powerpc/perf: Add support for the hv 24x7 interface
authorCody P Schafer <cody@linux.vnet.ibm.com>
Fri, 14 Mar 2014 05:00:42 +0000 (16:00 +1100)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Sun, 23 Mar 2014 22:48:32 +0000 (09:48 +1100)
This provides a basic interface between hv_24x7 and perf. Similar to
the one provided for gpci, it lacks transaction support and does not
list any events.

Example usage via perf tool:

perf stat -e 'hv_24x7/domain=2,offset=8,starting_index=0,lpar=0xffffffff/' -r 0 -C 0 -x ' ' sleep 0.1

Signed-off-by: Cody P Schafer <cody@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/perf/hv-24x7.c [new file with mode: 0644]

diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c
new file mode 100644 (file)
index 0000000..297c910
--- /dev/null
@@ -0,0 +1,510 @@
+/*
+ * 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(char page[static 4096],
+                                            u32 version, u32 index)
+{
+       WARN_ON(!IS_ALIGNED((unsigned long)page, 4096));
+       return plpar_hcall_norets(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;
+       uint32_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 = be32_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=%d, 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)be32_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_info("not a virtualized system, not enabling\n");
+               return -ENODEV;
+       }
+
+       hret = hv_perf_caps_get(&caps);
+       if (hret) {
+               pr_info("could not obtain capabilities, error 0x%80lx, not enabling\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);