From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail137.messagelabs.com (mail137.messagelabs.com [216.82.249.19]) by kanga.kvack.org (Postfix) with SMTP id 12A976B00E8 for ; Tue, 28 Jun 2011 13:10:13 -0400 (EDT) Date: Tue, 28 Jun 2011 20:09:53 +0300 From: Gleb Natapov Subject: Re: [PATCH 2/2] mm: Document handle_mm_fault() Message-ID: <20110628170953.GA3482@redhat.com> References: <20110628164750.281686775@goodmis.org> <20110628165303.010143380@goodmis.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20110628165303.010143380@goodmis.org> Sender: owner-linux-mm@kvack.org List-ID: To: Steven Rostedt Cc: linux-kernel@vger.kernel.org, linux-mm@kvack.org, Andrew Morton , Linus Torvalds , Russell King , Thomas Gleixner , Peter Zijlstra , Hugh Dickins , Rik van Riel , Michel Lespinasse , Avi Kivity , Marcelo Tosatti On Tue, Jun 28, 2011 at 12:47:52PM -0400, Steven Rostedt wrote: > From: Steven Rostedt > > The function handle_mm_fault() is long overdue for comments. > Adding a kernel doc header for the function and explaining the subtle > use of the flags with respect to mmap_sem will prove useful in the > future when others work with this code. > > Russell King noticed that the code in arch/x86/mm/fault.c looked > buggy as the do_page_fault() code would grab the mmap_sem multiple > times without letting it go. But it only did this when the > handle_mm_fault() would return VM_FAULT_RETRY. > > Examining the code and reading the git change logs, it was discovered > that commit d065bd810b6deb67d4897a14bfe21f8eb526ba99 > mm: retry page fault when blocking on disk transfer > added code to remove contention with the mmap_sem when the page_lock > was being held for IO. As waiting on IO holding the mmap_sem can > cause lots of contention between threads. The flag > FAULT_FLAG_ALLOW_RETRY was added to let handle_mm_fault() know > that it can safely release the mmap_sem. > > Adding to the confusion here with handle_mm_fault(), another > commit 318b275fbca1ab9ec0862de71420e0e92c3d1aa7 > mm: allow GUP to fail instead of waiting on a page > was added that would not release the mmap_sem, even if > FAULT_FLAG_ALLOW_RETRY was set and the page_lock was not taken > and VM_FAULT_RETRY was returned, if FAULT_FLAGS_RETRY_NOWAIT was > set. > > All of this is poorly documented and makes using or modifying > handle_mm_fault() fragile. Documenting all of these subtle changes > at the head of handle_mm_fault() should help future developers > understand what is happening. > > Reported-by: Russell King > Cc: Gleb Natapov > Cc: Linus Torvalds > Cc: Hugh Dickins > Cc: Rik van Riel > Cc: Michel Lespinasse > Cc: Avi Kivity > Cc: Marcelo Tosatti > Cc: Andrew Morton > Signed-off-by: Steven Rostedt > --- > mm/memory.c | 22 ++++++++++++++++++++-- > 1 files changed, 20 insertions(+), 2 deletions(-) > > diff --git a/mm/memory.c b/mm/memory.c > index 5371b5e..3cf30f6 100644 > --- a/mm/memory.c > +++ b/mm/memory.c > @@ -3378,8 +3378,26 @@ unlock: > return 0; > } > > -/* > - * By the time we get here, we already hold the mm semaphore > +/** > + * handle_mm_fault - main routine for handling page faults > + * @mm: the mm_struct of the target address space > + * @vma: vm_area_struct holding the applicable pages > + * @address: the address that took the fault > + * @flags: flags modifying lookup behaviour > + * > + * Must have @mm->mmap_sem held. > + * > + * Note: if @flags has FAULT_FLAG_ALLOW_RETRY set then the mmap_sem > + * may be released if it failed to arquire the page_lock. If the > + * mmap_sem is released then it will return VM_FAULT_RETRY set. > + * This is to keep the time mmap_sem is held when the page_lock > + * is taken for IO. > + * Exception: If FAULT_FLAG_RETRY_NOWAIT is set, then it will > + * not release the mmap_sem, but will still return VM_FAULT_RETRY > + * if it failed to acquire the page_lock. I wouldn't describe it like that. It tells handle_mm_fault() to start IO if needed, but do not wait for its completion. > + * This is for helping virtualization. See get_user_page_nowait(). The virtialization is the only user right now, but I wouldn't describe this flag as virtialization specific. Why should we put this in the comment? The comment will become outdated when other users arise and meanwhile a simple grep will reveal the above information anyway. > + * > + * Returns status flags based on the VM_FAULT_* flags in > */ > int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, > unsigned long address, unsigned int flags) > -- > 1.7.5.4 > -- Gleb. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: email@kvack.org