Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Sun, 26 May 2019 13:17:03 -0700 (PDT)
From:   Hugh Dickins <hughd@google.com>
To:     Sebastian Andrzej Siewior <bigeasy@linutronix.de>
cc:     Hugh Dickins <hughd@google.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Mike Rapoport <rppt@linux.ibm.com>,
        Andrea Arcangeli <aarcange@redhat.com>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org, Borislav Petkov <bp@suse.de>,
        Pavel Machek <pavel@ucw.cz>,
        Dave Hansen <dave.hansen@linux.intel.com>
Subject: Re: [PATCH] mm/gup: continue VM_FAULT_RETRY processing event for
 pre-faults
In-Reply-To: <20190526193651.spvm2vtrwxlhsjrv@linutronix.de>
Message-ID: <alpine.LSU.2.11.1905261250590.2394@eggly.anvils>
References: <1557844195-18882-1-git-send-email-rppt@linux.ibm.com> <20190522122113.a2edc8aba32f0fad189bae21@linux-foundation.org> <20190522194322.5k52docwgp5zkdcj@linutronix.de> <alpine.LSU.2.11.1905241429460.1141@eggly.anvils> <20190525084546.fap2wkefepeia22f@linutronix.de>
 <alpine.LSU.2.11.1905251033230.1112@eggly.anvils> <20190526193651.spvm2vtrwxlhsjrv@linutronix.de>
User-Agent: Alpine 2.11 (LSU 23 2013-08-11)
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Sun, 26 May 2019, Sebastian Andrzej Siewior wrote:
> 
> Okay. The GUP functions are not properly documented for my taste. There
> is no indication whether or not the mm_sem has to be acquired prior
> invoking it. Following the call chain of get_user_pages() I ended up in
> __get_user_pages_locked() `locked = NULL' indicated that mm_sem is no
> acquired and then I saw this:
> |                 if (!locked)
> |                         /* VM_FAULT_RETRY couldn't trigger, bypass */
> |                         return ret;
> 
> kind of suggesting that it is okay to invoke it without holding the
> mm_sem prefault. It passed a few tests and then
> 	https://lkml.kernel.org/r/1556657902.6132.13.camel@lca.pw
> 
> happened. After that, I switched to the locked variant and the problem
> disappeared (also I noticed that MPX code is invoked within ->mmap()).

Ah, I wasn't following what happened here while it was in linux-next.

I certainly agree that all the variants are very confusing, and the
matter of mmap_sem particularly so. Because it used to be a simple
rule that you needed to hold mmap_sem to call get_user_pages(); but
that can be so contended that get_user_pages_fast(), and VM_FAULT_RETRY,
optimizations came in, then interfaces like get_user_pages_locked() and
get_user_pages_unlocked().

I think when you say that you "switched to the locked variant", you're
writing of when you switched to using get_user_pages_unlocked(), which
takes and releases the mmap_sem itself: yes, using get_user_pages()
without holding mmap_sem would have been open to serious errors.

The documentation in comments above get_user_pages_locked() and
get_user_pages_unlocked() looks rather good to me, actually.

But this is all good reason to use the less challenging
fault_in_pages_writable() instead. (And also saves all those GUP
developers, who from time to time have to search through all users
of GUP in one form or another, to make this or that improvement,
from having to visit arch/x86/kernel/fpu/signal.c: they will
quietly thank you.)

Hugh