On Wednesday, 26 February 2020 15:08:24 CET Daniel P. Berrangé
wrote:
> On Wed, Feb 26, 2020 at 02:39:04PM +0100, Pino Toscano wrote:
> > select() has a maximum value for the FDs it can monitor, and since
> > the libguestfs library can be used in other applications, this limit
> > may be hit by users in case lots of FDs are opened.
> >
> > As solution, switch to poll(): it has a slightly better interface to
> > check what changed and for which FD, and it does not have a limit in the
> > value of the FDs monitored.
> >
> > poll() is supported on the platforms we support, so there is no need to
> > use the gnulib module for it.
> > ---
> > lib/command.c | 54 ++++++++++++++++++++++++++++++++++-----------------
> > 1 file changed, 36 insertions(+), 18 deletions(-)
> >
> > diff --git a/lib/command.c b/lib/command.c
> > index f2161de9a..13b084934 100644
> > --- a/lib/command.c
> > +++ b/lib/command.c
> > @@ -89,9 +89,8 @@
> > #include <signal.h>
> > #include <errno.h>
> > #include <assert.h>
> > -#include <sys/types.h>
> > #include <sys/stat.h>
> > -#include <sys/select.h>
> > +#include <poll.h>
> >
> > #ifdef HAVE_SYS_TIME_H
> > #include <sys/time.h>
> > @@ -650,37 +649,41 @@ run_child (struct command *cmd)
> > static int
> > loop (struct command *cmd)
> > {
> > - fd_set rset, rset2;
> > - int maxfd = -1, r;
> > + struct pollfd fds[2];
> > + int r;
> > size_t nr_fds = 0;
> > CLEANUP_FREE char *buf = safe_malloc (cmd->g, BUFSIZ);
> > ssize_t n;
> >
> > - FD_ZERO (&rset);
> > + memset (&fds, 0, sizeof fds);
> > +
> > + fds[0].fd = cmd->errorfd;
> > + fds[1].fd = cmd->outfd;
> >
> > if (cmd->errorfd >= 0) {
> > - FD_SET (cmd->errorfd, &rset);
> > - maxfd = MAX (cmd->errorfd, maxfd);
> > + fds[0].events = POLLIN;
> > nr_fds++;
> > }
> >
> > if (cmd->outfd >= 0) {
> > - FD_SET (cmd->outfd, &rset);
> > - maxfd = MAX (cmd->outfd, maxfd);
> > + fds[1].events = POLLIN;
> > nr_fds++;
> > }
> >
> > while (nr_fds > 0) {
> > - rset2 = rset;
> > - r = select (maxfd+1, &rset2, NULL, NULL, NULL);
> > + r = poll (fds, 2, -1);
> > if (r == -1) {
> > if (errno == EINTR || errno == EAGAIN)
> > continue;
> > - perrorf (cmd->g, "select");
> > + perrorf (cmd->g, "poll");
> > + return -1;
> > + }
> > + if (fds[0].revents & POLLERR || fds[1].revents & POLLERR) {
> > + perrorf (cmd->g, "poll");
> > return -1;
> > }
> >
> > - if (cmd->errorfd >= 0 && FD_ISSET (cmd->errorfd,
&rset2)) {
> > + if (fds[0].revents & POLLIN) {
> > /* Read output and send it to the log. */
> > n = read (cmd->errorfd, buf, BUFSIZ);
> > if (n > 0)
> > @@ -689,20 +692,26 @@ loop (struct command *cmd)
> > else if (n == 0) {
> > if (close (cmd->errorfd) == -1)
> > perrorf (cmd->g, "close: errorfd");
> > - FD_CLR (cmd->errorfd, &rset);
> > + fds[0].fd = -1;
> > cmd->errorfd = -1;
> > nr_fds--;
> > }
> > else if (n == -1) {
> > perrorf (cmd->g, "read: errorfd");
> > close (cmd->errorfd);
> > - FD_CLR (cmd->errorfd, &rset);
> > + fds[0].fd = -1;
> > cmd->errorfd = -1;
> > nr_fds--;
> > }
> > }
> > + if (fds[0].revents & POLLHUP) {
> > + close (cmd->errorfd);
> > + fds[0].fd = -1;
> > + cmd->errorfd = -1;
> > + nr_fds--;
> > + }
>
> Now fds[0] is -1, and fds[1] is still an open file
> descriptor.
>
> Next time around the loop it will call poll() on 'fds' array
> which contains this FD == -1. In theory that results in POLLNVAL,
> but on OS-X at least it causes poll() itself to return an error.
Hmm in the documentations there are these bits:
http://man7.org/linux/man-pages/man2/poll.2.html
"The field fd contains a file descriptor for an open file. If this
field is negative, then the corresponding events field is ignored and
the revents field returns zero. (This provides an easy way of ignoring
a file descriptor for a single poll() call: simply negate the fd field.
Note, however, that this technique can't be used to ignore file
descriptor 0.)"
https://linux.die.net/man/3/poll
"If the value of fd is less than 0, events shall be ignored, and revents
shall be set to 0 in that entry on return from poll()."
So to my understanding it is fine, and indeed I did not get POLLNVAL
in my testing.
In fact libnbd & nbdkit also rely on this behaviour. I don't believe
we've ever tried either one on macOS though. It seems to be mandated
by POSIX too:
"If the value of fd is less than 0, events shall be ignored, and
revents shall be set to 0 in that entry on return from poll()."
Rich.
--
Richard Jones, Virtualization Group, Red Hat
libguestfs lets you edit virtual machines. Supports shell scripting,
bindings from many languages.