Kalle Valo <kvalo@xxxxxxxxxx> writes:
> Kalle Valo <kvalo@xxxxxxxxxx> writes:
>
>> Baochen Qiang <quic_bqiang@xxxxxxxxxxx> writes:
>>
>>> Currently during resume, QMI target memory is not properly handled, resulting
>>> in kernel crash in case DMA remap is not supported:
>>>
>>> BUG: Bad page state in process kworker/u16:54 pfn:36e80
>>> page: refcount:1 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x36e80
>>> page dumped because: nonzero _refcount
>>> Call Trace:
>>> bad_page
>>> free_page_is_bad_report
>>> __free_pages_ok
>>> __free_pages
>>> dma_direct_free
>>> dma_free_attrs
>>> ath12k_qmi_free_target_mem_chunk
>>> ath12k_qmi_msg_mem_request_cb
>>>
>>> The reason is:
>>> Once ath12k module is loaded, firmware sends memory request to host. In case
>>> DMA remap not supported, ath12k refuses the first request due to failure in
>>> allocating with large segment size:
>>>
>>> ath12k_pci 0000:04:00.0: qmi firmware request memory request
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 7077888
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 8454144
>>> ath12k_pci 0000:04:00.0: qmi dma allocation failed (7077888 B type 1), will try later with small size
>>> ath12k_pci 0000:04:00.0: qmi delays mem_request 2
>>> ath12k_pci 0000:04:00.0: qmi firmware request memory request
>>>
>>> Later firmware comes back with more but small segments and allocation
>>> succeeds:
>>>
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 262144
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 524288
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 4 size 65536
>>> ath12k_pci 0000:04:00.0: qmi mem seg type 1 size 524288
>>>
>>> Now ath12k is working. If suspend is triggered, firmware will be reloaded
>>> during resume. As same as before, firmware requests two large segments at
>>> first. In ath12k_qmi_msg_mem_request_cb() segment count and size are
>>> assigned:
>>>
>>> ab->qmi.mem_seg_count == 2
>>> ab->qmi.target_mem[0].size == 7077888
>>> ab->qmi.target_mem[1].size == 8454144
>>>
>>> Then allocation failed like before and ath12k_qmi_free_target_mem_chunk()
>>> is called to free all allocated segments. Note the first segment is skipped
>>> because its v.addr is cleared due to allocation failure:
>>>
>>> chunk->v.addr = dma_alloc_coherent()
>>>
>>> Also note that this leaks that segment because it has not been freed.
>>>
>>> While freeing the second segment, a size of 8454144 is passed to
>>> dma_free_coherent(). However remember that this segment is allocated at
>>> the first time firmware is loaded, before suspend. So its real size is
>>> 524288, much smaller than 8454144. As a result kernel found we are freeing
>>> some memory which is in use and thus crashed.
>>>
>>> So one possible fix would be to free those segments during suspend. This
>>> works because with them freed, ath12k_qmi_free_target_mem_chunk() does
>>> nothing: all segment addresses are NULL so dma_free_coherent() is not called.
>>>
>>> But note that ath11k has similar logic but never hits this issue. Reviewing
>>> code there shows the luck comes from QMI memory reuse logic. So the decision
>>> is to port it to ath12k. Like in ath11k, the crash is avoided by adding
>>> prev_size to target_mem_chunk structure and caching real segment size in it,
>>> then prev_size instead of current size is passed to dma_free_coherent(),
>>> no unexpected memory is freed now.
>>>
>>> Also reuse m3 buffer.
>>>
>>> Tested-on: WCN7850 hw2.0 PCI WLAN.HMT.1.0-03427-QCAHMTSWPL_V1.0_V2.0_SILICONZ-1.15378.4
>>> Tested-on: WCN7850 hw2.0 PCI WLAN.HMT.1.0.c5-00481-QCAHMTSWPL_V1.0_V2.0_SILICONZ-3
>>>
>>> Fixes: 64430ddfb132 ("wifi: ath12k: support suspend/resume")
>
> ERROR: Commit id not found from current branch: Fixes: 64430ddfb132
> ("wifi: ath12k: support suspend/resume")
>
> Do note that commit ids in the pending branch are not stable so fixes
> tags cannot be used for those commits.
>
>>> Signed-off-by: Baochen Qiang <quic_bqiang@xxxxxxxxxxx>
>>
>> To keep git bisect clean should this patch be applied before applying
>> the suspend patchset? No need resend or anything, just checking what
>> should be the correct order.
>
> Ah, I just checked and it looks like this patch depends on the suspend
> patchset. I now applied this patch to pendign branch
> (ath-pending-202404190433).
To avoid breaking git bisect I now reshuffled the pending branch so that
'wifi: ath12k: fix kernel crash during resume' is before the suspend
patchset and my plan is to commit the patches in same order. Does this
make sense for everyone?
I had a simple conflict in ath12k_core_reset() in these commits:
https://git.kernel.org/pub/scm/linux/kernel/git/kvalo/ath.git/commit/?h=pending&id=052a6a7c74ae4f20a31a632184f4439edcb0f40c
https://git.kernel.org/pub/scm/linux/kernel/git/kvalo/ath.git/commit/?h=pending&id=6bc740619993535cb64c4dcebb83c66acd506763
Please check my changes.
--
https://patchwork.kernel.org/project/linux-wireless/list/
https://wireless.wiki.kernel.org/en/developers/documentation/submittingpatches
