/* test_wal — iteration 9 Task 2: typed WAL + boot replay. * Round-trip through a replay, torn-tail drop, reopen-overwrites-tear, * and the commit-then-kill crash battery: a forked child inserts rows and * acks each COMMITTED id over a pipe; SIGKILL lands mid-stream; the parent * verifies with the offline oracle and a replay that every acked id is * present with the right contents. */ #define _POSIX_C_SOURCE 200809L #include #include #include #include #include #include #include #include "gc.h" #include "obj.h" #include "t.h" #include "table.h" #include "wal.h" /* class 0: Row { n: scalar, label: Text } */ static const uint8_t row_kinds[] = {WO_K_SCALAR, WO_K_TEXT}; static const wo_classdesc CLASSES[] = { {.name = 0, .flags = 0, .field_cnt = 2, .kinds = row_kinds}, }; static char g_dir[64]; static void test_roundtrip_replay(void) { char path[128]; snprintf(path, sizeof path, "%s/basic.wal", g_dir); wo_rt rt; T_EQ(wo_rt_init(&rt, 1 << 20, CLASSES, 1), 0); wo_db db; T_EQ(wo_db_init(&db, CLASSES, 1, 0, 1), 0); wo_wal w; T_EQ(wo_wal_open(&w, path, 1 << 16), 0); const char *msg = ""; /* three inserts and one remove, RAM first, WAL second, one commit */ uint64_t ids[3]; for (int i = 0; i < 3; i++) { wo_str *s = wo_str_new(&rt, "abcXYZ" + i, 3); /* "abc","bcX","cXY" */ uint64_t vals[2] = {(uint64_t)(i * 10), (uint64_t)(uintptr_t)s}; ids[i] = wo_row_insert(&db, 0, vals, &msg, NULL); T_CHECK(ids[i] != 0); T_EQ(wo_wal_append_insert(&w, &db, 0, ids[i]), 0); wo_str_free(&rt, s); } T_EQ(wo_row_remove(&db, 0, ids[1]), 0); T_EQ(wo_wal_append_remove(&w, 0, ids[1]), 0); T_EQ(wo_wal_commit(&w), 0); wo_wal_close(&w); wo_db_destroy(&db); /* boot: fresh engine, replay, deep-compare */ wo_db db2; T_EQ(wo_db_init(&db2, CLASSES, 1, 0, 1), 0); T_EQ(wo_wal_replay(path, &db2), 4); uint64_t out[2]; T_EQ(wo_row_read(&db2, &rt, 0, ids[0], out, &msg), 0); T_EQ(out[0], 0); wo_str *s0 = (wo_str *)(uintptr_t)out[1]; T_CHECK(s0->len == 3 && memcmp(s0->data, "abc", 3) == 0); wo_str_free(&rt, s0); T_EQ(wo_row_read(&db2, &rt, 0, ids[1], out, &msg), -1); /* removed */ T_EQ(wo_row_read(&db2, &rt, 0, ids[2], out, &msg), 0); T_EQ(out[0], 20); wo_str_free(&rt, (wo_str *)(uintptr_t)out[1]); /* next_id advanced past the replayed ids: a fresh insert never collides */ uint64_t vals[2] = {99, 0}; uint64_t fresh = wo_row_insert(&db2, 0, vals, &msg, NULL); T_CHECK(fresh > ids[2]); wo_db_destroy(&db2); /* update record: re-log, replay replaces */ { char upath[128]; snprintf(upath, sizeof upath, "%s/upd.wal", g_dir); wo_db du; T_EQ(wo_db_init(&du, CLASSES, 1, 0, 1), 0); wo_wal wu; T_EQ(wo_wal_open(&wu, upath, 0), 0); wo_str *s1 = wo_str_new(&rt, "old", 3); uint64_t uv[2] = {7, (uint64_t)(uintptr_t)s1}; uint64_t uid = wo_row_insert(&du, 0, uv, &msg, NULL); T_EQ(wo_wal_append_insert(&wu, &du, 0, uid), 0); int ek = 0; wo_str *s2 = wo_str_new(&rt, "new!", 4); T_EQ(wo_row_update_field(&du, 0, uid, 1, (uint64_t)(uintptr_t)s2, &msg, &ek), 0); T_EQ(wo_row_update_field(&du, 0, uid, 0, 8, &msg, &ek), 0); T_EQ(wo_wal_append_update(&wu, &du, 0, uid), 0); T_EQ(wo_wal_commit(&wu), 0); wo_wal_close(&wu); wo_db_destroy(&du); wo_db db4; T_EQ(wo_db_init(&db4, CLASSES, 1, 0, 1), 0); T_EQ(wo_wal_replay(upath, &db4), 2); uint64_t uo[2]; T_EQ(wo_row_read(&db4, &rt, 0, uid, uo, &msg), 0); T_EQ(uo[0], 8); wo_str *us = (wo_str *)(uintptr_t)uo[1]; T_CHECK(us->len == 4 && memcmp(us->data, "new!", 4) == 0); wo_str_free(&rt, us); wo_drop_obj(&rt, (wo_hdr *)s1); wo_drop_obj(&rt, (wo_hdr *)s2); wo_db_destroy(&db4); } /* replay of a missing file is a fresh boot, not an error */ wo_db db3; T_EQ(wo_db_init(&db3, CLASSES, 1, 0, 1), 0); T_EQ(wo_wal_replay("/nonexistent/nope.wal", &db3), 0); wo_db_destroy(&db3); wo_rt_destroy(&rt); } static void test_torn_tail(void) { char path[128]; snprintf(path, sizeof path, "%s/torn.wal", g_dir); wo_db db; T_EQ(wo_db_init(&db, CLASSES, 1, 0, 1), 0); wo_wal w; T_EQ(wo_wal_open(&w, path, 0), 0); const char *msg = ""; for (int i = 0; i < 5; i++) { uint64_t vals[2] = {(uint64_t)i, 0}; uint64_t id = wo_row_insert(&db, 0, vals, &msg, NULL); T_EQ(wo_wal_append_insert(&w, &db, 0, id), 0); T_EQ(wo_wal_commit(&w), 0); } uint64_t intact_end = w.off; /* tear: append half a record's worth of a valid-looking header + junk */ uint32_t fake_len = 40, fake_crc = 0xDEAD; uint8_t junk[20] = {7, 7, 7}; T_CHECK(pwrite(w.fd, &fake_len, 4, (off_t)intact_end) == 4); T_CHECK(pwrite(w.fd, &fake_crc, 4, (off_t)(intact_end + 4)) == 4); T_CHECK(pwrite(w.fd, junk, sizeof junk, (off_t)(intact_end + 8)) == (ssize_t)sizeof junk); wo_wal_close(&w); wo_db_destroy(&db); /* the oracle sees exactly the intact prefix */ uint64_t at = 0; T_EQ(wo_wal_check(path, &at), 5); T_EQ(at, intact_end); /* replay drops the tear whole */ wo_db db2; T_EQ(wo_db_init(&db2, CLASSES, 1, 0, 1), 0); T_EQ(wo_wal_replay(path, &db2), 5); T_EQ(db2.tables[0].count, 5); wo_db_destroy(&db2); /* reopen positions AT the tear: the next commit overwrites it */ wo_db db3; T_EQ(wo_db_init(&db3, CLASSES, 1, 0, 1), 0); T_EQ(wo_wal_replay(path, &db3), 5); wo_wal w2; T_EQ(wo_wal_open(&w2, path, 0), 0); T_EQ(w2.off, intact_end); uint64_t vals[2] = {100, 0}; uint64_t id = wo_row_insert(&db3, 0, vals, &msg, NULL); T_EQ(wo_wal_append_insert(&w2, &db3, 0, id), 0); T_EQ(wo_wal_commit(&w2), 0); wo_wal_close(&w2); T_EQ(wo_wal_check(path, NULL), 6); /* tear gone, record in its place */ wo_db_destroy(&db3); } /* ---- the crash battery -------------------------------------------------- */ /* Child: insert forever — RAM, WAL, COMMIT, and only then ack the id down * the pipe. Killed mid-stream by the parent. */ static void battery_child(const char *path, int ack_fd) { wo_rt rt; wo_db db; wo_wal w; if (wo_rt_init(&rt, 1 << 20, CLASSES, 1) != 0) _exit(9); if (wo_db_init(&db, CLASSES, 1, 0, 1) != 0) _exit(9); if (wo_wal_open(&w, path, 1 << 20) != 0) _exit(9); const char *msg = ""; for (uint64_t i = 0;; i++) { char label[32]; int n = snprintf(label, sizeof label, "row-%llu", (unsigned long long)i); wo_str *s = wo_str_new(&rt, label, (uint32_t)n); uint64_t vals[2] = {i * 3 + 1, (uint64_t)(uintptr_t)s}; uint64_t id = wo_row_insert(&db, 0, vals, &msg, NULL); wo_str_free(&rt, s); if (!id) _exit(9); if (wo_wal_append_insert(&w, &db, 0, id) != 0) _exit(9); if (wo_wal_commit(&w) != 0) _exit(9); /* durable BEFORE the ack */ ssize_t wr = write(ack_fd, &id, 8); if (wr != 8) _exit(0); /* parent went away */ } } static void test_crash_battery(void) { for (int round = 0; round < 5; round++) { char path[128]; snprintf(path, sizeof path, "%s/crash-%d.wal", g_dir, round); int pipefd[2]; T_EQ(pipe(pipefd), 0); pid_t pid = fork(); T_CHECK(pid >= 0); if (pid == 0) { close(pipefd[0]); battery_child(path, pipefd[1]); _exit(0); } close(pipefd[1]); /* collect acks for a few ms, then kill mid-stream — no sync with the child's commit loop, which is the point */ struct timespec ts = {0, (20 + round * 13) * 1000000L}; while (nanosleep(&ts, &ts) != 0) {} kill(pid, SIGKILL); int status; waitpid(pid, &status, 0); /* drain every ack that made it into the pipe */ uint64_t acked[65536]; size_t n_acked = 0; for (;;) { uint64_t id; ssize_t n = read(pipefd[0], &id, 8); if (n != 8) break; if (n_acked < 65536) acked[n_acked++] = id; } close(pipefd[0]); T_CHECK(n_acked > 0); /* the child got at least one commit out */ /* offline oracle: the file's intact prefix covers every ack */ int64_t intact = wo_wal_check(path, NULL); T_CHECK(intact >= (int64_t)n_acked); /* replay and verify: every acked id present, contents exact */ wo_rt rt; T_EQ(wo_rt_init(&rt, 1 << 22, CLASSES, 1), 0); wo_db db; T_EQ(wo_db_init(&db, CLASSES, 1, 0, 1), 0); int64_t applied = wo_wal_replay(path, &db); T_CHECK(applied >= (int64_t)n_acked); const char *msg = ""; int bad = 0; for (size_t i = 0; i < n_acked; i++) { uint64_t out[2]; if (wo_row_read(&db, &rt, 0, acked[i], out, &msg) != 0) { bad++; continue; } /* id = i+1 (shard 0 of 1), field 0 = i*3+1, label = "row-i" */ char want[32]; int wl = snprintf(want, sizeof want, "row-%llu", (unsigned long long)(acked[i] - 1)); wo_str *s = (wo_str *)(uintptr_t)out[1]; if (out[0] != (acked[i] - 1) * 3 + 1 || s->len != (uint32_t)wl || memcmp(s->data, want, (size_t)wl) != 0) bad++; wo_str_free(&rt, s); } T_EQ(bad, 0); /* zero acked-but-missing, zero acked-but-wrong */ wo_db_destroy(&db); wo_rt_destroy(&rt); } } /* iteration 19: a Float column and a Bytes column survive a WAL round trip * BIT-EXACT. Bit-exact is the whole assertion — the durability path must not * render a float as decimal anywhere, or NaN, the infinities and -0.0 would * each come back as something else. Bytes goes through the same length- * prefixed blob a Text does and must come back as a Bytes, not a Text. */ static const uint8_t fb_kinds[] = {WO_K_FLOAT, WO_K_BYTES}; static const wo_classdesc FB_CLASSES[] = { {.name = 0, .flags = 0, .field_cnt = 2, .kinds = fb_kinds}, }; static void test_float_bytes_replay(void) { char path[128]; snprintf(path, sizeof path, "%s/floatbytes.wal", g_dir); wo_rt rt; T_EQ(wo_rt_init(&rt, 1 << 20, FB_CLASSES, 1), 0); wo_db db; T_EQ(wo_db_init(&db, FB_CLASSES, 1, 0, 1), 0); wo_wal w; T_EQ(wo_wal_open(&w, path, 1 << 16), 0); const char *msg = ""; /* the values a decimal round trip would destroy, plus a NUL-bearing blob * that a NUL-terminated string path would truncate */ const double vals_f[] = {9.99, 0.0 / 0.0, 1.0 / 0.0, -1.0 / 0.0, -0.0, 1e308}; const char blob[] = {'a', '\0', 'b'}; enum { N = sizeof vals_f / sizeof vals_f[0] }; uint64_t ids[N]; for (int i = 0; i < N; i++) { wo_str *b = wo_bytes_new(&rt, blob, sizeof blob); T_CHECK(b != NULL); uint64_t vals[2] = {wo_bits(vals_f[i]), (uint64_t)(uintptr_t)b}; ids[i] = wo_row_insert(&db, 0, vals, &msg, NULL); T_CHECK(ids[i] != 0); T_EQ(wo_wal_append_insert(&w, &db, 0, ids[i]), 0); wo_str_free(&rt, b); } T_EQ(wo_wal_commit(&w), 0); wo_wal_close(&w); wo_db_destroy(&db); wo_db db2; T_EQ(wo_db_init(&db2, FB_CLASSES, 1, 0, 1), 0); T_EQ(wo_wal_replay(path, &db2), N); for (int i = 0; i < N; i++) { uint64_t out[2]; T_EQ(wo_row_read(&db2, &rt, 0, ids[i], out, &msg), 0); /* BITS, not value: NaN != NaN and -0.0 == 0.0, so a value comparison * would pass while silently having lost the payload or the sign */ T_EQ(out[0], wo_bits(vals_f[i])); wo_str *b = (wo_str *)(uintptr_t)out[1]; T_CHECK(b != NULL); T_EQ(b->h.class_id, WO_CLS_BYTES); /* a Bytes column yields a Bytes */ T_CHECK(b->len == sizeof blob && memcmp(b->data, blob, sizeof blob) == 0); wo_str_free(&rt, b); } wo_db_destroy(&db2); wo_rt_destroy(&rt); } /* databasev2 2: offset capture. wo_wal_next_offset must name exactly where a * record lands, so a resident:keys table can read it back by that offset * later. A wrong offset is the worst possible bug here: it reads a * NEIGHBOURING record, which passes its own CRC and returns the wrong row * silently. So this asserts the recovered id per record, not just that a * record parses. * * Covers the two awkward cases the design called out: records straddling a * buffer growth (stage() doubles from 4096, so 400 rows with Text payloads * cross it repeatedly), and a batch spanning several commits. */ static void test_offset_capture(void) { char path[128]; snprintf(path, sizeof path, "%s/offsets.wal", g_dir); wo_rt rt; T_EQ(wo_rt_init(&rt, 1 << 20, CLASSES, 1), 0); wo_db db; T_EQ(wo_db_init(&db, CLASSES, 1, 0, 1), 0); wo_wal w; T_EQ(wo_wal_open(&w, path, 1 << 16), 0); const char *msg = ""; enum { N = 400 }; uint64_t ids[N], offs[N]; /* commit in uneven batches so offsets are exercised both mid-buffer and * immediately after a flush reset len to 0 */ for (int i = 0; i < N; i++) { char lbl[32]; int ln = snprintf(lbl, sizeof lbl, "label-%d-padding", i); wo_str *s = wo_str_new(&rt, lbl, (uint32_t)ln); uint64_t vals[2] = {(uint64_t)i, (uint64_t)(uintptr_t)s}; ids[i] = wo_row_insert(&db, 0, vals, &msg, NULL); T_CHECK(ids[i] != 0); /* BEFORE the append: this is the contract */ offs[i] = wo_wal_next_offset(&w); T_EQ(wo_wal_append_insert(&w, &db, 0, ids[i]), 0); wo_str_free(&rt, s); if (i % 7 == 6) T_EQ(wo_wal_commit(&w), 0); } T_EQ(wo_wal_commit(&w), 0); /* offsets must be strictly increasing and inside the written region */ for (int i = 1; i < N; i++) T_CHECK(offs[i] > offs[i - 1]); /* read each record back BY ITS REPORTED OFFSET and check the id matches: * payload is [kind u8][class u32][id u64], after the 8-byte len+crc head */ int checked = 0; for (int i = 0; i < N; i++) { uint8_t head[8], body[13]; T_EQ((int)pread(w.fd, head, 8, (off_t)offs[i]), 8); T_EQ((int)pread(w.fd, body, 13, (off_t)(offs[i] + 8)), 13); T_EQ(body[0], WO_WAL_INSERT); uint32_t cid; uint64_t rid; memcpy(&cid, body + 1, 4); memcpy(&rid, body + 5, 8); T_EQ(cid, 0u); T_EQ(rid, ids[i]); checked++; } T_EQ(checked, N); wo_wal_close(&w); wo_db_destroy(&db); wo_rt_destroy(&rt); } /* databasev2 2: an offset reported for a record whose commit FAILED must * never be trusted. Simulated by closing the fd under the wal so pwrite * fails: the offset accessor must not have advanced past the durable tail, * so a later successful commit reuses the same place. */ static void test_offset_after_failed_commit(void) { char path[128]; snprintf(path, sizeof path, "%s/offfail.wal", g_dir); wo_rt rt; T_EQ(wo_rt_init(&rt, 1 << 20, CLASSES, 1), 0); wo_db db; T_EQ(wo_db_init(&db, CLASSES, 1, 0, 1), 0); wo_wal w; T_EQ(wo_wal_open(&w, path, 1 << 16), 0); const char *msg = ""; uint64_t vals[2] = {7u, 0u}; uint64_t id1 = wo_row_insert(&db, 0, vals, &msg, NULL); T_CHECK(id1 != 0); uint64_t at1 = wo_wal_next_offset(&w); T_EQ(wo_wal_append_insert(&w, &db, 0, id1), 0); /* break the fd, so the commit cannot succeed */ int saved = dup(w.fd); T_CHECK(saved >= 0); close(w.fd); w.fd = -1; T_CHECK(wo_wal_commit(&w) != 0); /* The DURABLE TAIL is what must not move. `next_offset` legitimately * points PAST the still-staged record (off unchanged, len still holding * it) — asserting otherwise was this test's own first mistake. The * invariant that matters: off is untouched, so the record still lands at * the offset already reported for it. */ T_EQ(w.off, at1); /* restore and commit for real: the record lands exactly where promised */ w.fd = saved; T_EQ(wo_wal_commit(&w), 0); uint8_t body[13]; T_EQ((int)pread(w.fd, body, 13, (off_t)(at1 + 8)), 13); uint64_t rid; memcpy(&rid, body + 5, 8); T_EQ(rid, id1); wo_wal_close(&w); wo_db_destroy(&db); wo_rt_destroy(&rt); } /* databasev2 2 (5b): read rows back BY OFFSET and deep-compare. * * The point is not that a record parses — test_offset_capture already showed * the offsets are right. The point is that the VALUES come back intact, * including a nil Text, and that the two refusal paths refuse instead of * handing back something plausible. */ static void test_read_row_at(void) { char path[128]; snprintf(path, sizeof path, "%s/readat.wal", g_dir); wo_rt rt; T_EQ(wo_rt_init(&rt, 1 << 20, CLASSES, 1), 0); wo_db db; T_EQ(wo_db_init(&db, CLASSES, 1, 0, 1), 0); wo_wal w; T_EQ(wo_wal_open(&w, path, 1 << 16), 0); const char *msg = ""; enum { N = 24 }; uint64_t ids[N], offs[N]; const char *labels[N]; for (int i = 0; i < N; i++) { /* every third row has a NIL Text, so the nil path is covered */ wo_str *s = NULL; if (i % 3 != 0) { char lbl[24]; int ln = snprintf(lbl, sizeof lbl, "row-%d", i); s = wo_str_new(&rt, lbl, (uint32_t)ln); T_CHECK(s != NULL); } uint64_t vals[2] = {(uint64_t)(i * 3 + 1), (uint64_t)(uintptr_t)s}; ids[i] = wo_row_insert(&db, 0, vals, &msg, NULL); T_CHECK(ids[i] != 0); labels[i] = (i % 3 != 0) ? "set" : "nil"; offs[i] = wo_wal_next_offset(&w); T_EQ(wo_wal_append_insert(&w, &db, 0, ids[i]), 0); if (s) wo_str_free(&rt, s); } T_EQ(wo_wal_commit(&w), 0); /* read each row back by offset and compare field by field */ for (int i = 0; i < N; i++) { uint64_t got[2] = {0, 0}; uint32_t cid = 0xFFFFFFFFu; uint64_t id = 0; T_EQ(wo_wal_read_row_at(&w, &db, &rt, offs[i], &cid, &id, got, &msg), 0); T_EQ(cid, 0u); T_EQ(id, ids[i]); T_EQ(got[0], (uint64_t)(i * 3 + 1)); if (labels[i][0] == 'n') { T_EQ(got[1], 0u); /* nil Text stays nil through the round trip */ } else { wo_str *back = (wo_str *)(uintptr_t)got[1]; T_CHECK(back != NULL); char want[24]; int wl = snprintf(want, sizeof want, "row-%d", i); T_EQ((int)back->len, wl); T_EQ(memcmp(back->data, want, (size_t)wl), 0); wo_str_free(&rt, back); /* out-gate: the VM value is ours to free */ } } /* refusal 1: a tombstone is refused, not decoded as a live row */ T_EQ(wo_row_remove(&db, 0, ids[0]), 0); uint64_t tomb_off = wo_wal_next_offset(&w); T_EQ(wo_wal_append_remove(&w, 0, ids[0]), 0); T_EQ(wo_wal_commit(&w), 0); { uint64_t got[2] = {0, 0}; T_EQ(wo_wal_read_row_at(&w, &db, &rt, tomb_off, NULL, NULL, got, &msg), -1); } /* refusal 2: a wrong offset (mid-record) refuses rather than returning a * neighbouring row -- the silent-wrong-row failure this guards */ { uint64_t got[2] = {0, 0}; T_EQ(wo_wal_read_row_at(&w, &db, &rt, offs[5] + 3u, NULL, NULL, got, &msg), -1); } /* refusal 3: past the end of the intact prefix */ { uint64_t got[2] = {0, 0}; T_EQ(wo_wal_read_row_at(&w, &db, &rt, w.off + 4096u, NULL, NULL, got, &msg), -1); } wo_wal_close(&w); wo_db_destroy(&db); wo_rt_destroy(&rt); } int main(void) { snprintf(g_dir, sizeof g_dir, "/tmp/wo-wal-test-XXXXXX"); if (!mkdtemp(g_dir)) return 1; test_roundtrip_replay(); test_torn_tail(); test_float_bytes_replay(); test_offset_capture(); test_offset_after_failed_commit(); test_read_row_at(); test_crash_battery(); /* leave the dir for a failed run's forensics only */ if (!t_fail) { char cmd[128]; snprintf(cmd, sizeof cmd, "rm -rf %s", g_dir); if (system(cmd) != 0) fprintf(stderr, "cleanup failed, kept %s\n", g_dir); } else { fprintf(stderr, "kept %s\n", g_dir); } return t_report("test_wal"); }