Defer taskrun changes
cover Defer taskrun changes
This series adds support to liburing for the IORING_SETUP_DEFER_TASKRUN flag.
The second half of the series is a bit mixed and includes some documentation fixes, overflow cleanups and test cleanups. I sent these a couple of months ago and forgot about it, but now it does depend on the new API so it needs to be ordered.
I can send it separately if you like.
Patches:
1 copies the definition from the kernel include file 2 introduces new APIs required for this feature 3/4/5 add tests for IORING_SETUP_DEFER_TASKRUN
6/7/8 clean and update existing documentation to match upstream 9 exposes the overflow state to the application
10 uses this and tests overflow functionality
11 gives an explicit warning if there is a short read in file-verify
Dylan Yudaken (11):
Copy defer task run definition from kernel
add io_uring_submit_and_get_events and io_uring_get_events add a t_probe_defer_taskrun helper function for tests update existing tests for defer taskrun
add a defer-taskrun test
fix documentation shortenings
update io_uring_enter.2 docs for IORING_FEAT_NODROP add docs for overflow lost errors
expose CQ ring overflow state
overflow: add tests
file-verify test: log if short read
man/io_uring_enter.2 | 24 +++- man/io_uring_setup.2 | 11 +- man/io_uring_wait_cqe.3 | 2 +- man/io_uring_wait_cqe_nr.3 | 2 +- man/io_uring_wait_cqe_timeout.3 | 2 +- man/io_uring_wait_cqes.3 | 2 +- src/include/liburing.h | 11 ++ src/include/liburing/io_uring.h | 7 + src/queue.c | 28 ++-- test/Makefile | 1 + test/cq-overflow.c | 248 +++++++++++++++++++++++++++++++- test/defer-taskrun.c | 217 ++++++++++++++++++++++++++++ test/eventfd-disable.c | 32 ++++- test/file-verify.c | 4 + test/helpers.c | 12 ++ test/helpers.h | 2 + test/iopoll.c | 16 ++- test/multicqes_drain.c | 49 ++++++- test/poll-mshot-overflow.c | 39 ++++- test/recv-multishot.c | 32 +++-- test/rsrc_tags.c | 10 +- 21 files changed, 697 insertions(+), 54 deletions(-) create mode 100644 test/defer-taskrun.c
base-commit: 15cc446cb8b0caa5c939423d31182eba66141539
Signature · 1 line
2.30.2
PATCH 1/11 Copy defer task run definition from kernel
Copy the flag from upstream
--- src/include/liburing/io_uring.h | 7 +++++++ 1 file changed, 7 insertions(+)
src/include/liburing/io_uring.h +7 -0
@@ -156,6 +156,13 @@ enum {*/ #define IORING_SETUP_SINGLE_ISSUER (1U << 12) +/* +* Defer running task work to get events. +* Rather than running bits of task work whenever the task transitions +* try to do it just before it is needed. +*/ +#define IORING_SETUP_DEFER_TASKRUN (1U << 13) +enum io_uring_op {IORING_OP_NOP ,IORING_OP_READV ,
Signature · 1 line
2.30.2
PATCH 2/11 add io_uring_submit_and_get_events and io_uring_get_events
In reply to Dylan YudakenWith deferred task running, we would like to be able to combine submit with get events (regardless of if there are CQE's available), or if there is nothing to submit then simply do an enter with IORING_ENTER_GETEVENTS set, in order to process any available work.
Expose these APIs
--- src/include/liburing.h | 2 ++ src/queue.c | 28 +++++++++++++++++++--------- 2 files changed, 21 insertions(+), 9 deletions(-)
src/include/liburing.h +2 -0
@@ -192,6 +192,8 @@ int io_uring_register_file_alloc_range(struct io_uring *ring,int io_uring_register_notifications (struct io_uring *ring,unsigned nr,struct io_uring_notification_slot *slots);int io_uring_unregister_notifications (struct io_uring *ring); +int io_uring_get_events (struct io_uring *ring); +int io_uring_submit_and_get_events (struct io_uring *ring);/* * Helper for the peek/wait single cqe functions. Exported because of that,
src/queue.c +19 -9
@@ -124,6 +124,15 @@ int __io_uring_get_cqe(struct io_uring *ring, struct io_uring_cqe **cqe_ptr,return _io_uring_get_cqe (ring, cqe_ptr, &data); } +int io_uring_get_events (struct io_uring *ring) +{ +int flags =IORING_ENTER_GETEVENTS ; + +if (ring->int_flags &INT_FLAG_REG_RING ) + flags |=IORING_ENTER_REGISTERED_RING ; +return __sys_io_uring_enter (ring->enter_ring_fd,0 ,0 , flags,NULL ); +} +/* * Fill in an array of IO completions up to count, if any are available. * Returns the amount of IO completions filled.
@@ -158,11 +167,7 @@ again: goto done;if (cq_ring_needs_flush (ring)) { -int flags =IORING_ENTER_GETEVENTS ; - -if (ring->int_flags &INT_FLAG_REG_RING ) - flags |=IORING_ENTER_REGISTERED_RING ; -__sys_io_uring_enter (ring->enter_ring_fd,0 ,0 , flags,NULL ); +io_uring_get_events (ring); overflow_checked = true; goto again; }
@@ -348,14 +353,14 @@ int io_uring_wait_cqe_timeout(struct io_uring *ring,* Returns number of sqes submitted */ static int __io_uring_submit (struct io_uring *ring,unsigned submitted, -unsigned wait_nr) +unsigned wait_nr,bool getevents) {unsigned flags;int ret; flags =0 ; -if (sq_ring_needs_enter (ring, &flags) || wait_nr) { -if (wait_nr || (ring->flags &IORING_SETUP_IOPOLL )) +if (getevents ||sq_ring_needs_enter (ring, &flags) || wait_nr) { +if (getevents || wait_nr || (ring->flags &IORING_SETUP_IOPOLL )) flags |=IORING_ENTER_GETEVENTS ;if (ring->int_flags &INT_FLAG_REG_RING ) flags |=IORING_ENTER_REGISTERED_RING ;
@@ -370,7 +375,7 @@ static int __io_uring_submit(struct io_uring *ring, unsigned submitted,static int __io_uring_submit_and_wait (struct io_uring *ring,unsigned wait_nr) { -return __io_uring_submit (ring,__io_uring_flush_sq (ring), wait_nr); +return __io_uring_submit (ring,__io_uring_flush_sq (ring), wait_nr, false); }/*
@@ -393,6 +398,11 @@ int io_uring_submit_and_wait(struct io_uring *ring, unsigned wait_nr)return __io_uring_submit_and_wait (ring, wait_nr); } +int io_uring_submit_and_get_events (struct io_uring *ring) +{ +return __io_uring_submit (ring,__io_uring_flush_sq (ring),0 , true); +} +#ifdef LIBURING_INTERNAL struct io_uring_sqe *io_uring_get_sqe (struct io_uring *ring) {
Signature · 1 line
2.30.2
PATCH 3/11 add a t_probe_defer_taskrun helper function for tests
In reply to Dylan YudakenHelp tests to determine if they can use IORING_SETUP_DEFER_TASKRUN
--- test/helpers.c | 12 ++++++++++++ test/helpers.h | 2 ++ 2 files changed, 14 insertions(+)
test/helpers.c +12 -0
@@ -235,3 +235,15 @@ errno_cleanup:close (fd[1 ]);return ret; } + +bool t_probe_defer_taskrun (void ) +{ +struct io_uring ring; +int ret; + + ret =io_uring_queue_init (1 , &ring,IORING_SETUP_DEFER_TASKRUN ); +if (ret <0 ) +return false; +io_uring_queue_exit (&ring); +return true; +}
test/helpers.h +2 -0
@@ -75,6 +75,8 @@ enum t_setup_ret t_register_buffers(struct io_uring *ring,const struct iovec *iovecs, unsigned nr_iovecs); +bool t_probe_defer_taskrun (void ); +#define ARRAY_SIZE (x) (sizeof(x) / sizeof((x)[0]))#ifdef __cplusplus
Signature · 1 line
2.30.2
PATCH 4/11 update existing tests for defer taskrun
In reply to Dylan YudakenAdd defer_taskrun to a few choice tests that can expose some bad behaviour.
This requires adding some io_uring_get_events calls to make sure deferred tasks are run
--- test/eventfd-disable.c | 32 ++++++++++++++++++++++--- test/iopoll.c | 16 +++++++++---- test/multicqes_drain.c | 49 +++++++++++++++++++++++++++++++++----- test/poll-mshot-overflow.c | 39 ++++++++++++++++++++++++++---- test/recv-multishot.c | 32 ++++++++++++++++--------- test/rsrc_tags.c | 10 ++++++-- 6 files changed, 147 insertions(+), 31 deletions(-)
test/eventfd-disable.c +29 -3
@@ -15,7 +15,7 @@#include "liburing.h" #include "helpers.h" -int main (int argc,char *argv[]) +static int test (bool defer) {struct io_uring_params p = {};struct io_uring_sqe *sqe;
@@ -28,8 +28,8 @@ int main(int argc, char *argv[]) };int ret, evfd, i; -if (argc >1 ) -return T_EXIT_SKIP ; +if (defer) + p.flags |=IORING_SETUP_DEFER_TASKRUN ; ret =io_uring_queue_init_params (64 , &ring, &p);if (ret) {
@@ -148,5 +148,31 @@ int main(int argc, char *argv[])io_uring_cqe_seen (&ring, cqe); } +io_uring_queue_exit (&ring); +close (evfd);return T_EXIT_PASS ; } + +int main (int argc,char *argv[]) +{ +int ret; + +if (argc >1 ) +return T_EXIT_SKIP ; + + ret =test (false); +if (ret !=T_EXIT_PASS ) { +fprintf (stderr,"%s: test(false) failed\n" , argv[0 ]); +return ret; + } + +if (t_probe_defer_taskrun ()) { + ret =test (true); +if (ret !=T_EXIT_PASS ) { +fprintf (stderr,"%s: test(true) failed\n" , argv[0 ]); +return ret; + } + } + +return ret; +}
test/iopoll.c +11 -5
@@ -274,7 +274,7 @@ ok: }static int test_io (const char *file,int write,int sqthread,int fixed, -int buf_select) +int buf_select,int defer) {struct io_uring ring;int ret, ring_flags =IORING_SETUP_IOPOLL ;
@@ -282,6 +282,9 @@ static int test_io(const char *file, int write, int sqthread, int fixed,if (no_iopoll)return 0 ; +if (defer) + ring_flags |=IORING_SETUP_DEFER_TASKRUN ; + ret =t_create_ring (64 , &ring, ring_flags);if (ret ==T_SETUP_SKIP )return 0 ;
@@ -337,19 +340,22 @@ int main(int argc, char *argv[]) vecs =t_create_buffers (BUFFERS ,BS ); - nr =16 ; + nr =32 ;if (no_buf_select) nr =8 ; +else if (!t_probe_defer_taskrun ()) + nr =16 ;for (i =0 ; i < nr; i++) {int write = (i &1 ) !=0 ;int sqthread = (i &2 ) !=0 ;int fixed = (i &4 ) !=0 ;int buf_select = (i &8 ) !=0 ; +int defer = (i &16 ) !=0 ; - ret =test_io (fname, write, sqthread, fixed, buf_select); + ret =test_io (fname, write, sqthread, fixed, buf_select, defer);if (ret) { -fprintf (stderr,"test_io failed %d/%d/%d/%d\n" , - write, sqthread, fixed, buf_select); +fprintf (stderr,"test_io failed %d/%d/%d/%d/%d\n" , + write, sqthread, fixed, buf_select, defer); goto err; }if (no_iopoll)
test/multicqes_drain.c +43 -6
92 lines not on this page · raw message
test/poll-mshot-overflow.c +35 -4
@@ -42,7 +42,7 @@ int check_final_cqe(struct io_uring *ring)return T_EXIT_PASS ; } -int main (int argc,char *argv[]) +static int test (bool defer_taskrun) {struct io_uring_cqe *cqe;struct io_uring_sqe *sqe;
@@ -50,9 +50,6 @@ int main(int argc, char *argv[])int pipe1[2 ];int ret, i; -if (argc >1 ) -return 0 ; -if (pipe (pipe1) !=0 ) {perror ("pipe" );return T_EXIT_FAIL ;
@@ -66,6 +63,9 @@ int main(int argc, char *argv[]) .cq_entries = 2 }; +if (defer_taskrun) + params.flags |=IORING_SETUP_DEFER_TASKRUN ; + ret =io_uring_queue_init_params (2 , &ring, ¶ms);if (ret)return T_EXIT_SKIP ;
@@ -113,6 +113,9 @@ int main(int argc, char *argv[])io_uring_cqe_seen (&ring, cqe); } +/* make sure everything is processed */ +io_uring_get_events (&ring); +/* now remove the poll */ sqe =io_uring_get_sqe (&ring);io_uring_prep_poll_remove (sqe,1 );
@@ -126,5 +129,33 @@ int main(int argc, char *argv[]) ret =check_final_cqe (&ring); +close (pipe1[0 ]); +close (pipe1[1 ]); +io_uring_queue_exit (&ring); + +return ret; +} + +int main (int argc,char *argv[]) +{ +int ret; + +if (argc >1 ) +return T_EXIT_SKIP ; + + ret =test (false); +if (ret !=T_EXIT_PASS ) { +fprintf (stderr,"%s: test(false) failed\n" , argv[0 ]); +return ret; + } + +if (t_probe_defer_taskrun ()) { + ret =test (true); +if (ret !=T_EXIT_PASS ) { +fprintf (stderr,"%s: test(true) failed\n" , argv[0 ]); +return ret; + } + } +return ret; }
test/recv-multishot.c +21 -11
@@ -29,6 +29,7 @@ struct args {bool wait_each;bool recvmsg;enum early_error_t early_error; +bool defer; };static int check_sockaddr (struct sockaddr_in *in)
@@ -76,19 +77,21 @@ static int test(struct args *args) .tv_sec = 1, };struct msghdr msg; +struct io_uring_params params = { }; +int n_sqe =32 ;memset (recv_buffs,0 ,sizeof (recv_buffs)); -if (args->early_error ==ERROR_EARLY_OVERFLOW ) { -struct io_uring_params params = { - .flags =IORING_SETUP_CQSIZE , - .cq_entries =N_CQE_OVERFLOW - }; +if (args->defer) + params.flags |=IORING_SETUP_DEFER_TASKRUN ; - ret =io_uring_queue_init_params (N_CQE_OVERFLOW , &ring, ¶ms); - }else { - ret =io_uring_queue_init (32 , &ring,0 ); +if (args->early_error ==ERROR_EARLY_OVERFLOW ) { + params.flags |=IORING_SETUP_CQSIZE ; + params.cq_entries =N_CQE_OVERFLOW ; + n_sqe =N_CQE_OVERFLOW ; } + + ret =io_uring_queue_init_params (n_sqe, &ring, ¶ms);if (ret) {fprintf (stderr,"queue init failed: %d\n" , ret);return ret;
@@ -457,23 +460,30 @@ int main(int argc, char *argv[])int ret;int loop;int early_error =0 ; +bool has_defer;if (argc >1 )return T_EXIT_SKIP ; -for (loop =0 ; loop <8 ; loop++) { + has_defer =t_probe_defer_taskrun (); + +for (loop =0 ; loop <16 ; loop++) {struct args a = { .stream = loop &0x01 , .wait_each = loop &0x2 , .recvmsg = loop &0x04 , + .defer = loop &0x08 , }; +if (a.defer && !has_defer) +continue ;for (early_error =0 ; early_error <ERROR_EARLY_LAST ; early_error++) { a.early_error = (enum early_error_t )early_error; ret =test (&a);if (ret) {fprintf (stderr, -"test stream=%d wait_each=%d recvmsg=%d early_error=%d failed\n" , - a.stream, a.wait_each, a.recvmsg, a.early_error); +"test stream=%d wait_each=%d recvmsg=%d early_error=%d " +" defer=%d failed\n" , + a.stream, a.wait_each, a.recvmsg, a.early_error, a.defer);return T_EXIT_FAIL ; }if (no_recv_mshot)
test/rsrc_tags.c +8 -2
@@ -401,7 +401,8 @@ static int test_notag(void)int main (int argc,char *argv[]) { -int ring_flags[] = {0 ,IORING_SETUP_IOPOLL ,IORING_SETUP_SQPOLL }; +int ring_flags[] = {0 ,IORING_SETUP_IOPOLL ,IORING_SETUP_SQPOLL , +IORING_SETUP_DEFER_TASKRUN };int i, ret;if (argc >1 )
@@ -423,7 +424,12 @@ int main(int argc, char *argv[]) }for (i =0 ; i <sizeof (ring_flags) /sizeof (ring_flags[0 ]); i++) { - ret =test_files (ring_flags[i]); +int flag = ring_flags[i]; + +if (flag &IORING_SETUP_DEFER_TASKRUN && !t_probe_defer_taskrun ()) +continue ; + + ret =test_files (flag);if (ret) {printf ("test_tag failed, type %i\n" , i);return ret;
Signature · 1 line
2.30.2
PATCH 5/11 add a defer-taskrun test
In reply to Dylan YudakenAdd a test specifically for IORING_SETUP_DEFER_TASKRUN
--- test/Makefile | 1 + test/defer-taskrun.c | 217 +++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 218 insertions(+) create mode 100644 test/defer-taskrun.c
test/Makefile +1 -0
@@ -62,6 +62,7 @@ test_srcs := \ d4ae271dfaae.c \ d77a67ed5f27.c \ defer.c \ + defer-taskrun.c \ double-poll-crash.c \ drop-submit.c \ eeed8b54e0df.c \
test/defer-taskrun.c new file +217 -0
217 lines not on this page · raw message
Signature · 1 line
2.30.2
PATCH 6/11 fix documentation shortenings
In reply to Dylan Yudaken'parm' probably was intended to be param, but parameter is more explicit, so change it.
--- man/io_uring_wait_cqe.3 | 2 +- man/io_uring_wait_cqe_nr.3 | 2 +- man/io_uring_wait_cqe_timeout.3 | 2 +- man/io_uring_wait_cqes.3 | 2 +- 4 files changed, 4 insertions(+), 4 deletions(-)
man/io_uring_wait_cqe.3 +1 -1
@@ -31,7 +31,7 @@ the application can retrieve the completion with .SH RETURN VALUE On success .BR io_uring_wait_cqe (3) -returns 0 and the cqe_ptr parm is filled in. On failure it returns +returns 0 and the cqe_ptr parameter is filled in. On failure it returns .BR -errno . The return value indicates the result of waiting for a CQE, and it has no relation to the CQE result itself.
man/io_uring_wait_cqe_nr.3 +1 -1
@@ -34,7 +34,7 @@ the application can retrieve the completion with .SH RETURN VALUE On success .BR io_uring_wait_cqe_nr (3) -returns 0 and the cqe_ptr parm is filled in. On failure it returns +returns 0 and the cqe_ptr parameter is filled in. On failure it returns .BR -errno . The return value indicates the result of waiting for a CQE, and it has no relation to the CQE result itself.
man/io_uring_wait_cqe_timeout.3 +1 -1
@@ -43,7 +43,7 @@ when waiting for a request. .SH RETURN VALUE On success .BR io_uring_wait_cqes (3) -returns 0 and the cqe_ptr parm is filled in. On failure it returns +returns 0 and the cqe_ptr parameter is filled in. On failure it returns .BR -errno . The return value indicates the result of waiting for a CQE, and it has no relation to the CQE result itself.
man/io_uring_wait_cqes.3 +1 -1
@@ -48,7 +48,7 @@ when waiting for a request. .SH RETURN VALUE On success .BR io_uring_wait_cqes (3) -returns 0 and the cqe_ptr parm is filled in. On failure it returns +returns 0 and the cqe_ptr parameter is filled in. On failure it returns .BR -errno . .SH SEE ALSO .BR io_uring_submit (3),
Signature · 1 line
2.30.2
PATCH 7/11 update io_uring_enter.2 docs for IORING_FEAT_NODROP
In reply to Dylan YudakenThe EBUSY docs are out of date, so update them for the IORING_FEAT_NODROP feature flag
--- man/io_uring_enter.2 | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-)
man/io_uring_enter.2 +12 -1
@@ -1327,7 +1327,18 @@ is a valid file descriptor, but the io_uring ring is not in the right state for details on how to enable the ring. .TP .B EBUSY -The application is attempting to overcommit the number of requests it can have +If the +.B IORING_FEAT_NODROP +feature flag is set, then +.B EBUSY +will be returned if there were overflow entries, +.B IORING_ENTER_GETEVENTS +flag is set and not all of the overflow entries were able to be flushed to +the CQ ring. + +Without +.B IORING_FEAT_NODROP +the application is attempting to overcommit the number of requests it can have pending. The application should wait for some completions and try again. May occur if the application tries to queue more requests than we have room for in the CQ ring, or if the application attempts to wait for more events without
Signature · 1 line
2.30.2
PATCH 8/11 add docs for overflow lost errors
In reply to Dylan YudakenAdd man docs for return values indicating a CQE was lost.
--- man/io_uring_enter.2 | 11 +++++++++++ man/io_uring_setup.2 | 11 +++++++++-- 2 files changed, 20 insertions(+), 2 deletions(-)
man/io_uring_enter.2 +11 -0
@@ -1326,6 +1326,17 @@ is a valid file descriptor, but the io_uring ring is not in the right state .BR io_uring_register (2) for details on how to enable the ring. .TP +.B EBADR +At least one CQE was dropped even with the +.B IORING_FEAT_NODROP +feature, and there are no otherwise available CQEs. This clears the error state +and so with no other changes the next call to +.BR io_uring_setup (2) +will not have this error. This error should be extremely rare and indicates the +machine is running critically low on memory and. It may be reasonable for the +application to terminate running unless it is able to safely handle any CQE +being lost. +.TP .B EBUSY If the .B IORING_FEAT_NODROP
man/io_uring_setup.2 +9 -2
@@ -279,7 +279,7 @@ call. The SQEs must still be allocated separately. This brings the necessary calls down from three to two. Available since kernel 5.4. .TP .B IORING_FEAT_NODROP -If this flag is set, io_uring supports never dropping completion events. +If this flag is set, io_uring supports almost never dropping completion events. If a completion event occurs and the CQ ring is full, the kernel stores the event internally until such a time that the CQ ring has room for more entries. If this overflow condition is entered, attempting to submit more
@@ -287,7 +287,14 @@ IO will fail with the .B -EBUSY error value, if it can't flush the overflown events to the CQ ring. If this happens, the application must reap events from the CQ ring and attempt the -submit again. Available since kernel 5.5. +submit again. If the kernel has no free memory to store the event internally +it will be visible by an increase in the overflow value on the cqring. +Available since kernel 5.5. Additionally +.BR io_uring_enter (2) +will return +.B -EBADR +the next time it would otherwise sleep waiting for completions (since kernel 5.19). + .TP .B IORING_FEAT_SUBMIT_STABLE If this flag is set, applications can be certain that any data for
Signature · 1 line
2.30.2
PATCH 9/11 expose CQ ring overflow state
In reply to Dylan YudakenAllow the application to easily view if the CQ ring is in overflow state in case it would like to explicitly flush using io_uring_get_events.
Explicit flushing can be useful for applications that prefer to have reduced latency on CQs than to process as many as possible.
--- src/include/liburing.h | 9 +++++++++ 1 file changed, 9 insertions(+)
src/include/liburing.h +9 -0
@@ -1064,6 +1064,15 @@ static inline unsigned io_uring_cq_ready(const struct io_uring *ring)return io_uring_smp_load_acquire (ring->cq.ktail) - *ring->cq.khead; } +/* +* Returns true if there are overflow entries waiting to be flushed onto +* the CQ ring +*/ +static inline bool io_uring_cq_has_overflow (const struct io_uring *ring) +{ +return IO_URING_READ_ONCE (*ring->sq.kflags) &IORING_SQ_CQ_OVERFLOW ; +} +/* * Returns true if the eventfd notification is currently enabled */
Signature · 1 line
2.30.2
PATCH 10/11 overflow: add tests
In reply to Dylan YudakenAdd tests that verify that overflow conditions behave appropriately. Specifically:
* if overflow is continually flushed, then CQEs should arrive mostly in order to prevent starvation of some completions
* if CQEs are dropped due to GFP_ATOMIC allocation failures it is possible to terminate cleanly. This is not tested by default as it requires debug kernel config, and also has system-wide effects
--- test/cq-overflow.c | 248 +++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 241 insertions(+), 7 deletions(-)
test/cq-overflow.c +241 -7
@@ -9,6 +9,7 @@#include <stdlib.h> #include <string.h> #include <fcntl.h> +#include <assert.h> #include "helpers.h" #include "liburing.h"
@@ -21,6 +22,32 @@ static struct iovec *vecs;#define ENTRIES 8 +/* +* io_uring has rare cases where CQEs are lost. +* This happens when there is no space in the CQ ring, and also there is no +* GFP_ATOMIC memory available. In reality this probably means that the process +* is about to be killed as many other things might start failing, but we still +* want to test that liburing and the kernel deal with this properly. The fault +* injection framework allows us to test this scenario. Unfortunately this +* requires some system wide changes and so we do not enable this by default. +* The tests in this file should work in both cases (where overflows are queued +* and where they are dropped) on recent kernels. +* +* In order to test dropped CQEs you should enable fault injection in the kernel +* config: +* +* CONFIG_FAULT_INJECTION=y +* CONFIG_FAILSLAB=y +* CONFIG_FAULT_INJECTION_DEBUG_FS=y +* +* and then run the test as follows: +* echo Y > /sys/kernel/debug/failslab/task-filter +* echo 100 > /sys/kernel/debug/failslab/probability +* echo 0 > /sys/kernel/debug/failslab/verbose +* echo 100000 > /sys/kernel/debug/failslab/times +* bash -c "echo 1 > /proc/self/make-it-fail && exec ./cq-overflow.t" +*/ +static int test_io (const char *file,unsigned long usecs,unsigned *drops,int fault) {struct io_uring_sqe *sqe;
@@ -29,6 +56,7 @@ static int test_io(const char *file, unsigned long usecs, unsigned *drops, int funsigned reaped, total;struct io_uring ring;int nodrop, i, fd, ret; +bool cqe_dropped = false; fd =open (file,O_RDONLY |O_DIRECT );if (fd <0 ) {
@@ -104,8 +132,8 @@ static int test_io(const char *file, unsigned long usecs, unsigned *drops, int f reap_it: reaped =0 ; do { - if (nodrop) { - /* nodrop should never lose events */ + if (nodrop && !cqe_dropped) { + /* nodrop should never lose events unless cqe_dropped */ if (reaped == total) break; } else {
@@ -113,7 +141,10 @@ reap_it:break ; } ret =io_uring_wait_cqe (&ring, &cqe); -if (ret) { +if (nodrop && ret == -EBADR ) { + cqe_dropped = true; +continue ; + }else if (ret) {fprintf (stderr,"wait_cqe=%d\n" , ret); goto err; }
@@ -133,7 +164,7 @@ reap_it: goto err; } -if (!nodrop) { +if (!nodrop || cqe_dropped) { *drops = *ring.cq.koverflow; }else if (*ring.cq.koverflow) {fprintf (stderr,"Found %u overflows\n" , *ring.cq.koverflow);
@@ -154,18 +185,29 @@ static int reap_events(struct io_uring *ring, unsigned nr_events, int do_wait) {struct io_uring_cqe *cqe;int i, ret =0 , seq =0 ; +unsigned int start_overflow = *ring->cq.koverflow; +bool dropped = false;for (i =0 ; i < nr_events; i++) {if (do_wait) ret =io_uring_wait_cqe (ring, &cqe);else ret =io_uring_peek_cqe (ring, &cqe); -if (ret) { +if (do_wait && ret == -EBADR ) { +unsigned int this_drop = *ring->cq.koverflow - + start_overflow; + + dropped = true; + start_overflow = *ring->cq.koverflow; +assert (this_drop >0 ); + i += (this_drop -1 ); +continue ; + }else if (ret) {if (ret != -EAGAIN )fprintf (stderr,"cqe peek failed: %d\n" , ret);break ; } -if (cqe->user_data != seq) { +if (!dropped && cqe->user_data != seq) {fprintf (stderr,"cqe sequence out-of-order\n" );fprintf (stderr,"got %d, wanted %d\n" , (int ) cqe->user_data, seq);
@@ -242,19 +284,211 @@ err:return 1 ; } + +static void submit_one_nop (struct io_uring *ring,int ud) +{ +struct io_uring_sqe *sqe; +int ret; + + sqe =io_uring_get_sqe (ring); +assert (sqe); +io_uring_prep_nop (sqe); + sqe->user_data = ud; + ret =io_uring_submit (ring); +assert (ret ==1 ); +} + +/* +* Create an overflow condition and ensure that SQEs are still processed +*/ +static int test_overflow_handling ( +bool batch, +int cqe_multiple, +bool poll, +bool defer) +{ +struct io_uring ring; +struct io_uring_params p; +int ret, i, j, ud, cqe_count; +unsigned int count; +int const N =8 ; +int const LOOPS =128 ; +int const QUEUE_LENGTH =1024 ; +int completions[N ]; +int queue[QUEUE_LENGTH ]; +int queued =0 ; +int outstanding =0 ; +bool cqe_dropped = false; + +memset (&completions,0 ,sizeof (int ) *N ); +memset (&p,0 ,sizeof (p)); + p.cq_entries =2 * cqe_multiple; + p.flags |=IORING_SETUP_CQSIZE ; + +if (poll) + p.flags |=IORING_SETUP_IOPOLL ; + +if (defer) + p.flags |=IORING_SETUP_DEFER_TASKRUN ; + + ret =io_uring_queue_init_params (2 , &ring, &p); +if (ret) { +fprintf (stderr,"io_uring_queue_init failed %d\n" , ret); +return 1 ; + } + +assert (p.cq_entries <N ); +/* submit N SQEs, some should overflow */ +for (i =0 ; i <N ; i++) { +submit_one_nop (&ring, i); + outstanding++; + } + +for (i =0 ; i <LOOPS ; i++) { +struct io_uring_cqe *cqes[N ]; + +if (io_uring_cq_has_overflow (&ring)) { +/* +* Flush any overflowed CQEs and process those. Actively +* flush these to make sure CQEs arrive in vague order +* of being sent. +*/ + ret =io_uring_get_events (&ring); +if (ret !=0 ) { +fprintf (stderr, +"io_uring_get_events returned %d\n" , + ret); + goto err; + } + }else if (!cqe_dropped) { +for (j =0 ; j < queued; j++) { +submit_one_nop (&ring, queue[j]); + outstanding++; + } + queued =0 ; + } + +/* We have lost some random cqes, stop if no remaining. */ +if (cqe_dropped && outstanding == *ring.cq.koverflow) +break ; + + ret =io_uring_wait_cqe (&ring, &cqes[0 ]); +if (ret == -EBADR ) { + cqe_dropped = true; +fprintf (stderr,"CQE dropped\n" ); +continue ; + }else if (ret !=0 ) { +fprintf (stderr,"io_uring_wait_cqes failed %d\n" , ret); + goto err; + } + cqe_count =1 ; +if (batch) { + ret =io_uring_peek_batch_cqe (&ring, &cqes[0 ],2 ); +if (ret <0 ) { +fprintf (stderr, +"io_uring_peek_batch_cqe failed %d\n" , + ret); + goto err; + } + cqe_count = ret; + } +for (j =0 ; j < cqe_count; j++) { +assert (cqes[j]->user_data <N ); + ud = cqes[j]->user_data; + completions[ud]++; +assert (queued <QUEUE_LENGTH ); + queue[queued++] = (int )ud; + } +io_uring_cq_advance (&ring, cqe_count); + outstanding -= cqe_count; + } + +/* See if there were any drops by flushing the CQ ring *and* overflow */ +do { +struct io_uring_cqe *cqe; + + ret =io_uring_get_events (&ring); +if (ret <0 ) { +if (ret == -EBADR ) { +fprintf (stderr,"CQE dropped\n" ); + cqe_dropped = true; +break ; + } + goto err; + } +if (outstanding && !io_uring_cq_ready (&ring)) + ret =io_uring_wait_cqe_timeout (&ring, &cqe,NULL ); + +if (ret && ret != -ETIME ) { +if (ret == -EBADR ) { +fprintf (stderr,"CQE dropped\n" ); + cqe_dropped = true; +break ; + } +fprintf (stderr,"wait_cqe_timeout = %d\n" , ret); + goto err; + } + count =io_uring_cq_ready (&ring); +io_uring_cq_advance (&ring, count); + outstanding -= count; + }while (count); + +io_uring_queue_exit (&ring); + +/* Make sure that completions come back in the same order they were +* sent. If they come back unfairly then this will concentrate on a +* couple of indices. +*/ +for (i =1 ; !cqe_dropped && i <N ; i++) { +if (abs (completions[i] - completions[i -1 ]) >1 ) { +fprintf ( + stderr, +"bad completion size %d %d\n" , + completions[i], + completions[i -1 ]); + goto err; + } + } +return 0 ; +err: +io_uring_queue_exit (&ring); +return 1 ; +} +int main (int argc,char *argv[]) {const char *fname =".cq-overflow" ;unsigned iters, drops;unsigned long usecs;int ret; +int i; +bool can_defer;if (argc >1 )return T_EXIT_SKIP ; + can_defer =t_probe_defer_taskrun (); +for (i =0 ; i <16 ; i++) { +bool batch = i &1 ; +int mult = (i &2 ) ?1 :2 ; +bool poll = i &4 ; +bool defer = i &8 ; + +if (defer && !can_defer) +continue ; + + ret =test_overflow_handling (batch, mult, poll, defer); +if (ret) { +fprintf (stderr,"test_overflow_handling(" +"batch=%d, mult=%d, poll=%d, defer=%d) failed\n" , + batch, mult, poll, defer); + goto err; + } + } + ret =test_overflow ();if (ret) { -printf ("test_overflow failed\n" ); +fprintf (stderr,"test_overflow failed\n" );return ret; }
Signature · 1 line
2.30.2
PATCH 11/11 file-verify test: log if short read
In reply to Dylan Yudakenthis test assumes any success is a full read. for now we can just log if it's short and fail, and if in future this wants to be handled that shouldn't be too much work.
--- test/file-verify.c | 4 ++++ 1 file changed, 4 insertions(+)
test/file-verify.c +4 -0
@@ -430,6 +430,10 @@ static int test(struct io_uring *ring, const char *fname, int buffered,fprintf (stderr,"bad read %d, read %d\n" , cqe->res, i); goto err; } +if (cqe->res <CHUNK_SIZE ) { +fprintf (stderr,"short read %d, read %d\n" , cqe->res, i); + goto err; + }if (cqe->flags &IORING_CQE_F_BUFFER ) index = cqe->flags >>16 ;else
Signature · 1 line
2.30.2