[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <CAGt9f=RffRM=gvvEFvmhWMk7PKDAp3bHhRFFpgaDWq0aV=AKyg@mail.gmail.com>
Date: Tue, 7 Jan 2025 15:53:38 +0530
From: Rohit Ner <rohitner@...gle.com>
To: Bart Van Assche <bvanassche@....org>
Cc: Can Guo <quic_cang@...cinc.com>, Bean Huo <beanhuo@...ron.com>,
Stanley Chu <stanley.chu@...iatek.com>,
"Martin K. Petersen" <martin.petersen@...cle.com>, Jaegeuk Kim <jaegeuk@...nel.org>,
Avri Altman <avri.altman@....com>, linux-scsi@...r.kernel.org,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH] scsi: ufs: core: Fix setup_xfer_req invocation
On Thu, Feb 22, 2024 at 8:56 PM Bart Van Assche <bvanassche@....org> wrote:
>
> On 2/22/24 00:27, Rohit Ner wrote:
> > Can we stick to the current approach of moving the .setup_xfer_req()
> > up, keeping in mind the following pros?
> > 1. Avoid redundant callbacks for setting up transfers
> > 2. Trim the duration for which hba->outstanding_lock is acquired unnecessarily
>
> No, we can't. The Exynos implementation of the .setup_xfer_req() callback
> is not thread-safe and relies on serialization by the caller. This patch
> breaks the Exynos driver. A better title for this patch would be "Break
> the setup_xfer_req() invocation".
It would be inaccurate to tag this patch as breaking as Can did
mention a vops use case in the hotpath for UFSHCI compliant drivers.
Having two different setup_xfer_req functions for mcq/lsdb mode just
because a particular vendor driver relies on serialization by the
caller defeats the purpose of having vops as the core logic is still
burdened with quirks.
>
> Bart.
>
Rohit.
Powered by blists - more mailing lists