On Sat, Mar 2, 2024 at 1:49 AM Andrew Jones <ajones@xxxxxxxxxxxxxxxx> wrote:
On Wed, Feb 28, 2024 at 05:01:23PM -0800, Atish Patra wrote:
PMU Snapshot function allows to minimize the number of traps when the
guest access configures/access the hpmcounters. If the snapshot feature
is enabled, the hypervisor updates the shared memory with counter
data and state of overflown counters. The guest can just read the
shared memory instead of trap & emulate done by the hypervisor.
This patch doesn't implement the counter overflow yet.
Reviewed-by: Anup Patel <anup@xxxxxxxxxxxxxx>
Signed-off-by: Atish Patra <atishp@xxxxxxxxxxxx>
---
arch/riscv/include/asm/kvm_vcpu_pmu.h | 7 ++
arch/riscv/kvm/vcpu_pmu.c | 120 +++++++++++++++++++++++++-
arch/riscv/kvm/vcpu_sbi_pmu.c | 3 +
drivers/perf/riscv_pmu_sbi.c | 2 +-
4 files changed, 129 insertions(+), 3 deletions(-)
diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
index 395518a1664e..586bab84be35 100644
--- a/arch/riscv/include/asm/kvm_vcpu_pmu.h
+++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
@@ -50,6 +50,10 @@ struct kvm_pmu {
bool init_done;
/* Bit map of all the virtual counter used */
DECLARE_BITMAP(pmc_in_use, RISCV_KVM_MAX_COUNTERS);
+ /* The address of the counter snapshot area (guest physical address) */
+ gpa_t snapshot_addr;
+ /* The actual data of the snapshot */
+ struct riscv_pmu_snapshot_data *sdata;
};
#define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu_context)
@@ -85,6 +89,9 @@ int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_ba
int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
struct kvm_vcpu_sbi_return *retdata);
void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu);
+int kvm_riscv_vcpu_pmu_setup_snapshot(struct kvm_vcpu *vcpu, unsigned long saddr_low,
+ unsigned long saddr_high, unsigned long flags,
+ struct kvm_vcpu_sbi_return *retdata);
I prefer to name this function
kvm_riscv_vcpu_pmu_snapshot_set_shmem
Sure.
void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu);
void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu);
diff --git a/arch/riscv/kvm/vcpu_pmu.c b/arch/riscv/kvm/vcpu_pmu.c
index 29bf4ca798cb..74865e6050a1 100644
--- a/arch/riscv/kvm/vcpu_pmu.c
+++ b/arch/riscv/kvm/vcpu_pmu.c
@@ -311,6 +311,81 @@ int kvm_riscv_vcpu_pmu_read_hpm(struct kvm_vcpu *vcpu, unsigned int csr_num,
return ret;
}
+static void kvm_pmu_clear_snapshot_area(struct kvm_vcpu *vcpu)
+{
+ struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
+ int snapshot_area_size = sizeof(struct riscv_pmu_snapshot_data);
+
+ if (kvpmu->sdata) {
+ memset(kvpmu->sdata, 0, snapshot_area_size);
+ if (kvpmu->snapshot_addr != INVALID_GPA)
It's a KVM bug if we have non-null sdata but snapshot_addr is INVALID_GPA,
right? Maybe we should warn if we see that. We can also move the memset
inside the if block.
Added a warning.
+ kvm_vcpu_write_guest(vcpu, kvpmu->snapshot_addr,
+ kvpmu->sdata, snapshot_area_size);
+ kfree(kvpmu->sdata);
+ kvpmu->sdata = NULL;
+ }
+ kvpmu->snapshot_addr = INVALID_GPA;
+}
+
+int kvm_riscv_vcpu_pmu_setup_snapshot(struct kvm_vcpu *vcpu, unsigned long saddr_low,
+ unsigned long saddr_high, unsigned long flags,
+ struct kvm_vcpu_sbi_return *retdata)
+{
+ struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
+ int snapshot_area_size = sizeof(struct riscv_pmu_snapshot_data);
+ int sbiret = 0;
+ gpa_t saddr;
+ unsigned long hva;
+ bool writable;
+
+ if (!kvpmu) {
+ sbiret = SBI_ERR_INVALID_PARAM;
+ goto out;
+ }
Need to check that flags is zero or return SBI_ERR_INVALID_PARAM.
Fixed.
+
+ if (saddr_low == -1 && saddr_high == -1) {
We introduced SBI_STA_SHMEM_DISABLE for these magic -1's for STA. Since
SBI is using the -1 approach for all its shmem, then maybe we should
rename SBI_STA_SHMEM_DISABLE to SBI_SHMEM_DISABLE and then use them here
too.
Fixed
+ kvm_pmu_clear_snapshot_area(vcpu);
+ return 0;
+ }
+
+ saddr = saddr_low;
+
+ if (saddr_high != 0) {
+ if (IS_ENABLED(CONFIG_32BIT))
+ saddr |= ((gpa_t)saddr << 32);
+ else
+ sbiret = SBI_ERR_INVALID_ADDRESS;
+ goto out;
+ }
+
+ if (kvm_is_error_gpa(vcpu->kvm, saddr)) {
+ sbiret = SBI_ERR_INVALID_PARAM;
+ goto out;
+ }
Does the check above provide anything more than what the check below does?
You are correct. I have removed the check
+
+ hva = kvm_vcpu_gfn_to_hva_prot(vcpu, saddr >> PAGE_SHIFT, &writable);
+ if (kvm_is_error_hva(hva) || !writable) {
+ sbiret = SBI_ERR_INVALID_ADDRESS;
+ goto out;
+ }
+
+ kvpmu->snapshot_addr = saddr;
+ kvpmu->sdata = kzalloc(snapshot_area_size, GFP_ATOMIC);
+ if (!kvpmu->sdata)
Should reset snapshot_addr to INVALID_GPA here on error. Or maybe we
should just set snapshot_addr to saddr at the bottom of this function if
we make it.
Done.
+ return -ENOMEM;
+
+ if (kvm_vcpu_write_guest(vcpu, saddr, kvpmu->sdata, snapshot_area_size)) {
+ kfree(kvpmu->sdata);
+ kvpmu->snapshot_addr = INVALID_GPA;
+ sbiret = SBI_ERR_FAILURE;
I agree we should return this SBI error for this case, but unfortunately
the spec is missing the
SBI_ERR_FAILED - The request failed for unspecified or unknown other reasons.
that we have for other SBI functions. I guess we should keep the code like
this and open a PR to the spec.
I have created a blanket github issue for now. I will send a PR.
+ }
+
+out:
+ retdata->err_val = sbiret;
+
+ return 0;
+}
+
int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu,
struct kvm_vcpu_sbi_return *retdata)
{
@@ -344,20 +419,33 @@ int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
int i, pmc_index, sbiret = 0;
struct kvm_pmc *pmc;
int fevent_code;
+ bool snap_flag_set = flags & SBI_PMU_START_FLAG_INIT_FROM_SNAPSHOT;
This function should confirm no undefined bits are set in flags and the
spec should specify that the reserved flags must be zero otherwise an
invalid param will be returned.
Also here would should confirm that only one of the two flags is set,
otherwise return invalid param, as they've specified to be mutually
exclusive.
That makes sense. Update the same github issue.
(https://github.com/riscv-non-isa/riscv-sbi-doc/issues/145)
I will make the necessary changes in a separate series after the spec is merged.
Regarding the spec, the note about the counter value not being modified
unless SBI_PMU_START_SET_INIT_VALUE is set should be modified to state
unless either of the two flags are set (so I think we need another spec
PR).
(The same flags checking/specifying comments apply to the other functions
with flags too.)
Noted (https://github.com/riscv-non-isa/riscv-sbi-doc/issues/146).
if (kvm_pmu_validate_counter_mask(kvpmu, ctr_base, ctr_mask) < 0) {
sbiret = SBI_ERR_INVALID_PARAM;
goto out;
}
+ if (snap_flag_set && kvpmu->snapshot_addr == INVALID_GPA) {
+ sbiret = SBI_ERR_NO_SHMEM;
+ goto out;
+ }
+
/* Start the counters that have been configured and requested by the guest */
for_each_set_bit(i, &ctr_mask, RISCV_MAX_COUNTERS) {
pmc_index = i + ctr_base;
if (!test_bit(pmc_index, kvpmu->pmc_in_use))
continue;
pmc = &kvpmu->pmc[pmc_index];
- if (flags & SBI_PMU_START_FLAG_SET_INIT_VALUE)
+ if (flags & SBI_PMU_START_FLAG_SET_INIT_VALUE) {
pmc->counter_val = ival;
+ } else if (snap_flag_set) {
+ kvm_vcpu_read_guest(vcpu, kvpmu->snapshot_addr, kvpmu->sdata,
+ sizeof(struct riscv_pmu_snapshot_data));
The snapshot read should be outside the for_each_set_bit() loop and we
should warn and abort the counter starting if the read fails.
Fixed. This should also fall under the SBI_ERR_FAILURE category.
+ /* The counter index in the snapshot are relative to the counter base */
+ pmc->counter_val = kvpmu->sdata->ctr_values[i];
+ }
+
if (pmc->cinfo.type == SBI_PMU_CTR_TYPE_FW) {
fevent_code = get_event_code(pmc->event_idx);
if (fevent_code >= SBI_PMU_FW_MAX) {
@@ -398,14 +486,21 @@ int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
{
struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
int i, pmc_index, sbiret = 0;
+ u64 enabled, running;
struct kvm_pmc *pmc;
int fevent_code;
+ bool snap_flag_set = flags & SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT;
- if (kvm_pmu_validate_counter_mask(kvpmu, ctr_base, ctr_mask) < 0) {
+ if ((kvm_pmu_validate_counter_mask(kvpmu, ctr_base, ctr_mask) < 0)) {
Added unnecessary () here.
Fixed.
sbiret = SBI_ERR_INVALID_PARAM;
goto out;
}
+ if (snap_flag_set && kvpmu->snapshot_addr == INVALID_GPA) {
+ sbiret = SBI_ERR_NO_SHMEM;
+ goto out;
+ }
+
/* Stop the counters that have been configured and requested by the guest */
for_each_set_bit(i, &ctr_mask, RISCV_MAX_COUNTERS) {
pmc_index = i + ctr_base;
@@ -438,9 +533,28 @@ int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
} else {
sbiret = SBI_ERR_INVALID_PARAM;
}
+
+ if (snap_flag_set && !sbiret) {
+ if (pmc->cinfo.type == SBI_PMU_CTR_TYPE_FW)
+ pmc->counter_val = kvpmu->fw_event[fevent_code].value;
+ else if (pmc->perf_event)
+ pmc->counter_val += perf_event_read_value(pmc->perf_event,
+ &enabled, &running);
+ /* TODO: Add counter overflow support when sscofpmf support is added */
+ kvpmu->sdata->ctr_values[i] = pmc->counter_val;
+ kvm_vcpu_write_guest(vcpu, kvpmu->snapshot_addr, kvpmu->sdata,
+ sizeof(struct riscv_pmu_snapshot_data));
Should just set a boolean here saying that the snapshot needs an update
and then do the update outside the for_each_set_bit loop.
Done.
+ }
+
if (flags & SBI_PMU_STOP_FLAG_RESET) {
pmc->event_idx = SBI_PMU_EVENT_IDX_INVALID;
clear_bit(pmc_index, kvpmu->pmc_in_use);
+ if (snap_flag_set) {
+ /* Clear the snapshot area for the upcoming deletion event */
+ kvpmu->sdata->ctr_values[i] = 0;
+ kvm_vcpu_write_guest(vcpu, kvpmu->snapshot_addr, kvpmu->sdata,
+ sizeof(struct riscv_pmu_snapshot_data));
The spec isn't clear on this (so we should clarify it), but I'd expect
that a caller who set both the reset and the snapshot flag would want
the snapshot from before the reset when this call completes and then
assume that when they start counting again, and look at the snapshot
again, that those new counts would be from the reset values. Or maybe
not :-) Maybe they want to do a reset and take a snapshot in order to
look at the snapshot and confirm the reset happened? Either way, it
seems we should only do one of the two here. Either update the snapshot
before resetting, and not again after reset, or reset and then update
the snapshot (with no need to update before).
The reset call should happen when the event is deleted by the perf
framework in supervisor.
If we don't clear the values, the shared memory may have stale data of
last read counters
which is not ideal. That's why, I am clearing it upon resetting.
