Bug 75571
| Summary: | man page for sx(9) is misleading | ||
|---|---|---|---|
| Product: | Documentation | Reporter: | Darern Reed <darrenr> |
| Component: | Books & Articles | Assignee: | Giorgos Keramidas <keramida> |
| Status: | Closed FIXED | ||
| Severity: | Affects Only Me | ||
| Priority: | Normal | ||
| Version: | Latest | ||
| Hardware: | Any | ||
| OS: | Any | ||
|
Description
Darern Reed
2004-12-28 14:00:47 UTC
On 2004-12-28 13:55, Darren Reed <darrenr@FreeBSD.ORG> wrote: > > According to discussion on freebsd mailing lists, it is not possible > to hold an sx lock when you want a mtx lock. This should be documented. As far as I can tell, by looking at kern_sx.c and sys/sx.h, this is because the sx lock initialization uses an mtxpool for the mutex used to serialize access to the internal sx lock data. Leaf locks may be used in operations that msleep() but there can be only one of them in each lock path and no other lock can be obtained after them. This is sort of implied by the SEE ALSO reference of mtx_pool(9), but we should probably state it explicitly in CONTEXT. %%% Index: sx.9 =================================================================== RCS file: /home/ncvs/src/share/man/man9/sx.9,v retrieving revision 1.29 diff -u -r1.29 sx.9 --- sx.9 11 Jul 2004 16:08:25 -0000 1.29 +++ sx.9 28 Dec 2004 23:28:22 -0000 @@ -196,6 +196,11 @@ A thread may hold a shared or exclusive lock on an .Nm lock while sleeping. +The +.Nm +locks are implemented using +.Xr mtxpool 9 +shared leaf locks, so they should always be the last lock obtained. .Sh SEE ALSO .Xr condvar 9 , .Xr mtx_pool 9 , %%% What you wrote is not right. You cannot acquire an sx(9) lock while holding an mtx(9) lock. Darren On 2004-12-29 13:34, John Baldwin <jhb@freebsd.org> wrote: >On Wednesday 29 December 2004 03:40 am, Giorgos Keramidas wrote: >>On 2004-12-28 13:55, Darren Reed <darrenr@FreeBSD.ORG> wrote: >>> According to discussion on freebsd mailing lists, it is not possible >>> to hold an sx lock when you want a mtx lock. This should be documented. >> >> As far as I can tell, by looking at kern_sx.c and sys/sx.h, this is >> because the sx lock initialization uses an mtxpool for the mutex used to >> serialize access to the internal sx lock data. [...] > > The reason is largely because they can be held across a sleep, e.g.: > > sx_xlock(&foo->sx); > bar = malloc(sizeof(*bar), M_FOO, M_WAITOK); > TAILQ_INSERT_TAIL(&foo->head, bar, link); > sx_xunlock(&foo->sx); > > This is intentional and that is what should be documented. Basically, it > needs a paragraph to the effect of: > > .Pp > An > .Nm > lock may not be acquired while holding a mutex. > Since threads are allowed to sleep while holding an > .NM > lock, > a thread that acquired a mutex and then blocked on an > .Nm > lock would end up sleeping while holding a mutex which is not allowed. Nice :-) Thanks for putting this in words. Should I commit this? %%% Index: sx.9 =================================================================== RCS file: /home/ncvs/src/share/man/man9/sx.9,v retrieving revision 1.29 diff -u -5 -r1.29 sx.9 --- sx.9 11 Jul 2004 16:08:25 -0000 1.29 +++ sx.9 3 Jan 2005 23:08:40 -0000 @@ -194,10 +194,19 @@ attempting to do so will result in deadlock. .Sh CONTEXT A thread may hold a shared or exclusive lock on an .Nm lock while sleeping. +As a result, an +.Nm +lock may not be acquired while holding a mutex. +Since threads are allowed to sleep while holding an +.Nm +lock, +a thread that acquired a mutex and then blocked on an +.Nm +lock would end up sleeping while holding a mutex which is not allowed. .Sh SEE ALSO .Xr condvar 9 , .Xr mtx_pool 9 , .Xr mutex 9 , .Xr panic 9 , %%% State Changed From-To: open->patched Committed the last explanation of John to HEAD. Responsible Changed From-To: freebsd-doc->keramida I'll handle the MFC of this in a few days. State Changed From-To: patched->closed The sx(9) manpage in RELENG_5 is now in sync with HEAD. |