Re: 6.1-rc1 regression: bisected to 57365a04c921 iommu: Move bus setup to IOMMU device registration

From: Robin Murphy
Date: Thu Oct 20 2022 - 06:03:04 EST


On 2022-10-20 00:25, Brian Norris wrote:
Hi all,

I'm testing out asynchronous probe (that's, kernel cmdline
'driver_async_probe=*' or similar), and I've identified a regression in
v6.1-rc1 due to this:

commit 57365a04c92126525a58bf7a1599ddfa832415e9
Author: Robin Murphy <robin.murphy@xxxxxxx>
Date: Mon Aug 15 17:20:06 2022 +0100

iommu: Move bus setup to IOMMU device registration

In particular, I'm testing a Rockchip RK3399 system with
'driver_async_probe=rk_iommu', and finding I crash like this:

[ 0.180480] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000030
...
[ 0.180583] CPU: 2 PID: 49 Comm: kworker/u12:1 Not tainted 6.1.0-rc1 #57
[ 0.180593] Hardware name: Google Scarlet (DT)
[ 0.180602] Workqueue: events_unbound async_run_entry_fn
[ 0.180622] pstate: 20000005 (nzCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[ 0.180632] pc : dev_iommu_free+0x24/0x54
[ 0.180644] lr : __iommu_probe_device+0x110/0x180
...
[ 0.180785] Call trace:
[ 0.180791] dev_iommu_free+0x24/0x54
[ 0.180800] __iommu_probe_device+0x110/0x180
[ 0.180807] probe_iommu_group+0x40/0x58
[ 0.180816] bus_for_each_dev+0x8c/0xd8
[ 0.180829] bus_iommu_probe+0x5c/0x2d0
[ 0.180840] iommu_device_register+0xbc/0x104
[ 0.180851] rk_iommu_probe+0x260/0x354
[ 0.180861] platform_probe+0xb4/0xd4
[ 0.180872] really_probe+0xfc/0x284
[ 0.180884] __driver_probe_device+0xc0/0xec
[ 0.180894] driver_probe_device+0x4c/0xd4
[ 0.180905] __driver_attach_async_helper+0x3c/0x60
[ 0.180915] async_run_entry_fn+0x34/0xd4
[ 0.180926] process_one_work+0x1e0/0x3b4
[ 0.180936] worker_thread+0x120/0x404
[ 0.180942] [drm] Initialized vgem 1.0.0 20120112 for vgem on minor 0
[ 0.180944] kthread+0xf4/0x14c
[ 0.180953] ret_from_fork+0x10/0x20
[ 0.180968] Code: f9000bf3 910003fd f9416c13 f9016c1f (f9401a68)
[ 0.180981] ---[ end trace 0000000000000000 ]---

Hmm, on the face of it that callstack doesn't even make much sense -
dev_iommu_free+0x24 presumably has to be the dereference of
param->fwspec, but __iommu_probe_device can only hit that failure path
after a successful dev_iommu_get(), so dev->iommu should not be NULL.

I find if I revert the above commit (and 29e932295bfa ("iommu: Clean up
bus_set_iommu()"), to keep the reverts clean), things start working again.

I haven't worked out exactly what's going wrong, but the patch looks like it
isn't async safe at all, due to the way each device is poking (without locking)
at the global iommu_buses[].

We shouldn't really need locking for iommu_buses itself, but I guess to
support async probe we would need per-device locking in
iommu_probe_device() to prevent multiple threads trying to probe the
same device at once, which must be what's happening in your case to
cause a double-dev_iommu_free(). I'll see what I can do ASAP, since I
think that's worthwhile. In the meantime, as to why you're hitting the
failure path at all, I think that's another subtle oversight on my part,
does something like the diff below help?

Thanks,
Robin.

----->8-----
diff --git a/drivers/iommu/rockchip-iommu.c b/drivers/iommu/rockchip-iommu.c
index 4f2039789897..00fc57c86ea4 100644
--- a/drivers/iommu/rockchip-iommu.c
+++ b/drivers/iommu/rockchip-iommu.c
@@ -1163,6 +1163,14 @@ static struct iommu_group *rk_iommu_device_group(struct device *dev)
iommu = rk_iommu_from_dev(dev);
+ /*
+ * Use the first registered IOMMU device for domain to use with DMA
+ * API, since a domain might not physically correspond to a single
+ * IOMMU device..
+ */
+ if (!dma_dev)
+ dma_dev = iommu->dev;
+
return iommu_group_ref_get(iommu->group);
}
@@ -1293,14 +1301,6 @@ static int rk_iommu_probe(struct platform_device *pdev)
if (err)
goto err_remove_sysfs;
- /*
- * Use the first registered IOMMU device for domain to use with DMA
- * API, since a domain might not physically correspond to a single
- * IOMMU device..
- */
- if (!dma_dev)
- dma_dev = &pdev->dev;
-
pm_runtime_enable(dev);
for (i = 0; i < iommu->num_irq; i++) {