> >>>> + buffer = kmalloc(ALIGN(msg->len, BYTES_PER_FIFO_WORD) +
> >>>> + I2C_PACKET_HEADER_SIZE,
> >>>> GFP_KERNEL);
> >>>> + if (!buffer)
> >>>> + return -ENOMEM;
> >>>
> >>> Isn't it possible to avoid "buffer" allocation / copying overhead and
> >>> keep code as-is for the PIO mode?
> >>>
> >>
> >> Keeping PIO mode code as is and adding DMA mode will have redundant code
> >> for header generation and msg complete timeout sections.
> >> Also in existing PIO mode implementation, TX FIFO is filled as soon as the
> >> word is ready but for DMA mode need to put all msg bytes along with header
> >> together to send thru DMA.
> >> So created common buffer to perform PIO/DMA to reduce redundancy.
> >>
> >
> > Okay, what about something like in this sketch:
> >
> > ...
> > if (msg->flags & I2C_M_RD)
> > xfer_size = msg->len;
> > else
> > xfer_size = ALIGN(msg->len, BYTES_PER_FIFO_WORD) +
> > I2C_PACKET_HEADER_SIZE;
> >
> > dma = ((xfer_size > I2C_PIO_MODE_MAX_LEN) &&
> > i2c_dev->tx_dma_chan && i2c_dev->rx_dma_chan);
> >
> > if (dma) {
> > buffer = kmalloc(ALIGN(msg->len, BYTES_PER_FIFO_WORD) +
> > I2C_PACKET_HEADER_SIZE, GFP_KERNEL);
> > i2c_dev->msg_buf = (u8 *)buffer;
> > }
> >
> > packet_header = (0 << PACKET_HEADER0_HEADER_SIZE_SHIFT) |
> > PACKET_HEADER0_PROTOCOL_I2C |
> > (i2c_dev->cont_id <<
> > PACKET_HEADER0_CONT_ID_SHIFT) |
> > (1 << PACKET_HEADER0_PACKET_ID_SHIFT);
> > i2c_writel(i2c_dev, packet_header, I2C_TX_FIFO);
> >
> > if (dma)
> > (*buffer++) = packet_header;
> > else
> > i2c_writel(i2c_dev, packet_header, I2C_TX_FIFO);
> >
> > ... and so on.
> >
> > Likely that kmalloc overhead will be bigger than the above variant. Please
> > consider to avoid kmalloc and copying for the PIO case in the next version.
> >
>
> Also, is the intermediate kmalloc "buffer" really needed at all? Why just not
> to write directly into the DMA buffer?
Yes, that’s what I am thinking too. Changing it. Will send revised patch..
