From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from 74-93-104-97-Washington.hfc.comcastbusiness.net ([74.93.104.97]:47912 "EHLO sunset.davemloft.net" rhost-flags-OK-FAIL-OK-OK) by vger.kernel.org with ESMTP id S1754756AbXKGEee (ORCPT ); Tue, 6 Nov 2007 23:34:34 -0500 Date: Tue, 06 Nov 2007 20:34:33 -0800 (PST) Message-Id: <20071106.203433.144763156.davem@davemloft.net> Subject: Fix for sparc64 cpu hangs. From: David Miller Mime-Version: 1.0 Content-Type: Text/Plain; charset=us-ascii Content-Transfer-Encoding: 7bit Sender: linux-arch-owner@vger.kernel.org To: torvalds@linux-foundation.org Cc: linux-kernel@vger.kernel.org, sparclinux@vger.kernel.org, linux-arch@vger.kernel.org, bernd@bzed.de, joy@entuzijast.net, fabbione@ubuntu.com, arnd@arndb.de List-ID: [ Bernd, Josip, and Fabio, I think I finally nailed this cpu hang bug we were all seeing on sparc64. ] [FUTEX]: Fix address computation in compat code. compat_exit_robust_list() computes a pointer to the futex entry in userspace as follows: (void __user *)entry + futex_offset 'entry' is a 'struct robust_list __user *', and 'futex_offset' is a 'compat_long_t' (typically a 's32'). Things explode if the 32-bit sign bit is set in futex_offset. Type promotion sign extends futex_offset to a 64-bit value before adding it to 'entry'. This triggered a problem on sparc64 running 32-bit applications which would lock up a cpu looping forever in the fault handling for the userspace load in handle_futex_death(). Compat userspace runs with address masking (wherein the cpu zeros out the top 32-bits of every effective address given to a memory operation instruction) so the sparc64 fault handler accounts for this by zero'ing out the top 32-bits of the fault address too. Since the kernel properly uses the compat_uptr interfaces, kernel side accesses to compat userspace work too since they will only use addresses with the top 32-bits clear. Because of this compat futex layer bug we get into the following loop when executing the get_user() load near the top of handle_futex_death(): 1) load from address '0xfffffffff7f16bd8', FAULT 2) fault handler clears upper 32-bits, processes fault for address '0xf7f16bd8' which succeeds 3) goto #1 I want to thank Bernd Zeimetz, Josip Rodin, and Fabio Massimo Di Nitto for their tireless efforts helping me track down this bug. Signed-off-by: David S. Miller diff --git a/kernel/futex_compat.c b/kernel/futex_compat.c index 00b5726..8089e7e 100644 --- a/kernel/futex_compat.c +++ b/kernel/futex_compat.c @@ -76,11 +76,16 @@ void compat_exit_robust_list(struct task_struct *curr) * A pending lock might already be on the list, so * dont process it twice: */ - if (entry != pending) - if (handle_futex_death((void __user *)entry + futex_offset, - curr, pi)) - return; + if (entry != pending) { + void __user *uaddr; + compat_uptr_t base; + + base = ptr_to_compat(entry); + uaddr = compat_ptr(base + futex_offset); + if (handle_futex_death(uaddr, curr, pi)) + return; + } if (rc) return; uentry = next_uentry;