lkml.org 
[lkml]   [2022]   [Jan]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: [PATCH v5 25/32] iommu/mtk: Migrate to aggregate driver
From
Date
On Tue, 2022-01-11 at 16:27 -0800, Stephen Boyd wrote:
> Quoting Yong Wu (2022-01-11 04:22:23)
> > Hi Stephen,
> >
> > Thanks for helping update here.
> >
> > On Thu, 2022-01-06 at 13:45 -0800, Stephen Boyd wrote:
> > > Use an aggregate driver instead of component ops so that we can
> > > get
> > > proper driver probe ordering of the aggregate device with respect
> > > to
> > > all
> > > the component devices that make up the aggregate device.
> > >
> > > Cc: Yong Wu <yong.wu@mediatek.com>
> > > Cc: Joerg Roedel <joro@8bytes.org>
> > > Cc: Will Deacon <will@kernel.org>
> > > Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
> > > Cc: "Rafael J. Wysocki" <rafael@kernel.org>
> > > Cc: Rob Clark <robdclark@gmail.com>
> > > Cc: Russell King <rmk+kernel@arm.linux.org.uk>
> > > Cc: Saravana Kannan <saravanak@google.com>
> > > Signed-off-by: Stephen Boyd <swboyd@chromium.org>
> >
> > When I test this on mt8195 which have two IOMMU HWs(calling
> > component_aggregate_regsiter twice), it will abort like this. Then
> > what
> > should we do if we have two instances?
> >
>
> Thanks for testing it out. We can't register the struct driver more
> than
> once but this driver is calling the component_aggregate_register()
> function from the driver probe and there are two devices bound to the
> mtk-iommu driver so we try to register it more than once. Sigh!
>
> I see a couple options. One is to do a deep copy of the driver
> structure
> and change the driver name. Then it's a one to one relationship
> between
> device and driver. That's not very great because it leaves around
> junk
> so it should probably be avoided.
>
> Another option is to reference count the driver registration calls
> when
> component_aggregate_register() is called multiple times. Then we
> would
> only register the driver once and keep it pinned until the last
> unregister call is made, but still remove devices that are created
> for
> the match table.
>
> Can you try the attached patch? It is based on the next version of
> this
> patch series so the include part of the patch may not apply cleanly.
>
> ---8<---
> diff --git a/drivers/base/component.c b/drivers/base/component.c
> index 64ad7478c67a..97f253a41bdf 100644
> --- a/drivers/base/component.c
> +++ b/drivers/base/component.c
> @@ -492,15 +492,30 @@ static struct aggregate_device
> *__aggregate_find(struct device *parent)
> return dev ? to_aggregate_device(dev) : NULL;
> }
>
> +static DEFINE_MUTEX(aggregate_mutex);
> +
> static int aggregate_driver_register(struct aggregate_driver *adrv)
> {
> - adrv->driver.bus = &aggregate_bus_type;
> - return driver_register(&adrv->driver);
> + int ret = 0;
> +
> + mutex_lock(&aggregate_mutex);
> + if (!refcount_inc_not_zero(&adrv->count)) {
> + adrv->driver.bus = &aggregate_bus_type;
> + ret = driver_register(&adrv->driver);
> + if (!ret)
> + refcount_inc(&adrv->count);

This should be refcount_set(&adrv->count, 1)?

Otherwise, it will warning like this:

[ 2.654526] ------------[ cut here ]------------
[ 2.655558] refcount_t: addition on 0; use-after-free.
[ 2.656219] WARNING: CPU: 7 PID: 74 at ../v5.16-
rc1/kernel/mediatek/lib/refcount.c:25
refcount_warn_saturate+0x128/0x148
...
[ 2.672227] Call trace:
[ 2.672539] refcount_warn_saturate+0x128/0x148
[ 2.673118] component_aggregate_register+0x388/0x390
[ 2.673763] mtk_iommu_probe+0x638/0x690

[ 2.686467] ------------[ cut here ]------------
[ 2.687049] refcount_t: saturated; leaking memory.
[ 2.687666] WARNING: CPU: 5 PID: 74 at ../v5.16-
rc1/kernel/mediatek/lib/refcount.c:19 refcount_warn_saturate+0xfc/0x148

[ 2.703805] Call trace:
[ 2.704117] refcount_warn_saturate+0xfc/0x148
[ 2.704685] component_aggregate_register+0x1fc/0x390
[ 2.705330] mtk_iommu_probe+0x638/0x690

> + }
> + mutex_unlock(&aggregate_mutex);
> +
> + return ret;
> }
>
> static void aggregate_driver_unregister(struct aggregate_driver
> *adrv)
> {
> - driver_unregister(&adrv->driver);
> + if (refcount_dec_and_mutex_lock(&adrv->count,
> &aggregate_mutex)) {
> + driver_unregister(&adrv->driver);
> + mutex_unlock(&aggregate_mutex);
> + }
> }
>
> static struct aggregate_device *aggregate_device_add(struct device
> *parent,
> diff --git a/include/linux/component.h b/include/linux/component.h
> index 53d81203c095..b061341938aa 100644
> --- a/include/linux/component.h
> +++ b/include/linux/component.h
> @@ -4,6 +4,7 @@
>
> #include <linux/stddef.h>
> #include <linux/device.h>
> +#include <linux/refcount.h>
>
> struct aggregate_device;
>
> @@ -66,6 +67,7 @@ struct device *aggregate_device_parent(const struct
> aggregate_device *adev);
>
> /**
> * struct aggregate_driver - Aggregate driver (made up of other
> drivers)
> + * @count: driver registration refcount
> * @driver: device driver
> */
> struct aggregate_driver {
> @@ -101,6 +103,7 @@ struct aggregate_driver {
> */
> void (*shutdown)(struct aggregate_device *adev);
>
> + refcount_t count;
> struct device_driver driver;
> };

After this patch, the aggregate_driver flow looks ok. But our driver
still aborts like this:

[ 2.721316] Unable to handle kernel NULL pointer dereference at
virtual address 0000000000000000
...
[ 2.731658] pc : mtk_smi_larb_config_port_gen2_general+0xa4/0x138
[ 2.732434] lr : mtk_smi_larb_resume+0x54/0x98
...
[ 2.742457] Call trace:
[ 2.742768] mtk_smi_larb_config_port_gen2_general+0xa4/0x138
[ 2.743496] pm_generic_runtime_resume+0x2c/0x48
[ 2.744090] __genpd_runtime_resume+0x30/0xa8
[ 2.744648] genpd_runtime_resume+0x94/0x2c8
[ 2.745191] __rpm_callback+0x44/0x150
[ 2.745669] rpm_callback+0x6c/0x78
[ 2.746114] rpm_resume+0x314/0x558
[ 2.746559] __pm_runtime_resume+0x3c/0x88
[ 2.747080] pm_runtime_get_suppliers+0x7c/0x110
[ 2.747668] __driver_probe_device+0x4c/0xe8
[ 2.748212] driver_probe_device+0x44/0x130
[ 2.748745] __device_attach_driver+0x98/0xd0
[ 2.749300] bus_for_each_drv+0x68/0xd0
[ 2.749787] __device_attach+0xec/0x148
[ 2.750277] device_attach+0x14/0x20
[ 2.750733] bus_rescan_devices_helper+0x50/0x90
[ 2.751319] bus_for_each_dev+0x7c/0xd8
[ 2.751806] bus_rescan_devices+0x20/0x30
[ 2.752315] __component_add+0x7c/0xa0
[ 2.752795] component_add+0x14/0x20
[ 2.753253] mtk_smi_larb_probe+0xe0/0x120

This is because the device runtime_resume is called before the bind
operation(In our case this detailed function is mtk_smi_larb_bind).
The issue doesn't happen without this patchset. I'm not sure the right
sequence. If we should fix in mediatek driver, the patch could be:


diff --git a/drivers/memory/mtk-smi.c b/drivers/memory/mtk-smi.c
index b883dcc0bbfa..288841555067 100644
--- a/drivers/memory/mtk-smi.c
+++ b/drivers/memory/mtk-smi.c
@@ -483,8 +483,9 @@ static int __maybe_unused
mtk_smi_larb_resume(struct device *dev)
if (ret < 0)
return ret;

- /* Configure the basic setting for this larb */
- larb_gen->config_port(dev);
+ /* Configure the basic setting for this larb after it binds
with iommu */
+ if (larb->mmu)
+ larb_gen->config_port(dev);

return 0;
}

Another nitpick, the title should be: iommu/mediatek: xxxx

\
 
 \ /
  Last update: 2022-01-12 10:11    [W:0.152 / U:0.056 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site