Fix spinning if EOF from client
PATCH 1/1 Fix spinning if EOF from client
I observed that Tabled sometimes starts spinning on CPU. It happens when it manages to receive a EOF from client (reads zero bytes from the socket).
The fix is to treat EOF as an error and simply dispose of the cli.
However, we need to know when there are no bytes to be read from a socket that is still open. To that end we modify the API of cli_read:
now it only returns zero in case of EOF, and passes -EAGAIN up.
In order to prevent busy spinning for slow clients, callers of cli_read return false to the dispatch loop in case of -EAGAIN.
Note that this patch assumes that zero bytes is always EOF, and that we receive -EAGAIN if there is nothing to read in the first place.
--- server/server.c | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-)
server/server.c +13 -5
@@ -671,6 +671,12 @@ size_t cli_wqueued(struct client *cli)return cli->write_cnt; } +/* +* Return: +* 0: progress was NOT made (EOF) +* >0: some data was gotten +* <0: an error happened (equals to system error * -1; includes -EAGAIN) +*/ static int cli_read (struct client *cli) {ssize_t rc;
@@ -682,8 +688,6 @@ do_read:if (rc <0 ) {if (errno ==EINTR ) goto do_read; -if (errno ==EAGAIN ) -return 0 ;return -errno; }
@@ -699,7 +703,7 @@ do_read:if (cli->req_used ==CLI_REQ_BUF_SZ )return -ENOSPC ; -return 0 ; +return rc !=0 ; }bool cli_cb_free (struct client *cli,void *cb_data,bool done)
@@ -1186,9 +1190,11 @@ static bool cli_evt_parse_hdr(struct client *cli, unsigned int events)static bool cli_evt_read_hdr (struct client *cli,unsigned int events) {int rc =cli_read (cli); -if (rc <0 ) { +if (rc <=0 ) {if (rc == -ENOSPC )return cli_err (cli, InvalidArgument); +if (rc == -EAGAIN ) +return false; cli->state = evt_dispose; }else
@@ -1268,9 +1274,11 @@ err_out:static bool cli_evt_read_req (struct client *cli,unsigned int events) {int rc =cli_read (cli); -if (rc <0 ) { +if (rc <=0 ) {if (rc == -ENOSPC )return cli_err (cli, InvalidArgument); +if (rc == -EAGAIN ) +return false; cli->state = evt_dispose; }else
Jeff Garzik
AppliedQuoted 9 lines · Pete Zaitcev
I observed that Tabled sometimes starts spinning on CPU. It happens when it manages to receive a EOF from client (reads zero bytes from the socket).
The fix is to treat EOF as an error and simply dispose of the cli.
However, we need to know when there are no bytes to be read from a socket that is still open. To that end we modify the API of
cli_read:
now it only returns zero in case of EOF, and passes-EAGAINup.
Quoted 7 lines
In order to prevent busy spinning for slow clients, callers of
cli_readreturn false to the dispatch loop in case of-EAGAIN.Note that this patch assumes that zero bytes is always EOF, and that we receive
-EAGAINif there is nothing to read in the first place.Signed-off-by Pete Zaitcev<[email protected]>
applied