[Qemu-devel] [PATCH v8 0/9] VT-d: vfio enablement and misc enhances

[Peter Xu <peterx@redhat.com>]

This is v8 of vt-d vfio enablement series.

v8
- remove patches 1-9 since merged already
- add David's r-b for all the patches
- add Aviv's s-o-b in the last patch
- rename iommu to iommu_dmar [Jason]
- rename last patch subject to "remote IOTLB" [Jason]
- pick up jason's two patches to fix vhost breakage
- let vhost leverage the new IOMMU notifier interface

v7:
- for the two traces patches: Change subjects. Remove vtd_err() and
  vtd_err_nonzero_rsvd() tracers, instead using standalone trace for
  each of the places. Don't remove any DPRINTF() if there is no
  replacement. [Jason]
- add r-b and a-b for Alex/David/Jason.
- in patch "intel_iommu: renaming gpa to iova where proper", convert
  one more place where I missed [Jason]
- fix the place where I should use "~0ULL" not "~0" [Jason]
- squash patch 16 into 18 [Jason]

v6:
- do unmap in all cases when replay [Jason]
- do global replay even if context entry is invalidated [Jason]
- when iommu reset, send unmap to all registered notifiers [Jason]
- use rcu read lock to protect the whole vfio_iommu_map_notify()
  [Alex, Paolo]

v5:
- fix patch 4 subject too long, and error spelling [Eric]
- add ack-by for alex in patch 1 [Alex]
- squashing patch 19/20 into patch 18 [Jason]
- fix comments in vtd_page_walk() [Jason]
- remove all error_report() [Jason]
- add comment for patch 18, mention about that enabled vhost without
  ATS as well [Jason]
- remove skipped debug thing during page walk [Jason]
- remove duplicated page walk trace [Jason]
- some tunings in vtd_address_space_unmap(), to provide correct iova
  and addr_mask. For this, I tuned this patch as well a bit:
  "memory: add section range info for IOMMU notifier"
  to loosen the range check

v4:
- convert all error_report()s into traces (in the two patches that did
  that)
- rebased to Jason's DMAR series (master + one more patch:
  "[PATCH V4 net-next] vhost_net: device IOTLB support")
- let vhost use the new api iommu_notifier_init() so it won't break
  vhost dmar [Jason]
- touch commit message of the patch:
  "intel_iommu: provide its own replay() callback"
  old replay is not a dead loop, but it will just consume lots of time
  [Jason]
- add comment for patch:
  "intel_iommu: do replay when context invalidate"
  telling why replay won't be a problem even without CM=1 [Jason]
- remove a useless comment line [Jason]
- remove dmar_enabled parameter for vtd_switch_address_space() and
  vtd_switch_address_space_all() [Mst, Jason]
- merged the vfio patches in, to support unmap of big ranges at the
  beginning ("[PATCH RFC 0/3] vfio: allow to notify unmap for very big
  region")
- using caching_mode instead of cache_mode_enabled, and "caching-mode"
  instead of "cache-mode" [Kevin]
- when receive context entry invalidation, we unmap the entire region
  first, then replay [Alex]
- fix commit message for patch:
  "intel_iommu: simplify irq region translation" [Kevin]
- handle domain/global invalidation, and notify where proper [Jason,
  Kevin]

v3:
- fix style error reported by patchew
- fix comment in domain switch patch: use "IOMMU address space" rather
  than "IOMMU region" [Kevin]
- add ack-by for Paolo in patch:
  "memory: add section range info for IOMMU notifier"
  (this is seperately collected besides this thread)
- remove 3 patches which are merged already (from Jason)
- rebase to master b6c0897

v2:
- change comment for "end" parameter in vtd_page_walk() [Tianyu]
- change comment for "a iova" to "an iova" [Yi]
- fix fault printed val for GPA address in vtd_page_walk_level (debug
  only)
- rebased to master (rather than Aviv's v6 series) and merged Aviv's
  series v6: picked patch 1 (as patch 1 in this series), dropped patch
  2, re-wrote patch 3 (as patch 17 of this series).
- picked up two more bugfix patches from Jason's DMAR series
- picked up the following patch as well:
  "[PATCH v3] intel_iommu: allow dynamic switch of IOMMU region"

This RFC series is a re-work for Aviv B.D.'s vfio enablement series
with vt-d:

  https://lists.gnu.org/archive/html/qemu-devel/2016-11/msg01452.html

Aviv has done a great job there, and what we still lack there are
mostly the following:

(1) VFIO got duplicated IOTLB notifications due to splitted VT-d IOMMU
    memory region.

(2) VT-d still haven't provide a correct replay() mechanism (e.g.,
    when IOMMU domain switches, things will broke).

This series should have solved the above two issues.

Online repo:

  https://github.com/xzpeter/qemu/tree/vtd-vfio-enablement-v8

I would be glad to hear about any review comments for above patches.

=========
Test Done
=========

Build test passed for x86_64/arm/ppc64.

Simply tested with x86_64, assigning two PCI devices to a single VM,
boot the VM using:

bin=x86_64-softmmu/qemu-system-x86_64
$bin -M q35,accel=kvm,kernel-irqchip=split -m 1G \
     -device intel-iommu,intremap=on,eim=off,caching-mode=on \
     -netdev user,id=net0,hostfwd=tcp::5555-:22 \
     -device virtio-net-pci,netdev=net0 \
     -device vfio-pci,host=03:00.0 \
     -device vfio-pci,host=02:00.0 \
     -trace events=".trace.vfio" \
     /var/lib/libvirt/images/vm1.qcow2

pxdev:bin [vtd-vfio-enablement]# cat .trace.vfio
vtd_page_walk*
vtd_replay*
vtd_inv_desc*

Then, in the guest, run the following tool:

  https://github.com/xzpeter/clibs/blob/master/gpl/userspace/vfio-bind-group/vfio-bind-group.c

With parameter:

  ./vfio-bind-group 00:03.0 00:04.0

Check host side trace log, I can see pages are replayed and mapped in
00:04.0 device address space, like:

...
vtd_replay_ce_valid replay valid context device 00:04.00 hi 0x401 lo 0x38fe1001
vtd_page_walk Page walk for ce (0x401, 0x38fe1001) iova range 0x0 - 0x8000000000
vtd_page_walk_level Page walk (base=0x38fe1000, level=3) iova range 0x0 - 0x8000000000
vtd_page_walk_level Page walk (base=0x35d31000, level=2) iova range 0x0 - 0x40000000
vtd_page_walk_level Page walk (base=0x34979000, level=1) iova range 0x0 - 0x200000
vtd_page_walk_one Page walk detected map level 0x1 iova 0x0 -> gpa 0x22dc3000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x1000 -> gpa 0x22e25000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x2000 -> gpa 0x22e12000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x3000 -> gpa 0x22e2d000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x4000 -> gpa 0x12a49000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x5000 -> gpa 0x129bb000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x6000 -> gpa 0x128db000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x7000 -> gpa 0x12a80000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x8000 -> gpa 0x12a7e000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0x9000 -> gpa 0x12b22000 mask 0xfff perm 3
vtd_page_walk_one Page walk detected map level 0x1 iova 0xa000 -> gpa 0x12b41000 mask 0xfff perm 3
...

=========
Todo List
=========

- error reporting for the assigned devices (as Tianyu has mentioned)

- per-domain address-space: A better solution in the future may be -
  we maintain one address space per IOMMU domain in the guest (so
  multiple devices can share a same address space if they are sharing
  the same IOMMU domains in the guest), rather than one address space
  per device (which is current implementation of vt-d). However that's
  a step further than this series, and let's see whether we can first
  provide a workable version of device assignment with vt-d
  protection.

- don't need to notify IOTLB (psi/gsi/global) invalidations to devices
  that with ATS enabled

- investigate when guest map page while mask contains existing mapped
  pages (e.g. map 12k-16k first, then map 0-12k)

- coalesce unmap during page walk (currently, we send it once per
  page)

- when do PSI for unmap, whether we can send one notify directly
  instead of walking over the page table?

- more to come...

Thanks,

Jason Wang (1):
  intel_iommu: use the correct memory region for device IOTLB
    notification

Peter Xu (8):
  memory: add section range info for IOMMU notifier
  memory: provide IOMMU_NOTIFIER_FOREACH macro
  memory: provide iommu_replay_all()
  memory: introduce memory_region_notify_one()
  memory: add MemoryRegionIOMMUOps.replay() callback
  intel_iommu: provide its own replay() callback
  intel_iommu: allow dynamic switch of IOMMU region
  intel_iommu: enable remote IOTLB

 hw/i386/intel_iommu.c          | 442 +++++++++++++++++++++++++++++++++++++++--
 hw/i386/intel_iommu_internal.h |   1 +
 hw/i386/trace-events           |  10 +-
 hw/vfio/common.c               |  12 +-
 hw/virtio/vhost.c              |  10 +-
 include/exec/memory.h          |  49 ++++-
 include/hw/i386/intel_iommu.h  |  10 +
 memory.c                       |  52 ++++-
 8 files changed, 552 insertions(+), 34 deletions(-)

-- 
2.7.4