forward port redboot rootfs partition space expansion patch to 2.6.24

[openwrt.git] / target / linux / etrax / files / drivers / spi / spi_crisv32_sser.c

/*
 * SPI port driver for ETRAX FS et al. using a synchronous serial
 * port, but simplified by using the spi_bitbang framework.
 *
 * Copyright (c) 2007 Axis Communications AB
 *
 * Author: Hans-Peter Nilsson, though copying parts of
 * spi_s3c24xx_gpio.c, hence also:
 * Copyright (c) 2006 Ben Dooks
 * Copyright (c) 2006 Simtec Electronics
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 as
 * published by the Free Software Foundation.
 *
 * This driver restricts frequency, polarity, "word" length and endian
 * much more than the hardware does.  I'm happy to unrestrict it, but
 * only with what I can test myself (at time of writing, just SD/MMC
 * SPI) and what people actually test and report.
 */

#include <linux/types.h>
#include <linux/device.h>
#include <linux/spi/spi.h>
#include <linux/spi/spi_bitbang.h>
#include <linux/delay.h>
#include <linux/platform_device.h>
#include <linux/interrupt.h>
#include <asm/io.h>
#include <asm/arch/board.h>
#include <asm/arch/hwregs/reg_map.h>
#include <asm/arch/hwregs/reg_rdwr.h>
#include <asm/arch/hwregs/sser_defs.h>
#include <asm/arch/dma.h>
#include <asm/arch/hwregs/dma.h>

/* A size "not much larger" than the max typical transfer size.  */
#define DMA_CHUNKSIZ 512

/*
 * For a transfer expected to take this long, we busy-wait instead of enabling
 * interrupts.
 */
#define IRQ_USAGE_THRESHOLD_NS 14000

/* A few register access macros to avoid verbiage and reduce typos.  */
#define REG_RD_DI(reg) REG_RD(dma, regi_dmain, reg)
#define REG_RD_DO(reg) REG_RD(dma, regi_dmaout, reg)
#define REG_RD_SSER(reg) REG_RD(sser, regi_sser, reg)
#define REG_WR_DI(reg, val) REG_WR(dma, regi_dmain, reg, val)
#define REG_WR_DO(reg, val) REG_WR(dma, regi_dmaout, reg, val)
#define REG_WR_SSER(reg, val) REG_WR(sser, regi_sser, reg, val)
#define REG_WRINT_DI(reg, val) REG_WR_INT(dma, regi_dmain, reg, val)
#define REG_WRINT_DO(reg, val) REG_WR_INT(dma, regi_dmaout, reg, val)
#define REG_WRINT_SSER(reg, val) REG_WR_INT(sser, regi_sser, reg, val)
#define REG_RDINT_DI(reg) REG_RD_INT(dma, regi_dmain, reg)
#define REG_RDINT_DO(reg) REG_RD_INT(dma, regi_dmaout, reg)
#define REG_RDINT_SSER(reg) REG_RD_INT(sser, regi_sser, reg)

#define DMA_WAIT_UNTIL_RESET(inst)                      \
  do {                                                  \
        reg_dma_rw_stat r;                              \
        do {                                            \
                r = REG_RD(dma, (inst), rw_stat);       \
        } while (r.mode != regk_dma_rst);               \
  } while (0)

#define DMA_BUSY(inst) (REG_RD(dma, inst, rw_stream_cmd)).busy

/* Our main driver state.  */
struct crisv32_spi_hw_info {
        struct crisv32_regi_n_int sser;
        struct crisv32_regi_n_int dmain;
        struct crisv32_regi_n_int dmaout;

        reg_sser_rw_cfg cfg;
        reg_sser_rw_frm_cfg frm_cfg;
        reg_sser_rw_tr_cfg tr_cfg;
        reg_sser_rw_rec_cfg rec_cfg;
        reg_sser_rw_extra extra;

        /* We store the speed in kHz, so we can have expressions
         * multiplying 100MHz by * 4 before dividing by it, and still
         * keep it in an u32. */
        u32 effective_speed_kHz;

        /*
         * The time in 10s of nanoseconds for half a cycles.
         * For convenience and performance; derived from the above.
         */
        u32 half_cycle_delay_ns;

        /* This should be overridable by a module parameter.  */
        u32 max_speed_Hz;

        /* Pre-computed timout for the max transfer chunk-size.  */
        u32 dma_timeout;

        struct completion dma_done;

        /*
         * If we get a timeout from wait_for_completion_timeout on the
         * above, first look at this before panicking.
         */
        u32 dma_actually_done;

        /*
         * Resources don't seem available at the remove call, so we
         * have to save information we get through them.
         */
        struct crisv32_spi_sser_controller_data *gc;
};

/*
 * The driver state hides behind the spi_bitbang state; we're
 * responsible for allocating that, so we can get a little something
 * for ourselves.
 */
struct crisv32_spi_sser_devdata {
        struct spi_bitbang bitbang;
        struct crisv32_spi_hw_info hw;
};

/* Our DMA descriptors that need alignment.  */
struct crisv32_spi_dma_descrs {
        dma_descr_context in_ctxt __attribute__ ((__aligned__(32)));
        dma_descr_context out_ctxt __attribute__ ((__aligned__(32)));

        /*
         * The code takes advantage of the fact that in_descr and
         * out_descr are on the same cache-line when working around
         * the cache-bug in TR 106.
         */
        dma_descr_data in_descr __attribute__ ((__aligned__(16)));
        dma_descr_data out_descr __attribute__ ((__aligned__(16)));
};

/*
 * Whatever needs DMA access is here, besides whatever DMA-able memory
 * comes in transfers.
 */
struct crisv32_spi_dma_cs {
        struct crisv32_spi_dma_descrs *descrp;

        /* Scratch-buffers when the original was non-DMA.  */
        u8 rx_buf[DMA_CHUNKSIZ];
        u8 tx_buf[DMA_CHUNKSIZ];
};

/*
 * Max speed.  If set, we won't go faster, promise.  May be useful
 * when dealing with weak hardware; misrouted signal paths or various
 * debug-situations.
 */
static ulong crisv32_spi_speed_limit_Hz = 0;

/* Helper function getting the driver state from a spi_device.  */

static inline struct crisv32_spi_hw_info *spidev_to_hw(struct spi_device *spi)
{
        struct crisv32_spi_sser_devdata *dd = spi_master_get_devdata(spi->master);
        return &dd->hw;
}

/* SPI-bitbang word transmit-function for non-DMA.  */

static u32 crisv32_spi_sser_txrx_mode3(struct spi_device *spi,
                                       unsigned nsecs, u32 word, u8 bits)
{
        struct crisv32_spi_hw_info *hw = spidev_to_hw(spi);
        u32 regi_sser = hw->sser.regi;
        reg_sser_rw_ack_intr ack_intr = { .trdy = 1, .rdav = 1 };
        reg_sser_r_intr intr = {0};
        reg_sser_rw_tr_data w_data = { .data = (u8) word };
        reg_sser_r_rec_data r_data;
        u32 i;

        /*
         * The timeout reflects one iteration per 10ns (impossible at
         * 200MHz clock even without the ndelay) and a wait for a full
         * byte.
         */
        u32 timeout = 1000000/10*8/hw->effective_speed_kHz;

        BUG_ON(bits != 8);

        intr = REG_RD_SSER(r_intr);

        /*
         * We should never get xruns when we control the transmitter
         * and receiver in register mode.  And if we don't have
         * transmitter-ready and data-ready on entry, something's
         * seriously fishy.
         */
        if (!intr.trdy || !intr.rdav || intr.orun || intr.urun)
                panic("sser hardware or SPI driver broken (1) 0x%x\n",
                      REG_TYPE_CONV(u32, reg_sser_r_intr, intr));

        REG_WR_SSER(rw_ack_intr, ack_intr);
        REG_WR_SSER(rw_tr_data, w_data);

        for (i = 0; i < timeout; i++) {
                intr = REG_RD_SSER(r_intr);
                /* Wait for received data.  */
                if (intr.rdav)
                        break;
                ndelay(10);
        }

        if (!(intr.trdy && intr.rdav) || intr.orun || intr.urun)
                panic("sser hardware or SPI driver broken (2) 0x%x\n",
                      REG_TYPE_CONV(u32, reg_sser_r_intr, intr));

        r_data = REG_RD_SSER(r_rec_data);
        return r_data.data & 0xff;
}

/*
 * Wait for 1/2 bit-time if the transmitter or receiver is enabled.
 * We need to do this as the data-available indications may arrive
 * right at the edge, with half the last cycle remaining.
 */
static void inline crisv32_spi_sser_wait_halfabit(struct crisv32_spi_hw_info
                                                  *hw)
{
        if (hw->cfg.en)
                ndelay(hw->half_cycle_delay_ns);
}

/*
 * Assert or de-assert chip-select.
 * We have two functions, with the active one assigned to the bitbang
 * slot at setup, to avoid a performance penalty (1% on reads).
 */
static void crisv32_spi_sser_chip_select_active_high(struct spi_device *spi,
                                                     int value)
{
        struct crisv32_spi_hw_info *hw = spidev_to_hw(spi);
        u32 regi_sser = hw->sser.regi;

        /*
         * We may have received data at the "last producing clock
         * edge".  Thus we delay for another half a clock cycle.
         */
        crisv32_spi_sser_wait_halfabit(hw);

        hw->frm_cfg.frame_pin_use
                = value == BITBANG_CS_ACTIVE ? regk_sser_gio1 : regk_sser_gio0;
        REG_WR_SSER(rw_frm_cfg, hw->frm_cfg);
}

static void crisv32_spi_sser_chip_select_active_low(struct spi_device *spi,
                                                    int value)
{
        struct crisv32_spi_hw_info *hw = spidev_to_hw(spi);
        u32 regi_sser = hw->sser.regi;

        crisv32_spi_sser_wait_halfabit(hw);
        hw->frm_cfg.frame_pin_use
                = value == BITBANG_CS_ACTIVE ? regk_sser_gio0 : regk_sser_gio1;
        REG_WR_SSER(rw_frm_cfg, hw->frm_cfg);
}

/* Set the transmission speed in Hz.  */

static int crisv32_spi_sser_set_speed_Hz(struct crisv32_spi_hw_info *hw,
                                         u32 Hz)
{
        u32 kHz;
        u32 ns_delay;
        u32 regi_sser = hw->sser.regi;

        if (Hz > hw->max_speed_Hz)
                /*
                 * Should we complain?  Return error?  Current caller
                 * sequences want just the max speed.
                 */
                Hz = hw->max_speed_Hz;

        kHz = Hz/1000;

        /*
         * If absolutely needed, we *could* change the base frequency
         * and go lower.  Usually, a frequency set higher than wanted
         * is a problem but lower isn't.
         */
        if (Hz < 100000000 / 65536 + 1) {
                printk(KERN_ERR "attempt to set invalid sser speed: %u Hz\n",
                       Hz);
                Hz = 100000000 / 65536 + 1;
        }

        pr_debug("setting sser speed to %u Hz\n", Hz);

        /*
         * Avoid going above the requested speed if there's a
         * remainder for the 100 MHz clock-divider calculation, but
         * don't unnecessarily go below if it's even.
         */
        hw->cfg.clk_div = 100000000/Hz - ((100000000 % Hz) == 0);

        /* Make sure there's no ongoing transmission. */
        crisv32_spi_sser_wait_halfabit(hw);

        /*
         * Wait for 3 times max of the old and the new clock before and after
         * changing the frequency.  Not because of documentation or empirical
         * need, but because it seems sane to do so.  The three-bit-times
         * value is because that's the documented time it takes for a reset to
         * take effect.
         */
        ns_delay = 1000000*3/(kHz > hw->effective_speed_kHz
                              ? kHz : hw->effective_speed_kHz);
        ndelay(ns_delay);
        REG_WR_SSER(rw_cfg, hw->cfg);
        ndelay(ns_delay);

        hw->effective_speed_kHz = kHz;

        /*
         * A timeout of twice the time for the largest chunk (not
         * counting DMA overhead) plus one jiffy, should be more than
         * enough for the transmission.
         */
        hw->dma_timeout = 1 + usecs_to_jiffies(1000*2*DMA_CHUNKSIZ*8/kHz);

        hw->half_cycle_delay_ns
                = 1000000/2/hw->effective_speed_kHz;

        pr_debug(".clk_div %d, half %d, eff %d\n",
                 hw->cfg.clk_div, hw->half_cycle_delay_ns,
                 hw->effective_speed_kHz);
        return 0;
}

/*
 * Set up transmitter and receiver for non-DMA access.
 * Unfortunately, it doesn't seem like hispeed works for this mode
 * (mea culpa), so we're stuck with lospeed-mode.  A little slower,
 * but that's what you get for not allocating DMA.
 */
static int crisv32_setup_spi_sser_for_reg_access(struct crisv32_spi_hw_info *hw)
{
        u32 regi_sser = hw->sser.regi;

        reg_sser_rw_cfg cfg = {0};
        reg_sser_rw_frm_cfg frm_cfg = {0};
        reg_sser_rw_tr_cfg tr_cfg = {0};
        reg_sser_rw_rec_cfg rec_cfg = {0};
        reg_sser_rw_intr_mask mask = {0};
        reg_sser_rw_extra extra = {0};
        reg_sser_rw_tr_data tr_data = {0};
        reg_sser_r_intr intr;

        cfg.en = 0;
        tr_cfg.tr_en = 1;
        rec_cfg.rec_en = 1;
        REG_WR_SSER(rw_cfg, cfg);
        REG_WR_SSER(rw_tr_cfg, tr_cfg);
        REG_WR_SSER(rw_rec_cfg, rec_cfg);
        REG_WR_SSER(rw_intr_mask, mask);

        /*
         * See 23.7.2 SPI in the hardware documentation.
         * Except our configuration uses bulk mode; MMC/SD-SPI
         * isn't isochronous in nature.
         * Step 1.
         */
        cfg.gate_clk = regk_sser_yes;
        cfg.clkgate_in = regk_sser_no;
        cfg.clkgate_ctrl = regk_sser_tr;

        /* Step 2.  */
        cfg.out_clk_pol = regk_sser_pos;
        cfg.out_clk_src = regk_sser_intern_clk;

        /* Step 3.  */
        tr_cfg.clk_src = regk_sser_intern;
        rec_cfg.clk_src = regk_sser_intern;
        frm_cfg.clk_src = regk_sser_intern;

        /* Step 4.  */
        tr_cfg.clk_pol = regk_sser_neg;
        rec_cfg.clk_pol = regk_sser_pos;
        frm_cfg.clk_pol = regk_sser_neg;

        /*
         * Step 5: frame pin (PC03 or PD03) is frame; the status pin
         * (PC02, PD02) is configured as input.
         */
        frm_cfg.frame_pin_dir = regk_sser_out;

        /*
         * Contrary to the doc example, we don't generate the frame
         * signal "automatically".  This setting of the frame pin as
         * constant 1, reflects an inactive /CS setting, for just idle
         * clocking.  When we need to transmit or receive data, we
         * change it.
         */
        frm_cfg.frame_pin_use = regk_sser_gio1;
        frm_cfg.status_pin_dir = regk_sser_in;

        /*
         * Step 6.  This is probably not necessary, as we don't
         * generate the frame signal automatically.  Nevertheless,
         * modified for bulk transmission.
         */
        frm_cfg.out_on = regk_sser_tr;
        frm_cfg.out_off = regk_sser_tr;

        /* Step 7.  Similarly, maybe not necessary.  */
        frm_cfg.type = regk_sser_level;
        frm_cfg.level = regk_sser_neg_lo;

        /* Step 8.  These we have to set according to the bulk mode,
         * which for tr_delay is the same as for iso; a value of 1
         * means in sync with the frame signal.  For rec_delay, we
         * start it at the same time as the transmitter.  See figure
         * 23.7 in the hw documentation.  */
        frm_cfg.tr_delay = 1;
        frm_cfg.rec_delay = 0;

        /* Step 9.  */
        tr_cfg.sample_size = 7;
        rec_cfg.sample_size = 7;

        /* Step 10.  */
        frm_cfg.wordrate = 7;

        /* Step 11 (but for bulk).  */
        tr_cfg.rate_ctrl = regk_sser_bulk;

        /*
         * Step 12.  Similarly, maybe not necessary; still, modified
         * for bulk.
         */
        tr_cfg.frm_src = regk_sser_intern;
        rec_cfg.frm_src = regk_sser_tx_bulk;

        /* Step 13.  */
        tr_cfg.mode = regk_sser_lospeed;
        rec_cfg.mode = regk_sser_lospeed;

        /* Step 14.  */
        tr_cfg.sh_dir = regk_sser_msbfirst;
        rec_cfg.sh_dir = regk_sser_msbfirst;

        /*
         * Extra step for bulk-specific settings and other general
         * settings not specified in the SPI config example.
         * It's uncertain whether all of these are needed.
         */
        tr_cfg.bulk_wspace = 1;
        tr_cfg.use_dma = 0;

        tr_cfg.urun_stop = 1;
        rec_cfg.orun_stop = 1;
        rec_cfg.use_dma = 0;

        rec_cfg.fifo_thr = regk_sser_inf;
        frm_cfg.early_wend = regk_sser_yes;

        cfg.clk_dir = regk_sser_out;
        tr_cfg.data_pin_use = regk_sser_dout;
        cfg.base_freq = regk_sser_f100;

        /* Setup for the initial frequency given to us.  */
        hw->cfg = cfg;
        crisv32_spi_sser_set_speed_Hz(hw, hw->max_speed_Hz);
        cfg = hw->cfg;

        /*
         * Write it all, except cfg which is already written by
         * crisv32_spi_sser_set_speed_Hz.
         */
        REG_WR_SSER(rw_frm_cfg, frm_cfg);
        REG_WR_SSER(rw_tr_cfg, tr_cfg);
        REG_WR_SSER(rw_rec_cfg, rec_cfg);
        REG_WR_SSER(rw_extra, extra);

        /*
         * The transmit-register needs to be written before the
         * transmitter is enabled, and to get a valid trdy signal
         * waiting for us when we want to transmit a byte.  Because
         * the "frame event" is that the transmitter is written, this
         * will cause a dummy 0xff-byte to be transmitted, but that's
         * ok, because /CS is inactive.
         */
        tr_data.data = 0xffff;
        REG_WR_SSER(rw_tr_data, tr_data);

        /*
         * We ack everything interrupt-wise; left-over indicators don't have
         * to come from *this* code.
         */
        REG_WRINT_SSER(rw_ack_intr, -1);

        /*
         * Wait 3 cycles before enabling, after the transmit register
         * has been written.  (This'll be just a few microseconds for
         * e.g. 400 KHz.)
         */
        ndelay(3 * 2 * hw->half_cycle_delay_ns);
        cfg.en = 1;

        REG_WR_SSER(rw_cfg, cfg);

        /*
         * Now wait for 8 + 3 cycles.  The 0xff byte should now have
         * been transmitted and dummy data received.
         */
        ndelay((8 + 3) * 2 * hw->half_cycle_delay_ns);

        /*
         * Sanity-check that we have data-available and the
         * transmitter is ready to send new data.
         */
        intr = REG_RD_SSER(r_intr);
        if (!intr.rdav || !intr.trdy)
                panic("sser hw or SPI driver broken (3) 0x%x",
                      REG_TYPE_CONV(u32, reg_sser_r_intr, intr));

        hw->frm_cfg = frm_cfg;
        hw->tr_cfg = tr_cfg;
        hw->rec_cfg = rec_cfg;
        hw->extra = extra;
        hw->cfg = cfg;
        return 0;
}

/* Initialization, maybe fault recovery.  */

static void crisv32_reset_dma_hw(u32 regi)
{
        REG_WR_INT(dma, regi, rw_intr_mask, 0);

        DMA_RESET(regi);
        DMA_WAIT_UNTIL_RESET(regi);
        DMA_ENABLE(regi);
        REG_WR_INT(dma, regi, rw_ack_intr, -1);

        DMA_WR_CMD(regi, regk_dma_set_w_size1);
}

/* Interrupt from SSER, for use with DMA when only the transmitter is used.  */

static irqreturn_t sser_interrupt(int irqno, void *arg)
{
        struct crisv32_spi_hw_info *hw = arg;
        u32 regi_sser = hw->sser.regi;
        reg_sser_r_intr intr = REG_RD_SSER(r_intr);

        if (intr.tidle == 0 && intr.urun == 0) {
                printk(KERN_ERR
                       "sser @0x%x: spurious sser intr, flags: 0x%x\n",
                       regi_sser, REG_TYPE_CONV(u32, reg_sser_r_intr, intr));
        } else if (intr.urun == 0) {
                hw->dma_actually_done = 1;
                complete(&hw->dma_done);
        } else {
                /*
                 * Make any reception time out and notice the error,
                 * which it might not otherwise do data was *received*
                 * successfully.
                 */
                u32 regi_dmain = hw->dmain.regi;

                /*
                 * Recommended practice before acking urun is to turn
                 * off sser.  That might not be enough to stop DMA-in
                 * from signalling success if the underrun was late in
                 * the transmission, so we disable the DMA-in
                 * interrupts too.
                 */
                REG_WRINT_SSER(rw_cfg, 0);
                REG_WRINT_DI(rw_intr_mask, 0);
                REG_WRINT_DI(rw_ack_intr, -1);
        }

        REG_WRINT_SSER(rw_intr_mask, 0);

        /*
         * We must at least ack urun together with tidle, but keep it
         * simple and ack them all.
         */
        REG_WRINT_SSER(rw_ack_intr, -1);

        return IRQ_HANDLED;
}

/*
 * Interrupt from receiver DMA connected to SSER, for use when the
 * receiver is used, with or without the transmitter.
 */
static irqreturn_t rec_dma_interrupt(int irqno, void *arg)
{
        struct crisv32_spi_hw_info *hw = arg;
        u32 regi_dmain = hw->dmain.regi;
        u32 regi_sser = hw->sser.regi;
        reg_dma_r_intr intr = REG_RD_DI(r_intr);

        if (intr.data == 0) {
                printk(KERN_ERR
                       "sser @0x%x: spurious rec dma intr, flags: 0x%x\n",
                       regi_dmain, REG_TYPE_CONV(u32, reg_dma_r_intr, intr));
        } else {
                hw->dma_actually_done = 1;
                complete(&hw->dma_done);
        }

        REG_WRINT_DI(rw_intr_mask, 0);

        /* Avoid false underrun indications; stop all sser interrupts.   */
        REG_WRINT_SSER(rw_intr_mask, 0);
        REG_WRINT_SSER(rw_ack_intr, -1);

        REG_WRINT_DI(rw_ack_intr, -1);
        return IRQ_HANDLED;
}

/*
 * Set up transmitter and receiver for DMA access.  We use settings
 * from the "Atmel fast flash" example.
 */
static int crisv32_setup_spi_sser_for_dma_access(struct crisv32_spi_hw_info
                                                 *hw)
{
        int ret;
        u32 regi_sser = hw->sser.regi;

        reg_sser_rw_cfg cfg = {0};
        reg_sser_rw_frm_cfg frm_cfg = {0};
        reg_sser_rw_tr_cfg tr_cfg = {0};
        reg_sser_rw_rec_cfg rec_cfg = {0};
        reg_sser_rw_intr_mask mask = {0};
        reg_sser_rw_extra extra = {0};

        cfg.en = 0;
        tr_cfg.tr_en = 1;
        rec_cfg.rec_en = 1;
        REG_WR_SSER(rw_cfg, cfg);
        REG_WR_SSER(rw_tr_cfg, tr_cfg);
        REG_WR_SSER(rw_rec_cfg, rec_cfg);
        REG_WR_SSER(rw_intr_mask, mask);

        /*
         * See 23.7.5.2 (Atmel fast flash) in the hardware documentation.
         * Step 1.
         */
        cfg.gate_clk = regk_sser_no;

        /* Step 2.  */
        cfg.out_clk_pol = regk_sser_pos;

        /* Step 3.  */
        cfg.out_clk_src = regk_sser_intern_clk;

        /* Step 4.  */
        tr_cfg.sample_size = 1;
        rec_cfg.sample_size = 1;

        /* Step 5.  */
        frm_cfg.wordrate = 7;

        /* Step 6.  */
        tr_cfg.clk_src = regk_sser_intern;
        rec_cfg.clk_src = regk_sser_intern;
        frm_cfg.clk_src = regk_sser_intern;
        tr_cfg.clk_pol = regk_sser_neg;
        frm_cfg.clk_pol = regk_sser_neg;

        /* Step 7.  */
        rec_cfg.clk_pol = regk_sser_pos;

        /* Step 8.  */
        frm_cfg.tr_delay = 1;

        /* Step 9.  */
        frm_cfg.rec_delay = 1;

        /* Step 10.  */
        tr_cfg.sh_dir = regk_sser_msbfirst;
        rec_cfg.sh_dir = regk_sser_msbfirst;

        /* Step 11.  */
        tr_cfg.frm_src = regk_sser_intern;
        rec_cfg.frm_src = regk_sser_intern;

        /* Step 12.  */
        tr_cfg.rate_ctrl = regk_sser_iso;

        /*
         * Step 13.  Note that 0 != tx_null, so we're good regarding
         * the descriptor .md field.
         */
        tr_cfg.eop_stop = 1;

        /* Step 14.  */
        frm_cfg.frame_pin_use = regk_sser_gio1;
        frm_cfg.frame_pin_dir = regk_sser_out;

        /* Step 15.  */
        extra.clkon_en = 1;
        extra.clkoff_en = 1;

        /* Step 16.  We'll modify this value for each "burst".  */
        extra.clkoff_cycles = 7;

        /* Step 17.  */
        cfg.prepare = 1;

        /*
         * Things left out from the documented startup procedure.
         * It's uncertain whether all of these are needed.
         */
        frm_cfg.status_pin_dir = regk_sser_in;
        tr_cfg.mode = regk_sser_hispeed;
        rec_cfg.mode = regk_sser_hispeed;
        frm_cfg.out_on = regk_sser_intern_tb;
        frm_cfg.out_off = regk_sser_rec;
        frm_cfg.type = regk_sser_level;
        tr_cfg.use_dma = 1;
        tr_cfg.urun_stop = 1;
        rec_cfg.orun_stop = 1;
        rec_cfg.use_dma = 1;
        rec_cfg.fifo_thr = regk_sser_inf;
        frm_cfg.early_wend = regk_sser_yes;
        cfg.clk_dir = regk_sser_out;

        tr_cfg.data_pin_use = regk_sser_dout;
        cfg.base_freq = regk_sser_f100;

        REG_WR_SSER(rw_frm_cfg, frm_cfg);
        REG_WR_SSER(rw_tr_cfg, tr_cfg);
        REG_WR_SSER(rw_rec_cfg, rec_cfg);
        REG_WR_SSER(rw_extra, extra);
        REG_WR_SSER(rw_cfg, cfg);
        hw->frm_cfg = frm_cfg;
        hw->tr_cfg = tr_cfg;
        hw->rec_cfg = rec_cfg;
        hw->extra = extra;
        hw->cfg = cfg;

        crisv32_spi_sser_set_speed_Hz(hw, hw->max_speed_Hz);

        ret = request_irq(hw->sser.irq, sser_interrupt, 0, "sser", hw);
        if (ret != 0)
                goto noirq;

        ret = request_irq(hw->dmain.irq, rec_dma_interrupt, 0, "sser rec", hw);
        if (ret != 0)
                goto free_outirq;

        crisv32_reset_dma_hw(hw->dmain.regi);
        crisv32_reset_dma_hw(hw->dmaout.regi);
        return 0;

  free_outirq:
        free_irq(hw->sser.irq, hw);
  noirq:
        return ret;
}

/* SPI-master setup function for non-DMA.  */

static int crisv32_spi_sser_regs_master_setup(struct spi_device *spi)
{
        struct crisv32_spi_hw_info *hw = spidev_to_hw(spi);
        struct spi_bitbang *bitbang = spi_master_get_devdata(spi->master);
        int ret = 0;

        /* Just do a little initial constraining checks.  */
        if (spi->bits_per_word == 0)
                spi->bits_per_word = 8;

        if (spi->bits_per_word != 8)
                return -EINVAL;

        bitbang->chipselect = (spi->mode & SPI_CS_HIGH) != 0
                ? crisv32_spi_sser_chip_select_active_high
                : crisv32_spi_sser_chip_select_active_low;

        if (hw->max_speed_Hz == 0) {
                u32 max_speed_Hz;

                /*
                 * At this time; at the first call to the SPI master
                 * setup function, spi->max_speed_hz reflects the
                 * board-init value.  It will be changed later on by
                 * the protocol master, but at the master setup call
                 * is the only time we actually get to see the hw max
                 * and thus a reasonable time to init the hw field.
                 */

                /* The module parameter overrides everything.  */
                if (crisv32_spi_speed_limit_Hz != 0)
                        max_speed_Hz = crisv32_spi_speed_limit_Hz;
                /*
                 * I never could get hispeed mode to work for non-DMA.
                 * We adjust the max speed here (where we could
                 * presumably fix it), not in the board info file.
                 */
                else if (spi->max_speed_hz > 16667000)
                        max_speed_Hz = 16667000;
                else
                        max_speed_Hz = spi->max_speed_hz;

                hw->max_speed_Hz = max_speed_Hz;
                spi->max_speed_hz = max_speed_Hz;

                /*
                 * We also do one-time initialization of the hardware at this
                 * point.  We could defer to the return to the probe-function
                 * from spi_bitbang_start, but other hardware setup (like
                 * subsequent calls to this function before that) would have
                 * to be deferred until then too.
                 */
                ret = crisv32_setup_spi_sser_for_reg_access(hw);
                if (ret != 0)
                        return ret;

                ret = spi_bitbang_setup(spi);
                if (ret != 0)
                        return ret;

                dev_info(&spi->dev,
                         "CRIS v32 SPI driver for sser%d\n",
                         spi->master->bus_num);
        }

        return 0;
}

/*
 * SPI-master setup_transfer-function used for both DMA and non-DMA
 * (single function for DMA, together with spi_bitbang_setup_transfer
 * for non-DMA).
 */

static int crisv32_spi_sser_common_setup_transfer(struct spi_device *spi,
                                                  struct spi_transfer *t)
{
        struct crisv32_spi_hw_info *hw = spidev_to_hw(spi);
        u8 bits_per_word;
        u32 hz;
        int ret = 0;

        if (t) {
                bits_per_word = t->bits_per_word;
                hz = t->speed_hz;
        } else {
                bits_per_word = 0;
                hz = 0;
        }

        if (bits_per_word == 0)
                bits_per_word = spi->bits_per_word;

        if (bits_per_word != 8)
                return -EINVAL;

        if (hz == 0)
                hz = spi->max_speed_hz;

        if (hz != hw->effective_speed_kHz*1000 && hz != 0)
                ret = crisv32_spi_sser_set_speed_Hz(hw, hz);

        return ret;
}

/* Helper for a SPI-master setup_transfer function for non-DMA.  */

static int crisv32_spi_sser_regs_setup_transfer(struct spi_device *spi,
                                                struct spi_transfer *t)
{
        int ret = crisv32_spi_sser_common_setup_transfer(spi, t);

        if (ret != 0)
                return ret;

        /* Set up the loop-over-buffer parts.  */
        return spi_bitbang_setup_transfer (spi, t);
}

/* SPI-master setup function for DMA.  */

static int crisv32_spi_sser_dma_master_setup(struct spi_device *spi)
{
        /*
         * As we don't dispatch to the spi_bitbang default function,
         * we need to do whatever tests it does; keep it in sync.  On
         * the bright side, we can use the spi->controller_state slot;
         * we use it for DMA:able memory for the descriptors and
         * temporary buffers to copy non-DMA:able transfers.
         */
        struct crisv32_spi_hw_info *hw = spidev_to_hw(spi);
        struct spi_bitbang *bitbang = spi_master_get_devdata(spi->master);
        struct crisv32_spi_dma_cs *cs;
        u32 dmasize;
        int ret = 0;

        if (hw->max_speed_Hz == 0) {
                struct crisv32_spi_dma_descrs *descrp;
                u32 descrp_dma;
                u32 max_speed_Hz;

                /* The module parameter overrides everything.  */
                if (crisv32_spi_speed_limit_Hz != 0)
                        max_speed_Hz = crisv32_spi_speed_limit_Hz;
                /*
                 * See comment at corresponding statement in
                 * crisv32_spi_sser_regs_master_setup.
                 */
                else
                        max_speed_Hz = spi->max_speed_hz;

                hw->max_speed_Hz = max_speed_Hz;
                spi->max_speed_hz = max_speed_Hz;

                ret = crisv32_setup_spi_sser_for_dma_access(hw);
                if (ret != 0)
                        return ret;

                /* Allocate some extra for necessary alignment.  */
                dmasize = sizeof *cs + 31
                        + sizeof(struct crisv32_spi_dma_descrs);

                cs = kzalloc(dmasize, GFP_KERNEL | GFP_DMA);
                if (cs == NULL)
                        return -ENOMEM;

                /*
                 * Make descriptors aligned within the allocated area,
                 * some-place after cs.
                 */
                descrp = (struct crisv32_spi_dma_descrs *)
                        (((u32) (cs + 1) + 31) & ~31);
                descrp_dma = virt_to_phys(descrp);

                /* Set up the "constant" parts of the descriptors.  */
                descrp->out_descr.eol = 1;
                descrp->out_descr.intr = 1;
                descrp->out_descr.out_eop = 1;
                descrp->out_ctxt.saved_data = (dma_descr_data *)
                  (descrp_dma
                   + offsetof(struct crisv32_spi_dma_descrs, out_descr));
                descrp->out_ctxt.next = 0;

                descrp->in_descr.eol = 1;
                descrp->in_descr.intr = 1;
                descrp->in_ctxt.saved_data = (dma_descr_data *)
                        (descrp_dma
                         + offsetof(struct crisv32_spi_dma_descrs, in_descr));
                descrp->in_ctxt.next = 0;

                cs->descrp = descrp;
                spi->controller_state = cs;

                init_completion(&hw->dma_done);

                dev_info(&spi->dev,
                         "CRIS v32 SPI driver for sser%d/DMA\n",
                         spi->master->bus_num);
        }

        /* Do our extra constraining checks.  */
        if (spi->bits_per_word == 0)
                spi->bits_per_word = 8;

        if (spi->bits_per_word != 8)
                return -EINVAL;

        /* SPI_LSB_FIRST deliberately left out, and we only support mode 3.  */
        if ((spi->mode & ~(SPI_TX_1|SPI_CS_HIGH)) != SPI_MODE_3)
                return -EINVAL;

        bitbang->chipselect = (spi->mode & SPI_CS_HIGH) != 0
                ? crisv32_spi_sser_chip_select_active_high
                : crisv32_spi_sser_chip_select_active_low;

        ret = bitbang->setup_transfer(spi, NULL);
        if (ret != 0)
                return ret;

        /* Remember to de-assert chip-select before the first transfer.  */
        spin_lock(&bitbang->lock);
        if (!bitbang->busy) {
                bitbang->chipselect(spi, BITBANG_CS_INACTIVE);
                ndelay(hw->half_cycle_delay_ns);
        }
        spin_unlock(&bitbang->lock);

        return 0;
}

/* SPI-master cleanup function for DMA.  */

static void crisv32_spi_sser_dma_cleanup(struct spi_device *spi)
{
        kfree(spi->controller_state);
        spi->controller_state = NULL;
}

/*
 * Set up DMA transmitter descriptors for a chunk of data.
 * The caller is responsible for working around TR 106.
 */
static void crisv32_spi_sser_setup_dma_descr_out(u32 regi,
                                                 struct crisv32_spi_dma_cs *cs,
                                                 u32 out_phys, u32 chunk_len)
{
        BUG_ON(chunk_len > DMA_CHUNKSIZ);
        struct crisv32_spi_dma_descrs *descrp = cs->descrp;
        u32 descrp_dma = virt_to_phys(descrp);

        descrp->out_descr.buf = (u8 *) out_phys;
        descrp->out_descr.after = (u8 *) out_phys + chunk_len;
        descrp->out_ctxt.saved_data_buf = (u8 *) out_phys;

        DMA_START_CONTEXT(regi,
                          descrp_dma
                          + offsetof(struct crisv32_spi_dma_descrs, out_ctxt));
}

/*
 * Set up DMA receiver descriptors for a chunk of data.
 * Also, work around TR 106.
 */
static void crisv32_spi_sser_setup_dma_descr_in(u32 regi_dmain,
                                                struct crisv32_spi_dma_cs *cs,
                                                u32 in_phys, u32 chunk_len)
{
        BUG_ON(chunk_len > DMA_CHUNKSIZ);
        struct crisv32_spi_dma_descrs *descrp = cs->descrp;
        u32 descrp_dma = virt_to_phys(descrp);

        descrp->in_descr.buf = (u8 *) in_phys;
        descrp->in_descr.after = (u8 *) in_phys + chunk_len;
        descrp->in_ctxt.saved_data_buf = (u8 *) in_phys;

        flush_dma_descr(&descrp->in_descr, 1);

        DMA_START_CONTEXT(regi_dmain,
                          descrp_dma
                          + offsetof(struct crisv32_spi_dma_descrs, in_ctxt));
}

/*
 * SPI-bitbang txrx_bufs function for DMA.
 * FIXME: We have SG DMA descriptors; use them.
 * (Requires abandoning the spi_bitbang framework if done reasonably.)
 */
static int crisv32_spi_sser_dma_txrx_bufs(struct spi_device *spi,
                                          struct spi_transfer *t)
{
        struct crisv32_spi_dma_cs *cs = spi->controller_state;
        struct crisv32_spi_hw_info *hw = spidev_to_hw(spi);
        u32 len = t->len;
        reg_sser_rw_cfg cfg = hw->cfg;
        reg_sser_rw_tr_cfg tr_cfg = hw->tr_cfg;
        reg_sser_rw_rec_cfg rec_cfg = hw->rec_cfg;
        reg_sser_rw_extra extra = hw->extra;
        u32 regi_sser = hw->sser.regi;
        u32 dmain = 0;
        u32 dmaout = 0;
        u32 regi_dmain = hw->dmain.regi;
        u8 *rx_buf = t->rx_buf;

        /*
         * Using IRQ+completion is measured to give an overhead of 14
         * us, so let's instead busy-wait for the time that would be
         * wasted anyway, and get back sooner.  We're not counting in
         * other overhead such as the DMA descriptor in the
         * time-expression, which causes us to use busy-wait for
         * data-lengths that actually take a bit longer than
         * IRQ_USAGE_THRESHOLD_NS.  Still, with IRQ_USAGE_THRESHOLD_NS
         * = 14000, the threshold is for 20 MHz => 35 bytes, 25 => 44
         * and 50 => 88 and the typical SPI transfer lengths for
         * SDcard are { 1, 2, 7, 512 } bytes so a more complicated
         * would likely give nothing but worse performance due to
         * complexity.
         */
        int use_irq = len * hw->half_cycle_delay_ns
                > IRQ_USAGE_THRESHOLD_NS / 8 / 2;

        if (len > DMA_CHUNKSIZ) {
                /*
                 * It should be quite easy to adjust the code if the need
                 * arises for something much larger than the preallocated
                 * buffers (which could themselves easily just be increased)
                 * but still what fits in extra.clkoff_cycles: kmalloc a
                 * temporary dmaable buffer in this function and free it at
                 * the end.  No need to optimize rare requests.  Until then,
                 * we'll keep the code as simple as performance allows.
                 * Alternatively or if we need to send even larger data,
                 * consider calling self with the required number of "faked"
                 * shorter transfers here.
                 */
                dev_err(&spi->dev,
                        "Trying to transfer %d > max %d bytes:"
                        " need to adjust the SPI driver\n",
                        len, DMA_CHUNKSIZ);
                return -EMSGSIZE;
        }

        /*
         * Need to separately tell the hispeed machinery the number of
         * bits in this transmission.
         */
        extra.clkoff_cycles = len * 8 - 1;

        if (t->tx_buf != NULL) {
                if (t->tx_dma == 0) {
                        memcpy(cs->tx_buf, t->tx_buf, len);
                        dmaout = virt_to_phys(cs->tx_buf);
                } else
                        dmaout = t->tx_dma;

                crisv32_spi_sser_setup_dma_descr_out(hw->dmaout.regi,
                                                     cs, dmaout,
                                                     len);

                /* No need to do anything for TR 106; this DMA only reads.  */
                tr_cfg.tr_en = 1;
                tr_cfg.data_pin_use = regk_sser_dout;
        } else {
                tr_cfg.data_pin_use = (spi->mode & SPI_TX_1)
                        ? regk_sser_gio1 : regk_sser_gio0;
                tr_cfg.tr_en = 0;
        }

        if (rx_buf != 0) {
                if (t->rx_dma == 0)
                        dmain = virt_to_phys(cs->rx_buf);
                else
                        dmain = t->rx_dma;

                crisv32_spi_sser_setup_dma_descr_in(regi_dmain, cs,
                                                    dmain, len);
                rec_cfg.rec_en = 1;

                REG_WRINT_SSER(rw_ack_intr, -1);
                REG_WRINT_DI(rw_ack_intr, -1);

                /*
                 * If we're receiving, use the rec data interrupt from DMA as
                 * a signal that the HW is done.
                 */
                if (use_irq) {
                        reg_sser_rw_intr_mask mask = { .urun = 1 };
                        reg_dma_rw_intr_mask dmask = { .data = 1 };

                        REG_WR_DI(rw_intr_mask, dmask);

                        /*
                         * Catch transmitter underruns too.  We don't
                         * have to conditionalize that on the
                         * transmitter being enabled; it's off when
                         * the transmitter is off.  Any overruns will
                         * be indicated by a timeout, so we don't have
                         * to check for that specifically.
                         */
                        REG_WR_SSER(rw_intr_mask, mask);
                }
        } else {
                rec_cfg.rec_en = 0;

                /*
                 * Ack previous overrun, underrun and tidle interrupts.  Or
                 * why not all.  We'll get orun and urun "normally" due to the
                 * way hispeed is (documented to) work and need to clear them,
                 * and we'll have a tidle from a previous transmit if we used
                 * to both receive and transmit, but now only transmit.
                 */
                REG_WRINT_SSER(rw_ack_intr, -1);

                if (use_irq) {
                        reg_sser_rw_intr_mask mask = { .urun = 1, .tidle = 1 };
                        REG_WR_SSER(rw_intr_mask, mask);
                }
        }

        REG_WR_SSER(rw_rec_cfg, rec_cfg);
        REG_WR_SSER(rw_tr_cfg, tr_cfg);
        REG_WR_SSER(rw_extra, extra);

        /*
         * Barriers are needed to make sure that the completion inits don't
         * migrate past the register writes due to gcc scheduling.
         */
        mb();
        hw->dma_actually_done = 0;
        INIT_COMPLETION(hw->dma_done);
        mb();

        /*
         * Wait until DMA tx FIFO has more than one byte (it reads one
         * directly then one "very quickly") before starting sser tx.
         */
        if (tr_cfg.tr_en) {
                u32 regi_dmaout = hw->dmaout.regi;
                u32 minlen = len > 2 ? 2 : len;
                while ((REG_RD_DO(rw_stat)).buf < minlen)
                        ;
        }

        /* Wait until DMA-in is finished reading the descriptors.  */
        if (rec_cfg.rec_en)
                while (DMA_BUSY(regi_dmain))
                        ;
        /*
         * Wait 3 cycles before enabling (with .prepare = 1).
         * FIXME: Can we cut this by some time already passed?
         */
        ndelay(3 * 2 * hw->half_cycle_delay_ns);
        cfg.en = 1;
        REG_WR_SSER(rw_cfg, cfg);

        /*
         * Wait 3 more cycles plus 30 ns before letting go.
         * FIXME: Can we do something else before but after the
         * previous cfg write and cut this by the time already passed?
         */
        cfg.prepare = 0;
        hw->cfg = cfg;
        ndelay(3 * 2 * hw->half_cycle_delay_ns + 30);

        REG_WR_SSER(rw_cfg, cfg);

        /*, We'll disable sser next the time we change the configuration.  */
        cfg.en = 0;
        cfg.prepare = 1;
        hw->cfg = cfg;

        if (!use_irq) {
                /*
                 * We use a timeout corresponding to one iteration per ns,
                 * which of course is at least five * insns / loop times as
                 * much as reality, but we'll avoid a need for reading hw
                 * timers directly.
                 */
                u32 countdown = IRQ_USAGE_THRESHOLD_NS;

                do
                        if (rec_cfg.rec_en == 0) {
                                /* Using the transmitter only.  */
                                reg_sser_r_intr intr = REG_RD_SSER(r_intr);

                                if (intr.tidle != 0) {
                                        /*
                                         * Almost done...  Just check if we
                                         * had a transmitter underrun too.
                                         */
                                        if (!intr.urun)
                                                goto transmission_done;

                                        /*
                                         * Fall over to the "time is up" case;
                                         * no need to provide a special path
                                         * for the error case.
                                         */
                                        countdown = 1;
                                }
                        } else {
                                /* Using at least the receiver.  */
                                if ((REG_RD_DI(r_intr)).data != 0) {
                                        if ((REG_RD_SSER(r_intr)).urun == 0)
                                                goto transmission_done;
                                        countdown = 1;
                                }
                        }
                while (--countdown != 0);

                /*
                 * The time is up.  Something might be wrong, or perhaps we've
                 * started using data lengths where the threshold was about a
                 * magnitude wrong.  Fall over to IRQ.  Remember not to ack
                 * interrupts here (but always above, before starting), else
                 * we'll have a race condition with the interrupt.
                 */
                if (!rec_cfg.rec_en) {
                        reg_sser_rw_intr_mask mask = { .urun = 1, .tidle = 1 };
                        REG_WR_SSER(rw_intr_mask, mask);
                } else {
                        reg_dma_rw_intr_mask dmask = { .data = 1 };
                        reg_sser_rw_intr_mask mask = { .urun = 1 };

                        /*
                         * Never mind checking for tr being disabled; urun
                         * won't happen then.
                         */
                        REG_WR_SSER(rw_intr_mask, mask);
                        REG_WR_DI(rw_intr_mask, dmask);
                }
        }

        if (!wait_for_completion_timeout(&hw->dma_done, hw->dma_timeout)
            /*
             * Have to keep track manually too, else we'll get a timeout
             * indication for being scheduled out too long, while the
             * completion will still have trigged.
             */
            && !hw->dma_actually_done) {
                u32 regi_dmaout = hw->dmaout.regi;

                /*
                 * Transfer timed out.  Should not happen for a
                 * working controller, except perhaps if the system is
                 * badly conditioned, causing DMA memory bandwidth
                 * starvation.  Not much to do afterwards, but perhaps
                 * reset DMA and sser and hope it works the next time.
                 */
                REG_WRINT_SSER(rw_cfg, 0);
                REG_WR_SSER(rw_cfg, cfg);
                REG_WRINT_SSER(rw_intr_mask, 0);
                REG_WRINT_DI(rw_intr_mask, 0);
                REG_WRINT_SSER(rw_ack_intr, -1);
                crisv32_reset_dma_hw(hw->dmain.regi);
                crisv32_reset_dma_hw(hw->dmaout.regi);

                dev_err(&spi->dev, "timeout %u bytes %u kHz\n",
                        len, hw->effective_speed_kHz);
                dev_err(&spi->dev, "sser=(%x,%x,%x,%x,%x)\n",
                        REG_RDINT_SSER(rw_cfg), REG_RDINT_SSER(rw_tr_cfg),
                        REG_RDINT_SSER(rw_rec_cfg), REG_RDINT_SSER(rw_extra),
                        REG_RDINT_SSER(r_intr));
                dev_err(&spi->dev, "tx=(%x,%x,%x,%x)\n",
                        dmaout, REG_RDINT_DO(rw_stat), REG_RDINT_DO(rw_data),
                        REG_RDINT_DO(r_intr));
                dev_err(&spi->dev, "rx=(%x,%x,%x,%x)\n",
                        dmain, REG_RDINT_DI(rw_stat), REG_RDINT_DI(rw_data),
                        REG_RDINT_DI(r_intr));
                return -EIO;
        }

 transmission_done:
        /* Wait for the last half-cycle of the last cycle.  */
        crisv32_spi_sser_wait_halfabit(hw);

        /* Reset for another call.  */
        REG_WR_SSER(rw_cfg, cfg);

        /*
         * If we had to use the temp DMAable rec buffer, copy it to the right
         * position.
         */
        if (t->rx_buf != 0 && t->rx_dma == 0)
                memcpy (t->rx_buf, cs->rx_buf, len);

        /*
         * All clear.  The interrupt function disabled the interrupt, we don't
         * have to do more.
         */
        return len;
}

/* Platform-device probe function.  */

static int __devinit crisv32_spi_sser_probe(struct platform_device *dev)
{
        struct spi_master *master;
        struct crisv32_spi_sser_devdata *dd;
        struct crisv32_spi_hw_info *hw;
        struct resource *res;
        struct crisv32_spi_sser_controller_data *gc;
        int ret;

        /*
         * We need to get the controller data as a hardware resource,
         * or else it wouldn't be available until *after* the
         * spi_bitbang_start call!
         */
        res = platform_get_resource_byname(dev, 0, "controller_data_ptr");
        if (res == NULL) {
                dev_err(&dev->dev,
                        "can't get controller_data resource at probe\n");
                return -EIO;
        }

        gc = (struct crisv32_spi_sser_controller_data *) res->start;

        master = spi_alloc_master(&dev->dev, sizeof *dd);
        if (master == NULL) {
                dev_err(&dev->dev, "failed to allocate spi master\n");
                ret = -ENOMEM;
                goto err;
        }

        dd = spi_master_get_devdata(master);
        platform_set_drvdata(dev, dd);

        /*
         * The device data asks for this driver, and holds the id
         * number, which must be unique among the same-type devices.
         * We use this as the number of this SPI bus.
         */
        master->bus_num = dev->id;

        /* Setup SPI bitbang adapter hooks.  */
        dd->bitbang.master = spi_master_get(master);
        dd->bitbang.chipselect = crisv32_spi_sser_chip_select_active_low;

        hw = &dd->hw;
        hw->gc = gc;

        /* Pre-spi_bitbang_start setup. */
        if (gc->using_dma) {
                /* Setup DMA and interrupts.  */
                ret = gc->iface_allocate(&hw->sser, &hw->dmain, &hw->dmaout);
                if (ret != 0)
                        goto err_no_regs;

                dd->bitbang.master->setup = crisv32_spi_sser_dma_master_setup;
                dd->bitbang.setup_transfer
                        = crisv32_spi_sser_common_setup_transfer;
                dd->bitbang.txrx_bufs = crisv32_spi_sser_dma_txrx_bufs;
                dd->bitbang.master->cleanup = crisv32_spi_sser_dma_cleanup;
        } else {
                /* Just registers, then.  */
                ret = gc->iface_allocate(&hw->sser, NULL, NULL);
                if (ret != 0)
                        goto err_no_regs;

                dd->bitbang.master->setup
                        = crisv32_spi_sser_regs_master_setup;
                dd->bitbang.setup_transfer
                        = crisv32_spi_sser_regs_setup_transfer;
                dd->bitbang.master->cleanup = spi_bitbang_cleanup;

                /*
                 * We can do all modes pretty simply, but I have no
                 * simple enough way to test them, so I won't.
                 */
                dd->bitbang.txrx_word[SPI_MODE_3]
                        = crisv32_spi_sser_txrx_mode3;
        }

        ret = spi_bitbang_start(&dd->bitbang);
        if (ret)
                goto err_no_bitbang;

        /*
         * We don't have a dev_info here, as initialization that may fail is
         * postponed to the first master->setup call.  It's called from
         * spi_bitbang_start (above), where the call-chain doesn't look too
         * close at error return values; we'll get here successfully anyway,
         * so emitting a separate message here is at most confusing.
         */
        dev_dbg(&dev->dev,
                "CRIS v32 SPI driver for sser%d%s present\n",
                master->bus_num,
                gc->using_dma ? "/DMA" : "");

        return 0;

 err_no_bitbang:
        gc->iface_free();

 err_no_regs:
        platform_set_drvdata(dev, NULL);
        spi_master_put(dd->bitbang.master);

 err:
        return ret;
}

/* Platform-device remove-function.  */

static int __devexit crisv32_spi_sser_remove(struct platform_device *dev)
{
        struct crisv32_spi_sser_devdata *dd = platform_get_drvdata(dev);
        struct crisv32_spi_hw_info *hw = &dd->hw;
        struct crisv32_spi_sser_controller_data *gc = hw->gc;
        int ret;

        /* We need to stop all bitbanging activity separately.  */
        ret = spi_bitbang_stop(&dd->bitbang);
        if (ret != 0)
                return ret;

        spi_master_put(dd->bitbang.master);

        /*
         * If we get here, the queue is empty and there's no activity;
         * it's safe to flip the switch on the interfaces.
         */
        if (gc->using_dma) {
                u32 regi_dmain = hw->dmain.regi;
                u32 regi_dmaout = hw->dmaout.regi;
                u32 regi_sser = hw->sser.regi;

                REG_WRINT_SSER(rw_intr_mask, 0);
                REG_WRINT_DI(rw_intr_mask, 0);
                REG_WRINT_DO(rw_intr_mask, 0);
                hw->cfg.en = 0;
                REG_WR_SSER(rw_cfg, hw->cfg);
                DMA_RESET(regi_dmain);
                DMA_RESET(regi_dmaout);
                free_irq(hw->sser.irq, hw);
                free_irq(hw->dmain.irq, hw);
        }

        gc->iface_free();

        platform_set_drvdata(dev, NULL);
        return 0;
}

/*
 * For the time being, there's no suspend/resume support to care
 * about, so those handlers default to NULL.
 */
static struct platform_driver crisv32_spi_sser_drv = {
        .probe          = crisv32_spi_sser_probe,
        .remove         = __devexit_p(crisv32_spi_sser_remove),
        .driver         = {
                .name   = "spi_crisv32_sser",
                .owner  = THIS_MODULE,
        },
};

/* Module init function.  */

static int __devinit crisv32_spi_sser_init(void)
{
        return platform_driver_register(&crisv32_spi_sser_drv);
}

/* Module exit function.  */

static void __devexit crisv32_spi_sser_exit(void)
{
        platform_driver_unregister(&crisv32_spi_sser_drv);
}

/* Setter function for speed limit.  */

static int crisv32_spi_speed_limit_Hz_setter(const char *val,
                                             struct kernel_param *kp)
{
        char *endp;
        ulong num = simple_strtoul(val, &endp, 0);
        if (endp == val
            || *endp != 0
            || num <= 0
            /*
             * We can't go above 100 MHz speed.  Actually we can't go
             * above 50 MHz using the sser support but it might make
             * sense trying.
             */
            || num > 100000000)
                return -EINVAL;
        *(ulong *) kp->arg = num;
        return 0;
}

module_param_call(crisv32_spi_max_speed_hz,
                  crisv32_spi_speed_limit_Hz_setter, param_get_ulong,
                  &crisv32_spi_speed_limit_Hz, 0644);

module_init(crisv32_spi_sser_init);
module_exit(crisv32_spi_sser_exit);

MODULE_DESCRIPTION("CRIS v32 SPI-SSER Driver");
MODULE_AUTHOR("Hans-Peter Nilsson, <hp@axis.com>");
MODULE_LICENSE("GPL");