OpenCloudOS-Kernel/arch/powerpc/perf/power10-pmu.c

628 lines
17 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-or-later
/*
* Performance counter support for POWER10 processors.
*
* Copyright 2020 Madhavan Srinivasan, IBM Corporation.
* Copyright 2020 Athira Rajeev, IBM Corporation.
*/
#define pr_fmt(fmt) "power10-pmu: " fmt
#include "isa207-common.h"
/*
* Raw event encoding for Power10:
*
* 60 56 52 48 44 40 36 32
* | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
* | | [ ] [ src_match ] [ src_mask ] | [ ] [ l2l3_sel ] [ thresh_ctl ]
* | | | | | |
* | | *- IFM (Linux) | | thresh start/stop -*
* | *- BHRB (Linux) | src_sel
* *- EBB (Linux) *invert_bit
*
* 28 24 20 16 12 8 4 0
* | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
* [ ] [ sample ] [ ] [ ] [ pmc ] [unit ] [ ] | m [ pmcxsel ]
* | | | | | | |
* | | | | | | *- mark
* | | | *- L1/L2/L3 cache_sel | |*-radix_scope_qual
* | | sdar_mode |
* | *- sampling mode for marked events *- combine
* |
* *- thresh_sel
*
* Below uses IBM bit numbering.
*
* MMCR1[x:y] = unit (PMCxUNIT)
* MMCR1[24] = pmc1combine[0]
* MMCR1[25] = pmc1combine[1]
* MMCR1[26] = pmc2combine[0]
* MMCR1[27] = pmc2combine[1]
* MMCR1[28] = pmc3combine[0]
* MMCR1[29] = pmc3combine[1]
* MMCR1[30] = pmc4combine[0]
* MMCR1[31] = pmc4combine[1]
*
* if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011
* MMCR1[20:27] = thresh_ctl
* else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001
* MMCR1[20:27] = thresh_ctl
* else
* MMCRA[48:55] = thresh_ctl (THRESH START/END)
*
* if thresh_sel:
* MMCRA[45:47] = thresh_sel
*
* if l2l3_sel:
* MMCR2[56:60] = l2l3_sel[0:4]
*
* MMCR1[16] = cache_sel[0]
* MMCR1[17] = cache_sel[1]
* MMCR1[18] = radix_scope_qual
*
* if mark:
* MMCRA[63] = 1 (SAMPLE_ENABLE)
* MMCRA[57:59] = sample[0:2] (RAND_SAMP_ELIG)
* MMCRA[61:62] = sample[3:4] (RAND_SAMP_MODE)
*
* if EBB and BHRB:
* MMCRA[32:33] = IFM
*
* MMCRA[SDAR_MODE] = sdar_mode[0:1]
*/
/*
* Some power10 event codes.
*/
#define EVENT(_name, _code) enum{_name = _code}
#include "power10-events-list.h"
#undef EVENT
/* MMCRA IFM bits - POWER10 */
#define POWER10_MMCRA_IFM1 0x0000000040000000UL
#define POWER10_MMCRA_IFM2 0x0000000080000000UL
#define POWER10_MMCRA_IFM3 0x00000000C0000000UL
#define POWER10_MMCRA_BHRB_MASK 0x00000000C0000000UL
extern u64 PERF_REG_EXTENDED_MASK;
/* Table of alternatives, sorted by column 0 */
static const unsigned int power10_event_alternatives[][MAX_ALT] = {
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
{ PM_INST_CMPL_ALT, PM_INST_CMPL },
powerpc/perf: Fix power10 event alternatives When scheduling a group of events, there are constraint checks done to make sure all events can go in a group. Example, one of the criteria is that events in a group cannot use the same PMC. But platform specific PMU supports alternative event for some of the event codes. During perf_event_open(), if any event group doesn't match constraint check criteria, further lookup is done to find alternative event. By current design, the array of alternatives events in PMU code is expected to be sorted by column 0. This is because in find_alternative() the return criteria is based on event code comparison. ie. "event < ev_alt[i][0])". This optimisation is there since find_alternative() can be called multiple times. In power10 PMU code, the alternative event array is not sorted properly and hence there is breakage in finding alternative event. To work with existing logic, fix the alternative event array to be sorted by column 0 for power10-pmu.c Results: In case where an alternative event is not chosen when we could, events will be multiplexed. ie, time sliced where it could actually run concurrently. Example, in power10 PM_INST_CMPL_ALT(0x00002) has alternative event, PM_INST_CMPL(0x500fa). Without the fix, if a group of events with PMC1 to PMC4 is used along with PM_INST_CMPL_ALT, it will be time sliced since all programmable PMC's are consumed already. But with the fix, when it picks alternative event on PMC5, all events will run concurrently. Before: # perf stat -e r00002,r100fc,r200fa,r300fc,r400fc Performance counter stats for 'system wide': 328668935 r00002 (79.94%) 56501024 r100fc (79.95%) 49564238 r200fa (79.95%) 376 r300fc (80.19%) 660 r400fc (79.97%) 4.039150522 seconds time elapsed With the fix, since alternative event is chosen to run on PMC6, events will be run concurrently. After: # perf stat -e r00002,r100fc,r200fa,r300fc,r400fc Performance counter stats for 'system wide': 23596607 r00002 4907738 r100fc 2283608 r200fa 135 r300fc 248 r400fc 1.664671390 seconds time elapsed Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com> Reviewed-by: Madhavan Srinivasan <maddy@linux.vnet.ibm.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20220419114828.89843-2-atrajeev@linux.vnet.ibm.com
2022-04-19 19:48:28 +08:00
{ PM_CYC_ALT, PM_CYC },
};
static int power10_get_alternatives(u64 event, unsigned int flags, u64 alt[])
{
int num_alt = 0;
num_alt = isa207_get_alternatives(event, alt,
ARRAY_SIZE(power10_event_alternatives), flags,
power10_event_alternatives);
return num_alt;
}
static int power10_check_attr_config(struct perf_event *ev)
{
u64 val;
u64 event = ev->attr.config;
val = (event >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK;
if (val == 0x10 || isa3XX_check_attr_config(ev))
return -EINVAL;
return 0;
}
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
GENERIC_EVENT_ATTR(cpu-cycles, PM_CYC);
GENERIC_EVENT_ATTR(instructions, PM_INST_CMPL);
GENERIC_EVENT_ATTR(branch-instructions, PM_BR_CMPL);
GENERIC_EVENT_ATTR(branch-misses, PM_BR_MPRED_CMPL);
GENERIC_EVENT_ATTR(cache-references, PM_LD_REF_L1);
GENERIC_EVENT_ATTR(cache-misses, PM_LD_MISS_L1);
GENERIC_EVENT_ATTR(mem-loads, MEM_LOADS);
GENERIC_EVENT_ATTR(mem-stores, MEM_STORES);
GENERIC_EVENT_ATTR(branch-instructions, PM_BR_FIN);
GENERIC_EVENT_ATTR(branch-misses, PM_MPRED_BR_FIN);
GENERIC_EVENT_ATTR(cache-misses, PM_LD_DEMAND_MISS_L1_FIN);
CACHE_EVENT_ATTR(L1-dcache-load-misses, PM_LD_MISS_L1);
CACHE_EVENT_ATTR(L1-dcache-loads, PM_LD_REF_L1);
CACHE_EVENT_ATTR(L1-dcache-prefetches, PM_LD_PREFETCH_CACHE_LINE_MISS);
CACHE_EVENT_ATTR(L1-dcache-store-misses, PM_ST_MISS_L1);
CACHE_EVENT_ATTR(L1-icache-load-misses, PM_L1_ICACHE_MISS);
CACHE_EVENT_ATTR(L1-icache-loads, PM_INST_FROM_L1);
CACHE_EVENT_ATTR(L1-icache-prefetches, PM_IC_PREF_REQ);
CACHE_EVENT_ATTR(LLC-load-misses, PM_DATA_FROM_L3MISS);
CACHE_EVENT_ATTR(LLC-loads, PM_DATA_FROM_L3);
CACHE_EVENT_ATTR(LLC-prefetches, PM_L3_PF_MISS_L3);
CACHE_EVENT_ATTR(LLC-store-misses, PM_L2_ST_MISS);
CACHE_EVENT_ATTR(LLC-stores, PM_L2_ST);
CACHE_EVENT_ATTR(branch-load-misses, PM_BR_MPRED_CMPL);
CACHE_EVENT_ATTR(branch-loads, PM_BR_CMPL);
CACHE_EVENT_ATTR(dTLB-load-misses, PM_DTLB_MISS);
CACHE_EVENT_ATTR(iTLB-load-misses, PM_ITLB_MISS);
static struct attribute *power10_events_attr_dd1[] = {
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
GENERIC_EVENT_PTR(PM_CYC),
GENERIC_EVENT_PTR(PM_INST_CMPL),
GENERIC_EVENT_PTR(PM_BR_CMPL),
GENERIC_EVENT_PTR(PM_BR_MPRED_CMPL),
GENERIC_EVENT_PTR(PM_LD_REF_L1),
GENERIC_EVENT_PTR(PM_LD_MISS_L1),
GENERIC_EVENT_PTR(MEM_LOADS),
GENERIC_EVENT_PTR(MEM_STORES),
CACHE_EVENT_PTR(PM_LD_MISS_L1),
CACHE_EVENT_PTR(PM_LD_REF_L1),
CACHE_EVENT_PTR(PM_LD_PREFETCH_CACHE_LINE_MISS),
CACHE_EVENT_PTR(PM_ST_MISS_L1),
CACHE_EVENT_PTR(PM_L1_ICACHE_MISS),
CACHE_EVENT_PTR(PM_INST_FROM_L1),
CACHE_EVENT_PTR(PM_IC_PREF_REQ),
CACHE_EVENT_PTR(PM_DATA_FROM_L3MISS),
CACHE_EVENT_PTR(PM_DATA_FROM_L3),
CACHE_EVENT_PTR(PM_BR_MPRED_CMPL),
CACHE_EVENT_PTR(PM_BR_CMPL),
CACHE_EVENT_PTR(PM_DTLB_MISS),
CACHE_EVENT_PTR(PM_ITLB_MISS),
NULL
};
static struct attribute *power10_events_attr[] = {
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
GENERIC_EVENT_PTR(PM_CYC),
GENERIC_EVENT_PTR(PM_INST_CMPL),
GENERIC_EVENT_PTR(PM_BR_FIN),
GENERIC_EVENT_PTR(PM_MPRED_BR_FIN),
GENERIC_EVENT_PTR(PM_LD_REF_L1),
GENERIC_EVENT_PTR(PM_LD_DEMAND_MISS_L1_FIN),
GENERIC_EVENT_PTR(MEM_LOADS),
GENERIC_EVENT_PTR(MEM_STORES),
CACHE_EVENT_PTR(PM_LD_MISS_L1),
CACHE_EVENT_PTR(PM_LD_REF_L1),
CACHE_EVENT_PTR(PM_LD_PREFETCH_CACHE_LINE_MISS),
CACHE_EVENT_PTR(PM_ST_MISS_L1),
CACHE_EVENT_PTR(PM_L1_ICACHE_MISS),
CACHE_EVENT_PTR(PM_INST_FROM_L1),
CACHE_EVENT_PTR(PM_IC_PREF_REQ),
CACHE_EVENT_PTR(PM_DATA_FROM_L3MISS),
CACHE_EVENT_PTR(PM_DATA_FROM_L3),
CACHE_EVENT_PTR(PM_L3_PF_MISS_L3),
CACHE_EVENT_PTR(PM_L2_ST_MISS),
CACHE_EVENT_PTR(PM_L2_ST),
CACHE_EVENT_PTR(PM_BR_MPRED_CMPL),
CACHE_EVENT_PTR(PM_BR_CMPL),
CACHE_EVENT_PTR(PM_DTLB_MISS),
CACHE_EVENT_PTR(PM_ITLB_MISS),
NULL
};
static const struct attribute_group power10_pmu_events_group_dd1 = {
.name = "events",
.attrs = power10_events_attr_dd1,
};
static const struct attribute_group power10_pmu_events_group = {
.name = "events",
.attrs = power10_events_attr,
};
PMU_FORMAT_ATTR(event, "config:0-59");
PMU_FORMAT_ATTR(pmcxsel, "config:0-7");
PMU_FORMAT_ATTR(mark, "config:8");
PMU_FORMAT_ATTR(combine, "config:10-11");
PMU_FORMAT_ATTR(unit, "config:12-15");
PMU_FORMAT_ATTR(pmc, "config:16-19");
PMU_FORMAT_ATTR(cache_sel, "config:20-21");
PMU_FORMAT_ATTR(sdar_mode, "config:22-23");
PMU_FORMAT_ATTR(sample_mode, "config:24-28");
PMU_FORMAT_ATTR(thresh_sel, "config:29-31");
PMU_FORMAT_ATTR(thresh_stop, "config:32-35");
PMU_FORMAT_ATTR(thresh_start, "config:36-39");
PMU_FORMAT_ATTR(l2l3_sel, "config:40-44");
PMU_FORMAT_ATTR(src_sel, "config:45-46");
PMU_FORMAT_ATTR(invert_bit, "config:47");
PMU_FORMAT_ATTR(src_mask, "config:48-53");
PMU_FORMAT_ATTR(src_match, "config:54-59");
PMU_FORMAT_ATTR(radix_scope, "config:9");
powerpc/perf: Adds support for programming of Thresholding in P10 Thresholding, a performance monitoring unit feature, can be used to identify marked instructions which take more than expected cycles between start event and end event. Threshold compare (thresh_cmp) bits are programmed in MMCRA register. In Power9, thresh_cmp bits were part of the event code. But in case of P10, thresh_cmp are not part of event code due to inclusion of MMCR3 bits. Patch here adds an option to use attr.config1 variable to be used to pass thresh_cmp value to be programmed in MMCRA register. A new ppmu flag called PPMU_HAS_ATTR_CONFIG1 has been added and this flag is used to notify the use of attr.config1 variable. Patch has extended the parameter list of 'compute_mmcr', to include power_pmu's 'flags' element and parameter list of get_constraint to include attr.config1 value. It also extend parameter list of power_check_constraints inorder to pass perf_event list. As stated by commit ef0e3b650f8d ("powerpc/perf: Fix Threshold Event Counter Multiplier width for P10"), constraint bits for thresh_cmp is also needed to be increased to 11 bits, which is handled as part of this patch. We added bit number 53 as part of constraint bits of thresh_cmp for power10 to make it an 11 bit field. Updated layout for p10: /* * Layout of constraint bits: * * 60 56 52 48 44 40 36 32 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | * [ fab_match ] [ thresh_cmp ] [ thresh_ctl ] [ ] * | | * [ thresh_cmp bits for p10] thresh_sel -* * * 28 24 20 16 12 8 4 0 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | * [ ] | [ ] | [ sample ] [ ] [6] [5] [4] [3] [2] [1] * | | | | | * BHRB IFM -* | | |*radix_scope | Count of events for each PMC. * EBB -* | | p1, p2, p3, p4, p5, p6. * L1 I/D qualifier -* | * nc - number of counters -* * * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints * we want the low bit of each field to be added to any existing value. * * Everything else is a value field. */ Result: command#: cat /sys/devices/cpu/format/thresh_cmp config1:0-17 ex. usage: command#: perf record -I --weight -d -e cpu/event=0x67340101EC,thresh_cmp=500/ ./ebizzy -S 2 -t 1 -s 4096 1826636 records/s real 2.00 s user 2.00 s sys 0.00 s [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.038 MB perf.data (61 samples) ] Signed-off-by: Kajol Jain <kjain@linux.ibm.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20210209095234.837356-1-kjain@linux.ibm.com
2021-02-09 17:52:34 +08:00
PMU_FORMAT_ATTR(thresh_cmp, "config1:0-17");
static struct attribute *power10_pmu_format_attr[] = {
&format_attr_event.attr,
&format_attr_pmcxsel.attr,
&format_attr_mark.attr,
&format_attr_combine.attr,
&format_attr_unit.attr,
&format_attr_pmc.attr,
&format_attr_cache_sel.attr,
&format_attr_sdar_mode.attr,
&format_attr_sample_mode.attr,
&format_attr_thresh_sel.attr,
&format_attr_thresh_stop.attr,
&format_attr_thresh_start.attr,
&format_attr_l2l3_sel.attr,
&format_attr_src_sel.attr,
&format_attr_invert_bit.attr,
&format_attr_src_mask.attr,
&format_attr_src_match.attr,
&format_attr_radix_scope.attr,
powerpc/perf: Adds support for programming of Thresholding in P10 Thresholding, a performance monitoring unit feature, can be used to identify marked instructions which take more than expected cycles between start event and end event. Threshold compare (thresh_cmp) bits are programmed in MMCRA register. In Power9, thresh_cmp bits were part of the event code. But in case of P10, thresh_cmp are not part of event code due to inclusion of MMCR3 bits. Patch here adds an option to use attr.config1 variable to be used to pass thresh_cmp value to be programmed in MMCRA register. A new ppmu flag called PPMU_HAS_ATTR_CONFIG1 has been added and this flag is used to notify the use of attr.config1 variable. Patch has extended the parameter list of 'compute_mmcr', to include power_pmu's 'flags' element and parameter list of get_constraint to include attr.config1 value. It also extend parameter list of power_check_constraints inorder to pass perf_event list. As stated by commit ef0e3b650f8d ("powerpc/perf: Fix Threshold Event Counter Multiplier width for P10"), constraint bits for thresh_cmp is also needed to be increased to 11 bits, which is handled as part of this patch. We added bit number 53 as part of constraint bits of thresh_cmp for power10 to make it an 11 bit field. Updated layout for p10: /* * Layout of constraint bits: * * 60 56 52 48 44 40 36 32 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | * [ fab_match ] [ thresh_cmp ] [ thresh_ctl ] [ ] * | | * [ thresh_cmp bits for p10] thresh_sel -* * * 28 24 20 16 12 8 4 0 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | * [ ] | [ ] | [ sample ] [ ] [6] [5] [4] [3] [2] [1] * | | | | | * BHRB IFM -* | | |*radix_scope | Count of events for each PMC. * EBB -* | | p1, p2, p3, p4, p5, p6. * L1 I/D qualifier -* | * nc - number of counters -* * * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints * we want the low bit of each field to be added to any existing value. * * Everything else is a value field. */ Result: command#: cat /sys/devices/cpu/format/thresh_cmp config1:0-17 ex. usage: command#: perf record -I --weight -d -e cpu/event=0x67340101EC,thresh_cmp=500/ ./ebizzy -S 2 -t 1 -s 4096 1826636 records/s real 2.00 s user 2.00 s sys 0.00 s [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.038 MB perf.data (61 samples) ] Signed-off-by: Kajol Jain <kjain@linux.ibm.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20210209095234.837356-1-kjain@linux.ibm.com
2021-02-09 17:52:34 +08:00
&format_attr_thresh_cmp.attr,
NULL,
};
static const struct attribute_group power10_pmu_format_group = {
.name = "format",
.attrs = power10_pmu_format_attr,
};
static const struct attribute_group *power10_pmu_attr_groups_dd1[] = {
&power10_pmu_format_group,
&power10_pmu_events_group_dd1,
NULL,
};
static const struct attribute_group *power10_pmu_attr_groups[] = {
&power10_pmu_format_group,
&power10_pmu_events_group,
NULL,
};
static int power10_generic_events_dd1[] = {
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
[PERF_COUNT_HW_CPU_CYCLES] = PM_CYC,
[PERF_COUNT_HW_INSTRUCTIONS] = PM_INST_CMPL,
[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = PM_BR_CMPL,
[PERF_COUNT_HW_BRANCH_MISSES] = PM_BR_MPRED_CMPL,
[PERF_COUNT_HW_CACHE_REFERENCES] = PM_LD_REF_L1,
[PERF_COUNT_HW_CACHE_MISSES] = PM_LD_MISS_L1,
};
static int power10_generic_events[] = {
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
[PERF_COUNT_HW_CPU_CYCLES] = PM_CYC,
[PERF_COUNT_HW_INSTRUCTIONS] = PM_INST_CMPL,
[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = PM_BR_FIN,
[PERF_COUNT_HW_BRANCH_MISSES] = PM_MPRED_BR_FIN,
[PERF_COUNT_HW_CACHE_REFERENCES] = PM_LD_REF_L1,
[PERF_COUNT_HW_CACHE_MISSES] = PM_LD_DEMAND_MISS_L1_FIN,
};
static u64 power10_bhrb_filter_map(u64 branch_sample_type)
{
u64 pmu_bhrb_filter = 0;
/* BHRB and regular PMU events share the same privilege state
* filter configuration. BHRB is always recorded along with a
* regular PMU event. As the privilege state filter is handled
* in the basic PMC configuration of the accompanying regular
* PMU event, we ignore any separate BHRB specific request.
*/
/* No branch filter requested */
if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY)
return pmu_bhrb_filter;
/* Invalid branch filter options - HW does not support */
if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
return -1;
if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL) {
pmu_bhrb_filter |= POWER10_MMCRA_IFM2;
return pmu_bhrb_filter;
}
if (branch_sample_type & PERF_SAMPLE_BRANCH_COND) {
pmu_bhrb_filter |= POWER10_MMCRA_IFM3;
return pmu_bhrb_filter;
}
if (branch_sample_type & PERF_SAMPLE_BRANCH_CALL)
return -1;
if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) {
pmu_bhrb_filter |= POWER10_MMCRA_IFM1;
return pmu_bhrb_filter;
}
/* Every thing else is unsupported */
return -1;
}
static void power10_config_bhrb(u64 pmu_bhrb_filter)
{
pmu_bhrb_filter &= POWER10_MMCRA_BHRB_MASK;
/* Enable BHRB filter in PMU */
mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter));
}
#define C(x) PERF_COUNT_HW_CACHE_##x
/*
* Table of generalized cache-related events.
* 0 means not supported, -1 means nonsensical, other values
* are event codes.
*/
static u64 power10_cache_events_dd1[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
[C(L1D)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_LD_REF_L1,
[C(RESULT_MISS)] = PM_LD_MISS_L1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS)] = PM_ST_MISS_L1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = PM_LD_PREFETCH_CACHE_LINE_MISS,
[C(RESULT_MISS)] = 0,
},
},
[C(L1I)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_INST_FROM_L1,
[C(RESULT_MISS)] = PM_L1_ICACHE_MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = PM_INST_FROM_L1MISS,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = PM_IC_PREF_REQ,
[C(RESULT_MISS)] = 0,
},
},
[C(LL)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_DATA_FROM_L3,
[C(RESULT_MISS)] = PM_DATA_FROM_L3MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = 0,
},
},
[C(DTLB)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS)] = PM_DTLB_MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
[C(ITLB)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS)] = PM_ITLB_MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
[C(BPU)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_BR_CMPL,
[C(RESULT_MISS)] = PM_BR_MPRED_CMPL,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
[C(NODE)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
};
static u64 power10_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
[C(L1D)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_LD_REF_L1,
[C(RESULT_MISS)] = PM_LD_MISS_L1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS)] = PM_ST_MISS_L1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = PM_LD_PREFETCH_CACHE_LINE_MISS,
[C(RESULT_MISS)] = 0,
},
},
[C(L1I)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_INST_FROM_L1,
[C(RESULT_MISS)] = PM_L1_ICACHE_MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = PM_INST_FROM_L1MISS,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = PM_IC_PREF_REQ,
[C(RESULT_MISS)] = 0,
},
},
[C(LL)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_DATA_FROM_L3,
[C(RESULT_MISS)] = PM_DATA_FROM_L3MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = PM_L2_ST,
[C(RESULT_MISS)] = PM_L2_ST_MISS,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = PM_L3_PF_MISS_L3,
[C(RESULT_MISS)] = 0,
},
},
[C(DTLB)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS)] = PM_DTLB_MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
[C(ITLB)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS)] = PM_ITLB_MISS,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
[C(BPU)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = PM_BR_CMPL,
[C(RESULT_MISS)] = PM_BR_MPRED_CMPL,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
[C(NODE)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS)] = -1,
},
},
};
#undef C
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
/*
* Set the MMCR0[CC56RUN] bit to enable counting for
* PMC5 and PMC6 regardless of the state of CTRL[RUN],
* so that we can use counters 5 and 6 as PM_INST_CMPL and
* PM_CYC.
*/
static int power10_compute_mmcr(u64 event[], int n_ev,
unsigned int hwc[], struct mmcr_regs *mmcr,
struct perf_event *pevents[], u32 flags)
{
int ret;
ret = isa207_compute_mmcr(event, n_ev, hwc, mmcr, pevents, flags);
if (!ret)
mmcr->mmcr0 |= MMCR0_C56RUN;
return ret;
}
static struct power_pmu power10_pmu = {
.name = "POWER10",
.n_counter = MAX_PMU_COUNTERS,
.add_fields = ISA207_ADD_FIELDS,
.test_adder = ISA207_TEST_ADDER,
.group_constraint_mask = CNST_CACHE_PMC4_MASK,
.group_constraint_val = CNST_CACHE_PMC4_VAL,
powerpc/perf: Fix cycles/instructions as PM_CYC/PM_INST_CMPL in power10 On power9 and earlier platforms, the default event used for cyles and instructions is PM_CYC (0x0001e) and PM_INST_CMPL (0x00002) respectively. These events use two programmable PMCs and by default will count irrespective of the run latch state (idle state). But since they use programmable PMCs, these events can lead to multiplexing with other events, because there are only 4 programmable PMCs. Hence in power10, performance monitoring unit (PMU) driver uses performance monitor counter 5 (PMC5) and performance monitor counter6 (PMC6) for counting instructions and cycles. Currently on power10, the event used for cycles is PM_RUN_CYC (0x600F4) and instructions uses PM_RUN_INST_CMPL (0x500fa). But counting of these events in idle state is controlled by the CC56RUN bit setting in Monitor Mode Control Register0 (MMCR0). If the CC56RUN bit is zero, PMC5/6 will not count when CTRL[RUN] (run latch) is zero. This could lead to missing some counts if a thread is in idle state during system wide profiling. To fix it, set the CC56RUN bit in MMCR0 for power10, which makes PMC5 and PMC6 count instructions and cycles regardless of the run latch state. Since this change make PMC5/6 count as PM_INST_CMPL/PM_CYC, rename the event code 0x600f4 as PM_CYC instead of PM_RUN_CYC and event code 0x500fa as PM_INST_CMPL instead of PM_RUN_INST_CMPL. The changes are only for PMC5/6 event codes and will not affect the behaviour of PM_RUN_CYC/PM_RUN_INST_CMPL if progammed in other PMC's. Fixes: a64e697cef23 ("powerpc/perf: power10 Performance Monitoring support") Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.cm> Reviewed-by: Madhavan Srinivasan <maddy@linux.ibm.com> [mpe: Tweak change log wording for style and consistency] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20211007075121.28497-1-atrajeev@linux.vnet.ibm.com
2021-10-07 15:51:21 +08:00
.compute_mmcr = power10_compute_mmcr,
.config_bhrb = power10_config_bhrb,
.bhrb_filter_map = power10_bhrb_filter_map,
.get_constraint = isa207_get_constraint,
.get_alternatives = power10_get_alternatives,
.get_mem_data_src = isa207_get_mem_data_src,
.get_mem_weight = isa207_get_mem_weight,
.disable_pmc = isa207_disable_pmc,
.flags = PPMU_HAS_SIER | PPMU_ARCH_207S |
powerpc/perf: Adds support for programming of Thresholding in P10 Thresholding, a performance monitoring unit feature, can be used to identify marked instructions which take more than expected cycles between start event and end event. Threshold compare (thresh_cmp) bits are programmed in MMCRA register. In Power9, thresh_cmp bits were part of the event code. But in case of P10, thresh_cmp are not part of event code due to inclusion of MMCR3 bits. Patch here adds an option to use attr.config1 variable to be used to pass thresh_cmp value to be programmed in MMCRA register. A new ppmu flag called PPMU_HAS_ATTR_CONFIG1 has been added and this flag is used to notify the use of attr.config1 variable. Patch has extended the parameter list of 'compute_mmcr', to include power_pmu's 'flags' element and parameter list of get_constraint to include attr.config1 value. It also extend parameter list of power_check_constraints inorder to pass perf_event list. As stated by commit ef0e3b650f8d ("powerpc/perf: Fix Threshold Event Counter Multiplier width for P10"), constraint bits for thresh_cmp is also needed to be increased to 11 bits, which is handled as part of this patch. We added bit number 53 as part of constraint bits of thresh_cmp for power10 to make it an 11 bit field. Updated layout for p10: /* * Layout of constraint bits: * * 60 56 52 48 44 40 36 32 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | * [ fab_match ] [ thresh_cmp ] [ thresh_ctl ] [ ] * | | * [ thresh_cmp bits for p10] thresh_sel -* * * 28 24 20 16 12 8 4 0 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | * [ ] | [ ] | [ sample ] [ ] [6] [5] [4] [3] [2] [1] * | | | | | * BHRB IFM -* | | |*radix_scope | Count of events for each PMC. * EBB -* | | p1, p2, p3, p4, p5, p6. * L1 I/D qualifier -* | * nc - number of counters -* * * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints * we want the low bit of each field to be added to any existing value. * * Everything else is a value field. */ Result: command#: cat /sys/devices/cpu/format/thresh_cmp config1:0-17 ex. usage: command#: perf record -I --weight -d -e cpu/event=0x67340101EC,thresh_cmp=500/ ./ebizzy -S 2 -t 1 -s 4096 1826636 records/s real 2.00 s user 2.00 s sys 0.00 s [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.038 MB perf.data (61 samples) ] Signed-off-by: Kajol Jain <kjain@linux.ibm.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> Link: https://lore.kernel.org/r/20210209095234.837356-1-kjain@linux.ibm.com
2021-02-09 17:52:34 +08:00
PPMU_ARCH_31 | PPMU_HAS_ATTR_CONFIG1,
.n_generic = ARRAY_SIZE(power10_generic_events),
.generic_events = power10_generic_events,
.cache_events = &power10_cache_events,
.attr_groups = power10_pmu_attr_groups,
.bhrb_nr = 32,
.capabilities = PERF_PMU_CAP_EXTENDED_REGS,
.check_attr_config = power10_check_attr_config,
};
int __init init_power10_pmu(void)
{
unsigned int pvr;
int rc;
/* Comes from cpu_specs[] */
if (!cur_cpu_spec->oprofile_cpu_type ||
strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power10"))
return -ENODEV;
pvr = mfspr(SPRN_PVR);
/* Add the ppmu flag for power10 DD1 */
if ((PVR_CFG(pvr) == 1))
power10_pmu.flags |= PPMU_P10_DD1;
/* Set the PERF_REG_EXTENDED_MASK here */
PERF_REG_EXTENDED_MASK = PERF_REG_PMU_MASK_31;
if ((PVR_CFG(pvr) == 1)) {
power10_pmu.generic_events = power10_generic_events_dd1;
power10_pmu.attr_groups = power10_pmu_attr_groups_dd1;
power10_pmu.cache_events = &power10_cache_events_dd1;
}
rc = register_power_pmu(&power10_pmu);
if (rc)
return rc;
/* Tell userspace that EBB is supported */
cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;
return 0;
}