casync: when encoding, use a thread pool for compressing and storing chunks

This is a low hanging optimization fruit: when we encode a stream, then
pass the chunk compression/storing to a pool of worker threads.

This doesn't speed up encoding as much as I hoped, but still:

The firefox 63 sources (2.2G) are encoded in 1m16,902s instead of
1m24,281s on my 4 cpu system. i.e. roughly a 10% speed-up.

Author: poettering

src/affinity-count.c

@@ -0,0 +1,39 @@
+#include "affinity-count.h"
+
+#include <errno.h>
+#include <sched.h>
+#include <sys/types.h>
+
+int cpus_in_affinity_mask(void) {
+        size_t n = 16;
+        int r;
+
+        for (;;) {
+                cpu_set_t *c;
+
+                c = CPU_ALLOC(n);
+                if (!c)
+                        return -ENOMEM;
+
+                if (sched_getaffinity(0, CPU_ALLOC_SIZE(n), c) >= 0) {
+                        int k;
+
+                        k = CPU_COUNT_S(CPU_ALLOC_SIZE(n), c);
+                        CPU_FREE(c);
+
+                        if (k <= 0)
+                                return -EINVAL;
+
+                        return k;
+                }
+
+                r = -errno;
+                CPU_FREE(c);
+
+                if (r != -EINVAL)
+                        return r;
+                if (n*2 < n)
+                        return -ENOMEM;
+                n *= 2;
+        }
+}

src/affinity-count.h

@@ -0,0 +1,8 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#ifndef fooaffinitycounthfoo
+#define fooaffinitycounthfoo
+
+int cpus_in_affinity_mask(void);
+
+#endif

src/castore.c

@@ -3,9 +3,12 @@
 #include <dirent.h>
 #include <fcntl.h>
 #include <lzma.h>
+#include <pthread.h>
+#include <sys/socket.h>
 #include <sys/stat.h>
 #include <unistd.h>
 
+#include "affinity-count.h"
 #include "castore.h"
 #include "def.h"
 #include "dirent-util.h"
@@ -19,6 +22,8 @@
 /* #undef EBADMSG */
 /* #define EBADMSG __LINE__ */
 
+#define WORKER_THREADS_MAX 64U
+
 struct CaStore {
         char *root;
         bool is_cache:1;
@@ -34,6 +39,10 @@ struct CaStore {
 
         uint64_t n_requests;
         uint64_t n_request_bytes;
+
+        pthread_t worker_threads[WORKER_THREADS_MAX];
+        size_t n_worker_threads, n_worker_threads_max;
+        int worker_thread_socket[2];
 };
 
 struct CaStoreIterator {
@@ -47,28 +56,36 @@ struct CaStoreIterator {
 CaStore* ca_store_new(void) {
         CaStore *store;
 
-        store = new0(CaStore, 1);
+        store = new(CaStore, 1);
         if (!store)
                 return NULL;
 
-        store->digest_type = _CA_DIGEST_TYPE_INVALID;
-
-        store->compression = CA_CHUNK_COMPRESSED;
-        store->compression_type = CA_COMPRESSION_DEFAULT;
+        *store = (CaStore) {
+                .digest_type = _CA_DIGEST_TYPE_INVALID,
+                .compression = CA_CHUNK_COMPRESSED,
+                .compression_type = CA_COMPRESSION_DEFAULT,
+                .worker_thread_socket = { -1, -1},
+                .n_worker_threads_max = (size_t) -1,
+        };
 
         return store;
 }
 
 CaStore *ca_store_new_cache(void) {
         CaStore *s;
 
-        s = new0(CaStore, 1);
+        s = new(CaStore, 1);
         if (!s)
                 return NULL;
 
-        s->is_cache = true;
-        s->compression = CA_CHUNK_AS_IS;
-        s->compression_type = CA_COMPRESSION_DEFAULT;
+        *s = (CaStore) {
+                .is_cache = true,
+                .compression = CA_CHUNK_AS_IS,
+                .compression_type = CA_COMPRESSION_DEFAULT,
+
+                .worker_thread_socket = { -1, -1 },
+                .n_worker_threads_max = (size_t) -1,
+        };
 
         return s;
 }
@@ -77,6 +94,8 @@ CaStore* ca_store_unref(CaStore *store) {
         if (!store)
                 return NULL;
 
+        (void) ca_store_finalize(store);
+
         if (store->is_cache && store->root)
                 (void) rm_rf(store->root, REMOVE_ROOT|REMOVE_PHYSICAL);
 
@@ -240,6 +259,203 @@ int ca_store_has(CaStore *store, const CaChunkID *chunk_id) {
         return ca_chunk_file_test(AT_FDCWD, store->root, chunk_id);
 }
 
+struct queue_entry {
+        CaChunkID chunk_id;
+        CaChunkCompression effective_compression;
+        void *data;
+        size_t size;
+};
+
+static void* worker_thread(void *p) {
+        CaStore *store = p;
+        int ret = 0, r;
+
+        assert(store);
+        assert(store->worker_thread_socket[1] >= 0);
+
+        (void) pthread_setname_np(pthread_self(), "worker-thread");
+
+        for (;;) {
+                struct queue_entry e;
+                ssize_t n;
+
+                n = recv(store->worker_thread_socket[0], &e, sizeof(e), 0);
+                if (n < 0) {
+                        if (errno == EINTR)
+                                continue;
+
+                        log_debug_errno(errno, "Failed to read from thread pool socket: %m");
+                        return INT_TO_PTR(errno);
+                }
+                if (n == 0) /* Either EOF or zero-sized datagram (Linux doesn't really allow us to
+                             * distinguish that), we take both as an indication to exit the worker thread. */
+                        break;
+
+                assert(n == sizeof(e));
+
+                r = ca_chunk_file_save(
+                                AT_FDCWD, store->root,
+                                &e.chunk_id,
+                                e.effective_compression, store->compression,
+                                store->compression_type,
+                                e.data, e.size);
+                free(e.data);
+
+                if (r < 0) {
+                        log_debug_errno(r, "Failed to store chunk in store: %m");
+
+                        if (r != -EEXIST)
+                                ret = r;
+                }
+        }
+
+        return INT_TO_PTR(ret);
+}
+
+static int determine_worker_threads_max(CaStore *store) {
+        const char *e;
+        int r;
+
+        assert(store);
+
+        if (store->n_worker_threads_max != (size_t) -1)
+                return 0;
+
+        e = getenv("CASYNC_WORKER_THREADS");
+        if (e) {
+                unsigned u;
+
+                r = safe_atou(e, &u);
+                if (r < 0)
+                        log_debug_errno(r, "Failed to parse $CASYNC_WORKER_THREADS, ignoring: %s", e);
+                else if (u > WORKER_THREADS_MAX) {
+                        log_debug("$CASYNC_WORKER_THREADS out of range, clamping to %zu: %s", (size_t) WORKER_THREADS_MAX, e);
+                        store->n_worker_threads_max = WORKER_THREADS_MAX;
+                } else {
+                        store->n_worker_threads_max = u;
+                        return 0;
+                }
+        }
+
+        r = cpus_in_affinity_mask();
+        if (r < 0)
+                return log_debug_errno(r, "Failed to determine CPUs in affinity mask: %m");
+
+        store->n_worker_threads_max = MIN((size_t) r, WORKER_THREADS_MAX);
+        return 0;
+}
+
+static int start_worker_thread(CaStore *store) {
+        int r;
+
+        assert(store);
+
+        r = determine_worker_threads_max(store);
+        if (r < 0)
+                return r;
+
+        if (store->n_worker_threads >= (size_t) store->n_worker_threads_max)
+                return 0;
+
+        if (store->worker_thread_socket[0] < 0)
+                if (socketpair(AF_UNIX, SOCK_SEQPACKET|SOCK_CLOEXEC, 0, store->worker_thread_socket) < 0)
+                        return -errno;
+
+        r = pthread_create(store->worker_threads + store->n_worker_threads, NULL, worker_thread, store);
+        if (r != 0)
+                return -r;
+
+        store->n_worker_threads++;
+
+        log_debug("Started store worker thread %zu.", store->n_worker_threads);
+        return 0;
+}
+
+static int submit_to_worker_thread(
+                CaStore *store,
+                const CaChunkID *chunkid,
+                CaChunkCompression effective_compression,
+                const void *p,
+                uint64_t l) {
+
+        struct queue_entry e;
+        void *copy = NULL;
+        ssize_t n;
+        int r;
+
+        assert(store);
+
+        /* If there's no need to compress/decompress, then let's do things client side, since the operation
+         * is likely IO bound, not CPU bound */
+        if (store->compression == CA_CHUNK_AS_IS ||
+            store->compression == effective_compression)
+                return -ENOANO;
+
+        /* Before we submit the chunk for compression, let's see if it exists already. If so, let's return
+         * -EEXIST right away, so that the caller can count reused chunks. Note that this is a bit racy
+         * currently, as submitted but not yet processed chunks are not considered. */
+        r = ca_store_has(store, chunkid);
+        if (r < 0)
+                return r;
+        if (r > 0)
+                return -EEXIST;
+
+        /* Let's start a new worker thread each time we have a new job to process, until we reached all
+         * worker threads we need */
+        (void) start_worker_thread(store);
+
+        /* If there are no worker threads, do things client side */
+        if (store->n_worker_threads <= 0 ||
+            store->worker_thread_socket[1] < 0)
+                return -ENETDOWN;
+
+        copy = memdup(p, l);
+        if (!copy)
+                return -ENOMEM;
+
+        e = (struct queue_entry) {
+                .chunk_id = *chunkid,
+                .effective_compression = effective_compression,
+                .data = copy,
+                .size = l,
+        };
+
+        n = send(store->worker_thread_socket[1], &e, sizeof(e), 0);
+        if (n < 0) {
+                free(copy);
+                return -errno;
+        }
+
+        assert(n == sizeof(e));
+        return 0;
+}
+
+int ca_store_finalize(CaStore *store) {
+        int ret = 0, r;
+        size_t i;
+
+        assert(store);
+
+        /* Trigger EOF in all worker threads */
+        store->worker_thread_socket[1] = safe_close(store->worker_thread_socket[1]);
+
+        for (i = 0; i < store->n_worker_threads; i++) {
+                void *p;
+                r = pthread_join(store->worker_threads[i], &p);
+                if (r != 0)
+                        ret = -r;
+                if (p != NULL)
+                        ret = -PTR_TO_INT(p);
+        }
+
+        store->n_worker_threads = 0;
+        store->worker_thread_socket[0] = safe_close(store->worker_thread_socket[0]);
+
+        /* Propagate errors we ran into while processing store requests. This is useful for callers to
+         * determine whether the worker threads ran into any problems. */
+        return ret;
+}
+
 int ca_store_put(
                 CaStore *store,
                 const CaChunkID *chunk_id,
@@ -273,6 +489,14 @@ int ca_store_put(
                 store->mkdir_done = true;
         }
 
+        r = submit_to_worker_thread(
+                        store,
+                        chunk_id,
+                        effective_compression,
+                        data, size);
+        if (r >= 0)
+                return 0;
+
         return ca_chunk_file_save(
                         AT_FDCWD, store->root,
                         chunk_id,

src/castore.h

@@ -25,6 +25,8 @@ int ca_store_get(CaStore *store, const CaChunkID *chunk_id, CaChunkCompression d
 int ca_store_has(CaStore *store, const CaChunkID *chunk_id);
 int ca_store_put(CaStore *store, const CaChunkID *chunk_id, CaChunkCompression effective_compression, const void *data, uint64_t size);
 
+int ca_store_finalize(CaStore *store);
+
 int ca_store_get_requests(CaStore *s, uint64_t *ret);
 int ca_store_get_request_bytes(CaStore *s, uint64_t *ret);
 

src/casync.c

@@ -2058,6 +2058,14 @@ static int ca_sync_step_encode(CaSync *s) {
                 if (r < 0)
                         return r;
 
+                if (s->wstore) {
+                        /* Make sure the store ends all worker threads and pick up any pending errors from
+                         * it */
+                        r = ca_store_finalize(s->wstore);
+                        if (r < 0)
+                                return r;
+                }
+
                 r = ca_sync_install_archive(s);
                 if (r < 0)
                         return r;

src/meson.build

@@ -10,6 +10,8 @@ util_sources = files('''
 '''.split())
 
 libshared_sources = files('''
+        affinity-count.c
+        affinity-count.h
         cacache.c
         cacache.h
         cachunk.c