Skip to contentSkip to conversation

KVM: Coalesced IO cleanup and test

PATCH Applied Sean Christopherson 2 participants 5 messages Aug 28, 2024 – Oct 1, 2024

Subsystems (3)
lore
Share & tools
lore

Sean Christopherson

author reviewer
#0 lore raw

v2 cover KVM: Coalesced IO cleanup and test

Add a regression test for the bug fixed by commit 92f6d4130497 ("KVM:
Fix coalesced_mmio_has_room() to avoid premature userspace exit"), and then do additional clean up on the offending KVM code. I wrote the test mainly so that I was confident I actually understood Ilias' fix.

This applies on the aforementioned commit, which is sitting in kvm-x86/generic.

Fully tested on x86 and arm64, compile tested on RISC-V.

Ilias, I deliberately didn't include your Reviewed-by for the selftest, just in case I botched something during the (minor, in theory) rework.

v2:
- Collect a review. [Ilias]
- Fix a typo in patch 2's changelog. [Ilias]
- Rework the selftest to containerize the coalesced I/O ring and its state, to hopefully make it easier to extend the test to validate using a ring other than KVM's built-in default. [Ilias]

Sean Christopherson (2):
KVM: selftests: Add a test for coalesced MMIO (and PIO on x86)
KVM: Clean up coalesced MMIO ring full check

 tools/testing/selftests/kvm/Makefile          |   3 +
 .../testing/selftests/kvm/coalesced_io_test.c | 236 ++++++++++++++++++
 .../testing/selftests/kvm/include/kvm_util.h  |  26 ++
 virt/kvm/coalesced_mmio.c                     |  29 +--
 4 files changed, 273 insertions(+), 21 deletions(-)
 create mode 100644 tools/testing/selftests/kvm/coalesced_io_test.c

base-commit: 728d17c2cb8cc5f9ac899173d0e9a67fb8887622

Signature · 1 line
2.46.0.295.g3b9ea8a38a-goog
author reviewer
#1 lore raw

PATCH v2 1/2 KVM: selftests: Add a test for coalesced MMIO (and PIO on x86)

Add a test to verify that KVM correctly exits (or not) when a vCPU's coalesced I/O ring is full (or isn't). Iterate over all legal starting points in the ring (with an empty ring), and verify that KVM doesn't exit until the ring is full.

Opportunistically verify that KVM exits immediately on non-coalesced I/O, either because the MMIO/PIO region was never registered, or because a previous region was unregistered.

This is a regression test for a KVM bug where KVM would prematurely exit due to bad math resulting in a false positive if the first entry in the ring was before the halfway mark. See commit 92f6d4130497 ("KVM: Fix coalesced_mmio_has_room() to avoid premature userspace exit").

Enable the test for x86, arm64, and risc-v, i.e. all architectures except s390, which doesn't have MMIO.

On x86, which has both MMIO and PIO, interleave MMIO and PIO into the same ring, as KVM shouldn't exit until a non-coalesced I/O is encountered, regardless of whether the ring is filled with MMIO, PIO, or both.

Lastly, wrap the coalesced I/O ring in a structure to prepare for a potential future where KVM supports multiple ring buffers beyond KVM's "default" built-in buffer.

---
 tools/testing/selftests/kvm/Makefile          |   3 +
 .../testing/selftests/kvm/coalesced_io_test.c | 236 ++++++++++++++++++
 .../testing/selftests/kvm/include/kvm_util.h  |  26 ++
 3 files changed, 265 insertions(+)
 create mode 100644 tools/testing/selftests/kvm/coalesced_io_test.c
tools/testing/selftests/kvm/Makefile +3 -0
index 48d32c5aa3eb..45cb70c048bb 100644
@@ -130,6 +130,7 @@ TEST_GEN_PROGS_x86_64 += x86_64/max_vcpuid_cap_test
 TEST_GEN_PROGS_x86_64 += x86_64/triple_fault_event_test
 TEST_GEN_PROGS_x86_64 += x86_64/recalc_apic_map_test
 TEST_GEN_PROGS_x86_64 += access_tracking_perf_test
+TEST_GEN_PROGS_x86_64 += coalesced_io_test
 TEST_GEN_PROGS_x86_64 += demand_paging_test
 TEST_GEN_PROGS_x86_64 += dirty_log_test
 TEST_GEN_PROGS_x86_64 += dirty_log_perf_test
@@ -165,6 +166,7 @@ TEST_GEN_PROGS_aarch64 += aarch64/vgic_lpi_stress
 TEST_GEN_PROGS_aarch64 += aarch64/vpmu_counter_access
 TEST_GEN_PROGS_aarch64 += access_tracking_perf_test
 TEST_GEN_PROGS_aarch64 += arch_timer
+TEST_GEN_PROGS_aarch64 += coalesced_io_test
 TEST_GEN_PROGS_aarch64 += demand_paging_test
 TEST_GEN_PROGS_aarch64 += dirty_log_test
 TEST_GEN_PROGS_aarch64 += dirty_log_perf_test
@@ -198,6 +200,7 @@ TEST_GEN_PROGS_s390x += kvm_binary_stats_test
 TEST_GEN_PROGS_riscv += riscv/sbi_pmu_test
 TEST_GEN_PROGS_riscv += riscv/ebreak_test
 TEST_GEN_PROGS_riscv += arch_timer
+TEST_GEN_PROGS_riscv += coalesced_io_test
 TEST_GEN_PROGS_riscv += demand_paging_test
 TEST_GEN_PROGS_riscv += dirty_log_test
 TEST_GEN_PROGS_riscv += get-reg-list
tools/testing/selftests/kvm/coalesced_io_test.c new file +236 -0

236 lines not on this page · raw message

tools/testing/selftests/kvm/include/kvm_util.h +26 -0
index 63c2aaae51f3..b297a84f7be5 100644
@@ -460,6 +460,32 @@ static inline uint32_t kvm_vm_reset_dirty_ring(struct kvm_vm *vm)
 	return __vm_ioctl(vm, KVM_RESET_DIRTY_RINGS, NULL);
 }
 
+static inline void kvm_vm_register_coalesced_io(struct kvm_vm *vm,
+						uint64_t address,
+						uint64_t size, bool pio)
+{
+	struct kvm_coalesced_mmio_zone zone = {
+		.addr = address,
+		.size = size,
+		.pio  = pio,
+	};
+
+	vm_ioctl(vm, KVM_REGISTER_COALESCED_MMIO, &zone);
+}
+
+static inline void kvm_vm_unregister_coalesced_io(struct kvm_vm *vm,
+						  uint64_t address,
+						  uint64_t size, bool pio)
+{
+	struct kvm_coalesced_mmio_zone zone = {
+		.addr = address,
+		.size = size,
+		.pio  = pio,
+	};
+
+	vm_ioctl(vm, KVM_UNREGISTER_COALESCED_MMIO, &zone);
+}
+
 static inline int vm_get_stats_fd(struct kvm_vm *vm)
 {
 	int fd = __vm_ioctl(vm, KVM_GET_STATS_FD, NULL);
Signature · 1 line
2.46.0.295.g3b9ea8a38a-goog
author reviewer
#2 lore raw

PATCH v2 2/2 KVM: Clean up coalesced MMIO ring full check

In reply to Sean Christopherson

Fold coalesced_mmio_has_room() into its sole caller, coalesced_mmio_write(), as it's really just a single line of code, has a goofy return value, and is unnecessarily brittle.

E.g. if coalesced_mmio_has_room() were to check ring->last directly, or the caller failed to use READ_ONCE(), KVM would be susceptible to TOCTOU attacks from userspace.

Opportunistically add a comment explaining why on earth KVM leaves one entry free, which may not be obvious to readers that aren't familiar with ring buffers.

No functional change intended.

---
 virt/kvm/coalesced_mmio.c | 29 ++++++++---------------------
 1 file changed, 8 insertions(+), 21 deletions(-)
virt/kvm/coalesced_mmio.c +8 -21
index 184c5c40c9c1..375d6285475e 100644
@@ -40,25 +40,6 @@ static int coalesced_mmio_in_range(struct kvm_coalesced_mmio_dev *dev,
 	return 1;
 }
 
-static int coalesced_mmio_has_room(struct kvm_coalesced_mmio_dev *dev, u32 last)
-{
-	struct kvm_coalesced_mmio_ring *ring;
-
-	/* Are we able to batch it ? */
-
-	/* last is the first free entry
-	 * check if we don't meet the first used entry
-	 * there is always one unused entry in the buffer
-	 */
-	ring = dev->kvm->coalesced_mmio_ring;
-	if ((last + 1) % KVM_COALESCED_MMIO_MAX == READ_ONCE(ring->first)) {
-		/* full */
-		return 0;
-	}
-
-	return 1;
-}
-
 static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 				struct kvm_io_device *this, gpa_t addr,
 				int len, const void *val)
@@ -72,9 +53,15 @@ static int coalesced_mmio_write(struct kvm_vcpu *vcpu,
 
 	spin_lock(&dev->kvm->ring_lock);
 
+	/*
+	 * last is the index of the entry to fill.  Verify userspace hasn't
+	 * set last to be out of range, and that there is room in the ring.
+	 * Leave one entry free in the ring so that userspace can differentiate
+	 * between an empty ring and a full ring.
+	 */
 	insert = READ_ONCE(ring->last);
-	if (!coalesced_mmio_has_room(dev, insert) ||
-	    insert >= KVM_COALESCED_MMIO_MAX) {
+	if (insert >= KVM_COALESCED_MMIO_MAX ||
+	    (insert + 1) % KVM_COALESCED_MMIO_MAX == READ_ONCE(ring->first)) {
 		spin_unlock(&dev->kvm->ring_lock);
 		return -EOPNOTSUPP;
 	}
Signature · 1 line
2.46.0.295.g3b9ea8a38a-goog
author reviewer Applied
#3 lore raw
In reply to Sean Christopherson
Quoted 9 lines · Sean Christopherson

Add a regression test for the bug fixed by commit 92f6d4130497 ("KVM:
Fix coalesced_mmio_has_room() to avoid premature userspace exit"), and then do additional clean up on the offending KVM code. I wrote the test mainly so that I was confident I actually understood Ilias' fix.

This applies on the aforementioned commit, which is sitting in kvm-x86/generic.

[...]

Applied to kvm-x86 generic, thanks!

[1/2] KVM: selftests: Add a test for coalesced MMIO (and PIO on x86)

      https://github.com/kvm-x86/linux/commit/215b3cb7a84f
[2/2] KVM: Clean up coalesced MMIO ring full check
      https://github.com/kvm-x86/linux/commit/e027ba1b83ad

--
https://github.com/kvm-x86/linux/tree/next

#4 lore raw
In reply to Sean Christopherson
This series was applied to riscv/linux.git (fixes) by Sean Christopherson <[email protected]>:

Hello:

This series was applied to riscv/linux.git (fixes) by Sean Christopherson <[email protected]>:

Quoted 9 lines · On Wed, 28 Aug 2024 11:14:44 -0700 you wrote:

Add a regression test for the bug fixed by commit 92f6d4130497 ("KVM:
Fix coalesced_mmio_has_room() to avoid premature userspace exit"), and then do additional clean up on the offending KVM code. I wrote the test mainly so that I was confident I actually understood Ilias' fix.

This applies on the aforementioned commit, which is sitting in kvm-x86/generic.

[...]

Here is the summary with links:
- [v2,1/2] KVM: selftests: Add a test for coalesced MMIO (and PIO on x86)

    https://git.kernel.org/riscv/c/215b3cb7a84f
  - [v2,2/2] KVM: Clean up coalesced MMIO ring full check
    https://git.kernel.org/riscv/c/e027ba1b83ad

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html