For RP2350 boards with PSRAM, buffers for use with DMA may be unpredictably allocated in PSRAM. When DMA attempts to access a buffer in PSRAM, a bus fault occurs. This will manifest as a CircuitPython core code crash.

In the present implementation, SRAM and PSRAM are added to the main tlsf heap as separate pools. tlsf does not differentiate between these pools. It happens that tlsf will allocate memory from the SRAM pool first because it was added first. Once SRAM is filled, allocations will come from the PSRAM pool. As the tlsf heap ages, it becomes unpredictable from which pool allocations will come.

The dma_capable argument of port_malloc() is ignored by its RP2 port specialization. Likewise, tlsf does not provide a means of specifying which pool, SRAM or PSRAM, to use for an allocation.

Similarly, it is not predictable whether a VM heap space allocation is in SRAM or PSRAM. There is no means of requesting a VM heap space allocation in SRAM.

An example solution is to implement the dma_capable argument of port_malloc and allocate all DMA buffers from the main tlsf heap rather than the VM heap.

A workaround may be to allocate all DMA buffers early in the life of the system.