ++-
 xen/arch/x86/hvm/vmx/vmx.c                         |  46 +-
 xen/arch/x86/hvm/vpmu.c                            | 265 -------
 xen/arch/x86/oprofile/op_model_ppro.c              |   8 +-
 xen/arch/x86/platform_hypercall.c                  |  32 +
 xen/arch/x86/traps.c                               |  54 +-
 xen/arch/x86/vpmu.c                                | 814 +++++++++++++++++++++
 xen/arch/x86/{hvm/svm/vpmu.c => vpmu_amd.c}        | 136 ++--
 .../x86/{hvm/vmx/vpmu_core2.c => vpmu_intel.c}     | 584 +++++++--------
 xen/arch/x86/x86_64/compat/entry.S                 |   4 +
 xen/arch/x86/x86_64/entry.S                        |   4 +
 xen/common/event_channel.c                         |   1 +
 xen/common/symbols.c                               |  54 ++
 xen/include/Makefile                               |   1 +
 xen/include/asm-x86/domain.h                       |   2 +
 xen/include/asm-x86/hvm/vcpu.h                     |   3 -
 xen/include/asm-x86/hvm/vmx/vmcs.h                 |  10 +-
 xen/include/asm-x86/hvm/vmx/vpmu_core2.h           |  51 --
 xen/include/asm-x86/{hvm => }/vpmu.h               |  72 +-
 xen/include/public/arch-arm.h                      |   3 +
 xen/include/public/arch-x86/pmu.h                  |  75 ++
 xen/include/public/platform.h                      |  19 +
 xen/include/public/pmu.h                           |  93 +++
 xen/include/public/xen.h                           |   2 +
 xen/include/xen/hypercall.h                        |   4 +
 xen/include/xen/softirq.h                          |   1 +
 xen/include/xen/symbols.h                          |   3 +
 xen/include/xlat.lst                               |   5 +
 36 files changed, 1733 insertions(+), 774 deletions(-)
 delete mode 100644 xen/arch/x86/hvm/vpmu.c
 create mode 100644 xen/arch/x86/vpmu.c
 rename xen/arch/x86/{hvm/svm/vpmu.c => vpmu_amd.c} (76%)
 rename xen/arch/x86/{hvm/vmx/vpmu_core2.c => vpmu_intel.c} (61%)
 delete mode 100644 xen/include/asm-x86/hvm/vmx/vpmu_core2.h
 rename xen/include/asm-x86/{hvm => }/vpmu.h (64%)
 create mode 100644 xen/include/public/arch-x86/pmu.h
 create mode 100644 xen/include/public/pmu.h

-- 
1.8.1.4


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel