On Thu, 15 Feb 2024 16:08:16 -0800
Samuel Holland <samuel.holland@xxxxxxxxxx> wrote:
> From: Eric Lin <eric.lin@xxxxxxxxxx>
>
> Add a driver for the PMU found in the SiFive Extensible Cache
> controller. This PMU provides a configurable number of counters and a
> variety of events. Events are grouped into sets. Each counter can count
> events from only one set at a time; however, it can count any number of
> events within that set simultaneously. The PMU hardware does not provide
> an overflow interrupt.
>
> The counter inhibit register is used to atomically start/stop/read a
> group of counters so their values can be usefully compared.
>
> Some events can be filtered further by client ID (e.g. CPU or external
> DMA master). That functionality is not supported by this driver.
>
> This driver further assumes that a single Extensible Cache instance is
> shared by all CPUs in the system.
>
> Example usage:
>
> $ perf stat -e sifive_ecache_pmu/inner_rd_request/,
> sifive_ecache_pmu/inner_wr_request/,
> sifive_ecache_pmu/inner_rd_request_hit/,
> sifive_ecache_pmu/inner_wr_request_hit/ ls
>
> Performance counter stats for 'system wide':
>
> 148001 sifive_ecache_pmu/inner_rd_request/
> 121064 sifive_ecache_pmu/inner_wr_request/
> 113124 sifive_ecache_pmu/inner_rd_request_hit/
> 120860 sifive_ecache_pmu/inner_wr_request_hit/
>
> 0.010643962 seconds time elapsed
>
> Example combining the read/write events together within each counter:
>
> $ perf stat -e sifive_ecache_pmu/event=0x601/,
> sifive_ecache_pmu/event=0xc001/ ls
>
> Performance counter stats for 'system wide':
>
> 262619 sifive_ecache_pmu/event=0x601/
> 224533 sifive_ecache_pmu/event=0xc001/
>
> 0.009794808 seconds time elapsed
>
> Signed-off-by: Eric Lin <eric.lin@xxxxxxxxxx>
> Co-developed-by: Samuel Holland <samuel.holland@xxxxxxxxxx>
> Signed-off-by: Samuel Holland <samuel.holland@xxxxxxxxxx>
Hi Samuel,
Some comments inline. Note this is a driver by review so not very
thorough!
Jonathan
> +
> +static u64 read_counter(const struct sifive_ecache_pmu *ecache_pmu, const struct hw_perf_event *hwc)
> +{
> + u64 value = 0;
> +
> + for (int i = 0; i < ecache_pmu->n_slices; i++) {
> + void __iomem *base = ecache_pmu->slice[i].base;
> +
> + value += readq(base + hwc->event_base);
Feels like this summing should be a userspace problem.
Knowing about slice imbalance is often useful (depending on the
micro architecture obviously!)
> + }
> +
> + return value;
> +}
> +
> +static int sifive_ecache_pmu_probe(struct platform_device *pdev)
> +{
> + struct device *dev = &pdev->dev;
> + struct device_node *ecache_node = dev_of_node(dev);
> + struct sifive_ecache_pmu *ecache_pmu;
> + struct device_node *slice_node;
> + u32 slice_counters;
> + int n_slices, ret;
> + int i = 0;
> +
> + n_slices = of_get_available_child_count(ecache_node);
fwnode_get_available_child_count(dev_fwnode(&pdev->dev));
Not sure why there isn't yet a device version of this (IIRC anyway).
> + if (!n_slices)
> + return -ENODEV;
> +
> + ecache_pmu = devm_kzalloc(dev, struct_size(ecache_pmu, slice, n_slices), GFP_KERNEL);
> + if (!ecache_pmu)
> + return -ENOMEM;
> +
> + platform_set_drvdata(pdev, ecache_pmu);
> +
> + ecache_pmu->pmu = (struct pmu) {
> + .parent = dev,
> + .attr_groups = sifive_ecache_pmu_attr_grps,
> + .capabilities = PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
> + .task_ctx_nr = perf_invalid_context,
> + .event_init = sifive_ecache_pmu_event_init,
> + .add = sifive_ecache_pmu_add,
> + .del = sifive_ecache_pmu_del,
> + .start = sifive_ecache_pmu_start,
> + .stop = sifive_ecache_pmu_stop,
> + .read = sifive_ecache_pmu_read,
> + .start_txn = sifive_ecache_pmu_start_txn,
> + .commit_txn = sifive_ecache_pmu_commit_txn,
> + .cancel_txn = sifive_ecache_pmu_cancel_txn,
> + };
> + ecache_pmu->cpu = nr_cpu_ids;
> + ecache_pmu->n_counters = ECACHE_PMU_MAX_COUNTERS;
> + ecache_pmu->n_slices = n_slices;
> +
> + for_each_available_child_of_node(ecache_node, slice_node) {
device_for_each_child_node() (generic handlers only provide the available version btw
which is non obvious from naming)
> + struct sifive_ecache_pmu_slice *slice = &ecache_pmu->slice[i++];
> +
> + slice->base = devm_of_iomap(dev, slice_node, 0, NULL);
> + if (IS_ERR(slice->base))
Leaked slice_node
FWIW https://lore.kernel.org/linux-iio/20240217164249.921878-1-jic23@xxxxxxxxxx/
adds device_for_each_child_node_scoped() which deals with this stuff using
cleanup.h magic.
> + return PTR_ERR(slice->base);
> +
> + /* Get number of counters from slice node */
> + ret = of_property_read_u32(slice_node, "sifive,perfmon-counters", &slice_counters);
Not sure on what perf maintainers want, but I'd go with
device_property_read etc as in the previous driver.
> + if (ret)
leaked slice_node
> + return dev_err_probe(dev, ret,
> + "Slice %pOF missing sifive,perfmon-counters property\n",
> + slice_node);
> +
> + ecache_pmu->n_counters = min_t(u32, slice_counters, ecache_pmu->n_counters);
> + }
> +
> + sifive_ecache_pmu_hw_init(ecache_pmu);
> +
> + ret = cpuhp_state_add_instance(CPUHP_AP_PERF_RISCV_SIFIVE_ECACHE_ONLINE, &ecache_pmu->node);
> + if (ret)
> + return dev_err_probe(dev, ret, "Failed to add CPU hotplug instance\n");
> +
> + ret = perf_pmu_register(&ecache_pmu->pmu, "sifive_ecache_pmu", -1);
> + if (ret) {
> + dev_err_probe(dev, ret, "Failed to register PMU\n");
> + goto err_remove_instance;
Comments from other review apply here as well so if you agree apply them in both drivers.
> + }
> +
> + return 0;
> +
> +err_remove_instance:
> + cpuhp_state_remove_instance(CPUHP_AP_PERF_RISCV_SIFIVE_ECACHE_ONLINE, &ecache_pmu->node);
> +
> + return ret;
> +}
