Using kqueue with aio_read/write

[Alan Somers]

On Wed, Jan 2, 2019 at 5:53 AM Willem Jan Withagen <wjw at digiware.nl> wrote:
>
> On 02/01/2019 03:23, Alan Somers wrote:
> > On Tue, Jan 1, 2019 at 6:56 PM Willem Jan Withagen <wjw at digiware.nl> wrote:
> >> On 28/12/2018 02:47, Alan Somers wrote:
> >>> On Thu, Dec 27, 2018 at 6:15 PM Willem Jan Withagen <wjw at digiware.nl> wrote:
> >>>> Hi,
> >>>>
> >>>> Im trying to understand why I cannot get so code to work.
> >>>> This is the smallest extract I can make to show my problem.
> >>>>
> >>>> I would expect the kevent() call to return every timeo tick.
> >>>> Even if I tell it NOT to time-out I get these spurts of errors
> >>>>
> >>>> Since there is nothing to trigger the AIO-event, I would expect kqueue
> >>>> to hold indefinitly.
> >>>>
> >>>> But it does not generate anything other than errors
> >>>> And instead it repeatedly complains that there is a permission error:
> >>>>      get_events_kevent: EV_Error(1) kevent(): Operation not permitted
> >>>>
> >>>> But I'm not getting where that would the case...
> >>>>
> >>>> Surely a pilot error, but I do overlook it al the time.
> >>>> So suggestions are welcome.
> >>>>
> >>>> Thanx,
> >>>> --WjW
> >>>>
> >>>> #include <aio.h>
> >>>> #include <errno.h>
> >>>> #include <fcntl.h>
> >>>> #include <stdio.h>
> >>>> #include <stdlib.h>
> >>>> #include <string.h>
> >>>> #include <sys/stat.h>
> >>>> #include <sys/event.h>
> >>>> #include <unistd.h>
> >>>>
> >>>> #define BUFFER_SIZE     512
> >>>> #define MAX_EVENTS 32
> >>>>
> >>>> #define FILENAME "/tmp/aio_test"
> >>>> char filename[256];
> >>>> int fd;
> >>>> int done = 0;
> >>>>
> >>>> void get_events_kevent(int fd, int kq)
> >>>> {
> >>>>        printf("get_events function fd = %d, kq = %d\n", fd, kq);
> >>>>        int i = 0, errcnt = 0, err, ret, reterr, rev;
> >>>>        int search = 1;
> >>>>
> >>>>        int timeout_ms = 10;
> >>>>        struct timespec timeo = {
> >>>>            timeout_ms / 1000,
> >>>>            (timeout_ms % 1000) * 1000 * 1000
> >>>>        };
> >>>>        struct kevent filter[16];
> >>>>        struct kevent changed[16];
> >>>>
> >>>>        EV_SET(&filter[0], fd, EVFILT_AIO,
> >>>>                EV_ADD,
> >>>>                0, 0, 0 );
> >>> This is the first problem.  There's no need to explicitly set
> >>> EVFILT_AIO on the kqueue.  It gets set by the aio_read(2) or similar
> >>> syscall.  And this invocation wouldn't be correct anyway, because for
> >>> AIO the ident field refers to the address of the struct aiocb, not the
> >>> file descriptor.  If the only events you care about are AIO, then you
> >>> can pass NULL as the filter argument to kevent.  I suspect this is the
> >>> cause of your problem.  The kernel probably thinks you're trying to
> >>> register for an aiocb that's outside of your address space or
> >>> something like that.
> >>>
> >>>
> >>>>        while (!done) {
> >>>>            printf("+");
> >>>>            rev = kevent(kq, filter, 1, changed, 16, 0); //&timeo);
> >>>>            if (rev < 0) {
> >>>>                perror("kevent error");
> >>>>            } else if (rev == 0) {
> >>>>                printf("T");
> >>>>            } else {
> >>>>                printf("rev(%d)\n", rev);
> >>>>                if (changed[0].flags == EV_ERROR) {
> >>>>                    errno = changed[0].data;
> >>>>                    printf( "%s: EV_Error(%d) kevent(): %s\n", __func__, errno,
> >>>>                        strerror(errno));
> >>>>                    memset(&changed[0], 0, sizeof(struct kevent));
> >>>>                } else {
> >>>>                    err = aio_error((struct aiocb*)changed[0].udata);
> >>> No need to call aio_error(2) after kevent(2) returns.  You can go
> >>> straight to aio_return.  aio_error shouldn't hurt, but it isn't
> >>> necessary.
> >> According to kevent(2) calling kevent can return errors on the called
> >> aio_calls.
> >> It then returns with EV_ERROR in flags, and errno is stored in the
> >> event.data.
> >>
> >> But what would be going on when the event's flag contains EV_ERROR but
> >> event's data is still 0???
> >>
> >> the udata field still seems to point to the aio data that was passed
> >> into the aio block when calling aio_read().
> >>
> >> Should I ignore this as a non-error?
> >>
> >> --WjW
> > Are you sure you bzero()ed your aiocb before initializing it?  Any
> > stack garbage that was present in its
> > aio_sigevent.sigev_notify_kevent_flags field will be dutifully copied
> > into the returned kevent.  And in any case, the definitive way to get
> > the final status of a completed aio operation is with aio_return.
> That seems to help in getting things clear...
>
>     -13> 2019-01-02 13:32:31.834 dc15a80  1 bdev:327 paio
> get_next_completed processing event i = 0 aio_return(22) (22) Invalid
> argument
>
> Disadvantage is that it is not clear yet which of the many arguments
> that is?
>
> --WjW

It could be failing for any reason that read(2) can fail.  Or, the
iocb might not be complete.  aio_error(2) would tell you.  One common
problem is if you allocate the iocb on the stack and accidentally move
it after calling aio_read().  That's bitten me before.
-Alan