Count instructions where timing cannot tell 3% apart (110)

016's second stage passed its criterion: six interleaved pairs against
ee21aec gave TPC-C (RAM, W=1) -0.6% at the median and the bank's
queries 0.0% to +2.2%, no run interfered. The ranks query was slower
all six times, and chasing it showed that on this Pi timing cannot
separate 1-3%: two servers asked in turns disagree with themselves
(ee21aec to e02bfc9 +1.6%, e02bfc9 to 481642f -1.2%, ee21aec to 481642f
+2.8%), and a query whose instructions grew 0.6% ran 3.5% faster.

Instructions do not move so: pcount.c counts the server's user-mode
instructions with perf_event_open (no perf here), and psample.c samples
where they go. ranks had grown 0.4% in instructions (the rest was where
code and data landed), per_branch 3.6% and tables 20 instructions a row,
which 07047c3 halved. The six pairs of 07047c3 were interfered (another
session's build, load 4-12) and are kept apart, unused.

1 files changed, 94 insertions, 0 deletions

diff --git a/research/experiments/025-i64-cost/psample.c b/research/experiments/025-i64-cost/psample.c
new file mode 100644
index 0000000..308e8ff
--- /dev/null
+++ b/research/experiments/025-i64-cost/psample.c
@@ -0,0 +1,94 @@
+/* psample PID PERIOD: sample the user-mode instruction pointer of every
+ * thread of PID once per PERIOD instructions, from "ready" (on stderr) until
+ * stdin closes; prints one hex address per sample, then "lost N". Like
+ * pcount, perf_event_open alone (the threads PID has when it starts). */
+#define _GNU_SOURCE
+#include <dirent.h>
+#include <linux/perf_event.h>
+#include <poll.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
+#include <sys/syscall.h>
+#include <unistd.h>
+
+enum { PAGES = 64, MAX_THREADS = 256 };
+static size_t page;
+static uint64_t lost;
+
+typedef struct { int fd; struct perf_event_mmap_page *meta; unsigned char *data; } Ring;
+
+static void drain(Ring *r, FILE *out) {
+    uint64_t head = __atomic_load_n(&r->meta->data_head, __ATOMIC_ACQUIRE), tail = r->meta->data_tail;
+    size_t size = PAGES * page;
+    while (tail < head) {
+        struct perf_event_header h;
+        unsigned char record[256];
+        for (size_t i = 0; i < sizeof(h); ++i) ((unsigned char *)&h)[i] = r->data[(tail + i) % size];
+        if (h.size > sizeof(record) || h.size < sizeof(h)) break;
+        for (size_t i = 0; i < h.size; ++i) record[i] = r->data[(tail + i) % size];
+        if (h.type == PERF_RECORD_SAMPLE) {
+            uint64_t ip;
+            memcpy(&ip, record + sizeof(h), sizeof(ip));
+            fprintf(out, "%llx\n", (unsigned long long)ip);
+        } else if (h.type == PERF_RECORD_LOST) {
+            uint64_t n;
+            memcpy(&n, record + sizeof(h) + 8, sizeof(n));
+            lost += n;
+        }
+        tail += h.size;
+    }
+    __atomic_store_n(&r->meta->data_tail, tail, __ATOMIC_RELEASE);
+}
+
+int main(int argc, char **argv) {
+    if (argc != 3) return 2;
+    page = (size_t)sysconf(_SC_PAGESIZE);
+    char path[64];
+    snprintf(path, sizeof(path), "/proc/%s/task", argv[1]);
+    DIR *dir = opendir(path);
+    if (!dir) { perror("opendir"); return 1; }
+    Ring rings[MAX_THREADS];
+    int n = 0;
+    struct dirent *e;
+    while ((e = readdir(dir)) && n < MAX_THREADS) {
+        if (e->d_name[0] == '.') continue;
+        struct perf_event_attr attr;
+        memset(&attr, 0, sizeof(attr));
+        attr.size = sizeof(attr);
+        attr.type = PERF_TYPE_HARDWARE;
+        attr.config = PERF_COUNT_HW_INSTRUCTIONS;
+        attr.sample_period = strtoull(argv[2], NULL, 10);
+        attr.sample_type = PERF_SAMPLE_IP;
+        attr.disabled = 1;
+        attr.exclude_kernel = 1;
+        attr.exclude_hv = 1;
+        attr.wakeup_events = 1;
+        int fd = (int)syscall(SYS_perf_event_open, &attr, atoi(e->d_name), -1, -1, 0);
+        if (fd < 0) { perror("perf_event_open"); return 1; }
+        void *map = mmap(NULL, (PAGES + 1) * page, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+        if (map == MAP_FAILED) { perror("mmap"); return 1; }
+        rings[n++] = (Ring){fd, map, (unsigned char *)map + page};
+    }
+    closedir(dir);
+    FILE *out = stdout;
+    for (int i = 0; i < n; ++i) ioctl(rings[i].fd, PERF_EVENT_IOC_ENABLE, 0);
+    fprintf(stderr, "ready\n");
+    struct pollfd fds[MAX_THREADS + 1];
+    for (;;) {
+        fds[0] = (struct pollfd){.fd = 0, .events = POLLIN};
+        for (int i = 0; i < n; ++i) fds[i + 1] = (struct pollfd){.fd = rings[i].fd, .events = POLLIN};
+        poll(fds, (nfds_t)n + 1, 100);
+        for (int i = 0; i < n; ++i) drain(&rings[i], out);
+        if (fds[0].revents) {
+            char buffer[64];
+            if (read(0, buffer, sizeof(buffer)) <= 0) break;
+        }
+    }
+    for (int i = 0; i < n; ++i) ioctl(rings[i].fd, PERF_EVENT_IOC_DISABLE, 0), drain(&rings[i], out);
+    fprintf(out, "lost %llu\n", (unsigned long long)lost);
+    return 0;
+}