Re: [PATCH v7 01/11] iommu: Add DMA ownership management interfaces

From: Robin Murphy
Date: Fri Mar 04 2022 - 09:11:05 EST

On 2022-03-04 13:55, Eric Auger wrote:
Hi Robin,

On 3/4/22 1:22 PM, Robin Murphy wrote:
On 2022-03-04 10:43, Lu Baolu wrote:
Hi Eric,

On 2022/3/4 18:34, Eric Auger wrote:
I hit a WARN_ON() when unbinding an e1000e driver just after boot:

sudo modprobe -v vfio-pci
echo vfio-pci | sudo tee -a
/sys/bus/pci/devices/0004:01:00.0/driver_override
vfio-pci
echo 0004:01:00.0 | sudo tee -a  /sys/bus/pci/drivers/e1000e/unbind


[  390.042811] ------------[ cut here ]------------
[  390.046468] WARNING: CPU: 42 PID: 5589 at drivers/iommu/iommu.c:3123
iommu_device_unuse_default_domain+0x68/0x100
[  390.056710] Modules linked in: vfio_pci vfio_pci_core vfio_virqfd
vfio_iommu_type1 vfio xt_CHECKSUM xt_MASQUERADE xt_conntrack ipt_REJECT
nf_reject_ipv4 nft_compat nft_chain_nat nf_nat nf_conntrack
nf_defrag_ipv6 nf_defrag_ipv4 nf_tables nfnetlink bridge stp llc rfkill
sunrpc vfat fat mlx5_ib ib_uverbs ib_core acpi_ipmi ipmi_ssif
ipmi_devintf ipmi_msghandler cppc_cpufreq drm xfs libcrc32c
mlx5_core sg
mlxfw crct10dif_ce tls ghash_ce sha2_ce sha256_arm64 sha1_ce sbsa_gwdt
e1000e psample sdhci_acpi ahci_platform sdhci libahci_platform
qcom_emac
mmc_core hdma hdma_mgmt dm_mirror dm_region_hash dm_log dm_mod fuse
[  390.110618] CPU: 42 PID: 5589 Comm: tee Kdump: loaded Not tainted
5.17.0-rc4-lu-v7-official+ #24
[  390.119384] Hardware name: WIWYNN QDF2400 Reference Evaluation
Platform CV90-LA115-P120/QDF2400 Customer Reference Board, BIOS
0ACJA570
11/05/2018
[  390.132492] pstate: a0400005 (NzCv daif +PAN -UAO -TCO -DIT -SSBS
BTYPE=--)
[  390.139436] pc : iommu_device_unuse_default_domain+0x68/0x100
[  390.145165] lr : iommu_device_unuse_default_domain+0x38/0x100
[  390.150894] sp : ffff80000fbb3bc0
[  390.154193] x29: ffff80000fbb3bc0 x28: ffff03c0cf6b2400 x27:
0000000000000000
[  390.161311] x26: 0000000000000000 x25: 0000000000000000 x24:
ffff03c0c7cc5720
[  390.168429] x23: ffff03c0c2b9d150 x22: ffffb4e61df223f8 x21:
ffffb4e61df223f8
[  390.175547] x20: ffff03c7c03c3758 x19: ffff03c7c03c3700 x18:
0000000000000000
[  390.182665] x17: 0000000000000000 x16: 0000000000000000 x15:
0000000000000000
[  390.189783] x14: 0000000000000000 x13: 0000000000000030 x12:
ffff03c0d519cd80
[  390.196901] x11: 7f7f7f7f7f7f7f7f x10: 0000000000000dc0 x9 :
ffffb4e620b54f8c
[  390.204019] x8 : ffff03c0cf6b3220 x7 : ffff4ef132bba000 x6 :
00000000000000ff
[  390.211137] x5 : ffff03c0c2b9f108 x4 : ffff03c0d51f6438 x3 :
0000000000000000
[  390.218255] x2 : ffff03c0cf6b2400 x1 : 0000000000000000 x0 :
0000000000000000
[  390.225374] Call trace:
[  390.227804]  iommu_device_unuse_default_domain+0x68/0x100
[  390.233187]  pci_dma_cleanup+0x38/0x44
[  390.236919]  __device_release_driver+0x1a8/0x260
[  390.241519]  device_driver_detach+0x50/0xd0
[  390.245686]  unbind_store+0xf8/0x120
[  390.249245]  drv_attr_store+0x30/0x44
[  390.252891]  sysfs_kf_write+0x50/0x60
[  390.256537]  kernfs_fop_write_iter+0x134/0x1cc
[  390.260964]  new_sync_write+0xf0/0x18c
[  390.264696]  vfs_write+0x230/0x2d0
[  390.268082]  ksys_write+0x74/0x100
[  390.271467]  __arm64_sys_write+0x28/0x3c
[  390.275373]  invoke_syscall.constprop.0+0x58/0xf0
[  390.280061]  el0_svc_common.constprop.0+0x160/0x164
[  390.284922]  do_el0_svc+0x34/0xcc
[  390.288221]  el0_svc+0x30/0x140
[  390.291346]  el0t_64_sync_handler+0xa4/0x130
[  390.295599]  el0t_64_sync+0x1a0/0x1a4
[  390.299245] ---[ end trace 0000000000000000 ]---


I put some traces in the code and I can see that
iommu_device_use_default_domain() effectively is called on
0004:01:00.0 e1000e device on pci_dma_configure() but at that time
the iommu group is NULL:
[   10.569427] e1000e 0004:01:00.0: ------ ENTRY pci_dma_configure
driver_managed_area=0
[   10.569431] e1000e 0004:01:00.0: ****
iommu_device_use_default_domain ENTRY
[   10.569433] e1000e 0004:01:00.0: ****
iommu_device_use_default_domain no group
[   10.569435] e1000e 0004:01:00.0: pci_dma_configure
iommu_device_use_default_domain returned 0
[   10.569492] e1000e 0004:01:00.0: Adding to iommu group 3

^^^the group is added after the
iommu_device_use_default_domain() call
So the group->owner_cnt is not incremented as expected.

Thank you for reporting this. Do you have any idea why the driver is
loaded before iommu_probe_device()?

Urgh, this is the horrible firmware-data-ordering thing again. The
stuff I've been saying about having to rework the whole .dma_configure
mechanism in the near future is to fix this properly.

The summary is that in patch #4, calling
iommu_device_use_default_domain() *before* {of,acpi}_dma_configure is
currently a problem. As things stand, the IOMMU driver ignored the
initial iommu_probe_device() call when the device was added, since at
that point it had no fwspec yet. In this situation,
{of,acpi}_iommu_configure() are retriggering iommu_probe_device()
after the IOMMU driver has seen the firmware data via .of_xlate to
learn that it it actually responsible for the given device.

thank you for providing the info. Hope this is something Lu can work around.

Hopefully it's just a case of flipping the calls around, so that iommu_use_default_domain() goes at the end, and calls arch_teardown_dma_ops() if it fails. From a quick skim I *think* that should still work out to the desired behaviour (or at least close enough that we can move forward without a circular dependency between fixes...)

Robin.