On 12/18/2018 08:03 AM, Myungho Jung wrote: > On Mon, Dec 17, 2018 at 03:58:58PM +0100, Ursula Braun wrote: >> > > Hi Ursula, > > Thank you for your suggestion. I have a question on your comment. > >> >> On 12/17/2018 06:21 AM, Myungho Jung wrote: >>> clcsock can be released while kernel_accept() references it in TCP >>> listen worker. Also, clcsock needs to wake up before released if TCP >>> fallback is used and the clcsock is blocked by accept. Add a lock to >>> safely release clcsock and call kernel_sock_shutdown() to wake up >>> clcsock from accept in smc_release(). >> >> Thanks for your effort to solve this problem. I have some minor >> improvement proposals: >> >>> >>> Reported-by: syzbot+0bf2e01269f1274b4b03@xxxxxxxxxxxxxxxxxxxxxxxxx >>> Reported-by: syzbot+e3132895630f957306bc@xxxxxxxxxxxxxxxxxxxxxxxxx >>> Signed-off-by: Myungho Jung <mhjungk@xxxxxxxxx> >>> --- >>> net/smc/af_smc.c | 14 ++++++++++++-- >>> net/smc/smc.h | 2 ++ >>> 2 files changed, 14 insertions(+), 2 deletions(-) >>> >>> diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c >>> index 5fbaf1901571..5d06fb1bbccf 100644 >>> --- a/net/smc/af_smc.c >>> +++ b/net/smc/af_smc.c >>> @@ -147,8 +147,14 @@ static int smc_release(struct socket *sock) >>> sk->sk_shutdown |= SHUTDOWN_MASK; >>> } >>> if (smc->clcsock) { >>> + if (smc->use_fallback && sk->sk_state == SMC_LISTEN) { >>> + /* wake up clcsock accept */ >>> + rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR); >>> + } >> >> This part is not needed, since an SMC socket in state SMC_LISTEN is never >> a use_fallback socket. > > In smc_sendmsg(), set use_fallback to true if SMC socket is SMC_INIT > state and the message has MSG_FASTOPEN flag. After this, smc_listen() > would trigger smc_tcp_listen_work(). Is this not an expected scenario? > Then, what is the reason for not skipping smc_sendmsg() in SMC_INIT > state? > You are right, I have not had the FASTOPEN case in mind, sorry. If we want to allow fallback in case of FASTOPEN, we need the kernel_sock_shutdown() here for proper cleanup. Nice! >> >>> + mutex_lock(&smc->clcsock_release_lock); >>> sock_release(smc->clcsock); >>> smc->clcsock = NULL; >>> + mutex_unlock(&smc->clcsock_release_lock); >>> } >>> if (smc->use_fallback) { >>> if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT) >>> @@ -205,6 +211,7 @@ static struct sock *smc_sock_alloc(struct net *net, struct socket *sock, >>> spin_lock_init(&smc->conn.send_lock); >>> sk->sk_prot->hash(sk); >>> sk_refcnt_debug_inc(sk); >>> + mutex_init(&smc->clcsock_release_lock); >>> >>> return sk; >>> } >>> @@ -821,7 +828,7 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc) >>> struct socket *new_clcsock = NULL; >>> struct sock *lsk = &lsmc->sk; >>> struct sock *new_sk; >>> - int rc; >>> + int rc = 0; >> >> Without clcsock the good path should not be executed. Thus I suggest >> to initialize with something negative like -EINVAL. >> >>> >>> release_sock(lsk); >>> new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol); >>> @@ -834,7 +841,10 @@ static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc) >>> } >>> *new_smc = smc_sk(new_sk); >>> >>> - rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0); >>> + mutex_lock(&lsmc->clcsock_release_lock); >>> + if (lsmc->clcsock) >>> + rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0); >>> + mutex_unlock(&lsmc->clcsock_release_lock); >>> lock_sock(lsk); >>> if (rc < 0) >>> lsk->sk_err = -rc; >>> diff --git a/net/smc/smc.h b/net/smc/smc.h >>> index 08786ace6010..9a2795cf5d30 100644 >>> --- a/net/smc/smc.h >>> +++ b/net/smc/smc.h >>> @@ -219,6 +219,8 @@ struct smc_sock { /* smc sock container */ >>> * started, waiting for unsent >>> * data to be sent >>> */ >>> + struct mutex clcsock_release_lock; >>> + /* protects clcsock */ >> >> I suggest to be more precise: "protects clcsock of a listen socket" >> >>> }; >>> >>> static inline struct smc_sock *smc_sk(const struct sock *sk) >>> >> >