drm/amdkfd: Add memory exception handling

[linux/fpc-iii.git] / drivers / gpu / drm / msm / mdp / mdp5 / mdp5_smp.c

/*
 * Copyright (c) 2014, The Linux Foundation. All rights reserved.
 * Copyright (C) 2013 Red Hat
 * Author: Rob Clark <robdclark@gmail.com>
 *
 * 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 program is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
 * more details.
 *
 * You should have received a copy of the GNU General Public License along with
 * this program.  If not, see <http://www.gnu.org/licenses/>.
 */


#include "mdp5_kms.h"
#include "mdp5_smp.h"


/* SMP - Shared Memory Pool
 *
 * These are shared between all the clients, where each plane in a
 * scanout buffer is a SMP client.  Ie. scanout of 3 plane I420 on
 * pipe VIG0 => 3 clients: VIG0_Y, VIG0_CB, VIG0_CR.
 *
 * Based on the size of the attached scanout buffer, a certain # of
 * blocks must be allocated to that client out of the shared pool.
 *
 * In some hw, some blocks are statically allocated for certain pipes
 * and CANNOT be re-allocated (eg: MMB0 and MMB1 both tied to RGB0).
 *
 * For each block that can be dynamically allocated, it can be either
 * free, or pending/in-use by a client. The updates happen in three steps:
 *
 *  1) mdp5_smp_request():
 *     When plane scanout is setup, calculate required number of
 *     blocks needed per client, and request.  Blocks not inuse or
 *     pending by any other client are added to client's pending
 *     set.
 *
 *  2) mdp5_smp_configure():
 *     As hw is programmed, before FLUSH, MDP5_MDP_SMP_ALLOC registers
 *     are configured for the union(pending, inuse)
 *
 *  3) mdp5_smp_commit():
 *     After next vblank, copy pending -> inuse.  Optionally update
 *     MDP5_SMP_ALLOC registers if there are newly unused blocks
 *
 * On the next vblank after changes have been committed to hw, the
 * client's pending blocks become it's in-use blocks (and no-longer
 * in-use blocks become available to other clients).
 *
 * btw, hurray for confusing overloaded acronyms!  :-/
 *
 * NOTE: for atomic modeset/pageflip NONBLOCK operations, step #1
 * should happen at (or before)? atomic->check().  And we'd need
 * an API to discard previous requests if update is aborted or
 * (test-only).
 *
 * TODO would perhaps be nice to have debugfs to dump out kernel
 * inuse and pending state of all clients..
 */

struct mdp5_smp {
        struct drm_device *dev;

        int blk_cnt;
        int blk_size;

        spinlock_t state_lock;
        mdp5_smp_state_t state; /* to track smp allocation amongst pipes: */

        struct mdp5_client_smp_state client_state[MAX_CLIENTS];
};

static inline
struct mdp5_kms *get_kms(struct mdp5_smp *smp)
{
        struct msm_drm_private *priv = smp->dev->dev_private;

        return to_mdp5_kms(to_mdp_kms(priv->kms));
}

static inline u32 pipe2client(enum mdp5_pipe pipe, int plane)
{
#define CID_UNUSED      0

        if (WARN_ON(plane >= pipe2nclients(pipe)))
                return CID_UNUSED;

        /*
         * Note on SMP clients:
         * For ViG pipes, fetch Y/Cr/Cb-components clients are always
         * consecutive, and in that order.
         *
         * e.g.:
         * if mdp5_cfg->smp.clients[SSPP_VIG0] = N,
         *      Y  plane's client ID is N
         *      Cr plane's client ID is N + 1
         *      Cb plane's client ID is N + 2
         */

        return mdp5_cfg->smp.clients[pipe] + plane;
}

/* step #1: update # of blocks pending for the client: */
static int smp_request_block(struct mdp5_smp *smp,
                u32 cid, int nblks)
{
        struct mdp5_kms *mdp5_kms = get_kms(smp);
        const struct mdp5_cfg_hw *hw_cfg;
        struct mdp5_client_smp_state *ps = &smp->client_state[cid];
        int i, ret, avail, cur_nblks, cnt = smp->blk_cnt;
        int reserved;
        unsigned long flags;

        hw_cfg = mdp5_cfg_get_hw_config(mdp5_kms->cfg);
        reserved = hw_cfg->smp.reserved[cid];

        spin_lock_irqsave(&smp->state_lock, flags);

        if (reserved) {
                nblks = max(0, nblks - reserved);
                DBG("%d MMBs allocated (%d reserved)", nblks, reserved);
        }

        avail = cnt - bitmap_weight(smp->state, cnt);
        if (nblks > avail) {
                dev_err(mdp5_kms->dev->dev, "out of blks (req=%d > avail=%d)\n",
                                nblks, avail);
                ret = -ENOSPC;
                goto fail;
        }

        cur_nblks = bitmap_weight(ps->pending, cnt);
        if (nblks > cur_nblks) {
                /* grow the existing pending reservation: */
                for (i = cur_nblks; i < nblks; i++) {
                        int blk = find_first_zero_bit(smp->state, cnt);
                        set_bit(blk, ps->pending);
                        set_bit(blk, smp->state);
                }
        } else {
                /* shrink the existing pending reservation: */
                for (i = cur_nblks; i > nblks; i--) {
                        int blk = find_first_bit(ps->pending, cnt);
                        clear_bit(blk, ps->pending);
                        /* don't clear in global smp_state until _commit() */
                }
        }

fail:
        spin_unlock_irqrestore(&smp->state_lock, flags);
        return 0;
}

static void set_fifo_thresholds(struct mdp5_smp *smp,
                enum mdp5_pipe pipe, int nblks)
{
        struct mdp5_kms *mdp5_kms = get_kms(smp);
        u32 smp_entries_per_blk = smp->blk_size / (128 / BITS_PER_BYTE);
        u32 val;

        /* 1/4 of SMP pool that is being fetched */
        val = (nblks * smp_entries_per_blk) / 4;

        mdp5_write(mdp5_kms, REG_MDP5_PIPE_REQPRIO_FIFO_WM_0(pipe), val * 1);
        mdp5_write(mdp5_kms, REG_MDP5_PIPE_REQPRIO_FIFO_WM_1(pipe), val * 2);
        mdp5_write(mdp5_kms, REG_MDP5_PIPE_REQPRIO_FIFO_WM_2(pipe), val * 3);
}

/*
 * NOTE: looks like if horizontal decimation is used (if we supported that)
 * then the width used to calculate SMP block requirements is the post-
 * decimated width.  Ie. SMP buffering sits downstream of decimation (which
 * presumably happens during the dma from scanout buffer).
 */
int mdp5_smp_request(struct mdp5_smp *smp, enum mdp5_pipe pipe, u32 fmt, u32 width)
{
        struct mdp5_kms *mdp5_kms = get_kms(smp);
        struct drm_device *dev = mdp5_kms->dev;
        int rev = mdp5_cfg_get_hw_rev(mdp5_kms->cfg);
        int i, hsub, nplanes, nlines, nblks, ret;

        nplanes = drm_format_num_planes(fmt);
        hsub = drm_format_horz_chroma_subsampling(fmt);

        /* different if BWC (compressed framebuffer?) enabled: */
        nlines = 2;

        for (i = 0, nblks = 0; i < nplanes; i++) {
                int n, fetch_stride, cpp;

                cpp = drm_format_plane_cpp(fmt, i);
                fetch_stride = width * cpp / (i ? hsub : 1);

                n = DIV_ROUND_UP(fetch_stride * nlines, smp->blk_size);

                /* for hw rev v1.00 */
                if (rev == 0)
                        n = roundup_pow_of_two(n);

                DBG("%s[%d]: request %d SMP blocks", pipe2name(pipe), i, n);
                ret = smp_request_block(smp, pipe2client(pipe, i), n);
                if (ret) {
                        dev_err(dev->dev, "Cannot allocate %d SMP blocks: %d\n",
                                        n, ret);
                        return ret;
                }

                nblks += n;
        }

        set_fifo_thresholds(smp, pipe, nblks);

        return 0;
}

/* Release SMP blocks for all clients of the pipe */
void mdp5_smp_release(struct mdp5_smp *smp, enum mdp5_pipe pipe)
{
        int i, nblks;

        for (i = 0, nblks = 0; i < pipe2nclients(pipe); i++)
                smp_request_block(smp, pipe2client(pipe, i), 0);
        set_fifo_thresholds(smp, pipe, 0);
}

static void update_smp_state(struct mdp5_smp *smp,
                u32 cid, mdp5_smp_state_t *assigned)
{
        struct mdp5_kms *mdp5_kms = get_kms(smp);
        int cnt = smp->blk_cnt;
        u32 blk, val;

        for_each_set_bit(blk, *assigned, cnt) {
                int idx = blk / 3;
                int fld = blk % 3;

                val = mdp5_read(mdp5_kms, REG_MDP5_MDP_SMP_ALLOC_W_REG(0, idx));

                switch (fld) {
                case 0:
                        val &= ~MDP5_MDP_SMP_ALLOC_W_REG_CLIENT0__MASK;
                        val |= MDP5_MDP_SMP_ALLOC_W_REG_CLIENT0(cid);
                        break;
                case 1:
                        val &= ~MDP5_MDP_SMP_ALLOC_W_REG_CLIENT1__MASK;
                        val |= MDP5_MDP_SMP_ALLOC_W_REG_CLIENT1(cid);
                        break;
                case 2:
                        val &= ~MDP5_MDP_SMP_ALLOC_W_REG_CLIENT2__MASK;
                        val |= MDP5_MDP_SMP_ALLOC_W_REG_CLIENT2(cid);
                        break;
                }

                mdp5_write(mdp5_kms, REG_MDP5_MDP_SMP_ALLOC_W_REG(0, idx), val);
                mdp5_write(mdp5_kms, REG_MDP5_MDP_SMP_ALLOC_R_REG(0, idx), val);
        }
}

/* step #2: configure hw for union(pending, inuse): */
void mdp5_smp_configure(struct mdp5_smp *smp, enum mdp5_pipe pipe)
{
        int cnt = smp->blk_cnt;
        mdp5_smp_state_t assigned;
        int i;

        for (i = 0; i < pipe2nclients(pipe); i++) {
                u32 cid = pipe2client(pipe, i);
                struct mdp5_client_smp_state *ps = &smp->client_state[cid];

                bitmap_or(assigned, ps->inuse, ps->pending, cnt);
                update_smp_state(smp, cid, &assigned);
        }
}

/* step #3: after vblank, copy pending -> inuse: */
void mdp5_smp_commit(struct mdp5_smp *smp, enum mdp5_pipe pipe)
{
        int cnt = smp->blk_cnt;
        mdp5_smp_state_t released;
        int i;

        for (i = 0; i < pipe2nclients(pipe); i++) {
                u32 cid = pipe2client(pipe, i);
                struct mdp5_client_smp_state *ps = &smp->client_state[cid];

                /*
                 * Figure out if there are any blocks we where previously
                 * using, which can be released and made available to other
                 * clients:
                 */
                if (bitmap_andnot(released, ps->inuse, ps->pending, cnt)) {
                        unsigned long flags;

                        spin_lock_irqsave(&smp->state_lock, flags);
                        /* clear released blocks: */
                        bitmap_andnot(smp->state, smp->state, released, cnt);
                        spin_unlock_irqrestore(&smp->state_lock, flags);

                        update_smp_state(smp, CID_UNUSED, &released);
                }

                bitmap_copy(ps->inuse, ps->pending, cnt);
        }
}

void mdp5_smp_destroy(struct mdp5_smp *smp)
{
        kfree(smp);
}

struct mdp5_smp *mdp5_smp_init(struct drm_device *dev, const struct mdp5_smp_block *cfg)
{
        struct mdp5_smp *smp = NULL;
        int ret;

        smp = kzalloc(sizeof(*smp), GFP_KERNEL);
        if (unlikely(!smp)) {
                ret = -ENOMEM;
                goto fail;
        }

        smp->dev = dev;
        smp->blk_cnt = cfg->mmb_count;
        smp->blk_size = cfg->mmb_size;

        /* statically tied MMBs cannot be re-allocated: */
        bitmap_copy(smp->state, cfg->reserved_state, smp->blk_cnt);
        spin_lock_init(&smp->state_lock);

        return smp;
fail:
        if (smp)
                mdp5_smp_destroy(smp);

        return ERR_PTR(ret);
}