Date: Fri, 11 Apr 2003 02:30:27 -0600 (CST) From: Stephen Hurd <admin@tw2.thebbs.org> To: FreeBSD-gnats-submit@FreeBSD.org Cc: Rob Swindell <rob@synchro.net> Subject: kern/50827: [PATCH] no sane record locking on *nix. (More types needed) Message-ID: <200304110830.h3B8UR8J000899@tw2.thebbs.org> Resent-Message-ID: <200304111600.h3BG0PpL024743@freefall.freebsd.org>
next in thread | raw e-mail | index | archive | help
>Number: 50827 >Category: kern >Synopsis: [PATCH] no sane record locking on *nix. (More types needed) >Confidential: no >Severity: serious >Priority: medium >Responsible: freebsd-bugs >State: open >Quarter: >Keywords: >Date-Required: >Class: change-request >Submitter-Id: current-users >Arrival-Date: Fri Apr 11 09:00:24 PDT 2003 >Closed-Date: >Last-Modified: >Originator: Stephen Hurd >Release: FreeBSD 4.7-RELEASE i386 >Organization: >Environment: System: FreeBSD sharon.kingcole.local 4.7-RELEASE FreeBSD 4.7-RELEASE #33: Fri Jan 31 00:06:43 CST 2003 admin@sharon.kingcole.local:/usr/src/sys/compile/SHARON i386 >Description: Record locking in POSIX systems is in a sad state. Due to some poor design choices when fcntl() was developed, our only methods of record locking are (to quote the man page) "completely stupid". There are times when fcntl() locking is utterly unmanageable, and flock() locks are just too much. Currently, the record locking capabilities of DOS 3.0+ are more usefull than those in any POSIX compliant *nix. After leading the battle for sane locks with the implementation of flock() in 4.2BSD, the time has come for BSD to take up the job once again and implement sane record locking. >How-To-Repeat: 1) Write a threaded multi-instance server. 2) Do record locking. 3) Stop swearing at fcntl() (Optional) >Fix: The following patch adds the F_SANEWRLCK, F_SANERDLCK, F_SANEUNLCK, F_SANEWRLCKNO, and F_SANERDLCKNO lock types to fcntl(). The types with NO (No Overlap) will cause locks to conflict regardless of owner (ie: You cannot get an exclusive lock that is contained inside of a shared lock you currently hold) Sane locks are closed with the file descriptor, not the file and you can get an exclusive lock without opening the file with write access and a shared lock without opening the file with read access. --- sanelock.patch begins here --- diff -c /sys/kern.old/kern_descrip.c /sys/kern/kern_descrip.c *** /sys/kern.old/kern_descrip.c Fri Apr 11 01:47:51 2003 --- /sys/kern/kern_descrip.c Fri Apr 11 01:51:28 2003 *************** *** 328,333 **** --- 328,359 ---- error = VOP_ADVLOCK(vp, (caddr_t)p->p_leader, F_UNLCK, &fl, F_POSIX); break; + case F_SANEWRLCKNO: + flg |= F_NOOVRLP; + case F_SANEWRLCK: + fl.l_type=F_WRLCK; + flg &= ~F_POSIX; + flg |= F_FLOCK; + fp->f_flag |= FHASLOCK; + error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, + &fl, flg); + break; + case F_SANERDLCKNO: + flg |= F_NOOVRLP; + case F_SANERDLCK: + fl.l_type=F_RDLCK; + flg &= ~F_POSIX; + flg |= F_FLOCK; + fp->f_flag |= FHASLOCK; + error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, + &fl, flg); + break; + case F_SANEUNLCK: + flg &= ~F_POSIX; + flg |= F_FLOCK; + error = VOP_ADVLOCK(vp, (caddr_t)fp, F_UNLCK, + &fl, F_FLOCK); + break; default: error = EINVAL; break; diff -c /sys/kern.old/kern_lockf.c /sys/kern/kern_lockf.c *** /sys/kern.old/kern_lockf.c Fri Apr 11 01:47:51 2003 --- /sys/kern/kern_lockf.c Fri Apr 11 01:48:03 2003 *************** *** 578,585 **** start = lock->lf_start; end = lock->lf_end; while (lf != NOLOCKF) { ! if (((type & SELF) && lf->lf_id != lock->lf_id) || ! ((type & OTHERS) && lf->lf_id == lock->lf_id)) { *prev = &lf->lf_next; *overlap = lf = lf->lf_next; continue; --- 578,586 ---- start = lock->lf_start; end = lock->lf_end; while (lf != NOLOCKF) { ! if ((!(lock->lf_flags & F_NOOVRLP)) && ! (((type & SELF) && lf->lf_id != lock->lf_id) || ! ((type & OTHERS) && lf->lf_id == lock->lf_id))) { *prev = &lf->lf_next; *overlap = lf = lf->lf_next; continue; diff -c /sys/sys.old/fcntl.h /sys/sys/fcntl.h *** /sys/sys.old/fcntl.h Fri Apr 11 01:48:32 2003 --- /sys/sys/fcntl.h Fri Apr 11 01:48:11 2003 *************** *** 167,176 **** --- 167,196 ---- #define F_RDLCK 1 /* shared or read lock */ #define F_UNLCK 2 /* unlock */ #define F_WRLCK 3 /* exclusive or write lock */ + #ifndef _POSIX_SOURCE + /* + * The following lock types do NOT follow the completely stupid POSIX + * fcntl() semantics. Locks are per file descriptor not per file, and + * you can request an exclusive lock on a file opened for read as well as + * a read lock on a file opened for write. + */ + #define F_SANERDLCK 4 /* sane shared or read lock */ + #define F_SANEUNLCK 5 /* unlock sane locks */ + #define F_SANEWRLCK 6 /* sane exclusive or write lock */ + + /* + * These lock types are sane locks that fail if there is ANY lock in the region + * they are locking that would conflict (ie: process conflicts with itself as + * well as other processes. + */ + #define F_SANERDLCKNO 7 /* don't up/downgrade or merge locks */ + #define F_SANEWRLCKNO 8 + #endif #ifdef _KERNEL #define F_WAIT 0x010 /* Wait until lock is granted */ #define F_FLOCK 0x020 /* Use flock(2) semantics for lock */ #define F_POSIX 0x040 /* Use POSIX semantics for lock */ + #define F_NOOVRLP 0x080 /* Don't allow overlapping locks */ #endif /* --- sanelock.patch ends here --- >Release-Note: >Audit-Trail: >Unformatted:
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200304110830.h3B8UR8J000899>