I don't have a question but am rather posting the solution in case it helps someone else out who had the same problem as me.
With SDK v9/SoftDevice S110 v8, our application is too big to use the dual bank bootloader, so we're using the single bank bootloader. I kept running into an issue where DFU using Master Control Panel v3.10.0 would sometimes fail part way through the download process with the message "Error during firmware upload. Disconnected from device while waiting for notification from device.". Eventually I traced the problem back to this segment in dfu_transport_ble.c:
err_code = hci_mem_pool_rx_produce(length, (void **) &mp_rx_buffer);
if (err_code != NRF_SUCCESS)
{
//bl_printf("Failed to rx_produce\r\n", length);
dfu_error_notify(p_dfu, err_code);
return;
}
My settings in hci_mem_pool_internal.h were initially the same as what was in the dfu_dual_bank_ble_s110_pca10028 example:
#define RX_BUF_SIZE 32u /**< RX buffer size in bytes. */
#define RX_BUF_QUEUE_SIZE 8u /**< RX buffer element size. */
And it seems to work much better with the following settings instead. I could probably crank down RX_BUF_SIZE as the largest request I ever saw was for 20 bytes, and have updated the comments to reflect what those #defines actually represent.
#define RX_BUF_SIZE 32u /**< Size of a single RX buffer element in bytes. */
#define RX_BUF_QUEUE_SIZE 16u /**< Number of RX buffer elements in the queue. NOTE: MUST BE A POWER OF TWO */
The key change here is RX_BUF_QUEUE_SIZE, and the reason why this works has to deal with the architecture of the DFU spec. According to the message sequence diagram here, the DFU host first tells the DFU target "I want you to send me a notification every N data packets", and then sends data packets in bunches of N. The problem is that if N is greater than RX_BUF_QUEUE_SIZE, then hci_mem_pool_rx_produce() can run out of memory (since only RX_BUF_QUEUE_SIZE buffers are available) before the notification is sent. The same situation applies once pstorage writes the packet to flash; pstorage has a queue of outstanding commands (denoted by PSTORAGE_CMD_QUEUE_SIZE), and if the queue fills up before all outstanding commands have completed, cmd_queue_enqueue() will also return NRF_ERROR_NO_MEM. Therefore it is also necessary to ensure that PSTORAGE_CMD_QUEUE_SIZE is greater than N, so I made the following change in pstorage_platform.h:
#define PSTORAGE_CMD_QUEUE_SIZE 16
These settings happen to work with Master Control Panel v3.10.0 because MCP only requests a notification every 10 packets. If future versions of MCP changed how many packets were sent at a time, this could break compatibility with existing bootloaders. As a future enhancement, perhaps Nordic can add a mechanism for the DFU target to tell the host, "I'm sorry, but I can only accept a total of N packets at a time".
