On Thu, 4 Aug 2016 21:17:27 -0300 Mauricio Faria de Oliveira <mauri...@linux.vnet.ibm.com> wrote:
> > [snip] An alternative (and more idiomatic) fix would be to > > change the blk_rq_map_sg() interface to permit passing down some > > foo_NOWARN flag and propagating that down the stack into > > ppc_iommu_map_sg(). Was this approach evaluated? I suspect it might > > be messy. > > I see; I haven't evaluated that, but agree with you it might be messy. > > As far as I can see, in order to pass something to blk_rq_map_sg() and > have it eventually make into ppc_iommu_map_sg(), that something should > be present in the scatterlist -- which seems to be what's common/passed > to both blk_rq_map_sg() (the interface point proposed) and dma_map_sg() > (which is the function which reaches ppc_iommu_map_sg() down the chain). > > It seems a bit hidden, and (if I got the suggestion right), it doesn't > seem to be in the scope of scatterlist to contain such a flag. > > One point of the patches is make the attribute visible/explicit; I see > it can be inconvenient sometimes, but it allows for a clear / evident > difference between dma_map_sg() calls which are (not) OK with failures. > > (for example, the 2 calls in nvme_map_data() - they can return either > BLK_MQ_RQ_QUEUE_BUSY or BLK_MQ_RQ_QUEUE_ERROR - so the former is OK.) Of course, the alternative is to just delete the damn warnings from ppc_iommu_map_sg(). Imagine that! Have they ever been of any use to anyone?
