@@ -7,7 +7,7 @@ all_targets += poll poll-cancel ring-leak fsync io_uring_setup io_uring_register
917257daa0fe-test b19062a56726-test eeed8b54e0df-test link \
send_recvmsg a4c0b3decb33-test 500f9fbadef8-test timeout \
sq-space_left stdout cq-ready cq-peek-batch file-register \
- cq-size 8a9973408177-test a0908ae19763-test
+ cq-size 8a9973408177-test a0908ae19763-test timeout-overflow
include ../Makefile.quiet
@@ -22,7 +22,7 @@ test_srcs := poll.c poll-cancel.c ring-leak.c fsync.c io_uring_setup.c \
eeed8b54e0df-test.c link.c send_recvmsg.c a4c0b3decb33-test.c \
500f9fbadef8-test.c timeout.c sq-space_left.c stdout.c cq-ready.c\
cq-peek-batch.c file-register.c cq-size.c 8a9973408177-test.c \
- a0908ae19763-test.c
+ a0908ae19763-test.c timeout-overflow.c
test_objs := $(patsubst %.c,%.ol,$(test_srcs))
new file mode 100644
@@ -0,0 +1,184 @@
+/*
+ * Description: run timeout overflow test
+ *
+ */
+#include <errno.h>
+#include <stdio.h>
+#include <limits.h>
+#include <sys/time.h>
+
+#include "liburing.h"
+
+#define TIMEOUT_MSEC 1000
+static int not_supported;
+
+static int check_timeout_support()
+{
+ struct io_uring_sqe *sqe;
+ struct io_uring_cqe *cqe;
+ struct __kernel_timespec ts;
+ struct io_uring ring;
+ int ret;
+
+ ret = io_uring_queue_init(8, &ring, 0);
+ if (ret) {
+ printf("ring setup failed\n");
+ return 1;
+ }
+ sqe = io_uring_get_sqe(&ring);
+ ts.tv_sec = TIMEOUT_MSEC / 1000;
+ ts.tv_nsec = 0;
+ io_uring_prep_timeout(sqe, &ts, 1, 0);
+
+ ret = io_uring_submit(&ring);
+ if (ret < 0) {
+ printf("sqe submit failed: %d\n", ret);
+ goto err;
+ }
+
+ ret = io_uring_wait_cqe(&ring, &cqe);
+ if (ret < 0) {
+ printf("wait completion %d\n", ret);
+ goto err;
+ }
+
+ if (cqe->res == -EINVAL) {
+ not_supported = 1;
+ printf("Timeout not supported, ignored\n");
+ return 0;
+ }
+ io_uring_cqe_seen(&ring, cqe);
+
+ return 0;
+err:
+ return 1;
+
+}
+
+/*
+ * We first setup 4 timeout requests, which require a count value of 1, 1, 2,
+ * UINT_MAX, so the sequence is 1, 2, 4, 2. Before really timeout, this 4
+ * requests will not lead the change of cq_cached_tail, so as sq_dropped.
+ *
+ * And before this patch. The order of this four requests will be req1->req2->
+ * req4->req3. Actually, it should be req1->req2->req3->req4.
+ *
+ * Then, if there is 2 nop req. All timeout requests expect req4 will completed
+ * successful after the patch. And req1/req2 will completed successful with
+ * req3/req4 return -ETIME without this patch!
+ */
+static int test_timeout_overflow()
+{
+ struct io_uring_sqe *sqe;
+ struct io_uring_cqe *cqe;
+ struct __kernel_timespec ts;
+ struct io_uring ring;
+ int i, ret;
+
+ ret = io_uring_queue_init(16, &ring, 0);
+ if (ret) {
+ printf("ring setup failed\n");
+ return 1;
+ }
+
+ ts.tv_sec = TIMEOUT_MSEC / 1000;
+ ts.tv_nsec = 0;
+ for (i = 0; i < 4; i++) {
+ unsigned num;
+ sqe = io_uring_get_sqe(&ring);
+ switch (i) {
+ case 0:
+ case 1:
+ num = 1;
+ break;
+ case 2:
+ num = 2;
+ break;
+ case 3:
+ num = UINT_MAX;
+ break;
+ }
+ io_uring_prep_timeout(sqe, &ts, num, 0);
+ }
+
+ for (i = 0; i < 2; i++) {
+ sqe = io_uring_get_sqe(&ring);
+ io_uring_prep_nop(sqe);
+ io_uring_sqe_set_data(sqe, (void *) 1);
+ }
+ ret = io_uring_submit(&ring);
+ if (ret < 0) {
+ printf("sqe submit failed: %d\n", ret);
+ goto err;
+ }
+
+ i = 0;
+ while (i < 6) {
+ ret = io_uring_wait_cqe(&ring, &cqe);
+ if (ret < 0) {
+ printf("wait completion %d\n", ret);
+ goto err;
+ }
+
+ /*
+ * cqe1: first nop req
+ * cqe2: first timeout req, because of cqe1
+ * cqe3: second timeout req because of cqe1 + cqe2
+ * cqe4: second nop req
+ * cqe5~cqe6: the left three timeout req
+ */
+ switch (i) {
+ case 0:
+ case 3:
+ if (io_uring_cqe_get_data(cqe) != (void *) 1) {
+ printf("nop not seen as 1 or 2\n");
+ goto err;
+ }
+ break;
+ case 1:
+ case 2:
+ case 4:
+ if (cqe->res == -ETIME) {
+ printf("expected not return -ETIME for the "
+ "%d'th timeout req\n", i - 1);
+ goto err;
+ }
+ break;
+ case 5:
+ if (cqe->res != -ETIME) {
+ printf("expected return -ETIME for the "
+ "%d'th timeout req\n", i - 1);
+ goto err;
+ }
+ break;
+ }
+ io_uring_cqe_seen(&ring, cqe);
+ i++;
+ }
+
+ return 0;
+err:
+ return 1;
+}
+
+int main(int argc, char *argv[])
+{
+ int ret;
+
+ ret = check_timeout_support();
+ if (ret) {
+ printf("ring setup failed\n");
+ return 1;
+ }
+
+ if (not_supported)
+ return 0;
+
+ ret = test_timeout_overflow();
+ if (ret) {
+ printf("test_timeout_overflow failed\n");
+ return 1;
+ }
+
+ return 0;
+}
Before 5da0fb1ab34c ("io_uring: consider the overflow of sequence for timeout req"). We can meet some situation like below: 1. setup 2. prepare 4 timeout req which expected count is 1,1,2,UINT_MAX, and the sequence of this 4 requests will be 1,2,4,2, this 4 requests will not lead the change of cached_cq_tail and sq_dropped until the timeout really happened. So the tail_index in io_timeout will still be 0. 3. based on the above and before this patch, the order of timeout_list will be req1->req2->req4->req3, which the right order should be req1->req2->req3->req4. 4. setup two nop requests. And the timeout requests will return correctly with the patch. Add this testcase to cover it. Signed-off-by: yangerkun <yangerkun@huawei.com> --- test/Makefile | 4 +- test/timeout-overflow.c | 184 ++++++++++++++++++++++++++++++++++++++++ 2 files changed, 186 insertions(+), 2 deletions(-) create mode 100644 test/timeout-overflow.c