From: Akhil R <akhilrajeev@nvidia.com>
To: Dmitry Osipenko <digetx@gmail.com>,
"dan.j.williams@intel.com" <dan.j.williams@intel.com>,
"devicetree@vger.kernel.org" <devicetree@vger.kernel.org>,
"dmaengine@vger.kernel.org" <dmaengine@vger.kernel.org>,
Jonathan Hunter <jonathanh@nvidia.com>,
Krishna Yarlagadda <kyarlagadda@nvidia.com>,
Laxman Dewangan <ldewangan@nvidia.com>,
"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
"linux-tegra@vger.kernel.org" <linux-tegra@vger.kernel.org>,
"p.zabel@pengutronix.de" <p.zabel@pengutronix.de>,
Rajesh Gumasta <rgumasta@nvidia.com>,
"robh+dt@kernel.org" <robh+dt@kernel.org>,
"thierry.reding@gmail.com" <thierry.reding@gmail.com>,
"vkoul@kernel.org" <vkoul@kernel.org>
Cc: Pavan Kunapuli <pkunapuli@nvidia.com>
Subject: RE: [PATCH v16 2/4] dmaengine: tegra: Add tegra gpcdma driver
Date: Thu, 27 Jan 2022 16:29:36 +0000 [thread overview]
Message-ID: <DM5PR12MB1850677D5F07065BDC08EE24C0219@DM5PR12MB1850.namprd12.prod.outlook.com> (raw)
In-Reply-To: <80a90a28-d9b8-1ba6-b79e-07fd49cc92b7@gmail.com>
> 23.01.2022 19:49, Akhil R пишет:
> >> 21.01.2022 19:24, Akhil R пишет:
> >>>>>>>>> +static int tegra_dma_terminate_all(struct dma_chan *dc) {
> >>>>>>>>> + struct tegra_dma_channel *tdc = to_tegra_dma_chan(dc);
> >>>>>>>>> + unsigned long flags;
> >>>>>>>>> + LIST_HEAD(head);
> >>>>>>>>> + int err;
> >>>>>>>>> +
> >>>>>>>>> + if (tdc->dma_desc) {
> >>>>>>>>
> >>>>>>>> Needs locking protection against racing with the interrupt handler.
> >>>>>>> tegra_dma_stop_client() waits for the in-flight transfer to
> >>>>>>> complete and prevents any additional transfer to start.
> >>>>>>> Wouldn't it manage the race? Do you see any potential issue there?
> >>>>>>
> >>>>>> You should consider interrupt handler like a process running in a
> >>>>>> parallel thread. The interrupt handler sets tdc->dma_desc to NULL,
> >>>>>> hence you'll get NULL dereference in tegra_dma_stop_client().
> >>>>>
> >>>>> Is it better if I remove the below part from tegra_dma_stop_client()
> >>>>> so that dma_desc is not accessed at all?
> >>>>>
> >>>>> + wcount = tdc_read(tdc, TEGRA_GPCDMA_CHAN_XFER_COUNT);
> >>>>> + tdc->dma_desc->bytes_transferred +=
> >>>>> + tdc->dma_desc->bytes_requested - (wcount * 4);
> >>>>>
> >>>>> Because I don't see a point in updating the value there. dma_desc is
> >>>>> set to NULL in the next step in terminate_all() anyway.
> >>>>
> >>>> That isn't going help you much because you also can't release DMA
> >>>> descriptor while interrupt handler still may be running and using
> >>>> that descriptor.
> >>>
> >>> Does the below functions look good to resolve the issue, provided
> >>> tegra_dma_stop_client() doesn't access dma_desc?
> >>
> >> Stop shall not race with the start.
> >>
> >>> +static int tegra_dma_terminate_all(struct dma_chan *dc) {
> >>> + struct tegra_dma_channel *tdc = to_tegra_dma_chan(dc);
> >>> + unsigned long flags;
> >>> + LIST_HEAD(head);
> >>> + int err;
> >>> +
> >>> + err = tegra_dma_stop_client(tdc);
> >>> + if (err)
> >>> + return err;
> >>> +
> >>> + tegra_dma_stop(tdc);
> >>> +
> >>> + spin_lock_irqsave(&tdc->vc.lock, flags);
> >>> + tegra_dma_sid_free(tdc);
> >>> + tdc->dma_desc = NULL;
> >>> +
> >>> + vchan_get_all_descriptors(&tdc->vc, &head);
> >>> + spin_unlock_irqrestore(&tdc->vc.lock, flags);
> >>> +
> >>> + vchan_dma_desc_free_list(&tdc->vc, &head);
> >>> +
> >>> + return 0;
> >>> +}
> >>>
> >>> +static irqreturn_t tegra_dma_isr(int irq, void *dev_id) {
> >>> + struct tegra_dma_channel *tdc = dev_id;
> >>> + struct tegra_dma_desc *dma_desc = tdc->dma_desc;
> >>> + struct tegra_dma_sg_req *sg_req;
> >>> + u32 status;
> >>> +
> >>> + /* Check channel error status register */
> >>> + status = tdc_read(tdc, TEGRA_GPCDMA_CHAN_ERR_STATUS);
> >>> + if (status) {
> >>> + tegra_dma_chan_decode_error(tdc, status);
> >>> + tegra_dma_dump_chan_regs(tdc);
> >>> + tdc_write(tdc, TEGRA_GPCDMA_CHAN_ERR_STATUS, 0xFFFFFFFF);
> >>> + }
> >>> +
> >>> + status = tdc_read(tdc, TEGRA_GPCDMA_CHAN_STATUS);
> >>> + if (!(status & TEGRA_GPCDMA_STATUS_ISE_EOC))
> >>> + return IRQ_HANDLED;
> >>> +
> >>> + tdc_write(tdc, TEGRA_GPCDMA_CHAN_STATUS,
> >>> + TEGRA_GPCDMA_STATUS_ISE_EOC);
> >>> +
> >>> + spin_lock(&tdc->vc.lock);
> >>> + if (!dma_desc)
> >> All checks and assignments must be done inside of critical section.
> >
> > Okay. So, the lock should be held throughout the function.
> > Do you think tegra_dma_pause should also hold a lock
> > and remove irq_synchronize? That function also writes
> > to CSR register.
>
> Interrupt handler shall not unpause channel in a case of race condition,
> it should handle completed transfer and check whether channel is paused
> before issuing next transfer.
> So yes, pause also needs a lock.
Thanks for the points. I collected more details on this scenario and found that
DMA can only be paused if there is an in-flight transfer (i.e no interrupt will be
pending). And there cannot be an interrupt after DMA is paused. Hence, there
is no case of ISR unpausing a paused transfer. So, I guess, the check or a lock
might not be required there.
One thing I am thinking of is to wait for DMA to be busy so that it can be paused.
I would add these changes and send out a new patch.
Thanks,
Akhil
next prev parent reply other threads:[~2022-01-27 16:29 UTC|newest]
Thread overview: 16+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-01-10 16:05 [PATCH v16 0/4] Add NVIDIA Tegra GPC-DMA driver Akhil R
2022-01-10 16:05 ` [PATCH v16 1/4] dt-bindings: dmaengine: Add doc for tegra gpcdma Akhil R
2022-01-10 16:05 ` [PATCH v16 2/4] dmaengine: tegra: Add tegra gpcdma driver Akhil R
2022-01-14 11:57 ` Dmitry Osipenko
2022-01-17 7:02 ` Akhil R
2022-01-17 15:41 ` Dmitry Osipenko
2022-01-18 5:36 ` Akhil R
2022-01-18 13:48 ` Dmitry Osipenko
2022-01-21 16:24 ` Akhil R
2022-01-22 18:31 ` Dmitry Osipenko
2022-01-23 16:49 ` Akhil R
2022-01-23 21:07 ` Dmitry Osipenko
2022-01-27 16:29 ` Akhil R [this message]
2022-01-27 23:15 ` Dmitry Osipenko
2022-01-10 16:05 ` [PATCH v16 3/4] arm64: defconfig: tegra: Enable GPCDMA Akhil R
2022-01-10 16:05 ` [PATCH v16 4/4] arm64: tegra: Add GPCDMA node for tegra186 and tegra194 Akhil R
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=DM5PR12MB1850677D5F07065BDC08EE24C0219@DM5PR12MB1850.namprd12.prod.outlook.com \
--to=akhilrajeev@nvidia.com \
--cc=dan.j.williams@intel.com \
--cc=devicetree@vger.kernel.org \
--cc=digetx@gmail.com \
--cc=dmaengine@vger.kernel.org \
--cc=jonathanh@nvidia.com \
--cc=kyarlagadda@nvidia.com \
--cc=ldewangan@nvidia.com \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-tegra@vger.kernel.org \
--cc=p.zabel@pengutronix.de \
--cc=pkunapuli@nvidia.com \
--cc=rgumasta@nvidia.com \
--cc=robh+dt@kernel.org \
--cc=thierry.reding@gmail.com \
--cc=vkoul@kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).