Hello Christoph, Sorry for the delay with response. On Mon, Sep 12, 2022 at 10:29:10AM +0200, Christoph Hellwig wrote: > On Sat, Sep 10, 2022 at 03:35:42PM +0300, Serge Semin wrote: > > Well, both approaches will solve the denoted problem. I am just > > wondering why do you think that the kmalloc-ed buffer is more > > preferable? > > Because it clearly documents the intent. Here is one buffer that is > just a data buffer, and here is one with kernel internal structure. > The concept of embedding on-disk / on-the-wire structures into internal > stuctures always seemed rather weird and unexpected to me, as we now > need to ensure that the alignment works right on both sides. With > the right annotations (as done in this series) this will work, but > it feels a little fragile to me. IMO both the approaches seem unclear if a reader doesn't know what they have been introduced for. Anyway do you insist on using the kmalloc-ed buffer here instead? If so I'll resubmit the series with this patch updated accordingly. -Sergey > > > What would be the best solution if we had a qualifier like this: > > #ifdef CONFIG_DMA_NONCOHERENT > > #define ____dma_buffer ____cacheline_aligned > > #else > > #define ____dma_buffer > > #endif > > and used it instead of the direct ____cacheline_aligned utilization. > > So independent of my preference for separate allocations, this suggested > additional would still be very useful for the places where we need > to use the alignment for performance or other reasons. I'd use > something like __dma_alligned or similar, though.
