KVM: arm64: Statistical Profiling Extension Tests
cover KVM: arm64: Statistical Profiling Extension Tests
This series implements tests exercising the Statistical Profiling Extensions.
This was tested with associated unmerged kernel [1] and QEMU [2] series.
Depending on the comments, I can easily add other tests checking more configs, additional events and testing migration too. I hope this can be useful when respinning both series.
All SPE tests can be launched with:
./run_tests.sh -g spe
Tests also can be launched individually. For example:
./arm-run arm/spe.flat -append 'spe-buffer'
The series can be found at:
https://github.com/eauger/kut/tree/spe_rfc
References:
[1] [PATCH v2 00/18] arm64: KVM: add SPE profiling support
[2] [PATCH 0/7] target/arm: Add vSPE support to KVM guest
Eric Auger (4):
arm64: Move get_id_aa64dfr0() in processor.h
spe: Probing and Introspection Test
spe: Add profiling buffer test
spe: Test Profiling Buffer Events
arm/Makefile.common | 1 + arm/pmu.c | 1 - arm/spe.c | 463 ++++++++++++++++++++++++++++++++++++++ arm/unittests.cfg | 24 ++ lib/arm64/asm/barrier.h | 1 + lib/arm64/asm/processor.h | 5 + 6 files changed, 494 insertions(+), 1 deletion(-) create mode 100644 arm/spe.c
Signature · 1 line
2.21.3
RFC 1/4 arm64: Move get_id_aa64dfr0() in processor.h
We plan to use get_id_aa64dfr0() for SPE tests.
So let's move this latter in processor.h header.
--- arm/pmu.c | 1 - lib/arm64/asm/processor.h | 5 +++++ 2 files changed, 5 insertions(+), 1 deletion(-)
arm/pmu.c +0 -1
@@ -167,7 +167,6 @@ static void test_overflow_interrupt(void) {}#define ID_DFR0_PMU_V3_8_5 0b0110#define ID_DFR0_PMU_IMPDEF 0b1111 -static inline uint32_t get_id_aa64dfr0 (void ) {return read_sysreg (id_aa64dfr0_el1); }static inline uint32_t get_pmcr (void ) {return read_sysreg (pmcr_el0); }static inline void set_pmcr (uint32_t v) {write_sysreg (v, pmcr_el0); }static inline uint64_t get_pmccntr (void ) {return read_sysreg (pmccntr_el0); }
lib/arm64/asm/processor.h +5 -0
@@ -88,6 +88,11 @@ static inline uint64_t get_mpidr(void)return read_sysreg (mpidr_el1); } +static inline uint64_t get_id_aa64dfr0 (void ) +{ +return read_sysreg (id_aa64dfr0_el1); +} +#define MPIDR_HWID_BITMASK 0xff00ffffffextern int mpidr_to_cpu (uint64_t mpidr);
Signature · 6 lines
2.21.3 _______________________________________________ kvmarm mailing list [email protected] https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
RFC 2/4 spe: Probing and Introspection Test
In reply to Eric AugerTest whether Statistical Profiling Extensions (SPE) are supported and in the positive collect dimensioning data from the IDR registers. The First test only validates those.
--- arm/Makefile.common | 1 + arm/spe.c | 163 ++++++++++++++++++++++++++++++++++++++++++++ arm/unittests.cfg | 8 +++ 3 files changed, 172 insertions(+) create mode 100644 arm/spe.c
arm/Makefile.common +1 -0
@@ -8,6 +8,7 @@ tests-common = $(TEST_DIR)/selftest.flat tests-common += $(TEST_DIR)/spinlock-test.flat tests-common += $(TEST_DIR)/pci-test.flat tests-common += $(TEST_DIR)/pmu.flat +tests-common += $(TEST_DIR)/spe.flat tests-common += $(TEST_DIR)/gic.flat tests-common += $(TEST_DIR)/psci.flat tests-common += $(TEST_DIR)/sieve.flat
arm/spe.c new file +163 -0
163 lines not on this page · raw message
arm/unittests.cfg +8 -0
@@ -134,6 +134,14 @@ extra_params = -append 'pmu-overflow-interrupt' #groups = pmu #accel = tcg +[spe-introspection] +file = spe.flat +groups = spe +arch = arm64 +extra_params = -append 'spe-introspection' +accel = kvm +arch = arm64 + # Test GIC emulation [gicv2-ipi] file = gic.flat
Signature · 6 lines
2.21.3 _______________________________________________ kvmarm mailing list [email protected] https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
RFC 3/4 spe: Add profiling buffer test
In reply to Eric AugerAdd the code to prepare for profiling at EL1. The code under profiling is a simple loop doing memory addresses. We simply check the profiling buffer write position increases, ie. the buffer gets filled. No event is expected.
---
To make sure no buffer full events is likely to be received, the number of to be collected events should be estimated. This needs to be done. Same for next patch. I tried to read PMSICR.COUNT after a single iteration but I get a value greated than the set interval so I wonder whether this is a bug or rather than reading this value gives unpredictable value.
--- arm/spe.c | 161 ++++++++++++++++++++++++++++++++++++++++ arm/unittests.cfg | 8 ++ lib/arm64/asm/barrier.h | 1 + 3 files changed, 170 insertions(+)
arm/spe.c +161 -0
@@ -14,9 +14,11 @@#include "errata.h" #include "asm/barrier.h" #include "asm/sysreg.h" +#include "asm/page.h" #include "asm/processor.h" #include "alloc_page.h" #include <bitops.h> +#include "alloc.h" struct spe {int min_interval;
@@ -27,6 +29,10 @@ struct spe {bool fe_cap;int align;void *buffer; +uint64_t pmbptr_el1; +uint64_t pmblimitr_el1; +uint64_t pmsirr_el1; +uint64_t pmscr_el1;bool unique_record_size; };
@@ -36,6 +42,7 @@ static struct spe spe;static bool spe_probe (void ) {return false; }static void test_spe_introspection (void ) { } +static void test_spe_buffer (void ) { }#else
@@ -59,10 +66,35 @@ static void test_spe_introspection(void) { }#define PMSIDR_EL1_COUNTSIZE_SHIFT 16#define PMSIDR_EL1_COUNTSIZE_MASK 0xFUL +#define PMSIRR_EL1_INTERVAL_SHIFT 8 +#define PMSIRR_EL1_INTERVAL_MASK 0xFFFFFF + +#define PMSFCR_EL1_FE 0x1 +#define PMSFCR_EL1_FT 0x2 +#define PMSFCR_EL1_FL 0x4 +#define PMSFCR_EL1_B 0x10000 +#define PMSFCR_EL1_LD 0x20000 +#define PMSFCR_EL1_ST 0x40000 + +#define PMSCR_EL1 sys_reg(3, 0, 9, 9, 0) +#define PMSICR_EL1 sys_reg(3, 0, 9, 9, 2) +#define PMSIRR_EL1 sys_reg(3, 0, 9, 9, 3) +#define PMSFCR_EL1 sys_reg(3, 0, 9, 9, 4) +#define PMSEVFR_EL1 sys_reg(3, 0, 9, 9, 5)#define PMSIDR_EL1 sys_reg(3, 0, 9, 9, 7) +#define PMBLIMITR_EL1 sys_reg(3, 0, 9, 10, 0) +#define PMBPTR_EL1 sys_reg(3, 0, 9, 10, 1) +#define PMBSR_EL1 sys_reg(3, 0, 9, 10, 3)#define PMBIDR_EL1 sys_reg(3, 0, 9, 10, 7) +#define PMBLIMITR_EL1_E 0x1 + +#define PMSCR_EL1_E1SPE 0x2 +#define PMSCR_EL1_PA 0x10 +#define PMSCR_EL1_TS 0x20 +#define PMSCR_EL1_PCT 0x40 +static int min_interval (uint8_t idr_bits) {switch (idr_bits) {
@@ -138,6 +170,131 @@ static void test_spe_introspection(void) "PMSIDR_EL1: Minimal sampling interval = %d", spe.min_interval); } +static void mem_access_loop (void *addr,int loop,uint64_t pmblimitr) +{ +asmvolatile ( + " msr_s " xstr(PMBLIMITR_EL1) ", %[pmblimitr]\n" +" isb\n" +" mov x10, %[loop]\n" +"1: sub x10, x10, #1\n" +" ldr x9, [%[addr]]\n" +" cmp x10, #0x0\n" +" b.gt 1b\n" +" bfxil %[pmblimitr], xzr, 0, 1\n" + " msr_s " xstr(PMBLIMITR_EL1) ", %[pmblimitr]\n" +" isb\n" + : + : [addr] "r" (addr), [pmblimitr] "r" (pmblimitr), [loop] "r" (loop) + : "x8", "x9", "x10", "cc"); +} + +char null_buff[PAGE_SIZE ] = {}; + +static void reset (void ) +{ +/* erase the profiling buffer, reset the start and limit addresses */ + spe.pmbptr_el1 = (uint64_t )spe.buffer; + spe.pmblimitr_el1 = (uint64_t )(spe.buffer +PAGE_SIZE ); +write_sysreg_s (spe.pmbptr_el1,PMBPTR_EL1 ); +write_sysreg_s (spe.pmblimitr_el1,PMBLIMITR_EL1 ); +isb (); + +/* Drain any buffered data */ +psb_csync (); +dsb (nsh); + +memset (spe.buffer,0 ,PAGE_SIZE ); + +/* reset the syndrome register */ +write_sysreg_s (0 ,PMBSR_EL1 ); + +/* SW must write 0 to PMSICR_EL1 before enabling sampling profiling */ +write_sysreg_s (0 ,PMSICR_EL1 ); + +/* Filtering disabled */ +write_sysreg_s (0 ,PMSFCR_EL1 ); + +/* Interval Reload Register */ + spe.pmsirr_el1 = (spe.min_interval &PMSIRR_EL1_INTERVAL_MASK ) <<PMSIRR_EL1_INTERVAL_SHIFT ; +write_sysreg_s (spe.pmsirr_el1,PMSIRR_EL1 ); + +/* Control Registrer */ + spe.pmscr_el1 =PMSCR_EL1_E1SPE |PMSCR_EL1_TS |PMSCR_EL1_PCT |PMSCR_EL1_PA ; +write_sysreg_s (spe.pmscr_el1,PMSCR_EL1 ); + +/* Make sure the syndrome register is void */ +write_sysreg_s (0 ,PMBSR_EL1 ); +} + +static inline void drain (void ) +{ +/* ensure profiling data are written */ +psb_csync (); +dsb (nsh); +} + +static void test_spe_buffer (void ) +{ +uint64_t pmbsr_el1, val1, val2; +void *addr =malloc (10 *PAGE_SIZE ); + +reset (); + + val1 =read_sysreg_s (PMBPTR_EL1 ); + val2 =read_sysreg_s (PMBLIMITR_EL1 ); +report (val1 == spe.pmbptr_el1 && val2 == spe.pmblimitr_el1, +"PMBPTR_EL1, PMBLIMITR_EL1: reset" ); + + val1 =read_sysreg_s (PMSIRR_EL1 ); +report (val1 == spe.pmsirr_el1,"PMSIRR_EL1: Sampling interval set to %d" , spe.min_interval); + + val1 =read_sysreg_s (PMSCR_EL1 ); +report (val1 == spe.pmscr_el1,"PMSCR_EL1: EL1 Statistical Profiling enabled" ); + + val1 =read_sysreg_s (PMSFCR_EL1 ); +report (!val1,"PMSFCR_EL1: No Filter Control" ); + +report (!memcmp (spe.buffer, null_buff,PAGE_SIZE ), +"Profiling buffer empty before profiling" ); + + val1 =read_sysreg_s (PMBSR_EL1 ); +report (!val1,"PMBSR_EL1: Syndrome Register void before profiling" ); + +mem_access_loop (addr,1 , spe.pmblimitr_el1 |PMBLIMITR_EL1_E ); +drain (); + val1 =read_sysreg_s (PMSICR_EL1 ); +/* +* TODO: the value read in PMSICR_EL1.count currently seems not consistent with +* programmed interval. Getting a good value would allow to estimate the number +* of records to be collected in next step. +*/ +report_info ("count for a single iteration: PMSICR_EL1.count=%lld interval=%d" , + val1 &GENMASK_ULL (31 ,0 ), spe.min_interval); + +/* Stuff to profile */ + +mem_access_loop (addr,1000000 , spe.pmblimitr_el1 |PMBLIMITR_EL1_E ); + +/* end of stuff to profile */ + +drain (); + +report (memcmp (spe.buffer, null_buff,PAGE_SIZE ),"Profiling buffer filled" ); + + val1 =read_sysreg_s (PMBPTR_EL1 ); + val2 = val1 - (uint64_t )spe.buffer; +report (val1 > (uint64_t )spe.buffer, +"PMBPTR_EL1: Current write position has increased: 0x%lx -> 0x%lx (%ld bytes)" , + (uint64_t )spe.buffer, val1, val2); +if (spe.unique_record_size) +report_info ("This corresponds to %ld record(s) of %d bytes" , + val2 / spe.maxsize, spe.maxsize); + pmbsr_el1 =read_sysreg_s (PMBSR_EL1 ); +report (!pmbsr_el1,"PMBSR_EL1: no event" ); + +free (addr); +} +#endif int main (int argc,char *argv[])
@@ -156,6 +313,10 @@ int main(int argc, char *argv[])report_prefix_push (argv[1 ]);test_spe_introspection ();report_prefix_pop (); + }else if (strcmp (argv[1 ],"spe-buffer" ) ==0 ) { +report_prefix_push (argv[1 ]); +test_spe_buffer (); +report_prefix_pop (); }else {report_abort ("Unknown sub-test '%s'" , argv[1 ]); }
arm/unittests.cfg +8 -0
@@ -142,6 +142,14 @@ extra_params = -append 'spe-introspection' accel = kvm arch = arm64 +[spe-buffer] +file = spe.flat +groups = spe +arch = arm64 +extra_params = -append 'spe-buffer' +accel = kvm +arch = arm64 + # Test GIC emulation [gicv2-ipi] file = gic.flat
lib/arm64/asm/barrier.h +1 -0
@@ -23,5 +23,6 @@#define smp_mb () dmb(ish)#define smp_rmb () dmb(ishld)#define smp_wmb () dmb(ishst) +#define psb_csync () asm volatile("hint #17" : : : "memory")#endif /* _ASMARM64_BARRIER_H_ */
Signature · 6 lines
2.21.3 _______________________________________________ kvmarm mailing list [email protected] https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
RFC 4/4 spe: Test Profiling Buffer Events
In reply to Eric AugerSetup the infrastructure to check the occurence of events. The test checks the Buffer Full event occurs when no space is available. The PPI is handled and we check the syndrome register against the expected event.
--- arm/spe.c | 141 +++++++++++++++++++++++++++++++++++++++++++++- arm/unittests.cfg | 8 +++ 2 files changed, 148 insertions(+), 1 deletion(-)
arm/spe.c +140 -1
@@ -19,6 +19,7 @@#include "alloc_page.h" #include <bitops.h> #include "alloc.h" +#include <asm/gic.h> struct spe {int min_interval;
@@ -36,13 +37,37 @@ struct spe {bool unique_record_size; }; +enum spe_event_exception_class { +EC_STAGE1_DATA_ABORT =0x24 , +EC_STAGE2_DATA_ABORT =0x25 , +EC_OTHER =0 , +}; + +struct spe_event { +enum spe_event_exception_class ec; +bool dl;/* data lost */ +bool ea;/* external abort */ +bool s;/* service */ +bool coll;/* collision */ +union { +bool buffer_filled;/* ec = other */ + } mss; +}; +static struct spe spe; +struct spe_stats { +struct spe_event observed; +bool unexpected; +}; +static struct spe_stats spe_stats; +#ifdef __arm__static bool spe_probe (void ) {return false; }static void test_spe_introspection (void ) { }static void test_spe_buffer (void ) { } +static void test_spe_events (void ) { }#else
@@ -95,6 +120,16 @@ static void test_spe_buffer(void) { }#define PMSCR_EL1_TS 0x20#define PMSCR_EL1_PCT 0x40 +#define PMBSR_EL1_COLL 0x10000 +#define PMBSR_EL1_S 0x20000 +#define PMBSR_EL1_EA 0x40000 +#define PMBSR_EL1_DL 0x80000 +#define PMBSR_EL1_EC_SHIFT 26 +#define PMBSR_EL1_EC_MASK 0x3F +#define PMBSR_EL1_MISS_MASK 0xFFFF + +#define SPE_PPI 21 +static int min_interval (uint8_t idr_bits) {switch (idr_bits) {
@@ -119,6 +154,44 @@ static int min_interval(uint8_t idr_bits) } } +static int decode_syndrome_register (uint64_t sr,struct spe_event *event,bool verbose) +{ +if (!sr) +return 0 ; + +if (sr &PMBSR_EL1_S ) + event->s = true; +if (sr &PMBSR_EL1_COLL ) + event->coll = true; +if (sr &PMBSR_EL1_EA ) + event->ea = true; +if (sr &PMBSR_EL1_DL ) + event->dl = true; +if (verbose) +report_info ("PMBSR_EL1: Service=%d Collision=%d External Fault=%d DataLost=%d" , + event->s, event->coll, event->ea, event->dl); + +switch ((sr >>PMBSR_EL1_EC_SHIFT ) &PMBSR_EL1_EC_MASK ) { +case EC_OTHER : + event->ec =EC_OTHER ; + event->mss.buffer_filled = sr &0x1 ; +if (verbose) +report_info ("PMBSR_EL1: EC = OTHER buffer filled=%d" , event->mss.buffer_filled); +break ; +case EC_STAGE1_DATA_ABORT : + event->ec =EC_STAGE1_DATA_ABORT ; +report_info ("PMBSR_EL1: EC = stage 1 data abort" ); +break ; +case EC_STAGE2_DATA_ABORT : + event->ec =EC_STAGE2_DATA_ABORT ; +report_info ("PMBSR_EL1: EC = stage 2 data abort" ); +break ; +default : +return -1 ; + } +return 0 ; +} +static bool spe_probe (void ) {uint64_t pmbidr_el1, pmsidr_el1;
@@ -224,6 +297,13 @@ static void reset(void)/* Make sure the syndrome register is void */ write_sysreg_s (0 ,PMBSR_EL1 ); + +memset (&spe_stats,0 ,sizeof (spe_stats)); +} + +inline bool event_match (struct spe_event *observed,struct spe_event *expected) +{ +return !memcmp (observed, expected,sizeof (struct spe_event )); }static inline void drain (void )
@@ -235,6 +315,7 @@ static inline void drain(void)static void test_spe_buffer (void ) { +struct spe_event observed = {}, expected = {};uint64_t pmbsr_el1, val1, val2;void *addr =malloc (10 *PAGE_SIZE );
@@ -290,7 +371,61 @@ static void test_spe_buffer(void)report_info ("This corresponds to %ld record(s) of %d bytes" , val2 / spe.maxsize, spe.maxsize); pmbsr_el1 =read_sysreg_s (PMBSR_EL1 ); -report (!pmbsr_el1,"PMBSR_EL1: no event" ); +report (!(decode_syndrome_register (pmbsr_el1, &observed, true)) && +event_match (&observed, &expected),"PMBSR_EL1: no event" ); + +free (addr); +} + +static void irq_handler (struct pt_regs *regs) +{ +uint32_t irqstat, irqnr; + + irqstat =gic_read_iar (); + irqnr =gic_iar_irqnr (irqstat); + +if (irqnr ==SPE_PPI ) { +uint64_t pmbsr_el1 =read_sysreg_s (PMBSR_EL1 ); + +if (decode_syndrome_register (pmbsr_el1, &spe_stats.observed, true)) + spe_stats.unexpected = true; +report_info ("SPE IRQ! SR=0x%lx" , pmbsr_el1); +write_sysreg_s (0 ,PMBSR_EL1 ); + }else { + spe_stats.unexpected = true; + } +gic_write_eoir (irqstat); +} + +static inline bool has_event_occurred (struct spe_event *expected) +{ +return (!spe_stats.unexpected &&event_match (&spe_stats.observed, expected)); +} + +static void test_spe_events (void ) +{ +struct spe_event expected = {.ec =EC_OTHER , .mss.buffer_filled = true, .s = true}; +void *addr =malloc (10 *PAGE_SIZE ); + +gic_enable_defaults (); +install_irq_handler (EL1H_IRQ , irq_handler); +local_irq_enable (); +gic_enable_irq (SPE_PPI ); + +reset (); + +/* Willingly set pmblimitr tp pmdptr */ + spe.pmblimitr_el1 = spe.pmbptr_el1; + +mem_access_loop (addr,100000 , spe.pmblimitr_el1 |PMBLIMITR_EL1_E ); +drain (); +report (has_event_occurred (&expected),"PMBSR_EL1: buffer full event" ); + +/* redo it once */ + +mem_access_loop (addr,100000 , spe.pmblimitr_el1 |PMBLIMITR_EL1_E ); +drain (); +report (has_event_occurred (&expected),"PMBSR_EL1: buffer full event" );free (addr); }
@@ -317,6 +452,10 @@ int main(int argc, char *argv[])report_prefix_push (argv[1 ]);test_spe_buffer ();report_prefix_pop (); + }else if (strcmp (argv[1 ],"spe-events" ) ==0 ) { +report_prefix_push (argv[1 ]); +test_spe_events (); +report_prefix_pop (); }else {report_abort ("Unknown sub-test '%s'" , argv[1 ]); }
arm/unittests.cfg +8 -0
@@ -150,6 +150,14 @@ extra_params = -append 'spe-buffer' accel = kvm arch = arm64 +[spe-events] +file = spe.flat +groups = spe +arch = arm64 +extra_params = -append 'spe-events' +accel = kvm +arch = arm64 + # Test GIC emulation [gicv2-ipi] file = gic.flat
Signature · 6 lines
2.21.3 _______________________________________________ kvmarm mailing list [email protected] https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Re: [kvm-unit-tests RFC 4/4] spe: Test Profiling Buffer Events
Hi,
Quoted 191 lines · Eric Auger
Setup the infrastructure to check the occurence of events. The test checks the Buffer Full event occurs when no space is available. The PPI is handled and we check the syndrome register against the expected event.
Signed-off-by Eric Auger <[email protected]>--- arm/spe.c | 141 +++++++++++++++++++++++++++++++++++++++++++++- arm/unittests.cfg | 8 +++ 2 files changed, 148 insertions(+), 1 deletion(-)arm/spe.c +122
-1@@ -19,6 +19,7 @@#include "alloc_page.h" #include <bitops.h> #include "alloc.h" +#include <asm/gic.h> struct spe {int min_interval;@@ -36,13 +37,37 @@ struct spe {bool unique_record_size; }; +enum spe_event_exception_class { +EC_STAGE1_DATA_ABORT =0x24 , +EC_STAGE2_DATA_ABORT =0x25 , +EC_OTHER =0 , +}; + +struct spe_event { +enum spe_event_exception_class ec; +bool dl;/* data lost */ +bool ea;/* external abort */ +bool s;/* service */ +bool coll;/* collision */ +union { +bool buffer_filled;/* ec = other */ + } mss; +}; +static struct spe spe; +struct spe_stats { +struct spe_event observed; +bool unexpected; +}; +static struct spe_stats spe_stats; +#ifdef __arm__static bool spe_probe (void ) {return false; }static void test_spe_introspection (void ) { }static void test_spe_buffer (void ) { } +static void test_spe_events (void ) { }#else @@ -95,6 +120,16 @@ static void test_spe_buffer(void) { }#define PMSCR_EL1_TS 0x20#define PMSCR_EL1_PCT 0x40 +#define PMBSR_EL1_COLL 0x10000 +#define PMBSR_EL1_S 0x20000 +#define PMBSR_EL1_EA 0x40000 +#define PMBSR_EL1_DL 0x80000 +#define PMBSR_EL1_EC_SHIFT 26 +#define PMBSR_EL1_EC_MASK 0x3F +#define PMBSR_EL1_MISS_MASK 0xFFFF + +#define SPE_PPI 21 +static int min_interval (uint8_t idr_bits) {switch (idr_bits) {@@ -119,6 +154,44 @@ static int min_interval(uint8_t idr_bits) } } +static int decode_syndrome_register (uint64_t sr,struct spe_event *event,bool verbose) +{ +if (!sr) +return 0 ; + +if (sr &PMBSR_EL1_S ) + event->s = true; +if (sr &PMBSR_EL1_COLL ) + event->coll = true; +if (sr &PMBSR_EL1_EA ) + event->ea = true; +if (sr &PMBSR_EL1_DL ) + event->dl = true; +if (verbose) +report_info ("PMBSR_EL1: Service=%d Collision=%d External Fault=%d DataLost=%d" , + event->s, event->coll, event->ea, event->dl); + +switch ((sr >>PMBSR_EL1_EC_SHIFT ) &PMBSR_EL1_EC_MASK ) { +case EC_OTHER : + event->ec =EC_OTHER ; + event->mss.buffer_filled = sr &0x1 ; +if (verbose) +report_info ("PMBSR_EL1: EC = OTHER buffer filled=%d" , event->mss.buffer_filled); +break ; +case EC_STAGE1_DATA_ABORT : + event->ec =EC_STAGE1_DATA_ABORT ; +report_info ("PMBSR_EL1: EC = stage 1 data abort" ); +break ; +case EC_STAGE2_DATA_ABORT : + event->ec =EC_STAGE2_DATA_ABORT ; +report_info ("PMBSR_EL1: EC = stage 2 data abort" ); +break ; +default : +return -1 ; + } +return 0 ; +} +static bool spe_probe (void ) {uint64_t pmbidr_el1, pmsidr_el1;@@ -224,6 +297,13 @@ static void reset(void)/* Make sure the syndrome register is void */ write_sysreg_s (0 ,PMBSR_EL1 ); + +memset (&spe_stats,0 ,sizeof (spe_stats)); +} + +inline bool event_match (struct spe_event *observed,struct spe_event *expected) +{ +return !memcmp (observed, expected,sizeof (struct spe_event )); }static inline void drain (void )@@ -235,6 +315,7 @@ static inline void drain(void)static void test_spe_buffer (void ) { +struct spe_event observed = {}, expected = {};uint64_t pmbsr_el1, val1, val2;void *addr =malloc (10 *PAGE_SIZE );@@ -290,7 +371,61 @@ static void test_spe_buffer(void)report_info ("This corresponds to %ld record(s) of %d bytes" , val2 / spe.maxsize, spe.maxsize); pmbsr_el1 =read_sysreg_s (PMBSR_EL1 ); -report (!pmbsr_el1,"PMBSR_EL1: no event" ); +report (!(decode_syndrome_register (pmbsr_el1, &observed, true)) && +event_match (&observed, &expected),"PMBSR_EL1: no event" ); + +free (addr); +} + +static void irq_handler (struct pt_regs *regs) +{ +uint32_t irqstat, irqnr; + + irqstat =gic_read_iar (); + irqnr =gic_iar_irqnr (irqstat); + +if (irqnr ==SPE_PPI ) { +uint64_t pmbsr_el1 =read_sysreg_s (PMBSR_EL1 ); + +if (decode_syndrome_register (pmbsr_el1, &spe_stats.observed, true)) + spe_stats.unexpected = true; +report_info ("SPE IRQ! SR=0x%lx" , pmbsr_el1); +write_sysreg_s (0 ,PMBSR_EL1 ); + }else { + spe_stats.unexpected = true; + } +gic_write_eoir (irqstat); +} + +static inline bool has_event_occurred (struct spe_event *expected) +{ +return (!spe_stats.unexpected &&event_match (&spe_stats.observed, expected)); +} + +static void test_spe_events (void ) +{ +struct spe_event expected = {.ec =EC_OTHER , .mss.buffer_filled = true, .s = true}; +void *addr =malloc (10 *PAGE_SIZE ); + +gic_enable_defaults (); +install_irq_handler (EL1H_IRQ , irq_handler); +local_irq_enable (); +gic_enable_irq (SPE_PPI ); +
arm/spe.c static void test_spe_buffer(void) · line 424
Quoted 10 lines
arm/spe.c +10
-0+reset (); + +/* Willingly set pmblimitr tp pmdptr */ + spe.pmblimitr_el1 = spe.pmbptr_el1; + +mem_access_loop (addr,100000 , spe.pmblimitr_el1 |PMBLIMITR_EL1_E ); +drain (); +report (has_event_occurred (&expected),"PMBSR_EL1: buffer full event" ); + +/* redo it once */
I noticed I must reset the stats here. I will fix that in next version.
Thanks
Eric
Quoted 38 lines
arm/spe.c +8
-0+ +mem_access_loop (addr,100000 , spe.pmblimitr_el1 |PMBLIMITR_EL1_E ); +drain (); +report (has_event_occurred (&expected),"PMBSR_EL1: buffer full event" );free (addr); }@@ -317,6 +452,10 @@ int main(int argc, char *argv[])report_prefix_push (argv[1 ]);test_spe_buffer ();report_prefix_pop (); + }else if (strcmp (argv[1 ],"spe-events" ) ==0 ) { +report_prefix_push (argv[1 ]); +test_spe_events (); +report_prefix_pop (); }else {report_abort ("Unknown sub-test '%s'" , argv[1 ]); }arm/unittests.cfg +8
-0@@ -150,6 +150,14 @@ extra_params = -append 'spe-buffer' accel = kvm arch = arm64 +[spe-events] +file = spe.flat +groups = spe +arch = arm64 +extra_params = -append 'spe-events' +accel = kvm +arch = arm64 + # Test GIC emulation [gicv2-ipi] file = gic.flat
Hi Eric,
These patches are extremely welcome! I took over the KVM SPE patches from Andrew Murray, and I was working on something similar to help with development.
The KVM series on the public mailing list work only by chance because it is impossible to reliably map the SPE buffer at EL2 when profiling triggers a stage 2 data abort. That's because the DABT is reported asynchronously via the buffer management interrupt and the faulting IPA is not reported anywhere. I'm trying to fix this issue in the next iteration of the series, and then I'll come back to your patches for review and testing.
Thanks,
Alex
Quoted 37 lines · Eric Auger
This series implements tests exercising the Statistical Profiling Extensions.
This was tested with associated unmerged kernel [1] and QEMU [2] series.
Depending on the comments, I can easily add other tests checking more configs, additional events and testing migration too. I hope this can be useful when respinning both series.
All SPE tests can be launched with:
./run_tests.sh -g spe
Tests also can be launched individually. For example:
./arm-run arm/spe.flat -append 'spe-buffer'The series can be found at:
https://github.com/eauger/kut/tree/spe_rfcReferences:
[1] [PATCH v2 00/18] arm64: KVM: add SPE profiling support
[2] [PATCH 0/7] target/arm: Add vSPE support to KVM guestEric Auger (4):
arm64: Moveget_id_aa64dfr0()inprocessor.h
spe: Probing and Introspection Test
spe: Add profiling buffer test
spe: Test Profiling Buffer Eventsarm/Makefile.common | 1 + arm/pmu.c | 1 - arm/spe.c | 463 ++++++++++++++++++++++++++++++++++++++ arm/unittests.cfg | 24 ++ lib/arm64/asm/barrier.h | 1 + lib/arm64/asm/processor.h | 5 + 6 files changed, 494 insertions(+), 1 deletion(-) create mode 100644 arm/spe.c
Hi Alexandru,
Quoted 4 lines · Alexandru Elisei
Hi Eric,
These patches are extremely welcome! I took over the KVM SPE patches from Andrew Murray, and I was working on something similar to help with development.
Cool.
Quoted 7 lines
The KVM series on the public mailing list work only by chance because it is impossible to reliably map the SPE buffer at EL2 when profiling triggers a stage 2 data abort. That's because the DABT is reported asynchronously via the buffer management interrupt and the faulting IPA is not reported anywhere. I'm trying to fix this issue in the next iteration of the series, and then I'll come back to your patches for review and testing.
Sure. Looking forward to reviewing your respin.
Thanks
Eric
Quoted 44 lines
Thanks,
Alex
Quoted 37 lines · Eric Auger
This series implements tests exercising the Statistical Profiling Extensions.
This was tested with associated unmerged kernel [1] and QEMU [2] series.
Depending on the comments, I can easily add other tests checking more configs, additional events and testing migration too. I hope this can be useful when respinning both series.
All SPE tests can be launched with:
./run_tests.sh -g spe
Tests also can be launched individually. For example:
./arm-run arm/spe.flat -append 'spe-buffer'The series can be found at:
https://github.com/eauger/kut/tree/spe_rfcReferences:
[1] [PATCH v2 00/18] arm64: KVM: add SPE profiling support
[2] [PATCH 0/7] target/arm: Add vSPE support to KVM guestEric Auger (4):
arm64: Moveget_id_aa64dfr0()inprocessor.h
spe: Probing and Introspection Test
spe: Add profiling buffer test
spe: Test Profiling Buffer Eventsarm/Makefile.common | 1 + arm/pmu.c | 1 - arm/spe.c | 463 ++++++++++++++++++++++++++++++++++++++ arm/unittests.cfg | 24 ++ lib/arm64/asm/barrier.h | 1 + lib/arm64/asm/processor.h | 5 + 6 files changed, 494 insertions(+), 1 deletion(-) create mode 100644 arm/spe.c
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm