CVS commit: src/sys/kern

["Taylor R Campbell" <riastradh%netbsd.org@localhost>]

Module Name:    src
Committed By:   riastradh
Date:           Sun Mar  5 14:40:32 UTC 2023

Modified Files:
        src/sys/kern: vfs_syscalls.c

Log Message:
open(2): Don't map ERESTART to EINTR.

If a file or device's open function returns ERESTART, respect that --
restart the syscall; don't pretend a signal has been delivered when
it was not.  If an SA_RESTART signal was delivered, POSIX does not
allow it to fail with EINTR:

    SA_RESTART
        This flag affects the behavior of interruptible functions;
        that is, those specified to fail with errno set to [EINTR].
        If set, and a function specified as interruptible is
        interrupted by this signal, the function shall restart and
        shall not fail with [EINTR] unless otherwise specified.  If
        an interruptible function which uses a timeout is restarted,
        the duration of the timeout following the restart is set to
        an unspecified value that does not exceed the original
        timeout value.  If the flag is not set, interruptible
        functions interrupted by this signal shall fail with errno
        set to [EINTR].

https://pubs.opengroup.org/onlinepubs/9699919799/functions/sigaction.html

Nothing in the POSIX definition of open specifies otherwise.

In 1990, Kirk McKusick added these lines with a mysterious commit
message:

Author: Kirk McKusick <mckusick>
Date:   Tue Apr 10 19:36:33 1990 -0800

    eliminate longjmp from the kernel (for karels)

diff --git a/sys/kern/vfs_syscalls.c b/sys/kern/vfs_syscalls.c
index 7bc7b39bbf..d572d3a32d 100644
--- a/sys/kern/vfs_syscalls.c
+++ b/sys/kern/vfs_syscalls.c
@@ -14,7 +14,7 @@
  * IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
  *
- *     @(#)vfs_syscalls.c      7.42 (Berkeley) 3/26/90
+ *     @(#)vfs_syscalls.c      7.43 (Berkeley) 4/10/90
  */

 #include "param.h"
@@ -530,8 +530,10 @@ copen(scp, fmode, cmode, ndp, resultfd)
        if (error = vn_open(ndp, fmode, (cmode & 07777) &~ S_ISVTX)) {
                crfree(fp->f_cred);
                fp->f_count--;
-               if (error == -1)        /* XXX from fdopen */
-                       return (0);     /* XXX from fdopen */
+               if (error == EJUSTRETURN)       /* XXX from fdopen */
+                       return (0);             /* XXX from fdopen */
+               if (error == ERESTART)
+                       error = EINTR;
                scp->sc_ofile[indx] = NULL;
                return (error);
        }

(found via this git import of the CSRG history:
https://github.com/robohack/ucb-csrg-bsd/commit/cce2869b7ae5d360921eb411005b328a29c4a3fe)

This change appears to have served two related purposes:

1. The fdopen function (the erstwhile open routine for /dev/fd/N)
   used to return -1 as a hack to mean it had just duplicated the fd;
   it was recently changed by Mike Karels, in kern_descrip.c 7.9, to
   return EJUSTRETURN, now defined to be -2, presumably to avoid a
   conflict with ERESTART, defined to be -1.  So this change finished
   part of the change by Mike Karels to use a different magic return
   code from fdopen.

   Of course, today we use still another disgusting hack, EDUPFD, for
   the same purpose, so none of this is relevant any more.

2. Prior to April 1990, the kernel handled signals during tsleep(9)
   by longjmping out to the system call entry point or similar.  In
   April 1990, Mike Karels worked to convert all of that into
   explicit unwind logic by passing through EINTR or ERESTART as
   appropriate, instead of setjmp at each entry point.

However, it's not clear to me why this setjmp/longjmp and
fdopen/-1/EJUSTRETURN renovation justifies unconditional logic to map
ERESTART to EINTR in open(2).  I suspect it was a mistake.

In 2013, the corresponding logic to map ERESTART to EINTR in open(2)
was removed from FreeBSD:

   r246472 | kib | 2013-02-07 14:53:33 +0000 (Thu, 07 Feb 2013) | 11 lines

   Stop translating the ERESTART error from the open(2) into EINTR.
   Posix requires that open(2) is restartable for SA_RESTART.

   For non-posix objects, in particular, devfs nodes, still disable
   automatic restart of the opens. The open call to a driver could have
   significant side effects for the hardware.

   Noted and reviewed by:  jilles
   Discussed with: bde
   MFC after:      2 weeks

Index: vfs_syscalls.c
===================================================================
--- vfs_syscalls.c      (revision 246471)
+++ vfs_syscalls.c      (revision 246472)
@@ -1106,8 +1106,6 @@
                                goto success;
                }

-               if (error == ERESTART)
-                       error = EINTR;
                goto bad;
        }
        td->td_dupfd = 0;

https://cgit.freebsd.org/src/commit/sys/kern/vfs_syscalls.c?id=2ca49983425886121b506cb5126b60a705afc38c

It's not clear to me that there's any reason to treat device nodes
specially here; in fact, if a driver's .d_open routine sleeps and is
woken by a concurrent revoke without a signal pending or with an
SA_RESTART signal pending, it is wrong for it to fail with EINTR.
But it MUST restart the whole system call rather than continue
sleeping in a loop or just exit the loop and continue to open,
because it is mandatory in the security model of revoke for open(2)
to retry the permissions check at that point.

PR kern/57260

XXX pullup-8
XXX pullup-9
XXX pullup-10


To generate a diff of this commit:
cvs rdiff -u -r1.556 -r1.557 src/sys/kern/vfs_syscalls.c

Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.