Messages in this thread |  | | Subject | Re: [PATCH 1/2] mm: Remove use of ALLOW_RETRY when RETRY_NOWAIT is set | | From | Steven Rostedt <> | | Date | Wed, 29 Jun 2011 08:43:59 -0400 |
| |
On Wed, 2011-06-29 at 02:38 -0700, Michel Lespinasse wrote: > On Tue, Jun 28, 2011 at 9:47 AM, Steven Rostedt <rostedt@goodmis.org> wrote: > > From: Steven Rostedt <srostedt@redhat.com> > > > > The only user of FAULT_FLAG_RETRY_NOWAIT also sets the > > FAULT_FLAG_ALLOW_RETRY flag. This makes the check in the > > __lock_page_or_retry redundant as it checks the RETRY_NOWAIT > > just after checking ALLOW_RETRY and then returns if it is > > set. The FAULT_FLAG_ALLOW_RETRY does not make any other > > difference in this path. > > > > Setting both and then ignoring one is quite confusing, > > especially since this code has very subtle locking issues > > when it comes to the mmap_sem. > > > > Only set the RETRY_WAIT flag and have that do the necessary > > work instead of confusing reviewers of this code by setting > > ALLOW_RETRY and not releasing the mmap_sem. > > > > --- a/include/linux/mm.h > > +++ b/include/linux/mm.h > > @@ -151,8 +151,8 @@ extern pgprot_t protection_map[16]; > > #define FAULT_FLAG_WRITE 0x01 /* Fault was a write access */ > > #define FAULT_FLAG_NONLINEAR 0x02 /* Fault was via a nonlinear mapping */ > > #define FAULT_FLAG_MKWRITE 0x04 /* Fault was mkwrite of existing pte */ > > -#define FAULT_FLAG_ALLOW_RETRY 0x08 /* Retry fault if blocking */ > > -#define FAULT_FLAG_RETRY_NOWAIT 0x10 /* Don't drop mmap_sem and wait when retrying */ > > +#define FAULT_FLAG_ALLOW_RETRY 0x08 /* Retry fault if blocking (drops mmap_sem) */ > > +#define FAULT_FLAG_RETRY_NOWAIT 0x10 /* Wait when retrying (don't drop mmap_sem) */ > > You want to say "DONT wait when retrying" here...
Yeah, I thought that was weird in the comment. Heh, rereading the original comment I see it meant "don't drop nor wait" where I thought it meant "Don't drop mmap_sem and then wait when retrying". That needs to be fixed :)
> > Also - you argued higher up that having both flags set at once is > confusing, but I find it equally confusing to pass a flag to specify > you don't want to wait on retry if the flag that allows retry is not > set.
Hmm, I understand. The main issue I have with all these flags is the locking. I really don't care about the semantics for retry and such. What I care about is this dropping of the mmap_sem. It is very confusing to understand when it is dropped or not.
> I think the confusion comes from the way the nowait semantics got > bolted on the retry code for virtualization, even though (if I > understand the virtualization use case correctly) they dont actually > want to retry there, they just want to give up without blocking.
Right.
> > > Would the following proposal make more sense to you ? > > FAULT_FLAG_ALLOW_ASYNC: allow returning a VM_FAULT_ASYNC error code if > the page can't be obtained immediately (major fault). > FAULT_FLAG_ASYNC_WAIT: before returning VM_FAULT_ASYNC, drop the > mmap_sem and wait for major fault to complete.
Ug, I think that's worse. ASYNC to me is something that you enable and it will get back to you when done. Think AIO.
How about...
FAULT_FLAG_ALLOW_DROP_MMAP_SEM ?
and it returns VM_FAULT_DROPPED_MMAP_SEM if it drops it.
We could also have another flag:
FAULT_FLAG_NOWAIT_ON_PAGE
and it can return VM_FAULT_NOWAIT_PAGE if it failed to get the page lock and returned.
This to me documents exactly what is happening, and doesn't confuse kernel developers when they see handle_mm_fault() called, and then the code not releasing the mmap_sem.
If I saw in arch/x86/mm/fault.c:
if (fault & VM_FAULT_DROPPED_MMAP_SEM) { /* Clear FAULT_FLAG_ALLOW_DROPPED_MMAP_SEM to avoid any risk * of starvation. */ flags &= ~FAULT_FLAG_ALLOW_DROPPED_MMAP_SEM; goto retry; }
There would have never been this confusion about why we are doing:
retry: down_read(&mm->mmap_sem);
Without ever doing a up_read(&mm->mmap_sem);
> > existing uses of FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT > become FAULT_FLAG_ASYNC > existing uses of FAULT_FLAG_ALLOW_RETRY alone become FAULT_FLAG_ASYNC > | FAULT_FLAG_ASYNC_WAIT > existing uses of VM_FAULT_RETRY become VM_FAULT_ASYNC > > This may also help your documentation proposal since the flags would > now work together rather than having one be an exception to the other.
I'm thinking they should still be separate. Just because they do things differently with the mmap_sem.
-- Steve
|  |