serial: qcom-geni: Pre-map RX DMA buffer at probe to avoid sleep-in-atomic

Browse source

geni_se_rx_dma_prep() calls dma_map_single() which can trigger IOMMU
page table allocations under GFP_KERNEL. This is unsafe when called
from qcom_geni_serial_start_rx_dma(), which runs in atomic context
producing a "sleeping function called from invalid context" splat:

    __might_resched+0x15c/0x17c
    __alloc_pages_noprof+0xe4/0x4c8
    qcom_io_pgtable_alloc_page+0x100/0x250
    __arm_lpae_map+0x2d0/0x870
    geni_se_rx_dma_prep+0xd8/0x158
    qcom_geni_serial_start_rx_dma+0x84/0x16c
    qcom_geni_serial_startup+0x70/0x104

Fix this by mapping the RX DMA buffer once during probe, where sleeping
is allowed, and keeping it mapped for the lifetime of the device.

Replace the geni_se_rx_dma_prep() / geni_se_rx_dma_unprep() calls in
the runtime paths with dma_sync_single_for_device() before initiating a
transfer and dma_sync_single_for_cpu() on completion, using the persistent
mapping. The buffer is unmapped in probe's error path and in remove().

Signed-off-by: Praveen Talari <praveen.talari@oss.qualcomm.com>
Link: https://patch.msgid.link/20260701-fix-sleep-in-atomic-context-during-rx-dma-setup-v1-1-95c208380c65@oss.qualcomm.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

...

This commit is contained in:

Praveen Talari 2026-07-01 09:53:10 +05:30 • committed by Greg Kroah-Hartman

parent 8dfea56f35

commit bafe277d6c

1 changed files with 33 additions and 22 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download patch file Download diff file Expand all files Collapse all files

55

drivers/tty/serial/qcom_geni_serial.c

View file