In order to convert this file to the dma_ API, 3 steps are proposed. The first one is purely mechanical. All the updated memory allocations can use GFP_KERNEL because of 'mpt_config()' calls. This function can sleep so it sounds like a reasonable assumption to use GFP_KERNEL. The 2nd patch may be discussed further. GFP_ATOMIC is needed because at least one of the caller might sleep. But if we want to avoid GFP_ATOMIC as much as possible, the 'sleepFlag' could be passed in order to use the "right" flag. All other callers would need to be invistigated to see if GFP_ATOMIC or GFP_KERNEL is needed. The 3rd patch is just a clean up. When 'dma_set_mask_and_coherent' can be used, it is less verbose than 'dma_set_mask()/dma_set_coherent_mask()'. In this patch, look at the comment below ---, because the code looks spurious to me, but as I know neither this hardware, nor the "1078 errata", I'll let anyone that understand this this code to send the correct fix, if needed. Christophe JAILLET (3): scsi: mptbase: switch from 'pci_' to 'dma_' API scsi: mptbase: switch from 'pci_' to 'dma_' API in 'mpt_alloc_fw_memory()' scsi: mptbase: use 'dma_set_mask_and_coherent()' to simplify code drivers/message/fusion/mptbase.c | 151 +++++++++++++++++-------------- 1 file changed, 82 insertions(+), 69 deletions(-) -- 2.30.2
