[CRIU] [PATCH 16/17] unix: Add support of ghost sockets
Andrei Vagin
avagin at virtuozzo.com
Tue Apr 3 22:18:33 MSK 2018
On Sun, Apr 01, 2018 at 11:07:42PM +0300, Cyrill Gorcunov wrote:
> Unix sockets may be connected via deleted socket name,
> moreover the name may be reused (ie same sun_addr but
> different inodes).
>
> To be able to handle them we do a few tricks:
>
> - when collecting sockets we figure out if "deleted"
> mark is present on the socket and if such we rename
> it into a new unique name
>
> - then we wait until all users are connected and
> remove the socket from the FS
>
> Signed-off-by: Cyrill Gorcunov <gorcunov at gmail.com>
> ---
> criu/cr-restore.c | 4 +
> criu/include/sockets.h | 1 +
> criu/sk-unix.c | 290 ++++++++++++++++++++++++++++++++++++++++++-------
> 3 files changed, 255 insertions(+), 40 deletions(-)
>
> diff --git a/criu/cr-restore.c b/criu/cr-restore.c
> index db913b2dae2e..ff1e4dcc34df 100644
> --- a/criu/cr-restore.c
> +++ b/criu/cr-restore.c
> @@ -388,6 +388,10 @@ static int root_prepare_shared(void)
> if (ret)
> goto err;
>
> + ret = unix_resolve_ghost_addr();
> + if (ret)
> + goto err;
> +
> show_saved_files();
> err:
> return ret;
> diff --git a/criu/include/sockets.h b/criu/include/sockets.h
> index db330428850c..23f5b11c1b58 100644
> --- a/criu/include/sockets.h
> +++ b/criu/include/sockets.h
> @@ -60,6 +60,7 @@ extern int netlink_receive_one(struct nlmsghdr *hdr, struct ns_id *ns, void *arg
>
> extern int unix_sk_id_add(unsigned int ino);
> extern int unix_sk_ids_parse(char *optarg);
> +extern int unix_resolve_ghost_addr(void);
>
> extern int do_dump_opt(int sk, int level, int name, void *val, int len);
> #define dump_opt(s, l, n, f) do_dump_opt(s, l, n, f, sizeof(*f))
> diff --git a/criu/sk-unix.c b/criu/sk-unix.c
> index 4feaa7722a50..c0861d17c8b1 100644
> --- a/criu/sk-unix.c
> +++ b/criu/sk-unix.c
> @@ -9,6 +9,7 @@
> #include <sys/un.h>
> #include <stdlib.h>
> #include <dlfcn.h>
> +#include <libgen.h>
>
> #include "libnetlink.h"
> #include "cr_options.h"
> @@ -31,6 +32,7 @@
> #include "fdstore.h"
> #include "fdinfo.h"
> #include "kerndat.h"
> +#include "rst-malloc.h"
>
> #include "protobuf.h"
> #include "images/sk-unix.pb-c.h"
> @@ -90,10 +92,14 @@ struct unix_sk_desc {
> };
>
> static LIST_HEAD(unix_sockets);
> +static LIST_HEAD(unix_ghost_addr);
>
> static int unix_resolve_name(int lfd, uint32_t id, struct unix_sk_desc *d,
> UnixSkEntry *ue, const struct fd_parms *p);
>
> +struct unix_sk_info;
> +static void unlink_sk(struct unix_sk_info *ui);
> +
> struct unix_sk_listen_icon {
> unsigned int peer_ino;
> struct unix_sk_desc *sk_desc;
> @@ -892,6 +898,12 @@ struct unix_sk_info {
> struct list_head connected; /* List of sockets, connected to me */
> struct list_head node; /* To link in peer's connected list */
> struct list_head scm_fles;
> + struct list_head ghost_node;
> + struct list_head ghost_wait_head;
> + struct list_head ghost_waiter;
> + atomic_t name_ref;
> + atomic_t name_rdy;
I think you try to reinvent a synchronisation scheme. We already know
when a peer socket is bound.
> + struct unix_sk_info *ghost_master;
We already have ->peer, do we really need ghost_master? What is
difference between them?
>
> /*
> * For DGRAM sockets with queues, we should only restore the queue
> @@ -916,6 +928,7 @@ struct scm_fle {
>
> #define USK_PAIR_MASTER 0x1
> #define USK_PAIR_SLAVE 0x2
> +#define USK_GHOST_NAME 0x4
>
> static struct unix_sk_info *find_unix_sk_by_ino(int ino)
> {
> @@ -1077,6 +1090,17 @@ static int wake_connected_sockets(struct unix_sk_info *ui)
> return 0;
> }
>
> +static void wake_ghost_waiters(struct unix_sk_info *ui)
> +{
> + struct fdinfo_list_entry *fle;
> + struct unix_sk_info *tmp;
> +
> + list_for_each_entry(tmp, &ui->ghost_wait_head, ghost_waiter) {
> + fle = file_master(&tmp->d);
> + set_fds_event(fle->pid);
> + }
> +}
> +
> static bool peer_is_not_prepared(struct unix_sk_info *peer)
> {
> if (peer->ue->state != TCP_LISTEN)
> @@ -1239,6 +1263,38 @@ static int prep_unix_sk_cwd(struct unix_sk_info *ui, int *prev_cwd_fd,
> return -1;
> }
>
> +static void drop_ghost_master(struct unix_sk_info *ui)
> +{
> + struct unix_sk_info *gm = ui->ghost_master;
> + if (gm) {
> + if (atomic_dec_and_test(&gm->name_ref)) {
> + pr_debug("ghost: Unlinking ghost master %s\n", gm->ue->name.data);
> + unlink_sk(gm);
> + }
> + }
> +}
> +
> +static void drop_deleted(struct unix_sk_info *ui)
> +{
> + if (ui->ue->has_deleted && ui->ue->deleted) {
> + if (atomic_dec_and_test(&ui->name_ref)) {
> + pr_debug("ghost: Unlinking regular %s\n", ui->ue->name.data);
> + unlink_sk(ui);
> + }
> + }
> +}
> +
> +static bool wait_ghost_master(struct unix_sk_info *ui)
> +{
> + /*
> + * If we're to bound to deleted wait socket,
> + * wait until master create it.
> + */
> + if (ui->ghost_master)
> + return !atomic_read(&ui->ghost_master->name_rdy);
> + return false;
> +}
> +
> static int post_open_standalone(struct file_desc *d, int fd)
> {
> struct unix_sk_info *ui;
> @@ -1286,6 +1342,7 @@ static int post_open_standalone(struct file_desc *d, int fd)
> ui->is_connected = true;
>
> revert_unix_sk_cwd(peer, &cwd_fd, &root_fd, &ns_fd);
> + drop_ghost_master(ui);
>
> restore_queue:
> if (peer->queuer == ui &&
> @@ -1298,46 +1355,60 @@ static int post_open_standalone(struct file_desc *d, int fd)
> return restore_sk_common(fd, ui);
> }
>
> -static int bind_deleted_unix_sk(int sk, struct unix_sk_info *ui,
> - struct sockaddr_un *addr)
> +/*
> + * When path where socket lives is deleted, we need to reconstruct
> + * it back up but allow caller to remove it after.
> + */
> +static int bind_on_deleted(int sk, struct unix_sk_info *ui)
> {
> - char temp[PATH_MAX];
> + char path[PATH_MAX], *pos;
> + struct sockaddr_un addr;
> int ret;
>
> - pr_info("found duplicate unix socket bound at %s\n", addr->sun_path);
> -
> - ret = snprintf(temp, sizeof(temp),
> - "%s-%s-%d", addr->sun_path, "criu-temp", getpid());
> - /* this shouldn't happen, since sun_addr is only 108 chars long */
> - if (ret < 0 || ret >= sizeof(temp)) {
> - pr_err("snprintf of %s failed?\n", addr->sun_path);
> - return -1;;
> + if (ui->ue->name.len >= sizeof(path)) {
> + pr_err("Too long name for socket\n");
> + return -ENOSPC;
> }
>
> - ret = rename(addr->sun_path, temp);
> - if (ret < 0) {
> - pr_perror("couldn't move socket for binding");
> - return -1;
> + memcpy(path, ui->name, ui->ue->name.len);
> + path[ui->ue->name.len] = '\0';
> +
> + for (pos = strrchr(path, '/'); pos;
> + pos = strrchr(path, '/')) {
> + *pos = '\0';
> +
> + ret = access(path, R_OK | W_OK | X_OK);
> + if (ret == 0)
> + break;
> +
> + if (errno != ENOENT) {
> + ret = -errno;
> + pr_perror("Can't access %s\n", path);
> + return ret;
> + }
> }
>
> - ret = bind(sk, (struct sockaddr *)addr,
> - sizeof(addr->sun_family) + ui->ue->name.len);
> - if (ret < 0) {
> - pr_perror("Can't bind socket after move");
> - return -1;;
> + memcpy(path, ui->name, ui->ue->name.len);
> + path[ui->ue->name.len] = '\0';
> +
> + pos = dirname(path);
> + ret = mkdirpat(AT_FDCWD, pos, 0755);
> + if (ret) {
> + pr_err("Can't create %s\n", pos);
> + return ret;
> }
>
> - ret = rename(temp, addr->sun_path);
> + memset(&addr, 0, sizeof(addr));
> + addr.sun_family = AF_UNIX;
> + memcpy(&addr.sun_path, ui->name, ui->ue->name.len);
> +
> + ret = bind(sk, (struct sockaddr *)&addr,
> + sizeof(addr.sun_family) + ui->ue->name.len);
> if (ret < 0) {
> - pr_perror("couldn't move socket back");
> - return -1;
> + pr_perror("Can't bind on socket %s", (char *)ui->ue->name.data);
> + return ret;
> }
>
> - /* we've handled the deleted-ness of this
> - * socket and we don't want to delete it later
> - * since it's not /this/ socket.
> - */
> - ui->ue->deleted = false;
> return 0;
> }
>
> @@ -1347,7 +1418,7 @@ static int bind_unix_sk(int sk, struct unix_sk_info *ui)
> int cwd_fd = -1, root_fd = -1, ns_fd = -1;
> int ret, exit_code = -1;
>
> - if (ui->ue->name.len == 0)
> + if (ui->ue->name.len == 0 || atomic_read(&ui->name_rdy))
> return 0;
>
> if ((ui->ue->type == SOCK_STREAM) && (ui->ue->state == TCP_ESTABLISHED)) {
> @@ -1371,16 +1442,13 @@ static int bind_unix_sk(int sk, struct unix_sk_info *ui)
> ret = bind(sk, (struct sockaddr *)&addr,
> sizeof(addr.sun_family) + ui->ue->name.len);
> if (ret < 0) {
> - if (ui->ue->has_deleted && ui->ue->deleted && errno == EADDRINUSE) {
> - if (bind_deleted_unix_sk(sk, ui, &addr))
> - goto done;
> - } else {
> - pr_perror("Can't bind socket");
> + if (ui->ue->has_deleted && ui->ue->deleted)
> + ret = bind_on_deleted(sk, ui);
> + if (ret)
> goto done;
> - }
> }
>
> - if (*ui->name && ui->ue->file_perms) {
> + if (ui->ue->file_perms) {
> FilePermsEntry *perms = ui->ue->file_perms;
> char fname[PATH_MAX];
>
> @@ -1403,19 +1471,20 @@ static int bind_unix_sk(int sk, struct unix_sk_info *ui)
> }
> }
>
> - if (ui->ue->deleted && unlink((char *)ui->ue->name.data) < 0) {
> - pr_perror("failed to unlink %s", ui->ue->name.data);
> - goto done;
> - }
> + atomic_inc(&ui->name_rdy);
> + pr_debug("name_rdy %#x\n", ui->ue->ino);
>
> if (ui->ue->state != TCP_LISTEN) {
> ui->bound = 1;
> wake_connected_sockets(ui);
> }
>
> + wake_ghost_waiters(ui);
> exit_code = 0;
> done:
> revert_unix_sk_cwd(ui, &cwd_fd, &root_fd, &ns_fd);
> + if (exit_code == 0)
> + drop_deleted(ui);
> return exit_code;
> }
>
> @@ -1501,6 +1570,9 @@ static int open_unixsk_pair_master(struct unix_sk_info *ui, int *new_fd)
> if (bind_unix_sk(sk[1], peer))
> return -1;
>
> + drop_ghost_master(ui);
> + drop_ghost_master(peer);
> +
> *new_fd = sk[0];
> return 1;
> }
> @@ -1556,6 +1628,10 @@ static int open_unixsk_standalone(struct unix_sk_info *ui, int *new_fd)
>
> fle = file_master(&ui->d);
> pr_info_opening("standalone", ui, fle);
> +
> + if (wait_ghost_master(ui))
> + return 1;
> +
> if (fle->stage == FLE_OPEN)
> return post_open_standalone(&ui->d, fle->fe->fd);
>
> @@ -1814,11 +1890,15 @@ static int init_unix_sk_info(struct unix_sk_info *ui, UnixSkEntry *ue)
> ui->flags = 0;
> ui->peer = NULL;
> ui->queuer = NULL;
> + ui->ghost_master = NULL;
> ui->bound = 0;
> ui->listen = 0;
> ui->is_connected = 0;
> ui->peer_queue_restored = 0;
>
> + atomic_set(&ui->name_ref, 1);
> + atomic_set(&ui->name_rdy, 0);
> +
> memzero(&ui->peer_resolve, sizeof(ui->peer_resolve));
> memzero(&ui->d, sizeof(ui->d));
>
> @@ -1826,6 +1906,127 @@ static int init_unix_sk_info(struct unix_sk_info *ui, UnixSkEntry *ue)
> INIT_LIST_HEAD(&ui->connected);
> INIT_LIST_HEAD(&ui->node);
> INIT_LIST_HEAD(&ui->scm_fles);
> + INIT_LIST_HEAD(&ui->ghost_node);
> + INIT_LIST_HEAD(&ui->ghost_wait_head);
> + INIT_LIST_HEAD(&ui->ghost_waiter);
> +
> + return 0;
> +}
> +
> +#define GHOST_NAME_FMT "~criu-%u"
> +#define GHOST_NAME_FMT_PREFIX 6 /* num of chars before counter */
> +
> +static int ghost_new_name(char *name, size_t namelen,
> + char **name_new, size_t *namelen_new)
I think we don't need to generate a new unique name for deleted sockets.
We can use the old scheme, but before deleting a file, we need to open
it with O_PATH and save a file descriptor in fdstore, then we can use it
to connect peer sockets.
> +{
> + char sname[64], *pos, *oldname = name;
> + static unsigned int cnt = 0;
> + size_t k;
> +
> + pr_debug("\tghost: handling name %s namelen %zu\n", name, namelen);
> +
> + for (pos = &name[namelen - 1]; pos > name; pos--) {
> + if (*pos == GHOST_NAME_FMT[0])
> + break;
> + }
> +
> + if (strncmp(pos, GHOST_NAME_FMT, GHOST_NAME_FMT_PREFIX) == 0) {
> + unsigned int __cnt;
> + char *__name;
> +
> + if (sscanf(pos, GHOST_NAME_FMT, &__cnt) == 1) {
> + pr_debug("\tghost: cnt %d detected\n", __cnt);
> + cnt = __cnt + 1;
> + }
> +
> + namelen = (pos - name);
> + __name = alloca(namelen + 1);
> + memcpy(__name, name, namelen);
> + __name[namelen++] = '\0';
> + name = __name;
> + pr_debug("\tghost: Name stipped to %s (namelen %zu)\n",
> + name, namelen);
> + }
> +
> + memzero(sname, sizeof(sname));
> + k = snprintf(sname, sizeof(sname), GHOST_NAME_FMT, cnt++);
> + *namelen_new = namelen + k;
> + if (*namelen_new > UNIX_PATH_MAX) {
> + pr_err("\tghost: New name for socket is too long\n");
> + return -1;
> + }
> +
> + *name_new = shmalloc(*namelen_new);
> + if (!*name_new) {
> + pr_err("\tghost: Can't allocate new name for socket\n");
> + return -ENOMEM;
> + }
> +
> + k = snprintf(*name_new, *namelen_new, "%s%s", name, sname);
> + if (k != (*namelen_new - 1)) {
> + pr_err("\tghost: Name generation failed (%s %d %d)\n",
> + *name_new, (int)k, (int)*namelen_new);
> + return -1;
> + }
> +
> + pr_debug("\tghost: name transition %s -> %s\n", oldname, *name_new);
> + return 0;
> +}
> +
> +int unix_resolve_ghost_addr(void)
> +{
> + struct unix_sk_info *ui, *t;
> +
> + pr_debug("ghost: Resolving addresses\n");
> +
> + /*
> + * Walk over ghost unix entries and find one
> + * which gonna be a master and won't unlink
> + * the name until all peers are connected to
> + * this designation.
> + */
> +
> + list_for_each_entry(ui, &unix_ghost_addr, ghost_node) {
> + size_t newnamelen;
> + char *newname;
> +
> + pr_debug("ghost: ino %#x peer %#x address %s\n",
> + ui->ue->ino, ui->peer ? ui->peer->ue->ino : 0,
> + ui->name);
> +
> + unlink_sk(ui);
> +
> + if (ghost_new_name(ui->name, ui->ue->name.len,
> + &newname, &newnamelen))
> + return -1;
> +
> + ui->name = newname;
> + ui->ue->name.len = newnamelen;
> + ui->ue->name.data = (void *)newname;
> + ui->flags |= USK_GHOST_NAME;
> +
> + unlink_sk(ui);
> +
> + /*
> + * Figure out who is connected to this peer,
> + * so the name will be removed from FS only
> + * when last one is connected.
> + */
> + list_for_each_entry(t, &unix_sockets, list) {
> + if (t->flags & USK_GHOST_NAME)
> + continue;
> + if (ui == t || t->peer != ui)
> + continue;
> +
> + pr_debug("\t\tghost: connected to us %#x -> %#x\n",
> + t->ue->ino, ui->ue->ino);
> +
> + t->flags |= USK_GHOST_NAME;
> + t->ghost_master = ui;
> + atomic_inc(&ui->name_ref);
> + list_add(&t->ghost_waiter, &ui->ghost_wait_head);
> + }
> + }
>
> return 0;
> }
> @@ -1873,6 +2074,15 @@ static int collect_one_unixsk(void *o, ProtobufCMessage *base, struct cr_img *i)
> add_post_prepare_cb(&ui->peer_resolve);
> }
>
> + if (ui->ue->deleted) {
> + if (!ui->name || !ui->ue->name.len || !ui->name[0]) {
> + pr_err("No name present, ino %#x\n", ui->ue->ino);
> + return -1;
> + }
> +
> + list_add_tail(&ui->ghost_node, &unix_ghost_addr);
> + }
> +
> list_add_tail(&ui->list, &unix_sockets);
> return file_desc_add(&ui->d, ui->ue->id, &unix_desc_ops);
> }
> --
> 2.14.3
>
> _______________________________________________
> CRIU mailing list
> CRIU at openvz.org
> https://lists.openvz.org/mailman/listinfo/criu
More information about the CRIU
mailing list