arbiter.c

Go to the documentation of this file.

/*
 * Memory arbiter functions. Allocates bandwidth through the
 * arbiter and sets up arbiter breakpoints.
 *
 * The algorithm first assigns slots to the clients that has specified
 * bandwidth (e.g. ethernet) and then the remaining slots are divided
 * on all the active clients.
 *
 * Copyright (c) 2004-2007 Axis Communications AB.
 *
 * The artpec-3 has two arbiters. The memory hierarchy looks like this:
 *
 *
 * CPU DMAs
 *  |   |
 *  |   |
 * --------------    ------------------
 * | foo arbiter|----| Internal memory|
 * --------------    ------------------
 *      |
 * --------------
 * | L2 cache   |
 * --------------
 *             |
 * h264 etc    |
 *    |        |
 *    |        |
 * --------------
 * | bar arbiter|
 * --------------
 *       |
 * ---------
 * | SDRAM |
 * ---------
 *
 */

#include <hwregs/reg_map.h>
#include <hwregs/reg_rdwr.h>
#include <hwregs/marb_foo_defs.h>
#include <hwregs/marb_bar_defs.h>
#include <arbiter.h>
#include <hwregs/intr_vect.h>
#include <linux/interrupt.h>
#include <linux/irq.h>
#include <linux/signal.h>
#include <linux/errno.h>
#include <linux/spinlock.h>
#include <asm/io.h>
#include <asm/irq_regs.h>

#define D(x)

struct crisv32_watch_entry {
  unsigned long instance;
  watch_callback *cb;
  unsigned long start;
  unsigned long end;
  int used;
};

#define NUMBER_OF_BP 4
#define SDRAM_BANDWIDTH 400000000
#define INTMEM_BANDWIDTH 400000000
#define NBR_OF_SLOTS 64
#define NBR_OF_REGIONS 2
#define NBR_OF_CLIENTS 15
#define ARBITERS 2
#define UNASSIGNED 100

struct arbiter {
  unsigned long instance;
  int nbr_regions;
  int nbr_clients;
  int requested_slots[NBR_OF_REGIONS][NBR_OF_CLIENTS];
  int active_clients[NBR_OF_REGIONS][NBR_OF_CLIENTS];
};

static struct crisv32_watch_entry watches[ARBITERS][NUMBER_OF_BP] =
{
  {
  {regi_marb_foo_bp0},
  {regi_marb_foo_bp1},
  {regi_marb_foo_bp2},
  {regi_marb_foo_bp3}
  },
  {
  {regi_marb_bar_bp0},
  {regi_marb_bar_bp1},
  {regi_marb_bar_bp2},
  {regi_marb_bar_bp3}
  }
};

struct arbiter arbiters[ARBITERS] =
{
  { /* L2 cache arbiter */
    .instance = regi_marb_foo,
    .nbr_regions = 2,
    .nbr_clients = 15
  },
  { /* DDR2 arbiter */
    .instance = regi_marb_bar,
    .nbr_regions = 1,
    .nbr_clients = 9
  }
};

static int max_bandwidth[NBR_OF_REGIONS] = {SDRAM_BANDWIDTH, INTMEM_BANDWIDTH};

DEFINE_SPINLOCK(arbiter_lock);

static irqreturn_t
crisv32_foo_arbiter_irq(int irq, void *dev_id);
static irqreturn_t
crisv32_bar_arbiter_irq(int irq, void *dev_id);

/*
 * "I'm the arbiter, I know the score.
 *  From square one I'll be watching all 64."
 * (memory arbiter slots, that is)
 *
 *  Or in other words:
 * Program the memory arbiter slots for "region" according to what's
 * in requested_slots[] and active_clients[], while minimizing
 * latency. A caller may pass a non-zero positive amount for
 * "unused_slots", which must then be the unallocated, remaining
 * number of slots, free to hand out to any client.
 */

static void crisv32_arbiter_config(int arbiter, int region, int unused_slots)
{
    int slot;
    int client;
    int interval = 0;

    /*
     * This vector corresponds to the hardware arbiter slots (see
     * the hardware documentation for semantics). We initialize
     * each slot with a suitable sentinel value outside the valid
     * range {0 .. NBR_OF_CLIENTS - 1} and replace them with
     * client indexes. Then it's fed to the hardware.
     */
    s8 val[NBR_OF_SLOTS];

    for (slot = 0; slot < NBR_OF_SLOTS; slot++)
        val[slot] = -1;

    for (client = 0; client < arbiters[arbiter].nbr_clients; client++) {
        int pos;
        /* Allocate the requested non-zero number of slots, but
         * also give clients with zero-requests one slot each
         * while stocks last. We do the latter here, in client
         * order. This makes sure zero-request clients are the
         * first to get to any spare slots, else those slots
         * could, when bandwidth is allocated close to the limit,
         * all be allocated to low-index non-zero-request clients
         * in the default-fill loop below. Another positive but
         * secondary effect is a somewhat better spread of the
         * zero-bandwidth clients in the vector, avoiding some of
         * the latency that could otherwise be caused by the
         * partitioning of non-zero-bandwidth clients at low
         * indexes and zero-bandwidth clients at high
         * indexes. (Note that this spreading can only affect the
         * unallocated bandwidth.)  All the above only matters for
         * memory-intensive situations, of course.
         */
        if (!arbiters[arbiter].requested_slots[region][client]) {
        /*
         * Skip inactive clients. Also skip zero-slot
         * allocations in this pass when there are no known
         * free slots.
         */
        if (!arbiters[arbiter].active_clients[region][client] ||
                unused_slots <= 0)
            continue;

        unused_slots--;

        /* Only allocate one slot for this client. */
        interval = NBR_OF_SLOTS;
        } else
        interval = NBR_OF_SLOTS /
            arbiters[arbiter].requested_slots[region][client];

        pos = 0;
        while (pos < NBR_OF_SLOTS) {
        if (val[pos] >= 0)
           pos++;
        else {
            val[pos] = client;
            pos += interval;
        }
        }
    }

    client = 0;
    for (slot = 0; slot < NBR_OF_SLOTS; slot++) {
        /*
         * Allocate remaining slots in round-robin
         * client-number order for active clients. For this
         * pass, we ignore requested bandwidth and previous
         * allocations.
         */
        if (val[slot] < 0) {
            int first = client;
            while (!arbiters[arbiter].active_clients[region][client]) {
                client = (client + 1) %
                    arbiters[arbiter].nbr_clients;
                if (client == first)
                   break;
            }
            val[slot] = client;
            client = (client + 1) % arbiters[arbiter].nbr_clients;
        }
        if (arbiter == 0) {
            if (region == EXT_REGION)
                REG_WR_INT_VECT(marb_foo, regi_marb_foo,
                    rw_l2_slots, slot, val[slot]);
            else if (region == INT_REGION)
                REG_WR_INT_VECT(marb_foo, regi_marb_foo,
                    rw_intm_slots, slot, val[slot]);
        } else {
            REG_WR_INT_VECT(marb_bar, regi_marb_bar,
                rw_ddr2_slots, slot, val[slot]);
        }
    }
}

extern char _stext, _etext;

static void crisv32_arbiter_init(void)
{
    static int initialized;

    if (initialized)
        return;

    initialized = 1;

    /*
     * CPU caches are always set to active, but with zero
     * bandwidth allocated. It should be ok to allocate zero
     * bandwidth for the caches, because DMA for other channels
     * will supposedly finish, once their programmed amount is
     * done, and then the caches will get access according to the
     * "fixed scheme" for unclaimed slots. Though, if for some
     * use-case somewhere, there's a maximum CPU latency for
     * e.g. some interrupt, we have to start allocating specific
     * bandwidth for the CPU caches too.
     */
    arbiters[0].active_clients[EXT_REGION][11] = 1;
    arbiters[0].active_clients[EXT_REGION][12] = 1;
    crisv32_arbiter_config(0, EXT_REGION, 0);
    crisv32_arbiter_config(0, INT_REGION, 0);
    crisv32_arbiter_config(1, EXT_REGION, 0);

    if (request_irq(MEMARB_FOO_INTR_VECT, crisv32_foo_arbiter_irq,
            IRQF_DISABLED, "arbiter", NULL))
        printk(KERN_ERR "Couldn't allocate arbiter IRQ\n");

    if (request_irq(MEMARB_BAR_INTR_VECT, crisv32_bar_arbiter_irq,
            IRQF_DISABLED, "arbiter", NULL))
        printk(KERN_ERR "Couldn't allocate arbiter IRQ\n");

#ifndef CONFIG_ETRAX_KGDB
    /* Global watch for writes to kernel text segment. */
    crisv32_arbiter_watch(virt_to_phys(&_stext), &_etext - &_stext,
        MARB_CLIENTS(arbiter_all_clients, arbiter_bar_all_clients),
                  arbiter_all_write, NULL);
#endif

    /* Set up max burst sizes by default */
    REG_WR_INT(marb_bar, regi_marb_bar, rw_h264_rd_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_h264_wr_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_ccd_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_vin_wr_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_vin_rd_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_sclr_rd_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_vout_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_sclr_fifo_burst, 3);
    REG_WR_INT(marb_bar, regi_marb_bar, rw_l2cache_burst, 3);
}

int crisv32_arbiter_allocate_bandwidth(int client, int region,
                      unsigned long bandwidth)
{
    int i;
    int total_assigned = 0;
    int total_clients = 0;
    int req;
    int arbiter = 0;

    crisv32_arbiter_init();

    if (client & 0xffff0000) {
        arbiter = 1;
        client >>= 16;
    }

    for (i = 0; i < arbiters[arbiter].nbr_clients; i++) {
        total_assigned += arbiters[arbiter].requested_slots[region][i];
        total_clients += arbiters[arbiter].active_clients[region][i];
    }

    /* Avoid division by 0 for 0-bandwidth requests. */
    req = bandwidth == 0
        ? 0 : NBR_OF_SLOTS / (max_bandwidth[region] / bandwidth);

    /*
     * We make sure that there are enough slots only for non-zero
     * requests. Requesting 0 bandwidth *may* allocate slots,
     * though if all bandwidth is allocated, such a client won't
     * get any and will have to rely on getting memory access
     * according to the fixed scheme that's the default when one
     * of the slot-allocated clients doesn't claim their slot.
     */
    if (total_assigned + req > NBR_OF_SLOTS)
       return -ENOMEM;

    arbiters[arbiter].active_clients[region][client] = 1;
    arbiters[arbiter].requested_slots[region][client] = req;
    crisv32_arbiter_config(arbiter, region, NBR_OF_SLOTS - total_assigned);

    /* Propagate allocation from foo to bar */
    if (arbiter == 0)
        crisv32_arbiter_allocate_bandwidth(8 << 16,
            EXT_REGION, bandwidth);
    return 0;
}

/*
 * Main entry for bandwidth deallocation.
 *
 * Strictly speaking, for a somewhat constant set of clients where
 * each client gets a constant bandwidth and is just enabled or
 * disabled (somewhat dynamically), no action is necessary here to
 * avoid starvation for non-zero-allocation clients, as the allocated
 * slots will just be unused. However, handing out those unused slots
 * to active clients avoids needless latency if the "fixed scheme"
 * would give unclaimed slots to an eager low-index client.
 */

void crisv32_arbiter_deallocate_bandwidth(int client, int region)
{
    int i;
    int total_assigned = 0;
    int arbiter = 0;

    if (client & 0xffff0000)
        arbiter = 1;

    arbiters[arbiter].requested_slots[region][client] = 0;
    arbiters[arbiter].active_clients[region][client] = 0;

    for (i = 0; i < arbiters[arbiter].nbr_clients; i++)
        total_assigned += arbiters[arbiter].requested_slots[region][i];

    crisv32_arbiter_config(arbiter, region, NBR_OF_SLOTS - total_assigned);
}

int crisv32_arbiter_watch(unsigned long start, unsigned long size,
              unsigned long clients, unsigned long accesses,
              watch_callback *cb)
{
    int i;
    int arbiter;
    int used[2];
    int ret = 0;

    crisv32_arbiter_init();

    if (start > 0x80000000) {
        printk(KERN_ERR "Arbiter: %lX doesn't look like a "
            "physical address", start);
        return -EFAULT;
    }

    spin_lock(&arbiter_lock);

    if (clients & 0xffff)
        used[0] = 1;
    if (clients & 0xffff0000)
        used[1] = 1;

    for (arbiter = 0; arbiter < ARBITERS; arbiter++) {
        if (!used[arbiter])
            continue;

        for (i = 0; i < NUMBER_OF_BP; i++) {
            if (!watches[arbiter][i].used) {
                unsigned intr_mask;
                if (arbiter)
                    intr_mask = REG_RD_INT(marb_bar,
                        regi_marb_bar, rw_intr_mask);
                else
                    intr_mask = REG_RD_INT(marb_foo,
                        regi_marb_foo, rw_intr_mask);

                watches[arbiter][i].used = 1;
                watches[arbiter][i].start = start;
                watches[arbiter][i].end = start + size;
                watches[arbiter][i].cb = cb;

                ret |= (i + 1) << (arbiter + 8);
                if (arbiter) {
                    REG_WR_INT(marb_bar_bp,
                        watches[arbiter][i].instance,
                        rw_first_addr,
                        watches[arbiter][i].start);
                    REG_WR_INT(marb_bar_bp,
                        watches[arbiter][i].instance,
                        rw_last_addr,
                        watches[arbiter][i].end);
                    REG_WR_INT(marb_bar_bp,
                        watches[arbiter][i].instance,
                        rw_op, accesses);
                    REG_WR_INT(marb_bar_bp,
                        watches[arbiter][i].instance,
                        rw_clients,
                        clients & 0xffff);
                } else {
                    REG_WR_INT(marb_foo_bp,
                        watches[arbiter][i].instance,
                        rw_first_addr,
                        watches[arbiter][i].start);
                    REG_WR_INT(marb_foo_bp,
                        watches[arbiter][i].instance,
                        rw_last_addr,
                        watches[arbiter][i].end);
                    REG_WR_INT(marb_foo_bp,
                        watches[arbiter][i].instance,
                        rw_op, accesses);
                    REG_WR_INT(marb_foo_bp,
                        watches[arbiter][i].instance,
                        rw_clients, clients >> 16);
                }

                if (i == 0)
                    intr_mask |= 1;
                else if (i == 1)
                    intr_mask |= 2;
                else if (i == 2)
                    intr_mask |= 4;
                else if (i == 3)
                    intr_mask |= 8;

                if (arbiter)
                    REG_WR_INT(marb_bar, regi_marb_bar,
                        rw_intr_mask, intr_mask);
                else
                    REG_WR_INT(marb_foo, regi_marb_foo,
                        rw_intr_mask, intr_mask);

                spin_unlock(&arbiter_lock);

                break;
            }
        }
    }
    spin_unlock(&arbiter_lock);
    if (ret)
        return ret;
    else
        return -ENOMEM;
}

int crisv32_arbiter_unwatch(int id)
{
    int arbiter;
    int intr_mask;

    crisv32_arbiter_init();

    spin_lock(&arbiter_lock);

    for (arbiter = 0; arbiter < ARBITERS; arbiter++) {
        int id2;

        if (arbiter)
            intr_mask = REG_RD_INT(marb_bar, regi_marb_bar,
                rw_intr_mask);
        else
            intr_mask = REG_RD_INT(marb_foo, regi_marb_foo,
                rw_intr_mask);

        id2 = (id & (0xff << (arbiter + 8))) >> (arbiter + 8);
        if (id2 == 0)
            continue;
        id2--;
        if ((id2 >= NUMBER_OF_BP) || (!watches[arbiter][id2].used)) {
            spin_unlock(&arbiter_lock);
            return -EINVAL;
        }

        memset(&watches[arbiter][id2], 0,
            sizeof(struct crisv32_watch_entry));

        if (id2 == 0)
            intr_mask &= ~1;
        else if (id2 == 1)
            intr_mask &= ~2;
        else if (id2 == 2)
            intr_mask &= ~4;
        else if (id2 == 3)
            intr_mask &= ~8;

        if (arbiter)
            REG_WR_INT(marb_bar, regi_marb_bar, rw_intr_mask,
                intr_mask);
        else
            REG_WR_INT(marb_foo, regi_marb_foo, rw_intr_mask,
                intr_mask);
    }

    spin_unlock(&arbiter_lock);
    return 0;
}

extern void show_registers(struct pt_regs *regs);


static irqreturn_t
crisv32_foo_arbiter_irq(int irq, void *dev_id)
{
    reg_marb_foo_r_masked_intr masked_intr =
        REG_RD(marb_foo, regi_marb_foo, r_masked_intr);
    reg_marb_foo_bp_r_brk_clients r_clients;
    reg_marb_foo_bp_r_brk_addr r_addr;
    reg_marb_foo_bp_r_brk_op r_op;
    reg_marb_foo_bp_r_brk_first_client r_first;
    reg_marb_foo_bp_r_brk_size r_size;
    reg_marb_foo_bp_rw_ack ack = {0};
    reg_marb_foo_rw_ack_intr ack_intr = {
        .bp0 = 1, .bp1 = 1, .bp2 = 1, .bp3 = 1
    };
    struct crisv32_watch_entry *watch;
    unsigned arbiter = (unsigned)dev_id;

    masked_intr = REG_RD(marb_foo, regi_marb_foo, r_masked_intr);

    if (masked_intr.bp0)
        watch = &watches[arbiter][0];
    else if (masked_intr.bp1)
        watch = &watches[arbiter][1];
    else if (masked_intr.bp2)
        watch = &watches[arbiter][2];
    else if (masked_intr.bp3)
        watch = &watches[arbiter][3];
    else
        return IRQ_NONE;

    /* Retrieve all useful information and print it. */
    r_clients = REG_RD(marb_foo_bp, watch->instance, r_brk_clients);
    r_addr = REG_RD(marb_foo_bp, watch->instance, r_brk_addr);
    r_op = REG_RD(marb_foo_bp, watch->instance, r_brk_op);
    r_first = REG_RD(marb_foo_bp, watch->instance, r_brk_first_client);
    r_size = REG_RD(marb_foo_bp, watch->instance, r_brk_size);

    printk(KERN_DEBUG "Arbiter IRQ\n");
    printk(KERN_DEBUG "Clients %X addr %X op %X first %X size %X\n",
           REG_TYPE_CONV(int, reg_marb_foo_bp_r_brk_clients, r_clients),
           REG_TYPE_CONV(int, reg_marb_foo_bp_r_brk_addr, r_addr),
           REG_TYPE_CONV(int, reg_marb_foo_bp_r_brk_op, r_op),
           REG_TYPE_CONV(int, reg_marb_foo_bp_r_brk_first_client, r_first),
           REG_TYPE_CONV(int, reg_marb_foo_bp_r_brk_size, r_size));

    REG_WR(marb_foo_bp, watch->instance, rw_ack, ack);
    REG_WR(marb_foo, regi_marb_foo, rw_ack_intr, ack_intr);

    printk(KERN_DEBUG "IRQ occurred at %X\n", (unsigned)get_irq_regs());

    if (watch->cb)
        watch->cb();

    return IRQ_HANDLED;
}

static irqreturn_t
crisv32_bar_arbiter_irq(int irq, void *dev_id)
{
    reg_marb_bar_r_masked_intr masked_intr =
        REG_RD(marb_bar, regi_marb_bar, r_masked_intr);
    reg_marb_bar_bp_r_brk_clients r_clients;
    reg_marb_bar_bp_r_brk_addr r_addr;
    reg_marb_bar_bp_r_brk_op r_op;
    reg_marb_bar_bp_r_brk_first_client r_first;
    reg_marb_bar_bp_r_brk_size r_size;
    reg_marb_bar_bp_rw_ack ack = {0};
    reg_marb_bar_rw_ack_intr ack_intr = {
        .bp0 = 1, .bp1 = 1, .bp2 = 1, .bp3 = 1
    };
    struct crisv32_watch_entry *watch;
    unsigned arbiter = (unsigned)dev_id;

    masked_intr = REG_RD(marb_bar, regi_marb_bar, r_masked_intr);

    if (masked_intr.bp0)
        watch = &watches[arbiter][0];
    else if (masked_intr.bp1)
        watch = &watches[arbiter][1];
    else if (masked_intr.bp2)
        watch = &watches[arbiter][2];
    else if (masked_intr.bp3)
        watch = &watches[arbiter][3];
    else
        return IRQ_NONE;

    /* Retrieve all useful information and print it. */
    r_clients = REG_RD(marb_bar_bp, watch->instance, r_brk_clients);
    r_addr = REG_RD(marb_bar_bp, watch->instance, r_brk_addr);
    r_op = REG_RD(marb_bar_bp, watch->instance, r_brk_op);
    r_first = REG_RD(marb_bar_bp, watch->instance, r_brk_first_client);
    r_size = REG_RD(marb_bar_bp, watch->instance, r_brk_size);

    printk(KERN_DEBUG "Arbiter IRQ\n");
    printk(KERN_DEBUG "Clients %X addr %X op %X first %X size %X\n",
           REG_TYPE_CONV(int, reg_marb_bar_bp_r_brk_clients, r_clients),
           REG_TYPE_CONV(int, reg_marb_bar_bp_r_brk_addr, r_addr),
           REG_TYPE_CONV(int, reg_marb_bar_bp_r_brk_op, r_op),
           REG_TYPE_CONV(int, reg_marb_bar_bp_r_brk_first_client, r_first),
           REG_TYPE_CONV(int, reg_marb_bar_bp_r_brk_size, r_size));

    REG_WR(marb_bar_bp, watch->instance, rw_ack, ack);
    REG_WR(marb_bar, regi_marb_bar, rw_ack_intr, ack_intr);

    printk(KERN_DEBUG "IRQ occurred at %X\n", (unsigned)get_irq_regs()->erp);

    if (watch->cb)
        watch->cb();

    return IRQ_HANDLED;
}