kernel/git/stable/linux.git/arch/powerpc/perf, branch linux-5.19.y

powerpc/perf: Optimize clearing the pending PMI and remove WARN_ON for PMI check in power_pmu_disable

2022-08-17T13:15:47+00:00

[ Upstream commit 890005a7d98f7452cfe86dcfb2aeeb7df01132ce ] commit 2c9ac51b850d ("powerpc/perf: Fix PMU callbacks to clear pending PMI before resetting an overflown PMC") added a new function "pmi_irq_pending" in hw_irq.h. This function is to check if there is a PMI marked as pending in Paca (PACA_IRQ_PMI).This is used in power_pmu_disable in a WARN_ON. The intention here is to provide a warning if there is PMI pending, but no counter is found overflown. During some of the perf runs, below warning is hit: WARNING: CPU: 36 PID: 0 at arch/powerpc/perf/core-book3s.c:1332 power_pmu_disable+0x25c/0x2c0 Modules linked in: ----- NIP [c000000000141c3c] power_pmu_disable+0x25c/0x2c0 LR [c000000000141c8c] power_pmu_disable+0x2ac/0x2c0 Call Trace: [c000000baffcfb90] [c000000000141c8c] power_pmu_disable+0x2ac/0x2c0 (unreliable) [c000000baffcfc10] [c0000000003e2f8c] perf_pmu_disable+0x4c/0x60 [c000000baffcfc30] [c0000000003e3344] group_sched_out.part.124+0x44/0x100 [c000000baffcfc80] [c0000000003e353c] __perf_event_disable+0x13c/0x240 [c000000baffcfcd0] [c0000000003dd334] event_function+0xc4/0x140 [c000000baffcfd20] [c0000000003d855c] remote_function+0x7c/0xa0 [c000000baffcfd50] [c00000000026c394] flush_smp_call_function_queue+0xd4/0x300 [c000000baffcfde0] [c000000000065b24] smp_ipi_demux_relaxed+0xa4/0x100 [c000000baffcfe20] [c0000000000cb2b0] xive_muxed_ipi_action+0x20/0x40 [c000000baffcfe40] [c000000000207c3c] __handle_irq_event_percpu+0x8c/0x250 [c000000baffcfee0] [c000000000207e2c] handle_irq_event_percpu+0x2c/0xa0 [c000000baffcff10] [c000000000210a04] handle_percpu_irq+0x84/0xc0 [c000000baffcff40] [c000000000205f14] generic_handle_irq+0x54/0x80 [c000000baffcff60] [c000000000015740] __do_irq+0x90/0x1d0 [c000000baffcff90] [c000000000016990] __do_IRQ+0xc0/0x140 [c0000009732f3940] [c000000bafceaca8] 0xc000000bafceaca8 [c0000009732f39d0] [c000000000016b78] do_IRQ+0x168/0x1c0 [c0000009732f3a00] [c0000000000090c8] hardware_interrupt_common_virt+0x218/0x220 This means that there is no PMC overflown among the active events in the PMU, but there is a PMU pending in Paca. The function "any_pmc_overflown" checks the PMCs on active events in cpuhw->n_events. Code snippet: <<>> if (any_pmc_overflown(cpuhw)) clear_pmi_irq_pending(); else WARN_ON(pmi_irq_pending()); <<>> Here the PMC overflown is not from active event. Example: When we do perf record, default cycles and instructions will be running on PMC6 and PMC5 respectively. It could happen that overflowed event is currently not active and pending PMI is for the inactive event. Debug logs from trace_printk: <<>> any_pmc_overflown: idx is 5: pmc value is 0xd9a power_pmu_disable: PMC1: 0x0, PMC2: 0x0, PMC3: 0x0, PMC4: 0x0, PMC5: 0xd9a, PMC6: 0x80002011 <<>> Here active PMC (from idx) is PMC5 , but overflown PMC is PMC6(0x80002011). When we handle PMI interrupt for such cases, if the PMC overflown is from inactive event, it will be ignored. Reference commit: commit bc09c219b2e6 ("powerpc/perf: Fix finding overflowed PMC in interrupt") Patch addresses two changes: 1) Fix 1 : Removal of warning ( WARN_ON(pmi_irq_pending()); ) We were printing warning if no PMC is found overflown among active PMU events, but PMI pending in PACA. But this could happen in cases where PMC overflown is not in active PMC. An inactive event could have caused the overflow. Hence the warning is not needed. To know pending PMI is from an inactive event, we need to loop through all PMC's which will cause more SPR reads via mfspr and increase in context switch. Also in existing function: perf_event_interrupt, already we ignore PMI's overflown when it is from an inactive PMC. 2) Fix 2: optimization in clearing pending PMI. Currently we check for any active PMC overflown before clearing PMI pending in Paca. This is causing additional SPR read also. From point 1, we know that if PMI pending in Paca from inactive cases, that is going to be ignored during replay. Hence if there is pending PMI in Paca, just clear it irrespective of PMC overflown or not. In summary, remove the any_pmc_overflown check entirely in power_pmu_disable. ie If there is a pending PMI in Paca, clear it, since we are in pmu_disable. There could be cases where PMI is pending because of inactive PMC ( which later when replayed also will get ignored ), so WARN_ON could give false warning. Hence removing it. Fixes: 2c9ac51b850d ("powerpc/perf: Fix PMU callbacks to clear pending PMI before resetting an overflown PMC") Signed-off-by: Athira Rajeev Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20220522142256.24699-1-atrajeev@linux.vnet.ibm.com Signed-off-by: Sasha Levin

powerpc/perf: Fix the threshold compare group constraint for power9

2022-05-22T05:58:30+00:00

Thresh compare bits for a event is used to program thresh compare field in Monitor Mode Control Register A (MMCRA: 9-18 bits for power9). When scheduling events as a group, all events in that group should match value in threshold bits (like thresh compare, thresh control, thresh select). Otherwise event open for the sibling events should fail. But in the current code, incase thresh compare bits are not valid, we are not failing in group_constraint function which can result in invalid group schduling. Fix the issue by returning -1 incase event is threshold and threshold compare value is not valid. Thresh control bits in the event code is used to program thresh_ctl field in Monitor Mode Control Register A (MMCRA: 48-55). In below example, the scheduling of group events PM_MRK_INST_CMPL (873534401e0) and PM_THRESH_MET (8734340101ec) is expected to fail as both event request different thresh control bits and invalid thresh compare value. Result before the patch changes: [command]# perf stat -e "{r8735340401e0,r8734340101ec}" sleep 1 Performance counter stats for 'sleep 1': 11,048 r8735340401e0 1,967 r8734340101ec 1.001354036 seconds time elapsed 0.001421000 seconds user 0.000000000 seconds sys Result after the patch changes: [command]# perf stat -e "{r8735340401e0,r8734340101ec}" sleep 1 Error: The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (r8735340401e0). /bin/dmesg | grep -i perf may provide additional information. Fixes: 78a16d9fc1206 ("powerpc/perf: Avoid FAB_*_MATCH checks for power9") Signed-off-by: Kajol Jain Reviewed-by: Athira Rajeev Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20220506061015.43916-2-kjain@linux.ibm.com

powerpc/perf: Fix the threshold compare group constraint for power10

2022-05-22T05:58:29+00:00

Thresh compare bits for a event is used to program thresh compare field in Monitor Mode Control Register A (MMCRA: 8-18 bits for power10). When scheduling events as a group, all events in that group should match value in threshold bits. Otherwise event open for the sibling events should fail. But in the current code, incase thresh compare bits are not valid, we are not failing in group_constraint function which can result in invalid group schduling. Fix the issue by returning -1 incase event is threshold and threshold compare value is not valid in group_constraint function. Patch also fixes the p10_thresh_cmp_val function to return -1, incase threshold bits are not valid and changes corresponding check in is_thresh_cmp_valid function to return false only when the thresh_cmp value is less then 0. Thresh control bits in the event code is used to program thresh_ctl field in Monitor Mode Control Register A (MMCRA: 48-55). In below example, the scheduling of group events PM_MRK_INST_CMPL (3534401e0) and PM_THRESH_MET (34340101ec) is expected to fail as both event request different thresh control bits. Result before the patch changes: [command]# perf stat -e "{r35340401e0,r34340101ec}" sleep 1 Performance counter stats for 'sleep 1': 8,482 r35340401e0 0 r34340101ec 1.001474838 seconds time elapsed 0.001145000 seconds user 0.000000000 seconds sys Result after the patch changes: [command]# perf stat -e "{r35340401e0,r34340101ec}" sleep 1 Performance counter stats for 'sleep 1': r35340401e0 r34340101ec 1.001499607 seconds time elapsed 0.000204000 seconds user 0.000760000 seconds sys Fixes: 82d2c16b350f7 ("powerpc/perf: Adds support for programming of Thresholding in P10") Signed-off-by: Kajol Jain Reviewed-by: Athira Rajeev Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20220506061015.43916-1-kjain@linux.ibm.com

powerpc: Add missing headers

2022-05-08T12:15:40+00:00

Don't inherit headers "by chances" from asm/prom.h, asm/mpc52xx.h, asm/pci.h etc... Include the needed headers, and remove asm/prom.h when it was needed exclusively for pulling necessary headers. Signed-off-by: Christophe Leroy Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/be8bdc934d152a7d8ee8d1a840d5596e2f7d85e0.1646767214.git.christophe.leroy@csgroup.eu

powerpc: fix typos in comments

2022-05-05T12:12:44+00:00

Various spelling mistakes in comments. Detected with the help of Coccinelle. Signed-off-by: Julia Lawall Reviewed-by: Joel Stanley Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20220430185654.5855-1-Julia.Lawall@inria.fr

Merge tag 'v5.18-rc4' into next

2022-05-05T12:09:35+00:00

Merge master into next, to bring in commit 5f24d5a579d1 ("mm, hugetlb: allow for "high" userspace addresses"), which is needed as a prerequisite for the series converting powerpc to the generic mmap logic.

powerpc/perf/24x7: use 'unsigned int' instead of 'unsigned'

2022-05-04T09:37:44+00:00

Replace the 'unsigned' with 'unsigned int' which is more accurate. Signed-off-by: Jason Wang [chleroy: Fixed parenthesis alignment] Signed-off-by: Christophe Leroy Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20210729115252.40895-1-wangborong@cdjrlc.com

powerpc/perf: Fix symbol undeclared warning

2022-05-02T13:02:15+00:00

Build kernel with `C=2`: arch/powerpc/perf/isa207-common.c:24:18: warning: symbol 'isa207_pmu_format_attr' was not declared. Should it be static? arch/powerpc/perf/power9-pmu.c:101:5: warning: symbol 'p9_dd21_bl_ev' was not declared. Should it be static? arch/powerpc/perf/power9-pmu.c:115:5: warning: symbol 'p9_dd22_bl_ev' was not declared. Should it be static? Those symbols are used only in the files that define them so we declare them as static to fix the warnings. Signed-off-by: Wang Wensheng Reviewed-by: Athira Rajeev Signed-off-by: Christophe Leroy Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20200923071453.2540-1-wangwensheng4@huawei.com

powerpc/perf: Fix 32bit compile

2022-04-21T13:26:47+00:00

The "read_bhrb" global symbol is only called under CONFIG_PPC64 of arch/powerpc/perf/core-book3s.c but it is compiled for both 32 and 64 bit anyway (and LLVM fails to link this on 32bit). This fixes it by moving bhrb.o to obj64 targets. Signed-off-by: Alexey Kardashevskiy Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20220421025756.571995-1-aik@ozlabs.ru

powerpc/perf: Fix power10 event alternatives

2022-04-21T13:25:33+00:00

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 Reviewed-by: Madhavan Srinivasan Signed-off-by: Michael Ellerman Link: https://lore.kernel.org/r/20220419114828.89843-2-atrajeev@linux.vnet.ibm.com