From: "Nicholas A. Bellinger" Subject: Re: [PATCH 01/12] libiscsi: Convert to host_lock less w/ interrupts disabled internally Date: Sun, 19 Dec 2010 17:22:06 -0800 Message-ID: <1292808126.20840.33.camel@haakon2.linux-iscsi.org> References: <1292793727-31957-1-git-send-email-nab@linux-iscsi.org> <1292793727-31957-2-git-send-email-nab@linux-iscsi.org> <20101219233835.GJ1263@parisc-linux.org> <1292807703.20840.29.camel@haakon2.linux-iscsi.org> Mime-Version: 1.0 Content-Type: text/plain Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <1292807703.20840.29.camel@haakon2.linux-iscsi.org> Sender: linux-kernel-owner@vger.kernel.org To: Matthew Wilcox Cc: linux-scsi , linux-kernel , James Bottomley , Jeff Garzik , Christoph Hellwig , FUJITA Tomonori , Hannes Reinecke , Mike Christie List-Id: linux-scsi@vger.kernel.org On Sun, 2010-12-19 at 17:15 -0800, Nicholas A. Bellinger wrote: > On Sun, 2010-12-19 at 16:38 -0700, Matthew Wilcox wrote: > > On Sun, Dec 19, 2010 at 01:21:56PM -0800, Nicholas A. Bellinger wrote: > > > This patch changes iscsi_queuecommand_lck() to a host_lock less > > > iscsi_queuecommand() that will internally disable interrupts using > > > session->lock and drop the now legacy host_lock unlock. > > > > I think this patch is buggy. Before, iscsi_queuecommand_lck is called with > > interrupts disabled. Now, it's called with interrupts disabled. Elsewhere, > > session->lock is acquired with bh's disabled: > > > > void iscsi_put_task(struct iscsi_task *task) > > { > > struct iscsi_session *session = task->conn->session; > > > > spin_lock_bh(&session->lock); > > > > So I think you need to convert the > > spin_lock(&session->lock); > > in iscsi_queuecommand_lck to at least a spin_lock_bh, and possibly > > a spin_lock_irq -- it's not clear to me whether it needs to exclude > > against interrupt context, or only BH context. In some places, it's > > taken with spin_lock_irqsave(). > > > > I'd play it safe and use _irq, but someone more confident with this code > > might choose to only use _bh. > > > > Hmmm, indeed.. I must have dropped this recently as the last patch for > libiscsi posted below still has been converted to spin_lock_irq() for > session->lock here: > > http://marc.info/?l=linux-scsi&m=128952081412461&w=2 > > Converting back to spin_lock_irq() usage for the moment.. Thanks alot > of noticing this.. > > MikeC and Hannes, do you think this is safe to use spin_lock_bh() as > well..? > Actually sorry, Mike Christie did already make a clarification on this subject here: http://marc.info/?l=linux-scsi&m=129010439421506&w=2 I had originally thought the same that session->lock should be using some flavour of spin_lock_irq*() as well, but apparently this is not the case for libiscsi. Best Regards, --nab