Hi Yunfei,
>> Since __alloc_and_insert_iova_range fail will set the current alloc
>> iova size to max32_alloc_size (iovad->max32_alloc_size = size),
>> when the retry is executed into the __alloc_and_insert_iova_range
>> function, the retry action will be blocked by the check condition
>> (size >= iovad->max32_alloc_size) and goto iova32_full directly,
>> causes the action of retry regular alloc iova in
>> __alloc_and_insert_iova_range to not actually be executed.
>>
>> Based on the above, so need reset max32_alloc_size before retry alloc
>> iova when alloc iova fail, that is set the initial dma_32bit_pfn value
>> of iovad to max32_alloc_size, so that the action of retry alloc iova
>> in __alloc_and_insert_iova_range can be executed.
>
> Have you observed this making any difference in practice?
>
> Given that both free_cpu_cached_iovas() and free_global_cached_iovas()
> call iova_magazine_free_pfns(), which calls remove_iova(), which calls
> __cached_rbnode_delete_update(), I'm thinking no...
>
> Robin.
>
Like Robin pointed out, if some cached iovas are freed by
free_global_cached_iovas()/free_cpu_cached_iovas(),
the max32_alloc_size should be reset to iovad->dma_32bit_pfn.
If no cached iova is freed, resetting max32_alloc_size before
the retry allocation only give us a retry. Is it possible that
other users free their iovas during the additional retry?
alloc_iova_fast()
retry:
alloc_iova() // failed, iovad->max32_alloc_size = size
free_cpu_cached_iovas()
iova_magazine_free_pfns()
remove_iova()
__cached_rbnode_delete_update()
iovad->max32_alloc_size = iovad->dma_32bit_pfn // reset
free_global_cached_iovas()
iova_magazine_free_pfns()
remove_iova()
__cached_rbnode_delete_update()
iovad->max32_alloc_size = iovad->dma_32bit_pfn // reset
goto retry;
thanks,
Miles
