From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([209.51.188.92]:54370) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1gxBym-0001zL-2M for qemu-devel@nongnu.org; Fri, 22 Feb 2019 09:39:10 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1gxByk-0006Zg-Hx for qemu-devel@nongnu.org; Fri, 22 Feb 2019 09:39:07 -0500 Received: from mx1.redhat.com ([209.132.183.28]:50144) by eggs.gnu.org with esmtps (TLS1.0:DHE_RSA_AES_256_CBC_SHA1:32) (Exim 4.71) (envelope-from ) id 1gxByk-0006OZ-5d for qemu-devel@nongnu.org; Fri, 22 Feb 2019 09:39:06 -0500 References: <28e6b4ed-9afd-3a79-6267-86c7385c23ce@redhat.com> From: Paolo Bonzini Message-ID: <01ad59eb-6548-27cd-f192-91819e929dbc@redhat.com> Date: Fri, 22 Feb 2019 15:38:44 +0100 MIME-Version: 1.0 In-Reply-To: Content-Type: text/plain; charset=utf-8 Content-Language: en-US Content-Transfer-Encoding: quoted-printable Subject: Re: [Qemu-devel] [Qemu-block] Guest unresponsive after Virtqueue size exceeded error List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Stefan Hajnoczi , =?UTF-8?Q?Fernando_Casas_Sch=c3=b6ssow?= Cc: Natanael Copa , Richard Henderson , qemu-devel On 22/02/19 15:04, Stefan Hajnoczi wrote: > On Fri, Feb 22, 2019 at 12:57 PM Fernando Casas Sch=C3=B6ssow > wrote: >=20 > I have CCed Natanael Copa, qemu package maintainer in Alpine Linux. >=20 > Fernando: Can you confirm that the bug occurs with an unmodified > Alpine Linux qemu binary? I can confirm that the memcpy calls are there, at least. Paolo > Richard: Commit 7db2145a6826b14efceb8dd64bfe6ad8647072eb ("bswap: Add > host endian unaligned access functions") introduced the unaligned > memory access functions in question here. Please see below for > details on the bug - basically QEMU code assumes they are atomic, but > that is not guaranteed :(. Any ideas for how to fix this? >=20 > Natanael: It seems likely that the qemu package in Alpine Linux > suffers from a compilation issue resulting in a broken QEMU. It may > be necessary to leave the compiler optimization flag alone in APKBUILD > to work around this problem. >=20 > Here are the details. QEMU relies on the compiler turning > memcpy(&dst, &src, 2) turning into a load instruction in > include/qemu/bswap.h:lduw_he_p() (explanation below): >=20 > /* Any compiler worth its salt will turn these memcpy into native unali= gned > operations. Thus we don't need to play games with packed attributes= , or > inline byte-by-byte stores. */ >=20 > static inline int lduw_he_p(const void *ptr) > { > uint16_t r; > memcpy(&r, ptr, sizeof(r)); > return r; > } >=20 > Here is the disassembly snippet of virtqueue_pop() from Fedora 29 that > shows the load instruction: >=20 > 398166: 0f b7 42 02 movzwl 0x2(%rdx),%eax > 39816a: 66 89 43 32 mov %ax,0x32(%rbx) >=20 > Here is the instruction sequence in the Alpine Linux binary: >=20 > 455562: ba 02 00 00 00 mov $0x2,%edx > 455567: e8 74 24 f3 ff callq 3879e0 > 45556c: 0f b7 44 24 42 movzwl 0x42(%rsp),%eax > 455571: 66 41 89 47 32 mov %ax,0x32(%r15) >=20 > It's calling memcpy instead of using a load instruction. >=20 > Fernando found that QEMU's virtqueue_pop() function sees bogus values > when loading a 16-bit guest RAM location. Paolo figured out that the > bogus value can be produced by memcpy() when another thread is > updating the 16-bit memory location simultaneously. This is a race > condition between one thread loading the 16-bit value and another > thread storing it (in this case a guest vcpu thread). Sometimes > memcpy() may load one old byte and one new byte, resulting in a bogus > value. >=20 > The symptom that Fernando experienced is a "Virtqueue size exceeded" > error message from QEMU and then the virtio-blk or virtio-scsi device > stops working. This issue potentially affects other device emulation > code in QEMU as well, not just virtio devices. >=20 > For the time being, I suggest tweaking the APKBUILD so the memcpy() is > not generated. Hopefully QEMU can make the code more portable in the > future so the compiler always does the expected thing, but this may > not be easily possible. >=20 > Stefan >=20