aboutsummaryrefslogtreecommitdiff
path: root/arch/powerpc/perf
diff options
context:
space:
mode:
Diffstat (limited to 'arch/powerpc/perf')
-rw-r--r--arch/powerpc/perf/Makefile2
-rw-r--r--arch/powerpc/perf/core-book3s.c201
-rw-r--r--arch/powerpc/perf/hv-24x7-catalog.h33
-rw-r--r--arch/powerpc/perf/hv-24x7.c523
-rw-r--r--arch/powerpc/perf/hv-24x7.h109
-rw-r--r--arch/powerpc/perf/hv-common.c39
-rw-r--r--arch/powerpc/perf/hv-common.h36
-rw-r--r--arch/powerpc/perf/hv-gpci.c294
-rw-r--r--arch/powerpc/perf/hv-gpci.h73
-rw-r--r--arch/powerpc/perf/power7-events-list.h10
-rw-r--r--arch/powerpc/perf/power8-pmu.c229
11 files changed, 1512 insertions, 37 deletions
diff --git a/arch/powerpc/perf/Makefile b/arch/powerpc/perf/Makefile
index 60d71eea919..f9c083a5652 100644
--- a/arch/powerpc/perf/Makefile
+++ b/arch/powerpc/perf/Makefile
@@ -11,5 +11,7 @@ obj32-$(CONFIG_PPC_PERF_CTRS) += mpc7450-pmu.o
obj-$(CONFIG_FSL_EMB_PERF_EVENT) += core-fsl-emb.o
obj-$(CONFIG_FSL_EMB_PERF_EVENT_E500) += e500-pmu.o e6500-pmu.o
+obj-$(CONFIG_HV_PERF_CTRS) += hv-24x7.o hv-gpci.o hv-common.o
+
obj-$(CONFIG_PPC64) += $(obj64-y)
obj-$(CONFIG_PPC32) += $(obj32-y)
diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
index 29b89e863d7..fe52db2eea6 100644
--- a/arch/powerpc/perf/core-book3s.c
+++ b/arch/powerpc/perf/core-book3s.c
@@ -78,6 +78,7 @@ static unsigned int freeze_events_kernel = MMCR0_FCS;
#define MMCR0_FC56 0
#define MMCR0_PMAO 0
#define MMCR0_EBE 0
+#define MMCR0_BHRBA 0
#define MMCR0_PMCC 0
#define MMCR0_PMCC_U6 0
@@ -120,6 +121,7 @@ static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
void power_pmu_flush_branch_stack(void) {}
static inline void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw) {}
+static void pmao_restore_workaround(bool ebb) { }
#endif /* CONFIG_PPC32 */
static bool regs_use_siar(struct pt_regs *regs)
@@ -483,7 +485,7 @@ static bool is_ebb_event(struct perf_event *event)
* check that the PMU supports EBB, meaning those that don't can still
* use bit 63 of the event code for something else if they wish.
*/
- return (ppmu->flags & PPMU_EBB) &&
+ return (ppmu->flags & PPMU_ARCH_207S) &&
((event->attr.config >> PERF_EVENT_CONFIG_EBB_SHIFT) & 1);
}
@@ -502,8 +504,11 @@ static int ebb_event_check(struct perf_event *event)
if (!leader->attr.pinned || !leader->attr.exclusive)
return -EINVAL;
- if (event->attr.inherit || event->attr.sample_period ||
- event->attr.enable_on_exec || event->attr.freq)
+ if (event->attr.freq ||
+ event->attr.inherit ||
+ event->attr.sample_type ||
+ event->attr.sample_period ||
+ event->attr.enable_on_exec)
return -EINVAL;
}
@@ -542,13 +547,21 @@ static unsigned long ebb_switch_in(bool ebb, unsigned long mmcr0)
if (!ebb)
goto out;
- /* Enable EBB and read/write to all 6 PMCs for userspace */
- mmcr0 |= MMCR0_EBE | MMCR0_PMCC_U6;
+ /* Enable EBB and read/write to all 6 PMCs and BHRB for userspace */
+ mmcr0 |= MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC_U6;
- /* Add any bits from the user reg, FC or PMAO */
+ /*
+ * Add any bits from the user MMCR0, FC or PMAO. This is compatible
+ * with pmao_restore_workaround() because we may add PMAO but we never
+ * clear it here.
+ */
mmcr0 |= current->thread.mmcr0;
- /* Be careful not to set PMXE if userspace had it cleared */
+ /*
+ * Be careful not to set PMXE if userspace had it cleared. This is also
+ * compatible with pmao_restore_workaround() because it has already
+ * cleared PMXE and we leave PMAO alone.
+ */
if (!(current->thread.mmcr0 & MMCR0_PMXE))
mmcr0 &= ~MMCR0_PMXE;
@@ -559,13 +572,94 @@ static unsigned long ebb_switch_in(bool ebb, unsigned long mmcr0)
out:
return mmcr0;
}
-#endif /* CONFIG_PPC64 */
-static void perf_event_interrupt(struct pt_regs *regs);
-
-void perf_event_print_debug(void)
+static void pmao_restore_workaround(bool ebb)
{
+ unsigned pmcs[6];
+
+ if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
+ return;
+
+ /*
+ * On POWER8E there is a hardware defect which affects the PMU context
+ * switch logic, ie. power_pmu_disable/enable().
+ *
+ * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
+ * by the hardware. Sometime later the actual PMU exception is
+ * delivered.
+ *
+ * If we context switch, or simply disable/enable, the PMU prior to the
+ * exception arriving, the exception will be lost when we clear PMAO.
+ *
+ * When we reenable the PMU, we will write the saved MMCR0 with PMAO
+ * set, and this _should_ generate an exception. However because of the
+ * defect no exception is generated when we write PMAO, and we get
+ * stuck with no counters counting but no exception delivered.
+ *
+ * The workaround is to detect this case and tweak the hardware to
+ * create another pending PMU exception.
+ *
+ * We do that by setting up PMC6 (cycles) for an imminent overflow and
+ * enabling the PMU. That causes a new exception to be generated in the
+ * chip, but we don't take it yet because we have interrupts hard
+ * disabled. We then write back the PMU state as we want it to be seen
+ * by the exception handler. When we reenable interrupts the exception
+ * handler will be called and see the correct state.
+ *
+ * The logic is the same for EBB, except that the exception is gated by
+ * us having interrupts hard disabled as well as the fact that we are
+ * not in userspace. The exception is finally delivered when we return
+ * to userspace.
+ */
+
+ /* Only if PMAO is set and PMAO_SYNC is clear */
+ if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
+ return;
+
+ /* If we're doing EBB, only if BESCR[GE] is set */
+ if (ebb && !(current->thread.bescr & BESCR_GE))
+ return;
+
+ /*
+ * We are already soft-disabled in power_pmu_enable(). We need to hard
+ * enable to actually prevent the PMU exception from firing.
+ */
+ hard_irq_disable();
+
+ /*
+ * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
+ * Using read/write_pmc() in a for loop adds 12 function calls and
+ * almost doubles our code size.
+ */
+ pmcs[0] = mfspr(SPRN_PMC1);
+ pmcs[1] = mfspr(SPRN_PMC2);
+ pmcs[2] = mfspr(SPRN_PMC3);
+ pmcs[3] = mfspr(SPRN_PMC4);
+ pmcs[4] = mfspr(SPRN_PMC5);
+ pmcs[5] = mfspr(SPRN_PMC6);
+
+ /* Ensure all freeze bits are unset */
+ mtspr(SPRN_MMCR2, 0);
+
+ /* Set up PMC6 to overflow in one cycle */
+ mtspr(SPRN_PMC6, 0x7FFFFFFE);
+
+ /* Enable exceptions and unfreeze PMC6 */
+ mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
+
+ /* Now we need to refreeze and restore the PMCs */
+ mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
+
+ mtspr(SPRN_PMC1, pmcs[0]);
+ mtspr(SPRN_PMC2, pmcs[1]);
+ mtspr(SPRN_PMC3, pmcs[2]);
+ mtspr(SPRN_PMC4, pmcs[3]);
+ mtspr(SPRN_PMC5, pmcs[4]);
+ mtspr(SPRN_PMC6, pmcs[5]);
}
+#endif /* CONFIG_PPC64 */
+
+static void perf_event_interrupt(struct pt_regs *regs);
/*
* Read one performance monitor counter (PMC).
@@ -645,6 +739,57 @@ static void write_pmc(int idx, unsigned long val)
}
}
+/* Called from sysrq_handle_showregs() */
+void perf_event_print_debug(void)
+{
+ unsigned long sdar, sier, flags;
+ u32 pmcs[MAX_HWEVENTS];
+ int i;
+
+ if (!ppmu->n_counter)
+ return;
+
+ local_irq_save(flags);
+
+ pr_info("CPU: %d PMU registers, ppmu = %s n_counters = %d",
+ smp_processor_id(), ppmu->name, ppmu->n_counter);
+
+ for (i = 0; i < ppmu->n_counter; i++)
+ pmcs[i] = read_pmc(i + 1);
+
+ for (; i < MAX_HWEVENTS; i++)
+ pmcs[i] = 0xdeadbeef;
+
+ pr_info("PMC1: %08x PMC2: %08x PMC3: %08x PMC4: %08x\n",
+ pmcs[0], pmcs[1], pmcs[2], pmcs[3]);
+
+ if (ppmu->n_counter > 4)
+ pr_info("PMC5: %08x PMC6: %08x PMC7: %08x PMC8: %08x\n",
+ pmcs[4], pmcs[5], pmcs[6], pmcs[7]);
+
+ pr_info("MMCR0: %016lx MMCR1: %016lx MMCRA: %016lx\n",
+ mfspr(SPRN_MMCR0), mfspr(SPRN_MMCR1), mfspr(SPRN_MMCRA));
+
+ sdar = sier = 0;
+#ifdef CONFIG_PPC64
+ sdar = mfspr(SPRN_SDAR);
+
+ if (ppmu->flags & PPMU_HAS_SIER)
+ sier = mfspr(SPRN_SIER);
+
+ if (ppmu->flags & PPMU_ARCH_207S) {
+ pr_info("MMCR2: %016lx EBBHR: %016lx\n",
+ mfspr(SPRN_MMCR2), mfspr(SPRN_EBBHR));
+ pr_info("EBBRR: %016lx BESCR: %016lx\n",
+ mfspr(SPRN_EBBRR), mfspr(SPRN_BESCR));
+ }
+#endif
+ pr_info("SIAR: %016lx SDAR: %016lx SIER: %016lx\n",
+ mfspr(SPRN_SIAR), sdar, sier);
+
+ local_irq_restore(flags);
+}
+
/*
* Check if a set of events can all go on the PMU at once.
* If they can't, this will look at alternative codes for the events
@@ -851,7 +996,22 @@ static void power_pmu_read(struct perf_event *event)
} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
local64_add(delta, &event->count);
- local64_sub(delta, &event->hw.period_left);
+
+ /*
+ * A number of places program the PMC with (0x80000000 - period_left).
+ * We never want period_left to be less than 1 because we will program
+ * the PMC with a value >= 0x800000000 and an edge detected PMC will
+ * roll around to 0 before taking an exception. We have seen this
+ * on POWER8.
+ *
+ * To fix this, clamp the minimum value of period_left to 1.
+ */
+ do {
+ prev = local64_read(&event->hw.period_left);
+ val = prev - delta;
+ if (val < 1)
+ val = 1;
+ } while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
}
/*
@@ -973,11 +1133,12 @@ static void power_pmu_disable(struct pmu *pmu)
}
/*
- * Set the 'freeze counters' bit, clear EBE/PMCC/PMAO/FC56.
+ * Set the 'freeze counters' bit, clear EBE/BHRBA/PMCC/PMAO/FC56
*/
val = mmcr0 = mfspr(SPRN_MMCR0);
val |= MMCR0_FC;
- val &= ~(MMCR0_EBE | MMCR0_PMCC | MMCR0_PMAO | MMCR0_FC56);
+ val &= ~(MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC | MMCR0_PMAO |
+ MMCR0_FC56);
/*
* The barrier is to make sure the mtspr has been
@@ -1144,9 +1305,17 @@ static void power_pmu_enable(struct pmu *pmu)
cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
out_enable:
+ pmao_restore_workaround(ebb);
+
+ if (ppmu->flags & PPMU_ARCH_207S)
+ mtspr(SPRN_MMCR2, 0);
+
mmcr0 = ebb_switch_in(ebb, cpuhw->mmcr[0]);
mb();
+ if (cpuhw->bhrb_users)
+ ppmu->config_bhrb(cpuhw->bhrb_filter);
+
write_mmcr0(cpuhw, mmcr0);
/*
@@ -1158,8 +1327,6 @@ static void power_pmu_enable(struct pmu *pmu)
}
out:
- if (cpuhw->bhrb_users)
- ppmu->config_bhrb(cpuhw->bhrb_filter);
local_irq_restore(flags);
}
@@ -1547,7 +1714,7 @@ static int power_pmu_event_init(struct perf_event *event)
if (has_branch_stack(event)) {
/* PMU has BHRB enabled */
- if (!(ppmu->flags & PPMU_BHRB))
+ if (!(ppmu->flags & PPMU_ARCH_207S))
return -EOPNOTSUPP;
}
diff --git a/arch/powerpc/perf/hv-24x7-catalog.h b/arch/powerpc/perf/hv-24x7-catalog.h
new file mode 100644
index 00000000000..21b19dd86d9
--- /dev/null
+++ b/arch/powerpc/perf/hv-24x7-catalog.h
@@ -0,0 +1,33 @@
+#ifndef LINUX_POWERPC_PERF_HV_24X7_CATALOG_H_
+#define LINUX_POWERPC_PERF_HV_24X7_CATALOG_H_
+
+#include <linux/types.h>
+
+/* From document "24x7 Event and Group Catalog Formats Proposal" v0.15 */
+
+struct hv_24x7_catalog_page_0 {
+#define HV_24X7_CATALOG_MAGIC 0x32347837 /* "24x7" in ASCII */
+ __be32 magic;
+ __be32 length; /* In 4096 byte pages */
+ __be64 version; /* XXX: arbitrary? what's the meaning/useage/purpose? */
+ __u8 build_time_stamp[16]; /* "YYYYMMDDHHMMSS\0\0" */
+ __u8 reserved2[32];
+ __be16 schema_data_offs; /* in 4096 byte pages */
+ __be16 schema_data_len; /* in 4096 byte pages */
+ __be16 schema_entry_count;
+ __u8 reserved3[2];
+ __be16 event_data_offs;
+ __be16 event_data_len;
+ __be16 event_entry_count;
+ __u8 reserved4[2];
+ __be16 group_data_offs; /* in 4096 byte pages */
+ __be16 group_data_len; /* in 4096 byte pages */
+ __be16 group_entry_count;
+ __u8 reserved5[2];
+ __be16 formula_data_offs; /* in 4096 byte pages */
+ __be16 formula_data_len; /* in 4096 byte pages */
+ __be16 formula_entry_count;
+ __u8 reserved6[2];
+} __packed;
+
+#endif
diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
new file mode 100644
index 00000000000..e0766b82e16
--- /dev/null
+++ b/arch/powerpc/perf/hv-24x7.c
@@ -0,0 +1,523 @@
+/*
+ * Hypervisor supplied "24x7" performance counter support
+ *
+ * Author: Cody P Schafer <cody@linux.vnet.ibm.com>
+ * Copyright 2014 IBM Corporation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#define pr_fmt(fmt) "hv-24x7: " fmt
+
+#include <linux/perf_event.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <asm/firmware.h>
+#include <asm/hvcall.h>
+#include <asm/io.h>
+
+#include "hv-24x7.h"
+#include "hv-24x7-catalog.h"
+#include "hv-common.h"
+
+/*
+ * TODO: Merging events:
+ * - Think of the hcall as an interface to a 4d array of counters:
+ * - x = domains
+ * - y = indexes in the domain (core, chip, vcpu, node, etc)
+ * - z = offset into the counter space
+ * - w = lpars (guest vms, "logical partitions")
+ * - A single request is: x,y,y_last,z,z_last,w,w_last
+ * - this means we can retrieve a rectangle of counters in y,z for a single x.
+ *
+ * - Things to consider (ignoring w):
+ * - input cost_per_request = 16
+ * - output cost_per_result(ys,zs) = 8 + 8 * ys + ys * zs
+ * - limited number of requests per hcall (must fit into 4K bytes)
+ * - 4k = 16 [buffer header] - 16 [request size] * request_count
+ * - 255 requests per hcall
+ * - sometimes it will be more efficient to read extra data and discard
+ */
+
+/*
+ * Example usage:
+ * perf stat -e 'hv_24x7/domain=2,offset=8,starting_index=0,lpar=0xffffffff/'
+ */
+
+/* u3 0-6, one of HV_24X7_PERF_DOMAIN */
+EVENT_DEFINE_RANGE_FORMAT(domain, config, 0, 3);
+/* u16 */
+EVENT_DEFINE_RANGE_FORMAT(starting_index, config, 16, 31);
+/* u32, see "data_offset" */
+EVENT_DEFINE_RANGE_FORMAT(offset, config, 32, 63);
+/* u16 */
+EVENT_DEFINE_RANGE_FORMAT(lpar, config1, 0, 15);
+
+EVENT_DEFINE_RANGE(reserved1, config, 4, 15);
+EVENT_DEFINE_RANGE(reserved2, config1, 16, 63);
+EVENT_DEFINE_RANGE(reserved3, config2, 0, 63);
+
+static struct attribute *format_attrs[] = {
+ &format_attr_domain.attr,
+ &format_attr_offset.attr,
+ &format_attr_starting_index.attr,
+ &format_attr_lpar.attr,
+ NULL,
+};
+
+static struct attribute_group format_group = {
+ .name = "format",
+ .attrs = format_attrs,
+};
+
+static struct kmem_cache *hv_page_cache;
+
+/*
+ * read_offset_data - copy data from one buffer to another while treating the
+ * source buffer as a small view on the total avaliable
+ * source data.
+ *
+ * @dest: buffer to copy into
+ * @dest_len: length of @dest in bytes
+ * @requested_offset: the offset within the source data we want. Must be > 0
+ * @src: buffer to copy data from
+ * @src_len: length of @src in bytes
+ * @source_offset: the offset in the sorce data that (src,src_len) refers to.
+ * Must be > 0
+ *
+ * returns the number of bytes copied.
+ *
+ * The following ascii art shows the various buffer possitioning we need to
+ * handle, assigns some arbitrary varibles to points on the buffer, and then
+ * shows how we fiddle with those values to get things we care about (copy
+ * start in src and copy len)
+ *
+ * s = @src buffer
+ * d = @dest buffer
+ * '.' areas in d are written to.
+ *
+ * u
+ * x w v z
+ * d |.........|
+ * s |----------------------|
+ *
+ * u
+ * x w z v
+ * d |........------|
+ * s |------------------|
+ *
+ * x w u,z,v
+ * d |........|
+ * s |------------------|
+ *
+ * x,w u,v,z
+ * d |..................|
+ * s |------------------|
+ *
+ * x u
+ * w v z
+ * d |........|
+ * s |------------------|
+ *
+ * x z w v
+ * d |------|
+ * s |------|
+ *
+ * x = source_offset
+ * w = requested_offset
+ * z = source_offset + src_len
+ * v = requested_offset + dest_len
+ *
+ * w_offset_in_s = w - x = requested_offset - source_offset
+ * z_offset_in_s = z - x = src_len
+ * v_offset_in_s = v - x = request_offset + dest_len - src_len
+ */
+static ssize_t read_offset_data(void *dest, size_t dest_len,
+ loff_t requested_offset, void *src,
+ size_t src_len, loff_t source_offset)
+{
+ size_t w_offset_in_s = requested_offset - source_offset;
+ size_t z_offset_in_s = src_len;
+ size_t v_offset_in_s = requested_offset + dest_len - src_len;
+ size_t u_offset_in_s = min(z_offset_in_s, v_offset_in_s);
+ size_t copy_len = u_offset_in_s - w_offset_in_s;
+
+ if (requested_offset < 0 || source_offset < 0)
+ return -EINVAL;
+
+ if (z_offset_in_s <= w_offset_in_s)
+ return 0;
+
+ memcpy(dest, src + w_offset_in_s, copy_len);
+ return copy_len;
+}
+
+static unsigned long h_get_24x7_catalog_page_(unsigned long phys_4096,
+ unsigned long version,
+ unsigned long index)
+{
+ pr_devel("h_get_24x7_catalog_page(0x%lx, %lu, %lu)",
+ phys_4096,
+ version,
+ index);
+ WARN_ON(!IS_ALIGNED(phys_4096, 4096));
+ return plpar_hcall_norets(H_GET_24X7_CATALOG_PAGE,
+ phys_4096,
+ version,
+ index);
+}
+
+static unsigned long h_get_24x7_catalog_page(char page[],
+ u64 version, u32 index)
+{
+ return h_get_24x7_catalog_page_(virt_to_phys(page),
+ version, index);
+}
+
+static ssize_t catalog_read(struct file *filp, struct kobject *kobj,
+ struct bin_attribute *bin_attr, char *buf,
+ loff_t offset, size_t count)
+{
+ unsigned long hret;
+ ssize_t ret = 0;
+ size_t catalog_len = 0, catalog_page_len = 0, page_count = 0;
+ loff_t page_offset = 0;
+ uint64_t catalog_version_num = 0;
+ void *page = kmem_cache_alloc(hv_page_cache, GFP_USER);
+ struct hv_24x7_catalog_page_0 *page_0 = page;
+ if (!page)
+ return -ENOMEM;
+
+ hret = h_get_24x7_catalog_page(page, 0, 0);
+ if (hret) {
+ ret = -EIO;
+ goto e_free;
+ }
+
+ catalog_version_num = be64_to_cpu(page_0->version);
+ catalog_page_len = be32_to_cpu(page_0->length);
+ catalog_len = catalog_page_len * 4096;
+
+ page_offset = offset / 4096;
+ page_count = count / 4096;
+
+ if (page_offset >= catalog_page_len)
+ goto e_free;
+
+ if (page_offset != 0) {
+ hret = h_get_24x7_catalog_page(page, catalog_version_num,
+ page_offset);
+ if (hret) {
+ ret = -EIO;
+ goto e_free;
+ }
+ }
+
+ ret = read_offset_data(buf, count, offset,
+ page, 4096, page_offset * 4096);
+e_free:
+ if (hret)
+ pr_err("h_get_24x7_catalog_page(ver=%lld, page=%lld) failed:"
+ " rc=%ld\n",
+ catalog_version_num, page_offset, hret);
+ kfree(page);
+
+ pr_devel("catalog_read: offset=%lld(%lld) count=%zu(%zu) catalog_len=%zu(%zu) => %zd\n",
+ offset, page_offset, count, page_count, catalog_len,
+ catalog_page_len, ret);
+
+ return ret;
+}
+
+#define PAGE_0_ATTR(_name, _fmt, _expr) \
+static ssize_t _name##_show(struct device *dev, \
+ struct device_attribute *dev_attr, \
+ char *buf) \
+{ \
+ unsigned long hret; \
+ ssize_t ret = 0; \
+ void *page = kmem_cache_alloc(hv_page_cache, GFP_USER); \
+ struct hv_24x7_catalog_page_0 *page_0 = page; \
+ if (!page) \
+ return -ENOMEM; \
+ hret = h_get_24x7_catalog_page(page, 0, 0); \
+ if (hret) { \
+ ret = -EIO; \
+ goto e_free; \
+ } \
+ ret = sprintf(buf, _fmt, _expr); \
+e_free: \
+ kfree(page); \
+ return ret; \
+} \
+static DEVICE_ATTR_RO(_name)
+
+PAGE_0_ATTR(catalog_version, "%lld\n",
+ (unsigned long long)be64_to_cpu(page_0->version));
+PAGE_0_ATTR(catalog_len, "%lld\n",
+ (unsigned long long)be32_to_cpu(page_0->length) * 4096);
+static BIN_ATTR_RO(catalog, 0/* real length varies */);
+
+static struct bin_attribute *if_bin_attrs[] = {
+ &bin_attr_catalog,
+ NULL,
+};
+
+static struct attribute *if_attrs[] = {
+ &dev_attr_catalog_len.attr,
+ &dev_attr_catalog_version.attr,
+ NULL,
+};
+
+static struct attribute_group if_group = {
+ .name = "interface",
+ .bin_attrs = if_bin_attrs,
+ .attrs = if_attrs,
+};
+
+static const struct attribute_group *attr_groups[] = {
+ &format_group,
+ &if_group,
+ NULL,
+};
+
+static bool is_physical_domain(int domain)
+{
+ return domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CHIP ||
+ domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CORE;
+}
+
+static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix,
+ u16 lpar, u64 *res,
+ bool success_expected)
+{
+ unsigned long ret;
+
+ /*
+ * request_buffer and result_buffer are not required to be 4k aligned,
+ * but are not allowed to cross any 4k boundary. Aligning them to 4k is
+ * the simplest way to ensure that.
+ */
+ struct reqb {
+ struct hv_24x7_request_buffer buf;
+ struct hv_24x7_request req;
+ } __packed __aligned(4096) request_buffer = {
+ .buf = {
+ .interface_version = HV_24X7_IF_VERSION_CURRENT,
+ .num_requests = 1,
+ },
+ .req = {
+ .performance_domain = domain,
+ .data_size = cpu_to_be16(8),
+ .data_offset = cpu_to_be32(offset),
+ .starting_lpar_ix = cpu_to_be16(lpar),
+ .max_num_lpars = cpu_to_be16(1),
+ .starting_ix = cpu_to_be16(ix),
+ .max_ix = cpu_to_be16(1),
+ }
+ };
+
+ struct resb {
+ struct hv_24x7_data_result_buffer buf;
+ struct hv_24x7_result res;
+ struct hv_24x7_result_element elem;
+ __be64 result;
+ } __packed __aligned(4096) result_buffer = {};
+
+ ret = plpar_hcall_norets(H_GET_24X7_DATA,
+ virt_to_phys(&request_buffer), sizeof(request_buffer),
+ virt_to_phys(&result_buffer), sizeof(result_buffer));
+
+ if (ret) {
+ if (success_expected)
+ pr_err_ratelimited("hcall failed: %d %#x %#x %d => 0x%lx (%ld) detail=0x%x failing ix=%x\n",
+ domain, offset, ix, lpar,
+ ret, ret,
+ result_buffer.buf.detailed_rc,
+ result_buffer.buf.failing_request_ix);
+ return ret;
+ }
+
+ *res = be64_to_cpu(result_buffer.result);
+ return ret;
+}
+
+static unsigned long event_24x7_request(struct perf_event *event, u64 *res,
+ bool success_expected)
+{
+ return single_24x7_request(event_get_domain(event),
+ event_get_offset(event),
+ event_get_starting_index(event),
+ event_get_lpar(event),
+ res,
+ success_expected);
+}
+
+static int h_24x7_event_init(struct perf_event *event)
+{
+ struct hv_perf_caps caps;
+ unsigned domain;
+ unsigned long hret;
+ u64 ct;
+
+ /* Not our event */
+ if (event->attr.type != event->pmu->type)
+ return -ENOENT;
+
+ /* Unused areas must be 0 */
+ if (event_get_reserved1(event) ||
+ event_get_reserved2(event) ||
+ event_get_reserved3(event)) {
+ pr_devel("reserved set when forbidden 0x%llx(0x%llx) 0x%llx(0x%llx) 0x%llx(0x%llx)\n",
+ event->attr.config,
+ event_get_reserved1(event),
+ event->attr.config1,
+ event_get_reserved2(event),
+ event->attr.config2,
+ event_get_reserved3(event));
+ return -EINVAL;
+ }
+
+ /* unsupported modes and filters */
+ if (event->attr.exclude_user ||
+ event->attr.exclude_kernel ||
+ event->attr.exclude_hv ||
+ event->attr.exclude_idle ||
+ event->attr.exclude_host ||
+ event->attr.exclude_guest ||
+ is_sampling_event(event)) /* no sampling */
+ return -EINVAL;
+
+ /* no branch sampling */
+ if (has_branch_stack(event))
+ return -EOPNOTSUPP;
+
+ /* offset must be 8 byte aligned */
+ if (event_get_offset(event) % 8) {
+ pr_devel("bad alignment\n");
+ return -EINVAL;
+ }
+
+ /* Domains above 6 are invalid */
+ domain = event_get_domain(event);
+ if (domain > 6) {
+ pr_devel("invalid domain %d\n", domain);
+ return -EINVAL;
+ }
+
+ hret = hv_perf_caps_get(&caps);
+ if (hret) {
+ pr_devel("could not get capabilities: rc=%ld\n", hret);
+ return -EIO;
+ }
+
+ /* PHYSICAL domains & other lpars require extra capabilities */
+ if (!caps.collect_privileged && (is_physical_domain(domain) ||
+ (event_get_lpar(event) != event_get_lpar_max()))) {
+ pr_devel("hv permisions disallow: is_physical_domain:%d, lpar=0x%llx\n",
+ is_physical_domain(domain),
+ event_get_lpar(event));
+ return -EACCES;
+ }
+
+ /* see if the event complains */
+ if (event_24x7_request(event, &ct, false)) {
+ pr_devel("test hcall failed\n");
+ return -EIO;
+ }
+
+ return 0;
+}
+
+static u64 h_24x7_get_value(struct perf_event *event)
+{
+ unsigned long ret;
+ u64 ct;
+ ret = event_24x7_request(event, &ct, true);
+ if (ret)
+ /* We checked this in event init, shouldn't fail here... */
+ return 0;
+
+ return ct;
+}
+
+static void h_24x7_event_update(struct perf_event *event)
+{
+ s64 prev;
+ u64 now;
+ now = h_24x7_get_value(event);
+ prev = local64_xchg(&event->hw.prev_count, now);
+ local64_add(now - prev, &event->count);
+}
+
+static void h_24x7_event_start(struct perf_event *event, int flags)
+{
+ if (flags & PERF_EF_RELOAD)
+ local64_set(&event->hw.prev_count, h_24x7_get_value(event));
+}
+
+static void h_24x7_event_stop(struct perf_event *event, int flags)
+{
+ h_24x7_event_update(event);
+}
+
+static int h_24x7_event_add(struct perf_event *event, int flags)
+{
+ if (flags & PERF_EF_START)
+ h_24x7_event_start(event, flags);
+
+ return 0;
+}
+
+static int h_24x7_event_idx(struct perf_event *event)
+{
+ return 0;
+}
+
+static struct pmu h_24x7_pmu = {
+ .task_ctx_nr = perf_invalid_context,
+
+ .name = "hv_24x7",
+ .attr_groups = attr_groups,
+ .event_init = h_24x7_event_init,
+ .add = h_24x7_event_add,
+ .del = h_24x7_event_stop,
+ .start = h_24x7_event_start,
+ .stop = h_24x7_event_stop,
+ .read = h_24x7_event_update,
+ .event_idx = h_24x7_event_idx,
+};
+
+static int hv_24x7_init(void)
+{
+ int r;
+ unsigned long hret;
+ struct hv_perf_caps caps;
+
+ if (!firmware_has_feature(FW_FEATURE_LPAR)) {
+ pr_debug("not a virtualized system, not enabling\n");
+ return -ENODEV;
+ }
+
+ hret = hv_perf_caps_get(&caps);
+ if (hret) {
+ pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
+ hret);
+ return -ENODEV;
+ }
+
+ hv_page_cache = kmem_cache_create("hv-page-4096", 4096, 4096, 0, NULL);
+ if (!hv_page_cache)
+ return -ENOMEM;
+
+ r = perf_pmu_register(&h_24x7_pmu, h_24x7_pmu.name, -1);
+ if (r)
+ return r;
+
+ return 0;
+}
+
+device_initcall(hv_24x7_init);
diff --git a/arch/powerpc/perf/hv-24x7.h b/arch/powerpc/perf/hv-24x7.h
new file mode 100644
index 00000000000..720ebce4b43
--- /dev/null
+++ b/arch/powerpc/perf/hv-24x7.h
@@ -0,0 +1,109 @@
+#ifndef LINUX_POWERPC_PERF_HV_24X7_H_
+#define LINUX_POWERPC_PERF_HV_24X7_H_
+
+#include <linux/types.h>
+
+struct hv_24x7_request {
+ /* PHYSICAL domains require enabling via phyp/hmc. */
+#define HV_24X7_PERF_DOMAIN_PHYSICAL_CHIP 0x01
+#define HV_24X7_PERF_DOMAIN_PHYSICAL_CORE 0x02
+#define HV_24X7_PERF_DOMAIN_VIRTUAL_PROCESSOR_HOME_CORE 0x03
+#define HV_24X7_PERF_DOMAIN_VIRTUAL_PROCESSOR_HOME_CHIP 0x04
+#define HV_24X7_PERF_DOMAIN_VIRTUAL_PROCESSOR_HOME_NODE 0x05
+#define HV_24X7_PERF_DOMAIN_VIRTUAL_PROCESSOR_REMOTE_NODE 0x06
+ __u8 performance_domain;
+ __u8 reserved[0x1];
+
+ /* bytes to read starting at @data_offset. must be a multiple of 8 */
+ __be16 data_size;
+
+ /*
+ * byte offset within the perf domain to read from. must be 8 byte
+ * aligned
+ */
+ __be32 data_offset;
+
+ /*
+ * only valid for VIRTUAL_PROCESSOR domains, ignored for others.
+ * -1 means "current partition only"
+ * Enabling via phyp/hmc required for non-"-1" values. 0 forbidden
+ * unless requestor is 0.
+ */
+ __be16 starting_lpar_ix;
+
+ /*
+ * Ignored when @starting_lpar_ix == -1
+ * Ignored when @performance_domain is not VIRTUAL_PROCESSOR_*
+ * -1 means "infinite" or all
+ */
+ __be16 max_num_lpars;
+
+ /* chip, core, or virtual processor based on @performance_domain */
+ __be16 starting_ix;
+ __be16 max_ix;
+} __packed;
+
+struct hv_24x7_request_buffer {
+ /* 0 - ? */
+ /* 1 - ? */
+#define HV_24X7_IF_VERSION_CURRENT 0x01
+ __u8 interface_version;
+ __u8 num_requests;
+ __u8 reserved[0xE];
+ struct hv_24x7_request requests[];
+} __packed;
+
+struct hv_24x7_result_element {
+ __be16 lpar_ix;
+
+ /*
+ * represents the core, chip, or virtual processor based on the
+ * request's @performance_domain
+ */
+ __be16 domain_ix;
+
+ /* -1 if @performance_domain does not refer to a virtual processor */
+ __be32 lpar_cfg_instance_id;
+
+ /* size = @result_element_data_size of cointaining result. */
+ __u8 element_data[];
+} __packed;
+
+struct hv_24x7_result {
+ __u8 result_ix;
+
+ /*
+ * 0 = not all result elements fit into the buffer, additional requests
+ * required
+ * 1 = all result elements were returned
+ */
+ __u8 results_complete;
+ __be16 num_elements_returned;
+
+ /* This is a copy of @data_size from the coresponding hv_24x7_request */
+ __be16 result_element_data_size;
+ __u8 reserved[0x2];
+
+ /* WARNING: only valid for first result element due to variable sizes
+ * of result elements */
+ /* struct hv_24x7_result_element[@num_elements_returned] */
+ struct hv_24x7_result_element elements[];
+} __packed;
+
+struct hv_24x7_data_result_buffer {
+ /* See versioning for request buffer */
+ __u8 interface_version;
+
+ __u8 num_results;
+ __u8 reserved[0x1];
+ __u8 failing_request_ix;
+ __be32 detailed_rc;
+ __be64 cec_cfg_instance_id;
+ __be64 catalog_version_num;
+ __u8 reserved2[0x8];
+ /* WARNING: only valid for the first result due to variable sizes of
+ * results */
+ struct hv_24x7_result results[]; /* [@num_results] */
+} __packed;
+
+#endif
diff --git a/arch/powerpc/perf/hv-common.c b/arch/powerpc/perf/hv-common.c
new file mode 100644
index 00000000000..47e02b366f5
--- /dev/null
+++ b/arch/powerpc/perf/hv-common.c
@@ -0,0 +1,39 @@
+#include <asm/io.h>
+#include <asm/hvcall.h>
+
+#include "hv-gpci.h"
+#include "hv-common.h"
+
+unsigned long hv_perf_caps_get(struct hv_perf_caps *caps)
+{
+ unsigned long r;
+ struct p {
+ struct hv_get_perf_counter_info_params params;
+ struct cv_system_performance_capabilities caps;
+ } __packed __aligned(sizeof(uint64_t));
+
+ struct p arg = {
+ .params = {
+ .counter_request = cpu_to_be32(
+ CIR_SYSTEM_PERFORMANCE_CAPABILITIES),
+ .starting_index = cpu_to_be32(-1),
+ .counter_info_version_in = 0,
+ }
+ };
+
+ r = plpar_hcall_norets(H_GET_PERF_COUNTER_INFO,
+ virt_to_phys(&arg), sizeof(arg));
+
+ if (r)
+ return r;
+
+ pr_devel("capability_mask: 0x%x\n", arg.caps.capability_mask);
+
+ caps->version = arg.params.counter_info_version_out;
+ caps->collect_privileged = !!arg.caps.perf_collect_privileged;
+ caps->ga = !!(arg.caps.capability_mask & CV_CM_GA);
+ caps->expanded = !!(arg.caps.capability_mask & CV_CM_EXPANDED);
+ caps->lab = !!(arg.caps.capability_mask & CV_CM_LAB);
+
+ return r;
+}
diff --git a/arch/powerpc/perf/hv-common.h b/arch/powerpc/perf/hv-common.h
new file mode 100644
index 00000000000..5d79cecbd73
--- /dev/null
+++ b/arch/powerpc/perf/hv-common.h
@@ -0,0 +1,36 @@
+#ifndef LINUX_POWERPC_PERF_HV_COMMON_H_
+#define LINUX_POWERPC_PERF_HV_COMMON_H_
+
+#include <linux/perf_event.h>
+#include <linux/types.h>
+
+struct hv_perf_caps {
+ u16 version;
+ u16 collect_privileged:1,
+ ga:1,
+ expanded:1,
+ lab:1,
+ unused:12;
+};
+
+unsigned long hv_perf_caps_get(struct hv_perf_caps *caps);
+
+
+#define EVENT_DEFINE_RANGE_FORMAT(name, attr_var, bit_start, bit_end) \
+PMU_FORMAT_ATTR(name, #attr_var ":" #bit_start "-" #bit_end); \
+EVENT_DEFINE_RANGE(name, attr_var, bit_start, bit_end)
+
+#define EVENT_DEFINE_RANGE(name, attr_var, bit_start, bit_end) \
+static u64 event_get_##name##_max(void) \
+{ \
+ BUILD_BUG_ON((bit_start > bit_end) \
+ || (bit_end >= (sizeof(1ull) * 8))); \
+ return (((1ull << (bit_end - bit_start)) - 1) << 1) + 1; \
+} \
+static u64 event_get_##name(struct perf_event *event) \
+{ \
+ return (event->attr.attr_var >> (bit_start)) & \
+ event_get_##name##_max(); \
+}
+
+#endif
diff --git a/arch/powerpc/perf/hv-gpci.c b/arch/powerpc/perf/hv-gpci.c
new file mode 100644
index 00000000000..c9d399a2df8
--- /dev/null
+++ b/arch/powerpc/perf/hv-gpci.c
@@ -0,0 +1,294 @@
+/*
+ * Hypervisor supplied "gpci" ("get performance counter info") performance
+ * counter support
+ *
+ * Author: Cody P Schafer <cody@linux.vnet.ibm.com>
+ * Copyright 2014 IBM Corporation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#define pr_fmt(fmt) "hv-gpci: " fmt
+
+#include <linux/init.h>
+#include <linux/perf_event.h>
+#include <asm/firmware.h>
+#include <asm/hvcall.h>
+#include <asm/io.h>
+
+#include "hv-gpci.h"
+#include "hv-common.h"
+
+/*
+ * Example usage:
+ * perf stat -e 'hv_gpci/counter_info_version=3,offset=0,length=8,
+ * secondary_index=0,starting_index=0xffffffff,request=0x10/' ...
+ */
+
+/* u32 */
+EVENT_DEFINE_RANGE_FORMAT(request, config, 0, 31);
+/* u32 */
+EVENT_DEFINE_RANGE_FORMAT(starting_index, config, 32, 63);
+/* u16 */
+EVENT_DEFINE_RANGE_FORMAT(secondary_index, config1, 0, 15);
+/* u8 */
+EVENT_DEFINE_RANGE_FORMAT(counter_info_version, config1, 16, 23);
+/* u8, bytes of data (1-8) */
+EVENT_DEFINE_RANGE_FORMAT(length, config1, 24, 31);
+/* u32, byte offset */
+EVENT_DEFINE_RANGE_FORMAT(offset, config1, 32, 63);
+
+static struct attribute *format_attrs[] = {
+ &format_attr_request.attr,
+ &format_attr_starting_index.attr,
+ &format_attr_secondary_index.attr,
+ &format_attr_counter_info_version.attr,
+
+ &format_attr_offset.attr,
+ &format_attr_length.attr,
+ NULL,
+};
+
+static struct attribute_group format_group = {
+ .name = "format",
+ .attrs = format_attrs,
+};
+
+#define HV_CAPS_ATTR(_name, _format) \
+static ssize_t _name##_show(struct device *dev, \
+ struct device_attribute *attr, \
+ char *page) \
+{ \
+ struct hv_perf_caps caps; \
+ unsigned long hret = hv_perf_caps_get(&caps); \
+ if (hret) \
+ return -EIO; \
+ \
+ return sprintf(page, _format, caps._name); \
+} \
+static struct device_attribute hv_caps_attr_##_name = __ATTR_RO(_name)
+
+static ssize_t kernel_version_show(struct device *dev,
+ struct device_attribute *attr,
+ char *page)
+{
+ return sprintf(page, "0x%x\n", COUNTER_INFO_VERSION_CURRENT);
+}
+
+static DEVICE_ATTR_RO(kernel_version);
+HV_CAPS_ATTR(version, "0x%x\n");
+HV_CAPS_ATTR(ga, "%d\n");
+HV_CAPS_ATTR(expanded, "%d\n");
+HV_CAPS_ATTR(lab, "%d\n");
+HV_CAPS_ATTR(collect_privileged, "%d\n");
+
+static struct attribute *interface_attrs[] = {
+ &dev_attr_kernel_version.attr,
+ &hv_caps_attr_version.attr,
+ &hv_caps_attr_ga.attr,
+ &hv_caps_attr_expanded.attr,
+ &hv_caps_attr_lab.attr,
+ &hv_caps_attr_collect_privileged.attr,
+ NULL,
+};
+
+static struct attribute_group interface_group = {
+ .name = "interface",
+ .attrs = interface_attrs,
+};
+
+static const struct attribute_group *attr_groups[] = {
+ &format_group,
+ &interface_group,
+ NULL,
+};
+
+#define GPCI_MAX_DATA_BYTES \
+ (1024 - sizeof(struct hv_get_perf_counter_info_params))
+
+static unsigned long single_gpci_request(u32 req, u32 starting_index,
+ u16 secondary_index, u8 version_in, u32 offset, u8 length,
+ u64 *value)
+{
+ unsigned long ret;
+ size_t i;
+ u64 count;
+
+ struct {
+ struct hv_get_perf_counter_info_params params;
+ uint8_t bytes[GPCI_MAX_DATA_BYTES];
+ } __packed __aligned(sizeof(uint64_t)) arg = {
+ .params = {
+ .counter_request = cpu_to_be32(req),
+ .starting_index = cpu_to_be32(starting_index),
+ .secondary_index = cpu_to_be16(secondary_index),
+ .counter_info_version_in = version_in,
+ }
+ };
+
+ ret = plpar_hcall_norets(H_GET_PERF_COUNTER_INFO,
+ virt_to_phys(&arg), sizeof(arg));
+ if (ret) {
+ pr_devel("hcall failed: 0x%lx\n", ret);
+ return ret;
+ }
+
+ /*
+ * we verify offset and length are within the zeroed buffer at event
+ * init.
+ */
+ count = 0;
+ for (i = offset; i < offset + length; i++)
+ count |= arg.bytes[i] << (i - offset);
+
+ *value = count;
+ return ret;
+}
+
+static u64 h_gpci_get_value(struct perf_event *event)
+{
+ u64 count;
+ unsigned long ret = single_gpci_request(event_get_request(event),
+ event_get_starting_index(event),
+ event_get_secondary_index(event),
+ event_get_counter_info_version(event),
+ event_get_offset(event),
+ event_get_length(event),
+ &count);
+ if (ret)
+ return 0;
+ return count;
+}
+
+static void h_gpci_event_update(struct perf_event *event)
+{
+ s64 prev;
+ u64 now = h_gpci_get_value(event);
+ prev = local64_xchg(&event->hw.prev_count, now);
+ local64_add(now - prev, &event->count);
+}
+
+static void h_gpci_event_start(struct perf_event *event, int flags)
+{
+ local64_set(&event->hw.prev_count, h_gpci_get_value(event));
+}
+
+static void h_gpci_event_stop(struct perf_event *event, int flags)
+{
+ h_gpci_event_update(event);
+}
+
+static int h_gpci_event_add(struct perf_event *event, int flags)
+{
+ if (flags & PERF_EF_START)
+ h_gpci_event_start(event, flags);
+
+ return 0;
+}
+
+static int h_gpci_event_init(struct perf_event *event)
+{
+ u64 count;
+ u8 length;
+
+ /* Not our event */
+ if (event->attr.type != event->pmu->type)
+ return -ENOENT;
+
+ /* config2 is unused */
+ if (event->attr.config2) {
+ pr_devel("config2 set when reserved\n");
+ return -EINVAL;
+ }
+
+ /* unsupported modes and filters */
+ if (event->attr.exclude_user ||
+ event->attr.exclude_kernel ||
+ event->attr.exclude_hv ||
+ event->attr.exclude_idle ||
+ event->attr.exclude_host ||
+ event->attr.exclude_guest ||
+ is_sampling_event(event)) /* no sampling */
+ return -EINVAL;
+
+ /* no branch sampling */
+ if (has_branch_stack(event))
+ return -EOPNOTSUPP;
+
+ length = event_get_length(event);
+ if (length < 1 || length > 8) {
+ pr_devel("length invalid\n");
+ return -EINVAL;
+ }
+
+ /* last byte within the buffer? */
+ if ((event_get_offset(event) + length) > GPCI_MAX_DATA_BYTES) {
+ pr_devel("request outside of buffer: %zu > %zu\n",
+ (size_t)event_get_offset(event) + length,
+ GPCI_MAX_DATA_BYTES);
+ return -EINVAL;
+ }
+
+ /* check if the request works... */
+ if (single_gpci_request(event_get_request(event),
+ event_get_starting_index(event),
+ event_get_secondary_index(event),
+ event_get_counter_info_version(event),
+ event_get_offset(event),
+ length,
+ &count)) {
+ pr_devel("gpci hcall failed\n");
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+static int h_gpci_event_idx(struct perf_event *event)
+{
+ return 0;
+}
+
+static struct pmu h_gpci_pmu = {
+ .task_ctx_nr = perf_invalid_context,
+
+ .name = "hv_gpci",
+ .attr_groups = attr_groups,
+ .event_init = h_gpci_event_init,
+ .add = h_gpci_event_add,
+ .del = h_gpci_event_stop,
+ .start = h_gpci_event_start,
+ .stop = h_gpci_event_stop,
+ .read = h_gpci_event_update,
+ .event_idx = h_gpci_event_idx,
+};
+
+static int hv_gpci_init(void)
+{
+ int r;
+ unsigned long hret;
+ struct hv_perf_caps caps;
+
+ if (!firmware_has_feature(FW_FEATURE_LPAR)) {
+ pr_debug("not a virtualized system, not enabling\n");
+ return -ENODEV;
+ }
+
+ hret = hv_perf_caps_get(&caps);
+ if (hret) {
+ pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
+ hret);
+ return -ENODEV;
+ }
+
+ r = perf_pmu_register(&h_gpci_pmu, h_gpci_pmu.name, -1);
+ if (r)
+ return r;
+
+ return 0;
+}
+
+device_initcall(hv_gpci_init);
diff --git a/arch/powerpc/perf/hv-gpci.h b/arch/powerpc/perf/hv-gpci.h
new file mode 100644
index 00000000000..b25f460c9cc
--- /dev/null
+++ b/arch/powerpc/perf/hv-gpci.h
@@ -0,0 +1,73 @@
+#ifndef LINUX_POWERPC_PERF_HV_GPCI_H_
+#define LINUX_POWERPC_PERF_HV_GPCI_H_
+
+#include <linux/types.h>
+
+/* From the document "H_GetPerformanceCounterInfo Interface" v1.07 */
+
+/* H_GET_PERF_COUNTER_INFO argument */
+struct hv_get_perf_counter_info_params {
+ __be32 counter_request; /* I */
+ __be32 starting_index; /* IO */
+ __be16 secondary_index; /* IO */
+ __be16 returned_values; /* O */
+ __be32 detail_rc; /* O, only needed when called via *_norets() */
+
+ /*
+ * O, size each of counter_value element in bytes, only set for version
+ * >= 0x3
+ */
+ __be16 cv_element_size;
+
+ /* I, 0 (zero) for versions < 0x3 */
+ __u8 counter_info_version_in;
+
+ /* O, 0 (zero) if version < 0x3. Must be set to 0 when making hcall */
+ __u8 counter_info_version_out;
+ __u8 reserved[0xC];
+ __u8 counter_value[];
+} __packed;
+
+/*
+ * counter info version => fw version/reference (spec version)
+ *
+ * 8 => power8 (1.07)
+ * [7 is skipped by spec 1.07]
+ * 6 => TLBIE (1.07)
+ * 5 => v7r7m0.phyp (1.05)
+ * [4 skipped]
+ * 3 => v7r6m0.phyp (?)
+ * [1,2 skipped]
+ * 0 => v7r{2,3,4}m0.phyp (?)
+ */
+#define COUNTER_INFO_VERSION_CURRENT 0x8
+
+/*
+ * These determine the counter_value[] layout and the meaning of starting_index
+ * and secondary_index.
+ *
+ * Unless otherwise noted, @secondary_index is unused and ignored.
+ */
+enum counter_info_requests {
+
+ /* GENERAL */
+
+ /* @starting_index: must be -1 (to refer to the current partition)
+ */
+ CIR_SYSTEM_PERFORMANCE_CAPABILITIES = 0X40,
+};
+
+struct cv_system_performance_capabilities {
+ /* If != 0, allowed to collect data from other partitions */
+ __u8 perf_collect_privileged;
+
+ /* These following are only valid if counter_info_version >= 0x3 */
+#define CV_CM_GA (1 << 7)
+#define CV_CM_EXPANDED (1 << 6)
+#define CV_CM_LAB (1 << 5)
+ /* remaining bits are reserved */
+ __u8 capability_mask;
+ __u8 reserved[0xE];
+} __packed;
+
+#endif
diff --git a/arch/powerpc/perf/power7-events-list.h b/arch/powerpc/perf/power7-events-list.h
index 687790a2c0b..64f13d9260a 100644
--- a/arch/powerpc/perf/power7-events-list.h
+++ b/arch/powerpc/perf/power7-events-list.h
@@ -546,3 +546,13 @@ EVENT(PM_MRK_DATA_FROM_RL2L3_SHR, 0x1d04c)
EVENT(PM_DTLB_MISS_16M, 0x4c05e)
EVENT(PM_LSU1_LMQ_LHR_MERGE, 0x0d09a)
EVENT(PM_IFU_FIN, 0x40066)
+EVENT(PM_1THRD_CON_RUN_INSTR, 0x30062)
+EVENT(PM_CMPLU_STALL_COUNT, 0x4000B)
+EVENT(PM_MEM0_PB_RD_CL, 0x30083)
+EVENT(PM_THRD_1_RUN_CYC, 0x10060)
+EVENT(PM_THRD_2_CONC_RUN_INSTR, 0x40062)
+EVENT(PM_THRD_2_RUN_CYC, 0x20060)
+EVENT(PM_THRD_3_CONC_RUN_INST, 0x10062)
+EVENT(PM_THRD_3_RUN_CYC, 0x30060)
+EVENT(PM_THRD_4_CONC_RUN_INST, 0x20062)
+EVENT(PM_THRD_4_RUN_CYC, 0x40060)
diff --git a/arch/powerpc/perf/power8-pmu.c b/arch/powerpc/perf/power8-pmu.c
index 2ee4a707f0d..639cd915658 100644
--- a/arch/powerpc/perf/power8-pmu.c
+++ b/arch/powerpc/perf/power8-pmu.c
@@ -10,6 +10,8 @@
* 2 of the License, or (at your option) any later version.
*/
+#define pr_fmt(fmt) "power8-pmu: " fmt
+
#include <linux/kernel.h>
#include <linux/perf_event.h>
#include <asm/firmware.h>
@@ -25,15 +27,48 @@
#define PM_BRU_FIN 0x10068
#define PM_BR_MPRED_CMPL 0x400f6
+/* All L1 D cache load references counted at finish, gated by reject */
+#define PM_LD_REF_L1 0x100ee
+/* Load Missed L1 */
+#define PM_LD_MISS_L1 0x3e054
+/* Store Missed L1 */
+#define PM_ST_MISS_L1 0x300f0
+/* L1 cache data prefetches */
+#define PM_L1_PREF 0x0d8b8
+/* Instruction fetches from L1 */
+#define PM_INST_FROM_L1 0x04080
+/* Demand iCache Miss */
+#define PM_L1_ICACHE_MISS 0x200fd
+/* Instruction Demand sectors wriittent into IL1 */
+#define PM_L1_DEMAND_WRITE 0x0408c
+/* Instruction prefetch written into IL1 */
+#define PM_IC_PREF_WRITE 0x0408e
+/* The data cache was reloaded from local core's L3 due to a demand load */
+#define PM_DATA_FROM_L3 0x4c042
+/* Demand LD - L3 Miss (not L2 hit and not L3 hit) */
+#define PM_DATA_FROM_L3MISS 0x300fe
+/* All successful D-side store dispatches for this thread */
+#define PM_L2_ST 0x17080
+/* All successful D-side store dispatches for this thread that were L2 Miss */
+#define PM_L2_ST_MISS 0x17082
+/* Total HW L3 prefetches(Load+store) */
+#define PM_L3_PREF_ALL 0x4e052
+/* Data PTEG reload */
+#define PM_DTLB_MISS 0x300fc
+/* ITLB Reloaded */
+#define PM_ITLB_MISS 0x400fc
+
/*
* Raw event encoding for POWER8:
*
* 60 56 52 48 44 40 36 32
* | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
- * | [ thresh_cmp ] [ thresh_ctl ]
- * | |
- * *- EBB (Linux) thresh start/stop OR FAB match -*
+ * | | [ ] [ thresh_cmp ] [ thresh_ctl ]
+ * | | | |
+ * | | *- IFM (Linux) thresh start/stop OR FAB match -*
+ * | *- BHRB (Linux)
+ * *- EBB (Linux)
*
* 28 24 20 16 12 8 4 0
* | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
@@ -83,9 +118,18 @@
* 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
+ *
*/
#define EVENT_EBB_MASK 1ull
+#define EVENT_EBB_SHIFT PERF_EVENT_CONFIG_EBB_SHIFT
+#define EVENT_BHRB_MASK 1ull
+#define EVENT_BHRB_SHIFT 62
+#define EVENT_WANTS_BHRB (EVENT_BHRB_MASK << EVENT_BHRB_SHIFT)
+#define EVENT_IFM_MASK 3ull
+#define EVENT_IFM_SHIFT 60
#define EVENT_THR_CMP_SHIFT 40 /* Threshold CMP value */
#define EVENT_THR_CMP_MASK 0x3ff
#define EVENT_THR_CTL_SHIFT 32 /* Threshold control value (start/stop) */
@@ -110,6 +154,12 @@
#define EVENT_IS_MARKED (EVENT_MARKED_MASK << EVENT_MARKED_SHIFT)
#define EVENT_PSEL_MASK 0xff /* PMCxSEL value */
+/* Bits defined by Linux */
+#define EVENT_LINUX_MASK \
+ ((EVENT_EBB_MASK << EVENT_EBB_SHIFT) | \
+ (EVENT_BHRB_MASK << EVENT_BHRB_SHIFT) | \
+ (EVENT_IFM_MASK << EVENT_IFM_SHIFT))
+
#define EVENT_VALID_MASK \
((EVENT_THRESH_MASK << EVENT_THRESH_SHIFT) | \
(EVENT_SAMPLE_MASK << EVENT_SAMPLE_SHIFT) | \
@@ -118,7 +168,7 @@
(EVENT_UNIT_MASK << EVENT_UNIT_SHIFT) | \
(EVENT_COMBINE_MASK << EVENT_COMBINE_SHIFT) | \
(EVENT_MARKED_MASK << EVENT_MARKED_SHIFT) | \
- (EVENT_EBB_MASK << PERF_EVENT_CONFIG_EBB_SHIFT) | \
+ EVENT_LINUX_MASK | \
EVENT_PSEL_MASK)
/* MMCRA IFM bits - POWER8 */
@@ -142,10 +192,11 @@
*
* 28 24 20 16 12 8 4 0
* | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
- * | [ ] [ sample ] [ ] [6] [5] [4] [3] [2] [1]
- * EBB -* | |
- * | | Count of events for each PMC.
- * L1 I/D qualifier -* | p1, p2, p3, p4, p5, p6.
+ * [ ] | [ ] [ sample ] [ ] [6] [5] [4] [3] [2] [1]
+ * | | | |
+ * BHRB IFM -* | | | 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
@@ -164,6 +215,9 @@
#define CNST_EBB_VAL(v) (((v) & EVENT_EBB_MASK) << 24)
#define CNST_EBB_MASK CNST_EBB_VAL(EVENT_EBB_MASK)
+#define CNST_IFM_VAL(v) (((v) & EVENT_IFM_MASK) << 25)
+#define CNST_IFM_MASK CNST_IFM_VAL(EVENT_IFM_MASK)
+
#define CNST_L1_QUAL_VAL(v) (((v) & 3) << 22)
#define CNST_L1_QUAL_MASK CNST_L1_QUAL_VAL(3)
@@ -199,6 +253,7 @@
#define MMCR1_UNIT_SHIFT(pmc) (60 - (4 * ((pmc) - 1)))
#define MMCR1_COMBINE_SHIFT(pmc) (35 - ((pmc) - 1))
#define MMCR1_PMCSEL_SHIFT(pmc) (24 - (((pmc) - 1)) * 8)
+#define MMCR1_FAB_SHIFT 36
#define MMCR1_DC_QUAL_SHIFT 47
#define MMCR1_IC_QUAL_SHIFT 46
@@ -209,6 +264,7 @@
#define MMCRA_THR_SEL_SHIFT 16
#define MMCRA_THR_CMP_SHIFT 32
#define MMCRA_SDAR_MODE_TLB (1ull << 42)
+#define MMCRA_IFM_SHIFT 30
static inline bool event_is_fab_match(u64 event)
@@ -233,20 +289,22 @@ static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long
pmc = (event >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK;
unit = (event >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK;
cache = (event >> EVENT_CACHE_SEL_SHIFT) & EVENT_CACHE_SEL_MASK;
- ebb = (event >> PERF_EVENT_CONFIG_EBB_SHIFT) & EVENT_EBB_MASK;
-
- /* Clear the EBB bit in the event, so event checks work below */
- event &= ~(EVENT_EBB_MASK << PERF_EVENT_CONFIG_EBB_SHIFT);
+ ebb = (event >> EVENT_EBB_SHIFT) & EVENT_EBB_MASK;
if (pmc) {
+ u64 base_event;
+
if (pmc > 6)
return -1;
- mask |= CNST_PMC_MASK(pmc);
- value |= CNST_PMC_VAL(pmc);
+ /* Ignore Linux defined bits when checking event below */
+ base_event = event & ~EVENT_LINUX_MASK;
- if (pmc >= 5 && event != 0x500fa && event != 0x600f4)
+ if (pmc >= 5 && base_event != 0x500fa && base_event != 0x600f4)
return -1;
+
+ mask |= CNST_PMC_MASK(pmc);
+ value |= CNST_PMC_VAL(pmc);
}
if (pmc <= 4) {
@@ -267,9 +325,10 @@ static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long
* HV writable, and there is no API for guest kernels to modify
* it. The solution is for the hypervisor to initialise the
* field to zeroes, and for us to only ever allow events that
- * have a cache selector of zero.
+ * have a cache selector of zero. The bank selector (bit 3) is
+ * irrelevant, as long as the rest of the value is 0.
*/
- if (cache)
+ if (cache & 0x7)
return -1;
} else if (event & EVENT_IS_L1) {
@@ -310,6 +369,15 @@ static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long
/* EBB events must specify the PMC */
return -1;
+ if (event & EVENT_WANTS_BHRB) {
+ if (!ebb)
+ /* Only EBB events can request BHRB */
+ return -1;
+
+ mask |= CNST_IFM_MASK;
+ value |= CNST_IFM_VAL(event >> EVENT_IFM_SHIFT);
+ }
+
/*
* All events must agree on EBB, either all request it or none.
* EBB events are pinned & exclusive, so this should never actually
@@ -388,8 +456,8 @@ static int power8_compute_mmcr(u64 event[], int n_ev,
* the threshold bits are used for the match value.
*/
if (event_is_fab_match(event[i])) {
- mmcr1 |= (event[i] >> EVENT_THR_CTL_SHIFT) &
- EVENT_THR_CTL_MASK;
+ mmcr1 |= ((event[i] >> EVENT_THR_CTL_SHIFT) &
+ EVENT_THR_CTL_MASK) << MMCR1_FAB_SHIFT;
} else {
val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK;
mmcra |= val << MMCRA_THR_CTL_SHIFT;
@@ -399,6 +467,11 @@ static int power8_compute_mmcr(u64 event[], int n_ev,
mmcra |= val << MMCRA_THR_CMP_SHIFT;
}
+ if (event[i] & EVENT_WANTS_BHRB) {
+ val = (event[i] >> EVENT_IFM_SHIFT) & EVENT_IFM_MASK;
+ mmcra |= val << MMCRA_IFM_SHIFT;
+ }
+
hwc[i] = pmc - 1;
}
@@ -556,6 +629,8 @@ static int power8_generic_events[] = {
[PERF_COUNT_HW_INSTRUCTIONS] = PM_INST_CMPL,
[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = PM_BRU_FIN,
[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 u64 power8_bhrb_filter_map(u64 branch_sample_type)
@@ -595,6 +670,116 @@ static void power8_config_bhrb(u64 pmu_bhrb_filter)
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 int power8_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_L1_PREF,
+ [ 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_L1_DEMAND_WRITE,
+ [ C(RESULT_MISS) ] = -1,
+ },
+ [ C(OP_PREFETCH) ] = {
+ [ C(RESULT_ACCESS) ] = PM_IC_PREF_WRITE,
+ [ 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_PREF_ALL,
+ [ 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_BRU_FIN,
+ [ 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
+
static struct power_pmu power8_pmu = {
.name = "POWER8",
.n_counter = 6,
@@ -607,9 +792,10 @@ static struct power_pmu power8_pmu = {
.get_constraint = power8_get_constraint,
.get_alternatives = power8_get_alternatives,
.disable_pmc = power8_disable_pmc,
- .flags = PPMU_HAS_SSLOT | PPMU_HAS_SIER | PPMU_BHRB | PPMU_EBB,
+ .flags = PPMU_HAS_SSLOT | PPMU_HAS_SIER | PPMU_ARCH_207S,
.n_generic = ARRAY_SIZE(power8_generic_events),
.generic_events = power8_generic_events,
+ .cache_events = &power8_cache_events,
.attr_groups = power8_pmu_attr_groups,
.bhrb_nr = 32,
};
@@ -629,6 +815,9 @@ static int __init init_power8_pmu(void)
/* Tell userspace that EBB is supported */
cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;
+ if (cpu_has_feature(CPU_FTR_PMAO_BUG))
+ pr_info("PMAO restore workaround active.\n");
+
return 0;
}
early_initcall(init_power8_pmu);
generated by cgit v1.2.3-70-g09d2 (git 2.46.0) at 2025-11-04 07:52:05 +0000