x86,kvm: move qemu/guest FPU switching out to vcpu_run

Currently, every time a VCPU is scheduled out, the host kernel will
first save the guest FPU/xstate context, then load the qemu userspace
FPU context, only to then immediately save the qemu userspace FPU
context back to memory. When scheduling in a VCPU, the same extraneous
FPU loads and saves are done.

This could be avoided by moving from a model where the guest FPU is
loaded and stored with preemption disabled, to a model where the
qemu userspace FPU is swapped out for the guest FPU context for
the duration of the KVM_RUN ioctl.

This is done under the VCPU mutex, which is also taken when other
tasks inspect the VCPU FPU context, so the code should already be
safe for this change. That should come as no surprise, given that
s390 already has this optimization.

This can fix a bug where KVM calls get_user_pages while owning the
FPU, and the file system ends up requesting the FPU again:

    [258270.527947]  __warn+0xcb/0xf0
    [258270.527948]  warn_slowpath_null+0x1d/0x20
    [258270.527951]  kernel_fpu_disable+0x3f/0x50
    [258270.527953]  __kernel_fpu_begin+0x49/0x100
    [258270.527955]  kernel_fpu_begin+0xe/0x10
    [258270.527958]  crc32c_pcl_intel_update+0x84/0xb0
    [258270.527961]  crypto_shash_update+0x3f/0x110
    [258270.527968]  crc32c+0x63/0x8a [libcrc32c]
    [258270.527975]  dm_bm_checksum+0x1b/0x20 [dm_persistent_data]
    [258270.527978]  node_prepare_for_write+0x44/0x70 [dm_persistent_data]
    [258270.527985]  dm_block_manager_write_callback+0x41/0x50 [dm_persistent_data]
    [258270.527988]  submit_io+0x170/0x1b0 [dm_bufio]
    [258270.527992]  __write_dirty_buffer+0x89/0x90 [dm_bufio]
    [258270.527994]  __make_buffer_clean+0x4f/0x80 [dm_bufio]
    [258270.527996]  __try_evict_buffer+0x42/0x60 [dm_bufio]
    [258270.527998]  dm_bufio_shrink_scan+0xc0/0x130 [dm_bufio]
    [258270.528002]  shrink_slab.part.40+0x1f5/0x420
    [258270.528004]  shrink_node+0x22c/0x320
    [258270.528006]  do_try_to_free_pages+0xf5/0x330
    [258270.528008]  try_to_free_pages+0xe9/0x190
    [258270.528009]  __alloc_pages_slowpath+0x40f/0xba0
    [258270.528011]  __alloc_pages_nodemask+0x209/0x260
    [258270.528014]  alloc_pages_vma+0x1f1/0x250
    [258270.528017]  do_huge_pmd_anonymous_page+0x123/0x660
    [258270.528021]  handle_mm_fault+0xfd3/0x1330
    [258270.528025]  __get_user_pages+0x113/0x640
    [258270.528027]  get_user_pages+0x4f/0x60
    [258270.528063]  __gfn_to_pfn_memslot+0x120/0x3f0 [kvm]
    [258270.528108]  try_async_pf+0x66/0x230 [kvm]
    [258270.528135]  tdp_page_fault+0x130/0x280 [kvm]
    [258270.528149]  kvm_mmu_page_fault+0x60/0x120 [kvm]
    [258270.528158]  handle_ept_violation+0x91/0x170 [kvm_intel]
    [258270.528162]  vmx_handle_exit+0x1ca/0x1400 [kvm_intel]

No performance changes were detected in quick ping-pong tests on
my 4 socket system, which is expected since an FPU+xstate load is
on the order of 0.1us, while ping-ponging between CPUs is on the
order of 20us, and somewhat noisy.

Cc: stable@vger.kernel.org
Signed-off-by: Rik van Riel <riel@redhat.com>
Suggested-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
[Fixed a bug where reset_vcpu called put_fpu without preceding load_fpu,
 which happened inside from KVM_CREATE_VCPU ioctl. - Radim]
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>

1 files changed, 13 insertions, 0 deletions

diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 977de5fb968b..62527e053ee4 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -536,7 +536,20 @@ struct kvm_vcpu_arch {
 	struct kvm_mmu_memory_cache mmu_page_cache;
 	struct kvm_mmu_memory_cache mmu_page_header_cache;
 
+	/*
+	 * QEMU userspace and the guest each have their own FPU state.
+	 * In vcpu_run, we switch between the user and guest FPU contexts.
+	 * While running a VCPU, the VCPU thread will have the guest FPU
+	 * context.
+	 *
+	 * Note that while the PKRU state lives inside the fpu registers,
+	 * it is switched out separately at VMENTER and VMEXIT time. The
+	 * "guest_fpu" state here contains the guest FPU context, with the
+	 * host PRKU bits.
+	 */
+	struct fpu user_fpu;
 	struct fpu guest_fpu;
+
 	u64 xcr0;
 	u64 guest_supported_xcr0;
 	u32 guest_xstate_size;