Re: [PATCH v2] iommu/qcom: fix NULL pointer dereference during probe deferral

From: Stephan Gerhold
Date: Tue Feb 18 2020 - 07:04:52 EST


Hi Robin,

On Mon, Jan 06, 2020 at 01:26:58PM +0000, Robin Murphy wrote:
> On 04/01/2020 12:20 am, Brian Masney wrote:
> > When attempting to load the qcom-iommu driver, and an -EPROBE_DEFER
> > error occurs, the following attempted NULL pointer deference occurs:
> >
> > Unable to handle kernel NULL pointer dereference at virtual address 00000014
> > pgd = (ptrval)
> > [00000014] *pgd=00000000
> > Internal error: Oops: 5 [#1] PREEMPT SMP ARM
> > Modules linked in:
> > CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.5.0-rc2-next-20191220-00010-gfb6b8e8bced6-dirty #3
> > Hardware name: Generic DT based system
> > PC is at qcom_iommu_domain_free (include/linux/pm_runtime.h:226
> > drivers/iommu/qcom_iommu.c:358)
> > LR is at release_iommu_mapping (arch/arm/mm/dma-mapping.c:2141)
> > pc : lr : psr: 60000013
> > sp : ee89dc48 ip : 00000000 fp : c13a6684
> > r10: c13a661c r9 : 00000000 r8 : c13a1240
> > r7 : fffffdfb r6 : 00000000 r5 : edc32f00 r4 : edc32f1c
> > r3 : 00000000 r2 : 00000001 r1 : 00000004 r0 : edc32f1c
> > Flags: nZCv IRQs on FIQs on Mode SVC_32 ISA ARM Segment none
> > Control: 10c5787d Table: 0020406a DAC: 00000051
> > Process swapper/0 (pid: 1, stack limit = 0x(ptrval))
> > Stack: (0xee89dc48 to 0xee89e000)
> > dc40: edc3d010 edc37020 00000000 c0316af8 edc3d010 edc37000
> > dc60: 00000000 c0319684 c14341ac edc3d010 00000000 c083bd88 edc3d010 c13a1240
> > dc80: c083c2e8 c13a6684 c13a661c c13a6508 c13a661c c083c134 c13a1240 ee89dcec
> > dca0: edc3d010 00000000 ee89dcec c083c2e8 c13a6684 c13a661c c13a6508 c13a661c
> > dcc0: c13a6684 c083a31c c13a6684 ee82a86c edc327b8 c1304e48 edc3d010 00000001
> > dce0: edc3d054 c083bc08 ee82a880 edc3d010 00000001 c1304e48 edc3d010 edc3d010
> > dd00: c13a69e8 c083b010 edc3d010 00000000 eea1fc10 c0837aac 00000200 00000000
> > dd20: 00000000 00000000 00000000 c1304e48 00000000 edc3d000 eea1fc10 00000000
> > dd40: 00000000 eeff42f4 00000000 00000001 00000000 c09e96e0 eeff42a4 00000000
> > dd60: 00000000 00000000 eea1fc10 c09e98bc 00000001 eea1fc10 00000000 eea1fc10
> > dd80: edc32c00 c1391580 eea1fc10 00000001 eea1fc10 c0850f90 c2706dec c14368c0
> > dda0: 60000013 c1304e48 00000106 eeff42a4 eeff3fa0 00000000 00000000 eea1fc10
> > ddc0: 00000001 c1248900 00000106 c09e9bd0 00000001 c0c2ee64 eea1fc00 eea1fc10
> > dde0: eea1fc10 00000000 c13a5b70 00000000 c1248900 c081496c c1023d84 00000000
> > de00: eeff3fa0 c2706e48 c2706e48 c1304e48 00000001 00000000 eea1fc10 c13a5b70
> > de20: 00000000 c13a5b70 00000000 c1248900 00000106 c083dfb8 c14341ac eea1fc10
> > de40: 00000000 c083be58 eea1fc10 c13a5b70 c13a5b70 c13a69e8 c12003ec c123a854
> > de60: c1248900 c083c134 c1248900 c09e6f3c c0d8d514 eea1fc10 00000000 c13a5b70
> > de80: c13a69e8 c12003ec c123a854 c1248900 00000106 c083c3e0 00000000 c13a5b70
> > dea0: eea1fc10 c083c440 00000000 c13a5b70 c083c3e8 c083a23c 00000106 ee82a858
> > dec0: eea052b4 c1304e48 c13a5b70 edc32b80 00000000 c083b270 c1043084 c121d1d8
> > dee0: ffffe000 c13a5b70 c121d1d8 ffffe000 00000000 c083cfcc c13ece60 c121d1d8
> > df00: ffffe000 c0302f90 00000106 c034407c 00000000 c10e3a00 c1044dd0 c12003ec
> > df20: 00000000 00000006 00000006 c0fbecac c0fada88 c0fada3c 00000000 efffcbf8
> > df40: efffcc0d c1304e48 00000000 00000006 c13f9500 c1304e48 c123a830 00000007
> > df60: c13f9500 c13f9500 c123a834 c1200f64 00000006 00000006 00000000 c12003ec
> > df80: c0c28194 00000000 c0c28194 00000000 00000000 00000000 00000000 00000000
> > dfa0: 00000000 c0c2819c 00000000 c03010e8 00000000 00000000 00000000 00000000
> > dfc0: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
> > dfe0: 00000000 00000000 00000000 00000000 00000013 00000000 00000000 00000000
> > (qcom_iommu_domain_free) from release_iommu_mapping
> > (arch/arm/mm/dma-mapping.c:2141)
> > (release_iommu_mapping) from arch_teardown_dma_ops
> > (include/linux/dma-mapping.h:271 arch/arm/mm/dma-mapping.c:2335)
> > (arch_teardown_dma_ops) from really_probe (drivers/base/dd.c:607)
> > (really_probe) from driver_probe_device (drivers/base/dd.c:721)
> > (driver_probe_device) from bus_for_each_drv (drivers/base/bus.c:431)
> > (bus_for_each_drv) from __device_attach (drivers/base/dd.c:896)
> > (__device_attach) from bus_probe_device (drivers/base/bus.c:491)
> > (bus_probe_device) from device_add (drivers/base/core.c:2488)
> > (device_add) from of_platform_device_create_pdata
> > (drivers/of/platform.c:189)
> > (of_platform_device_create_pdata) from of_platform_bus_create
> > (drivers/of/platform.c:393 drivers/of/platform.c:346)
> > (of_platform_bus_create) from of_platform_populate
> > (drivers/of/platform.c:486)
> > (of_platform_populate) from msm_pdev_probe
> > (drivers/gpu/drm/msm/msm_drv.c:1197 drivers/gpu/drm/msm/msm_drv.c:1281)
> > (msm_pdev_probe) from platform_drv_probe (drivers/base/platform.c:726)
> > (platform_drv_probe) from really_probe (drivers/base/dd.c:553)
> > (really_probe) from driver_probe_device (drivers/base/dd.c:721)
> > (driver_probe_device) from device_driver_attach (drivers/base/dd.c:995)
> > (device_driver_attach) from __driver_attach (drivers/base/dd.c:1074)
> > (__driver_attach) from bus_for_each_dev (drivers/base/bus.c:304)
> > (bus_for_each_dev) from bus_add_driver (drivers/base/bus.c:623)
> > (bus_add_driver) from driver_register (drivers/base/driver.c:172)
> > (driver_register) from do_one_initcall (include/linux/compiler.h:232
> > include/linux/jump_label.h:254 include/linux/jump_label.h:264
> > include/trace/events/initcall.h:48 init/main.c:941)
> > (do_one_initcall) from kernel_init_freeable (init/main.c:1013
> > init/main.c:1022 init/main.c:1039 init/main.c:1231)
> > (kernel_init_freeable) from kernel_init (init/main.c:1127)
> > (kernel_init) from ret_from_fork (arch/arm/kernel/entry-common.S:156)
> >
> > qcom_iommu_domain_free() has a WARN_ON() that checks to see if the value
> > of iommu is NULL and returns early, so iommu->dev will always be NULL.
> > qcom_iommu_detach_dev() is called prior to freeing the IOMMU domain and
> > is what sets the iommu member to NULL.
> >
> > Let's fix this by adding the 'struct dev' pointer to the
> > qcom_iommu_domain struct.
>
> Actually, it looks like the qcom_domain->iommu logic is fundamentally broken
> anyway - this sequence of calls is perfectly valid, but AFAICS will make
> qcom-iommu go horribly wrong:
>
> dom = iommu_domain_alloc(...);
> iommu_attach_device(dom, dev1);
> iommu_attach_device(dom, dev2);
> iommu_detach_device(dom, dev2);
> // dev1 still attached but dom->iommu now NULL
>
> Furthermore, even this should technically be valid:
>
> dom = iommu_domain_alloc(...);
> iommu_attach_device(dom, dev);
> iommu_map(dom, addr, ...);
> iommu_detach_device(dom, dev);
> iommu_unmap(dom, addr, ...); // oops, dereferences NULL again
>
> Does something like the diff below work?
>
> Robin.

Are you going to send a patch for the diff below?
AFAICT this problem still exists in 5.6-rc2.

Your patch also seems to fix a warning during probe deferral on arm64
that has been around for quite a while. (At least for me...)

(See https://lore.kernel.org/linux-iommu/CA+G9fYtScOpkLvx=__gP903uJ2v87RwZgkAuL6RpF9_DTDs9Zw@xxxxxxxxxxxxxx/)

Thanks,
Stephan

>
> ----->8-----
> diff --git a/drivers/iommu/qcom_iommu.c b/drivers/iommu/qcom_iommu.c
> index c31e7bc4ccbe..af6ee4bf1712 100644
> --- a/drivers/iommu/qcom_iommu.c
> +++ b/drivers/iommu/qcom_iommu.c
> @@ -345,21 +345,20 @@ static void qcom_iommu_domain_free(struct iommu_domain
> *domain)
> {
> struct qcom_iommu_domain *qcom_domain = to_qcom_iommu_domain(domain);
>
> - if (WARN_ON(qcom_domain->iommu)) /* forgot to detach? */
> - return;
> -
> iommu_put_dma_cookie(domain);
>
> - /* NOTE: unmap can be called after client device is powered off,
> - * for example, with GPUs or anything involving dma-buf. So we
> - * cannot rely on the device_link. Make sure the IOMMU is on to
> - * avoid unclocked accesses in the TLB inv path:
> - */
> - pm_runtime_get_sync(qcom_domain->iommu->dev);
> + if (qcom_domain->iommu) {
> + /* NOTE: unmap can be called after client device is powered off,
> + * for example, with GPUs or anything involving dma-buf. So we
> + * cannot rely on the device_link. Make sure the IOMMU is on to
> + * avoid unclocked accesses in the TLB inv path:
> + */
> + pm_runtime_get_sync(qcom_domain->iommu->dev);
>
> - free_io_pgtable_ops(qcom_domain->pgtbl_ops);
> + free_io_pgtable_ops(qcom_domain->pgtbl_ops);
>
> - pm_runtime_put_sync(qcom_domain->iommu->dev);
> + pm_runtime_put_sync(qcom_domain->iommu->dev);
> + }
>
> kfree(qcom_domain);
> }
> @@ -405,7 +404,7 @@ static void qcom_iommu_detach_dev(struct iommu_domain
> *domain, struct device *de
> struct qcom_iommu_domain *qcom_domain = to_qcom_iommu_domain(domain);
> unsigned i;
>
> - if (!qcom_domain->iommu)
> + if (WARN_ON(!qcom_domain->iommu))
> return;
>
> pm_runtime_get_sync(qcom_iommu->dev);
> @@ -418,8 +417,6 @@ static void qcom_iommu_detach_dev(struct iommu_domain
> *domain, struct device *de
> ctx->domain = NULL;
> }
> pm_runtime_put_sync(qcom_iommu->dev);
> -
> - qcom_domain->iommu = NULL;
> }
>
> static int qcom_iommu_map(struct iommu_domain *domain, unsigned long iova,