[v4,0/6] KVM: arm64: Add support for FEAT_TLBIRANGE

Message

Raghavendra Rao Ananta May 19, 2023, 12:52 a.m. UTC

In certain code paths, KVM/ARM currently invalidates the entire VM's
page-tables instead of just invalidating a necessary range. For example,
when collapsing a table PTE to a block PTE, instead of iterating over
each PTE and flushing them, KVM uses 'vmalls12e1is' TLBI operation to
flush all the entries. This is inefficient since the guest would have
to refill the TLBs again, even for the addresses that aren't covered
by the table entry. The performance impact would scale poorly if many
addresses in the VM is going through this remapping.

For architectures that implement FEAT_TLBIRANGE, KVM can replace such
inefficient paths by performing the invalidations only on the range of
addresses that are in scope. This series tries to achieve the same in
the areas of stage-2 map, unmap and write-protecting the pages.

Patch-1 refactors the core arm64's __flush_tlb_range() to be used by
other entities.

Patch-2 adds a range-based TLBI mechanism for KVM (VHE and nVHE).

Patch-3 implements the kvm_arch_flush_remote_tlbs_range() for arm64.

Patch-4 aims to flush only the memslot that undergoes a write-protect,
instead of the entire VM.

Patch-5 operates on stage2_try_break_pte() to use the range based
TLBI instructions when collapsing a table entry. The map path is the
immediate consumer of this when KVM remaps a table entry into a block.

Patch-6 modifies the stage-2 unmap path in which, if the system supports
FEAT_TLBIRANGE, the TLB invalidations are skipped during the page-table.
walk. Instead it's done in one go after the entire walk is finished.

The series is based off of upstream v6.4-rc2, and applied David
Matlack's common API for TLB invalidations[1] on top.

The performance evaluation was done on a hardware that supports
FEAT_TLBIRANGE, on a VHE configuration, using a modified
kvm_page_table_test.
The modified version updates the guest code in the ADJUST_MAPPINGS case
to not only access this page but also to access up to 512 pages
backwards
for every new page it iterates through. This is done to test the effect
of TLBI misses after KVM has handled a fault.

The series captures the impact in the map and unmap paths as described
above.

$ kvm_page_table_test -m 2 -v 128 -s anonymous_hugetlb_2mb -b $i

+--------+------------------------------+------------------------------+
| mem_sz |    ADJUST_MAPPINGS (s)       |      Unmap VM (s)            |
|  (GB)  | Baseline | Baseline + series | Baseline | Baseline + series |
+--------+----------|-------------------+------------------------------+
|   1    |   3.44   |   2.97            | 0.007     | 0.005            |
|   2    |   5.56   |   5.63            | 0.010     | 0.006            |
|   4    |  11.03   |  10.44            | 0.015     | 0.008            |
|   8    |  24.54   |  19.00            | 0.024     | 0.011            |
|  16    |  40.16   |  36.83            | 0.041     | 0.018            |
|  32    |  75.76   |  73.84            | 0.074     | 0.029            |
|  64    | 151.58   | 152.62            | 0.148     | 0.050            |
| 128    | 330.42   | 306.86            | 0.280     | 0.090            |
+--------+----------+-------------------+----------+-------------------+

$ kvm_page_table_test -m 2 -b 128G -s anonymous_hugetlb_2mb -v $i

+--------+------------------------------+
| vCPUs  |    ADJUST_MAPPINGS (s)       |
|        | Baseline | Baseline + series |
+--------+----------|-------------------+
|   1    | 138.69   | 135.58            |
|   2    | 138.77   | 137.54            |
|   4    | 162.57   | 135.82            |
|   8    | 154.92   | 143.67            |
|  16    | 122.02   | 118.86            |
|  32    | 119.99   | 118.81            |
|  64    | 190.70   | 169.36            |
| 128    | 330.42   | 306.86            |   
+--------+----------+-------------------+

For the ADJUST_MAPPINGS cases, which maps back the 4K table entries to
2M hugepages, the series sees an average improvement of ~7%. For
unmapping 2M hugepages, we see at least a 3x improvement.

$ kvm_page_table_test -m 2 -b $i

+--------+------------------------------+
| mem_sz |      Unmap VM (s)            |
|  (GB)  | Baseline | Baseline + series |
+--------+------------------------------+
|   1    |  0.52    |  0.13             |
|   2    |  1.03    |  0.25             |
|   4    |  2.04    |  0.47             |
|   8    |  4.05    |  0.94             |
|  16    |  8.11    |  1.82             |
|  32    | 16.11    |  3.69             |
|  64    | 32.35    |  7.22             |
| 128    | 64.66    | 14.69             |   
+--------+----------+-------------------+

The series sees an average gain of 4x when the guest backed by
PAGE_SIZE (4K) pages.

v4:
Thanks again, Oliver for all the comments
- Updated the __kvm_tlb_flush_vmid_range() implementation for
  nVHE to adjust with the modfied __tlb_switch_to_guest() that
  accepts a new 'bool nsh' arg.
- Renamed stage2_put_pte() to stage2_unmap_put_pte() and removed
  the 'skip_flush' argument.
- Defined stage2_unmap_defer_tlb_flush() to check if the PTE
  flushes can be deferred during the unmap table walk. It's
  being called from stage2_unmap_put_pte() and
  kvm_pgtable_stage2_unmap().
- Got rid of the 'struct stage2_unmap_data'.

v3:
https://lore.kernel.org/all/[e-mail]/
Thanks, Oliver for all the suggestions.
- The core flush API (__kvm_tlb_flush_vmid_range()) now checks if
  the system support FEAT_TLBIRANGE or not, thus elimiating the
  redundancy in the upper layers.
- If FEAT_TLBIRANGE is not supported, the implementation falls
  back to invalidating all the TLB entries with the VMID, instead
  of doing an iterative flush for the range.
- The kvm_arch_flush_remote_tlbs_range() doesn't return -EOPNOTSUPP
  if the system doesn't implement FEAT_TLBIRANGE. It depends on
  __kvm_tlb_flush_vmid_range() to do take care of the decisions
  and return 0 regardless of the underlying feature support.
- __kvm_tlb_flush_vmid_range() doesn't take 'level' as input to
  calculate the 'stride'. Instead, it always assumes PAGE_SIZE.
- Fast unmap path is eliminated. Instead, the existing unmap walker
  is modified to skip the TLBIs during the walk, and do it all at
  once after the walk, using the range-based instructions.

v2:
https://lore.kernel.org/all/[e-mail]/
- Rebased the series on top of David Matlack's series for common
  TLB invalidation API[1].
- Implement kvm_arch_flush_remote_tlbs_range() for arm64, by extending
  the support introduced by [1].
- Use kvm_flush_remote_tlbs_memslot() introduced by [1] to flush
  only the current memslot after write-protect.
- Modified the __kvm_tlb_flush_range() macro to accepts 'level' as an
  argument to calculate the 'stride' instead of just using PAGE_SIZE.
- Split the patch that introduces the range-based TLBI to KVM and the
  implementation of IPA-based invalidation into its own patches.
- Dropped the patch that tries to optimize the mmu notifiers paths.
- Rename the function kvm_table_pte_flush() to
  kvm_pgtable_stage2_flush_range(), and accept the range of addresses to
  flush. [Oliver]
- Drop the 'tlb_level' argument for stage2_try_break_pte() and directly
  pass '0' as 'tlb_level' to kvm_pgtable_stage2_flush_range(). [Oliver]

v1:
https://lore.kernel.org/all/[e-mail]/

Thank you.
Raghavendra

[1]:
https://lore.kernel.org/linux-arm-kernel/[e-mail]/

Raghavendra Rao Ananta (6):
  arm64: tlb: Refactor the core flush algorithm of __flush_tlb_range
  KVM: arm64: Implement  __kvm_tlb_flush_vmid_range()
  KVM: arm64: Implement kvm_arch_flush_remote_tlbs_range()
  KVM: arm64: Flush only the memslot after write-protect
  KVM: arm64: Invalidate the table entries upon a range
  KVM: arm64: Use TLBI range-based intructions for unmap

 arch/arm64/include/asm/kvm_asm.h   |   3 +
 arch/arm64/include/asm/kvm_host.h  |   3 +
 arch/arm64/include/asm/tlbflush.h  | 108 +++++++++++++++--------------
 arch/arm64/kvm/hyp/nvhe/hyp-main.c |  11 +++
 arch/arm64/kvm/hyp/nvhe/tlb.c      |  37 ++++++++++
 arch/arm64/kvm/hyp/pgtable.c       |  44 +++++++++---
 arch/arm64/kvm/hyp/vhe/tlb.c       |  35 ++++++++++
 arch/arm64/kvm/mmu.c               |  13 +++-
 8 files changed, 192 insertions(+), 62 deletions(-)