[RFC PATCH v2 2/2] lib: sbi_pmu: Add PMU snapshot implementation
Inochi Amaoto
inochiama at outlook.com
Wed Jan 10 16:35:33 PST 2024
SBI PMU snapshot is an opetional extension that allow supervisor to save
a snapshot about PMU counter status. which is useful for KVM snapshot.
Add the implementation of the PMU snapshot, and remove the related error.
Signed-off-by: Inochi Amaoto <inochiama at outlook.com>
---
include/sbi/sbi_pmu.h | 4 ++
include/sbi/sbi_shm.h | 6 +-
lib/sbi/sbi_ecall_pmu.c | 3 +-
lib/sbi/sbi_pmu.c | 143 ++++++++++++++++++++++++++++++++++++++--
4 files changed, 146 insertions(+), 10 deletions(-)
diff --git a/include/sbi/sbi_pmu.h b/include/sbi/sbi_pmu.h
index 7d32a4d..c755ed5 100644
--- a/include/sbi/sbi_pmu.h
+++ b/include/sbi/sbi_pmu.h
@@ -150,4 +150,8 @@ int sbi_pmu_ctr_cfg_match(unsigned long cidx_base, unsigned long cidx_mask,
int sbi_pmu_ctr_incr_fw(enum sbi_pmu_fw_event_code_id fw_id);
+int sbi_pmu_set_snapshot_shmem(unsigned long shmem_phys_lo,
+ unsigned long shmem_phys_hi,
+ unsigned long flags);
+
#endif
diff --git a/include/sbi/sbi_shm.h b/include/sbi/sbi_shm.h
index c50753b..69afe81 100644
--- a/include/sbi/sbi_shm.h
+++ b/include/sbi/sbi_shm.h
@@ -63,7 +63,7 @@ static inline int sbi_shmem_read_u16(struct sbi_shmem *shmem,
uint16_t* value)
{
uint16_t tmp;
- int ret = sbi_shmem_raw_write(shmem, offset, &tmp, 2);
+ int ret = sbi_shmem_raw_read(shmem, offset, &tmp, 2);
*value = le16_to_cpu(tmp);
@@ -75,7 +75,7 @@ static inline int sbi_shmem_read_u32(struct sbi_shmem *shmem,
uint32_t* value)
{
uint32_t tmp;
- int ret = sbi_shmem_raw_write(shmem, offset, &tmp, 4);
+ int ret = sbi_shmem_raw_read(shmem, offset, &tmp, 4);
*value = le32_to_cpu(tmp);
@@ -87,7 +87,7 @@ static inline int sbi_shmem_read_u64(struct sbi_shmem *shmem,
uint64_t* value)
{
uint64_t tmp;
- int ret = sbi_shmem_raw_write(shmem, offset, &tmp, 8);
+ int ret = sbi_shmem_raw_read(shmem, offset, &tmp, 8);
*value = le64_to_cpu(tmp);
diff --git a/lib/sbi/sbi_ecall_pmu.c b/lib/sbi/sbi_ecall_pmu.c
index 40a63a6..8e9a82e 100644
--- a/lib/sbi/sbi_ecall_pmu.c
+++ b/lib/sbi/sbi_ecall_pmu.c
@@ -74,7 +74,8 @@ static int sbi_ecall_pmu_handler(unsigned long extid, unsigned long funcid,
ret = sbi_pmu_ctr_stop(regs->a0, regs->a1, regs->a2);
break;
case SBI_EXT_PMU_SNAPSHOT_SET_SHMEM:
- /* fallthrough as OpenSBI doesn't support snapshot yet */
+ ret = sbi_pmu_set_snapshot_shmem(regs->a0, regs->a1, regs->a2);
+ break;
default:
ret = SBI_ENOTSUPP;
}
diff --git a/lib/sbi/sbi_pmu.c b/lib/sbi/sbi_pmu.c
index 6209ccc..5892537 100644
--- a/lib/sbi/sbi_pmu.c
+++ b/lib/sbi/sbi_pmu.c
@@ -16,6 +16,7 @@
#include <sbi/sbi_platform.h>
#include <sbi/sbi_pmu.h>
#include <sbi/sbi_scratch.h>
+#include <sbi/sbi_shm.h>
#include <sbi/sbi_string.h>
/** Information about hardware counters */
@@ -68,8 +69,22 @@ struct sbi_pmu_hart_state {
* and hence can optimally share the same memory.
*/
uint64_t fw_counters_data[SBI_PMU_FW_CTR_MAX];
+
+ /* Shared memory for pmu */
+ struct sbi_shmem shmem;
};
+#define PMU_SNAPSHOT_SHMEM_ALIGN 4096
+#define PMU_SNAPSHOT_SHMEM_SIZE 4096
+
+#define PMU_SNAPSHOT_COUNTER_OF_MAP_OFFSET 0x0000
+#define PMU_SNAPSHOT_COUNTER_OFFSET_BASE 0x0008
+#define PMU_SNAPSHOT_COUNTER_ENTRY_SIZE 8
+
+#define PMU_SNAPSHOT_COUNTER_OFFSET(i) \
+ (PMU_SNAPSHOT_COUNTER_OFFSET_BASE + \
+ PMU_SNAPSHOT_COUNTER_ENTRY_SIZE * (i))
+
/** Offset of pointer to PMU HART state in scratch space */
static unsigned long phs_ptr_offset;
@@ -333,6 +348,39 @@ static int pmu_ctr_enable_irq_hw(int ctr_idx)
return 0;
}
+static uint64_t pmu_ctr_read_hw_of(uint32_t cidx)
+{
+ struct sbi_scratch *scratch = sbi_scratch_thishart_ptr();
+ uint32_t overflow;
+
+ if (!sbi_hart_has_extension(scratch, SBI_HART_EXT_SSCOFPMF))
+ return 0;
+
+ if (cidx < (CSR_INSTRET - CSR_CYCLE))
+ return 0;
+#if __riscv_xlen == 32
+ overflow = csr_read_num(CSR_MHPMEVENT3H + cidx - 3) & MHPMEVENTH_OF;
+#else
+ overflow = csr_read_num(CSR_MHPMEVENT3 + cidx - 3) & MHPMEVENT_OF;
+#endif
+ if (overflow)
+ return 1;
+ else
+ return 0;
+}
+
+static uint64_t pmu_ctr_read_hw(uint32_t cidx)
+{
+ uint64_t cval;
+#if __riscv_xlen == 32
+ cval = csr_read_num(CSR_MCYCLE + cidx) |
+ csr_read_num(CSR_MCYCLEH + cidx) << BITS_PER_LONG;
+#else
+ cval = csr_read_num(CSR_MCYCLE + cidx);
+#endif
+ return cval;
+}
+
static void pmu_ctr_write_hw(uint32_t cidx, uint64_t ival)
{
#if __riscv_xlen == 32
@@ -428,33 +476,64 @@ static int pmu_ctr_start_fw(struct sbi_pmu_hart_state *phs,
return 0;
}
+static void pmu_ctrl_read_snapshot(struct sbi_pmu_hart_state *phs,
+ int i, int eid_type,
+ uint64_t of_val, uint64_t *ival)
+{
+ sbi_shmem_read_u64(&phs->shmem,
+ PMU_SNAPSHOT_COUNTER_OFFSET(i),
+ ival);
+
+ if (eid_type == SBI_PMU_EVENT_TYPE_FW)
+ return;
+
+ if (BIT(i) & of_val)
+ *ival |= MHPMEVENT_OF;
+}
+
int sbi_pmu_ctr_start(unsigned long cbase, unsigned long cmask,
unsigned long flags, uint64_t ival)
{
+ struct sbi_scratch *scratch = sbi_scratch_thishart_ptr();
struct sbi_pmu_hart_state *phs = pmu_thishart_state_ptr();
int event_idx_type;
uint32_t event_code;
int ret = SBI_EINVAL;
- bool bUpdate = false;
+ bool bUpdate = false, sUpdate = false;
int i, cidx;
- uint64_t edata;
+ uint64_t edata, of_val = 0;
if ((cbase + sbi_fls(cmask)) >= total_ctrs)
return ret;
- if (flags & SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT)
- return SBI_ENO_SHMEM;
+ if (flags & SBI_PMU_START_FLAG_INIT_FROM_SNAPSHOT)
+ sUpdate = true;
if (flags & SBI_PMU_START_FLAG_SET_INIT_VALUE)
bUpdate = true;
+ if (sUpdate) {
+ if (sbi_hart_has_extension(scratch, SBI_HART_EXT_SSCOFPMF))
+ sbi_shmem_read_u64(&phs->shmem,
+ PMU_SNAPSHOT_COUNTER_OF_MAP_OFFSET,
+ &of_val);
+
+ /* Snapshot initialization implicit value update */
+ bUpdate = true;
+ }
+
for_each_set_bit(i, &cmask, BITS_PER_LONG) {
cidx = i + cbase;
event_idx_type = pmu_ctr_validate(phs, cidx, &event_code);
if (event_idx_type < 0)
/* Continue the start operation for other counters */
continue;
- else if (event_idx_type == SBI_PMU_EVENT_TYPE_FW) {
+
+ if (sUpdate)
+ pmu_ctrl_read_snapshot(phs, i, event_idx_type,
+ of_val, &ival);
+
+ if (event_idx_type == SBI_PMU_EVENT_TYPE_FW) {
edata = (event_code == SBI_PMU_FW_PLATFORM) ?
phs->fw_counters_data[cidx - num_hw_ctrs]
: 0x0;
@@ -538,13 +617,16 @@ int sbi_pmu_ctr_stop(unsigned long cbase, unsigned long cmask,
int ret = SBI_EINVAL;
int event_idx_type;
uint32_t event_code;
+ bool record = false;
+ uint64_t counter_val;
+ uint64_t of_val = 0;
int i, cidx;
if ((cbase + sbi_fls(cmask)) >= total_ctrs)
return SBI_EINVAL;
if (flag & SBI_PMU_STOP_FLAG_TAKE_SNAPSHOT)
- return SBI_ENO_SHMEM;
+ record = true;
for_each_set_bit(i, &cmask, BITS_PER_LONG) {
cidx = i + cbase;
@@ -558,12 +640,30 @@ int sbi_pmu_ctr_stop(unsigned long cbase, unsigned long cmask,
else
ret = pmu_ctr_stop_hw(cidx);
+ if (record) {
+ if (event_idx_type == SBI_PMU_EVENT_TYPE_FW)
+ ret = sbi_pmu_ctr_fw_read(cidx, &counter_val);
+ else {
+ counter_val = pmu_ctr_read_hw(cidx);
+ of_val = pmu_ctr_read_hw_of(cidx) << cidx;
+ }
+
+ sbi_shmem_write_u64(&phs->shmem,
+ PMU_SNAPSHOT_COUNTER_OFFSET(i),
+ counter_val);
+ }
+
if (cidx > (CSR_INSTRET - CSR_CYCLE) && flag & SBI_PMU_STOP_FLAG_RESET) {
phs->active_events[cidx] = SBI_PMU_EVENT_IDX_INVALID;
pmu_reset_hw_mhpmevent(cidx);
}
}
+ if (record)
+ sbi_shmem_write_u64(&phs->shmem,
+ PMU_SNAPSHOT_COUNTER_OF_MAP_OFFSET,
+ of_val);
+
return ret;
}
@@ -946,6 +1046,37 @@ static void pmu_reset_event_map(struct sbi_pmu_hart_state *phs)
phs->fw_counters_started = 0;
}
+int sbi_pmu_set_snapshot_shmem(unsigned long shmem_phys_lo,
+ unsigned long shmem_phys_hi,
+ unsigned long flags)
+{
+ struct sbi_pmu_hart_state *phs = pmu_thishart_state_ptr();
+ int ret;
+
+ if (flags)
+ return SBI_EINVAL;
+
+ /* remove the stop flag */
+ if (shmem_phys_lo == -1 && shmem_phys_hi == -1) {
+ sbi_shmem_region_clear(&phs->shmem);
+ return 0;
+ }
+
+ ret = sbi_shmem_region_init(shmem_phys_lo, shmem_phys_hi,
+ PMU_SNAPSHOT_SHMEM_SIZE,
+ PMU_SNAPSHOT_SHMEM_ALIGN,
+ &phs->shmem);
+ if (ret)
+ return ret;
+
+ ret = sbi_shmem_region_check(&phs->shmem,
+ SBI_SHMEM_M_RW | SBI_SHMEM_S_RW);
+ if (ret)
+ return ret;
+
+ return 0;
+}
+
const struct sbi_pmu_device *sbi_pmu_get_device(void)
{
return pmu_dev;
--
2.43.0
More information about the opensbi
mailing list