[PATCH/RFC] driver core: Postpone DMA tear-down until after devres release

[Geert Uytterhoeven <geert+renesas@glider.be>]

When unbinding the (IOMMU-enabled) R-Car SATA device on Salvator-XS
(R-Car H3 ES2.0), in preparation of rebinding against vfio-platform for
device pass-through for virtualization:

    echo ee300000.sata > /sys/bus/platform/drivers/sata_rcar/unbind

the kernel crashes with:

    Unable to handle kernel paging request at virtual address ffffffbf029ffffc
    Mem abort info:
      ESR = 0x96000006
      Exception class = DABT (current EL), IL = 32 bits
      SET = 0, FnV = 0
      EA = 0, S1PTW = 0
    Data abort info:
      ISV = 0, ISS = 0x00000006
      CM = 0, WnR = 0
    swapper pgtable: 4k pages, 39-bit VAs, pgdp = 000000007e8c586c
    [ffffffbf029ffffc] pgd=000000073bfc6003, pud=000000073bfc6003, pmd=0000000000000000
    Internal error: Oops: 96000006 [#1] SMP
    Modules linked in:
    CPU: 0 PID: 1098 Comm: bash Not tainted 5.0.0-rc5-salvator-x-00452-g37596f884f4318ef #287
    Hardware name: Renesas Salvator-X 2nd version board based on r8a7795 ES2.0+ (DT)
    pstate: 60400005 (nZCv daif +PAN -UAO)
    pc : __free_pages+0x8/0x58
    lr : __dma_direct_free_pages+0x50/0x5c
    sp : ffffff801268baa0
    x29: ffffff801268baa0 x28: 0000000000000000
    x27: ffffffc6f9c60bf0 x26: ffffffc6f9c60bf0
    x25: ffffffc6f9c60810 x24: 0000000000000000
    x23: 00000000fffff000 x22: ffffff8012145000
    x21: 0000000000000800 x20: ffffffbf029fffc8
    x19: 0000000000000000 x18: ffffffc6f86c42c8
    x17: 0000000000000000 x16: 0000000000000070
    x15: 0000000000000003 x14: 0000000000000000
    x13: ffffff801103d7f8 x12: 0000000000000028
    x11: ffffff8011117604 x10: 0000000000009ad8
    x9 : ffffff80110126d0 x8 : ffffffc6f7563000
    x7 : 6b6b6b6b6b6b6b6b x6 : 0000000000000018
    x5 : ffffff8011cf3cc8 x4 : 0000000000004000
    x3 : 0000000000080000 x2 : 0000000000000001
    x1 : 0000000000000000 x0 : ffffffbf029fffc8
    Process bash (pid: 1098, stack limit = 0x00000000c38e3e32)
    Call trace:
     __free_pages+0x8/0x58
     __dma_direct_free_pages+0x50/0x5c
     arch_dma_free+0x1c/0x98
     dma_direct_free+0x14/0x24
     dma_free_attrs+0x9c/0xdc
     dmam_release+0x18/0x20
     release_nodes+0x25c/0x28c
     devres_release_all+0x48/0x4c
     device_release_driver_internal+0x184/0x1f0
     device_release_driver+0x14/0x1c
     unbind_store+0x70/0xb8
     drv_attr_store+0x24/0x34
     sysfs_kf_write+0x4c/0x64
     kernfs_fop_write+0x154/0x1c4
     __vfs_write+0x34/0x164
     vfs_write+0xb4/0x16c
     ksys_write+0x5c/0xbc
     __arm64_sys_write+0x14/0x1c
     el0_svc_common+0x98/0x114
     el0_svc_handler+0x1c/0x24
     el0_svc+0x8/0xc
    Code: d51b4234 17fffffa a9bf7bfd 910003fd (b9403404)
    ---[ end trace 8c564cdd3a1a840f ]---

While I've bisected this to commit e8e683ae9a736407 ("iommu/of: Fix
probe-deferral"), and reverting that commit on post-v5.0-rc4 kernels
does fix the problem, this turned out to be a red herring.

On arm64, arch_teardown_dma_ops() resets dev->dma_ops to NULL.
Hence if a driver has used a managed DMA allocation API, the allocated
DMA memory will be freed using the direct DMA ops, while it may have
been allocated using a custom DMA ops (iommu_dma_ops in this case).

Fix this by reversing the order of the calls to devres_release_all() and
arch_teardown_dma_ops().

Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
---
Question:
Is this safe on arm32, which calls arm_teardown_iommu_dma_ops() instead
of resetting dev->dma_ops?

---
Adding some debug code, and comparing before/after commit
e8e683ae9a736407:

     sata_rcar ee300000.sata: of_iommu_configure:227: err = -517
    -sata_rcar ee300000.sata: of_iommu_configure:230: calling iommu_probe_device()
    -sata_rcar ee300000.sata: iommu_probe_device:122: Calling ipmmu_add_device
    -sata_rcar ee300000.sata: ipmmu_add_device:893
    -sata_rcar ee300000.sata: of_iommu_configure:232: iommu_probe_device() returned -19
    -sata_rcar ee300000.sata: dma_alloc_attrs:257: size 2048, ops =           (null)
    -sata_rcar ee300000.sata: __dma_direct_alloc_pages:104: size 4096
    -sata_rcar ee300000.sata: __dma_direct_alloc_pages:115: trying dma_alloc_from_contiguous()
    -sata_rcar ee300000.sata: dma_alloc_from_contiguous:202: cma_alloc(1) returned page ffffffbf00d20e00
    -sata_rcar ee300000.sata: dma_alloc_attrs:271: allocated using dma_direct_alloc()
    -sata_rcar ee300000.sata: dmam_alloc_attrs:98: size 2048 vaddr ffffff8011e95000 dma_handle 0x0x000000007c040000 attrs 0x0
    -scsi host0: sata_rcar
    -ata1: SATA max UDMA/133 irq 172
    -ata1: SATA link up 1.5 Gbps (SStatus 113 SControl 300)

DMA memory used to be allocated from CMA, dma_map_ops = NULL.
The SATA driver was probed and initialized.

    +sata_rcar ee300000.sata: of_iommu_configure:234: skipping iommu_probe_device()

SATA driver initialization is now deferred.
Later, it is retried:

    +sata_rcar ee300000.sata: of_iommu_configure:227: err = 0
    +sata_rcar ee300000.sata: of_iommu_configure:230: calling iommu_probe_device()
    +sata_rcar ee300000.sata: iommu_probe_device:122: Calling ipmmu_add_device
    +sata_rcar ee300000.sata: ipmmu_add_device:893
    +sata_rcar ee300000.sata: Using IPMMU context 1
    +iommu: Adding device ee300000.sata to group 0
    +sata_rcar ee300000.sata: ipmmu_add_device:908: Success
    +sata_rcar ee300000.sata: of_iommu_configure:232: iommu_probe_device() returned 0
    +sata_rcar ee300000.sata: dma_alloc_attrs:257: size 2048, ops = iommu_dma_ops
    +sata_rcar ee300000.sata: dma_alloc_attrs:274: allocated using __iommu_alloc_attrs()
    +sata_rcar ee300000.sata: dmam_alloc_attrs:98: size 2048 vaddr ffffff8012135000 dma_handle 0x0x00000000fffff000 attrs 0x0
    +scsi host0: sata_rcar
    +ata1: SATA max UDMA/133 irq 172
    +ata1: SATA link up 1.5 Gbps (SStatus 113 SControl 300)

DMA memory is now allocated using __iommu_alloc_attrs(), as dma_map_ops
= iommu_dma_ops.

Unbind:

    -sata_rcar ee300000.sata: dmam_release:32: size 2048 vaddr ffffff8011e95000 dma_handle 0x0x000000007c040000 attrs 0x0
    +sata_rcar ee300000.sata: dmam_release:32: size 2048 vaddr ffffff8012135000 dma_handle 0x0x00000000fffff000 attrs 0x0
     sata_rcar ee300000.sata: dma_free_attrs:289: size 2048, ops =           (null)

dma_map_ops = NULL, both before/after!

     sata_rcar ee300000.sata: dma_free_attrs:308: calling dma_direct_free()
    -sata_rcar ee300000.sata: __dma_direct_free_pages:189: size 2048 page ffffffbf00d20e00
    -sata_rcar ee300000.sata: dma_release_from_contiguous:223: cma_release(1) freed
    -sata_rcar ee300000.sata: __dma_direct_free_pages:194: skipping __free_pages()

DMA memory was correctly freed using CMA.

    +sata_rcar ee300000.sata: __dma_direct_free_pages:189: size 2048 page ffffffbf029fffc8
    +sata_rcar ee300000.sata: dma_release_from_contiguous:225: cma_release(1) failed, cma area ffffff8011cf5530
    +sata_rcar ee300000.sata: __dma_direct_free_pages:191: calling __free_pages()
    +Unable to handle kernel paging request at virtual address ffffffbf029ffffc

Memory is incorrectly freed using the direct ops, as dma_map_ops = NULL.
Oops...

After reversing the order of the calls to arch_teardown_dma_ops() and
devres_release_all(), dma_map_ops is still valid, and the DMA memory is
now released using __iommu_free_attrs():

    +sata_rcar ee300000.sata: dmam_release:32: size 2048 vaddr ffffff8012145000 dma_handle 0x0x00000000fffff000 attrs 0x0
    +sata_rcar ee300000.sata: dma_free_attrs:289: size 2048, ops = iommu_dma_ops
    +sata_rcar ee300000.sata: dma_free_attrs:311: calling __iommu_free_attrs()
---
 drivers/base/dd.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/base/dd.c b/drivers/base/dd.c
index 8ac10af17c0043a3..d62487d024559620 100644
--- a/drivers/base/dd.c
+++ b/drivers/base/dd.c
@@ -968,9 +968,9 @@ static void __device_release_driver(struct device *dev, struct device *parent)
 			drv->remove(dev);
 
 		device_links_driver_cleanup(dev);
-		arch_teardown_dma_ops(dev);
 
 		devres_release_all(dev);
+		arch_teardown_dma_ops(dev);
 		dev->driver = NULL;
 		dev_set_drvdata(dev, NULL);
 		if (dev->pm_domain && dev->pm_domain->dismiss)
-- 
2.17.1