Re: [PATCH v2 net] octeontx2-af: allocate qmem from page allocator to avoid CMA

On 2026-08-25 at 10:26:16, Ratheesh Kannoth (rkannoth@marvell.com) wrote:

qmem_alloc() uses dma_alloc_attrs() with DMA_ATTR_FORCE_CONTIGUOUS so
CN10K LMTST regions get physically contiguous DMA memory across page
boundaries.  On platforms with CMA enabled, that attribute causes
allocations to be drawn from the CMA pool.  qmem is used for NIX and
NPA queue contexts, admin queues, and LMTST regions, so the footprint
grows with the number of enabled interfaces and is difficult to
provision in CMA.

Add otx2_dma_alloc_coherent() and otx2_dma_free_coherent() helpers
that allocate compound pages with __get_free_pages(__GFP_COMP),
DMA-map them with dma_map_page(), and retry with GFP_DMA32 when the
physical address exceeds the device DMA mask.  Switch qmem_alloc()
and qmem_free() to use them instead of
dma_alloc_attrs()/dma_free_attrs().

Fixes: 73d33dbc0723 ("octeontx2-af: Use DMA_ATTR_FORCE_CONTIGUOUS attribute in DMA alloc")
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>

Will address sashiko comments in v3

pw-bot: changes-requested