/*
 * Copyright 2018 Red Hat Inc.
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * 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.
 *
 * Authors: Jérôme Glisse <jglisse@redhat.com>
 */
#include <sys/types.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <strings.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <stdio.h>
#include <fcntl.h>

#include "xf86drm.h"
#include "libdrm/nouveau/nouveau.h"
#include "nvif/class.h"
#include "nvif/cl0080.h"
#include "nvif/if0008.h"
#include "nvif/if000c.h"

struct nvk_00f0_cp_desc {
   uint32_t unk0[8];
   uint32_t entry;
   uint32_t unk9[2];
   uint32_t unk11_0      : 30;
   uint32_t linked_tsc   : 1;
   uint32_t unk11_31     : 1;
   uint32_t griddim_x    : 31;
   uint32_t unk12        : 1;
   uint16_t griddim_y;
   uint16_t unk13;
   uint16_t griddim_z;
   uint16_t unk14;
   uint32_t unk15[2];
   uint32_t shared_size  : 18;
   uint32_t unk17        : 14;
   uint16_t unk18;
   uint16_t blockdim_x;
   uint16_t blockdim_y;
   uint16_t blockdim_z;
   uint32_t cb_mask      : 8;
   uint32_t unk20        : 24;
   uint32_t unk21[8];
   uint32_t local_size_p : 24;
   uint32_t unk29        : 3;
   uint32_t bar_alloc    : 5;
   uint32_t local_size_n : 24;
   uint32_t gpr_alloc    : 8;
   uint32_t cstack_size  : 24;
   uint32_t unk31        : 8;
   struct {
      uint32_t address_l;
      uint32_t address_h : 17;
      uint32_t reserved  : 2;
      uint32_t size_sh4  : 13;
   } cb[8];
   uint32_t unk48[16];
};


typedef struct {
    struct nouveau_client *nvclient;
    struct nouveau_device *nvdevice;
    struct nouveau_object *nvchannel;
    struct nouveau_object *nvcompute;
    struct nouveau_pushbuf *nvpushbuf;
    struct nouveau_drm *nvdrm;
    void *hole;
    int fd;
} moche_t;

#define DRM_FILE_PAGE_OFFSET (0x100000000ULL)

int moche_init_hmm(moche_t *moche)
{
    moche->hole = mmap((void *)(1UL << 30), (2UL << 30), PROT_NONE,
            MAP_PRIVATE, moche->fd, DRM_FILE_PAGE_OFFSET);
    printf("hmm init %p\n", moche->hole);
    return 0;
}

int moche_init(moche_t *moche)
{
    int r;

    moche->fd = drmOpen("nouveau", NULL);
    if (moche->fd < 0) {
        printf("EE: failed to open nouveau device file (%d)\n", moche->fd);
        return -1;
    }

    if ((r = nouveau_drm_new(moche->fd, &moche->nvdrm))) {
        printf("EE: failed to create nouveau drm (%d)\n", r);
        return r;
    }

    {
        struct nv_device_v0 arg = {0};
        uint32_t oclass, size;
        void *data;

        arg.device = ~0ULL;
        oclass = NV_DEVICE;
        size = sizeof(arg);
        data = &arg;
        if ((r = nouveau_device_new(&moche->nvdrm->client, oclass, data,
                                    size, &moche->nvdevice))) {
            printf("EE: failed to create nouveau device (%d)\n", r);
            goto nouveau_device;
        }
    }

    if ((r = nouveau_client_new(moche->nvdevice, &moche->nvclient))) {
        printf("EE: failed to create nouveau client (%d)\n", r);
        goto nouveau_client;
    }

    if ((r = moche_init_hmm(moche))) {
        printf("EE: failed to enbable HMM (%d)\n", r);
        goto nouveau_hmm;
    }

    {
        struct nve0_fifo arg = {0};
        uint32_t oclass, size;
        void *data;

        oclass = NOUVEAU_FIFO_CHANNEL_CLASS;
        arg.engine = 0x01 | 0x10 | 0x20; // NVA06F_V0_ENGINE_CE0 | NVA06F_V0_ENGINE_CE1
        size = sizeof(arg);
        data = &arg;
        if ((r = nouveau_object_new(&moche->nvdevice->object, 0, oclass,
                                    data, size, &moche->nvchannel))) {
            printf("EE: failed to create nouveau channel (%d)\n", r);
            goto nouveau_channel;
        }
    }

    {
        if ((r = nouveau_pushbuf_new(moche->nvclient, moche->nvchannel, 2,
                                     64 * 1024, 1, &moche->nvpushbuf))) {
            printf("EE: failed to create nouveau pushbuf (%d)\n", r);
            goto nouveau_pushbuf;
        }
    }

    {
        if ((r = nouveau_object_new(moche->nvchannel, 0xcafec1c0,
                                    0xc1c0, NULL, 0, &moche->nvcompute))) {
            printf("EE: failed to create nouveau channel (%d)\n", r);
            goto nouveau_compute;
        }
    }

    return 0;

nouveau_compute:
nouveau_pushbuf:
    nouveau_object_del(&moche->nvchannel);
nouveau_channel:
nouveau_hmm:
    nouveau_client_del(&moche->nvclient);
nouveau_client:
    nouveau_device_del(&moche->nvdevice);
nouveau_device:
    nouveau_drm_del(&moche->nvdrm);
    return r;
}

void moche_fini(moche_t *moche)
{
    nouveau_object_del(&moche->nvcompute);
    nouveau_pushbuf_del(&moche->nvpushbuf);
    nouveau_object_del(&moche->nvchannel);
    nouveau_client_del(&moche->nvclient);
    nouveau_device_del(&moche->nvdevice);
    nouveau_drm_del(&moche->nvdrm);
}

static inline void moche_push_data(moche_t *moche, uint32_t data)
{
    *moche->nvpushbuf->cur++ = data;
}

static inline int moche_push_kick(moche_t *moche)
{
    return nouveau_pushbuf_kick(moche->nvpushbuf, moche->nvchannel);
}

static inline void moche_push_refn(moche_t *moche,
        struct nouveau_bo *bo, uint32_t flags)
{
   struct nouveau_pushbuf_refn ref = { bo, flags };

   nouveau_pushbuf_refn(moche->nvpushbuf, &ref, 1);
}

static inline uint32_t nvk_sq_cmd(unsigned subc, unsigned method, unsigned len)
{
    return ((method >> 2) & 0x1fff) |
           ((len & 0xfff) << 16) |
           ((subc & 0x7) << 13) |
           (0x1 << 29);
}

static inline uint32_t nvk_ni_cmd(unsigned subc, unsigned method, unsigned len)
{
    return ((method >> 2) & 0x1fff) |
           ((len & 0xfff) << 16) |
           ((subc & 0x7) << 13) |
           (0x3 << 29);
}

static inline uint32_t nvk_addr_high(uint64_t offset)
{
    return (offset >> 32) & 0xffffffff;
}

static inline uint32_t nvk_addr_low(uint64_t offset)
{
    return offset & 0xffffffff;
}

static inline uint32_t nvk_size_high(uint64_t offset)
{
    return (offset >> 32) & 0xffffffff;
}

static inline uint32_t nvk_size_low(uint64_t offset)
{
    return offset & 0xffffffff;
}

int moche_pushbuf_test(moche_t *moche)
{
    struct nouveau_bo *bo;
    uint32_t *ptr;
    int r, i;

    if ((r = nouveau_bo_new(moche->nvdevice, NOUVEAU_BO_GART,
                            0, 4 * 1024, NULL, &bo))) {
        printf("EE: failed to create nouveau bo (%d)\n", r);
        goto nouveau_bo_new;
    }
    if ((r = nouveau_bo_map(bo, NOUVEAU_BO_WR, moche->nvclient))) {
        printf("EE: failed to map nouveau bo (%d)\n", r);
        goto nouveau_bo_map;
    }

    ptr = bo->map;
    ptr[0] = 0xcafedead;

    if ((r = nouveau_pushbuf_space(moche->nvpushbuf, 9, 1, 0))) {
        printf("EE: pushbuf full (%d)\n", r);
        goto nouveau_bo_new;
    }
    moche_push_refn(moche, bo, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
    moche_push_data(moche, nvk_sq_cmd(1, 0x0000, 1));
    moche_push_data(moche, moche->nvcompute->oclass);
    moche_push_data(moche, nvk_sq_cmd(1, 0x0110, 1));
    moche_push_data(moche, 0x00000000);
#if 1
    moche_push_data(moche, nvk_sq_cmd(1, 0x1b00, 4));
    moche_push_data(moche, nvk_addr_high(bo->offset));
    moche_push_data(moche, nvk_addr_low(bo->offset));
    moche_push_data(moche, 0xdeadcafe);
    moche_push_data(moche, 0x00000000);
#else
    moche_push_data(moche, nvk_sq_cmd(1, 0x0010, 4));
    moche_push_data(moche, nvk_addr_high(bo->offset));
    moche_push_data(moche, nvk_addr_low(bo->offset));
    moche_push_data(moche, 0xdeadcafe);
    moche_push_data(moche, 0x00000002);
#endif

    if ((r = moche_push_kick(moche))) {
        printf("EE: pushbuf kick (%d)\n", r);
        goto nouveau_push_kick;
    }

    for (i = 0; i < 10 && ptr[0] != 0xdeadcafe; ++i) {
        nouveau_bo_wait(bo, NOUVEAU_BO_RD, moche->nvclient);
    }
    if (ptr[0] == 0xdeadcafe) {
        printf("OK: pusbuf test 0x%08x\n", ptr[0]);
    } else {
        printf("EE: pusbuf test 0x%08x\n", ptr[0]);
    }

nouveau_push_kick:
nouveau_bo_map:
    nouveau_bo_ref(NULL, &bo);
nouveau_bo_new:
    return r;
}

int moche_compute_test(moche_t *moche, uint64_t doffset,
        uint32_t *rptr, unsigned nelem,
        struct nouveau_bo *bor)
{
    uint64_t code[] = {
        //testtx:
        // sched 0x7f1 0x207f9 0x7f9
        // mov $r2 $tid.x
        // lea 0x1 cc $r0 $r2 c0[0x0] 0x2
        // lea hi x 0x1 $r1 $r2 c0[0x4] 0x0 0x2
        // sched 0x7f9 0x7f5 0x7ff
        // stg e b32 ncg[$r0] $r2
        // exit
        // exit
        0x001fe440ff2007f1ul,
        0xf0c8000002170002ul,
        0x4bd7810000070200ul,
        0x1a177f8000170201ul,
        0x001ffc00fea007f9ul,
        0xeedc200000070002ul,
        0xe30000000007000ful,
        0xe30000000007000ful,
    };
    struct nouveau_bo *bom, *tls;
    struct nvk_00f0_cp_desc *desc;
    uint32_t *mptr;
    int r, i, tls_size;

    tls_size = 16 << 20;
    if ((r = nouveau_bo_new(moche->nvdevice, NOUVEAU_BO_GART,
                            0, tls_size, NULL, &tls))) {
        printf("EE: failed to create nouveau bo (%d)\n", r);
        goto nouveau_tls_new;
    }
    if ((r = nouveau_bo_new(moche->nvdevice, NOUVEAU_BO_GART,
                            0, 64 * 1024, NULL, &bom))) {
        printf("EE: failed to create nouveau bo (%d)\n", r);
        goto nouveau_bom_new;
    }
    if ((r = nouveau_bo_map(bom, NOUVEAU_BO_WR, moche->nvclient))) {
        printf("EE: failed to map nouveau bo (%d)\n", r);
        goto nouveau_bom_map;
    }

    mptr = bom->map;
    printf("bom offset 0x%lx\n", (unsigned long)bom->offset);

    // initialize result bo
    memset(rptr, 0, nelem * 4);

    // fence
    mptr[0] = 0xcafedead;
    memcpy(&mptr[1024], code, sizeof(code));

    // param
    mptr[64] = nvk_addr_low(doffset);
    mptr[65] = nvk_addr_high(doffset);

    // compute descriptor
    desc = (void *)&mptr[128];
    memset(desc, 0, sizeof(*desc));
    desc->griddim_x = nelem;
    desc->griddim_y = 1;
    desc->griddim_z = 1;
    desc->blockdim_x = 32;
    desc->blockdim_y = 1;
    desc->blockdim_z = 1;
    desc->entry = 0x0;
    desc->shared_size = 0;
    desc->local_size_p = 0;
    desc->bar_alloc = 0;
    desc->local_size_n = 0;
    desc->gpr_alloc = 8;
    desc->cstack_size = 0x1000;
    desc->cb_mask = 1;
    desc->cb[0].address_h = nvk_addr_high(bom->offset + 64 * 4);
    desc->cb[0].address_l = nvk_addr_low(bom->offset + 64 * 4);
    desc->cb[0].size_sh4 = 256 >> 4;
    desc->unk0[4]  = 0x40;
    desc->unk11_0  = 0x04014000;

    if ((r = nouveau_pushbuf_space(moche->nvpushbuf, 1024, 3, 0))) {
        printf("EE: pushbuf full (%d)\n", r);
        goto nouveau_push_space;
    }
    moche_push_refn(moche, bom, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
    if (bor)
        moche_push_refn(moche, bor, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
    moche_push_refn(moche, tls, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
    moche_push_data(moche, nvk_sq_cmd(1, 0x0000, 1));
    moche_push_data(moche, moche->nvcompute->oclass);
    moche_push_data(moche, nvk_sq_cmd(1, 0x0110, 1));
    moche_push_data(moche, 0x00000000);

    tls_size = tls_size / 16;
    moche_push_data(moche, nvk_sq_cmd(1, 0x0790, 2));
    moche_push_data(moche, nvk_addr_high(tls->offset));
    moche_push_data(moche, nvk_addr_low(tls->offset));
    moche_push_data(moche, nvk_sq_cmd(1, 0x02e4, 3));
    moche_push_data(moche, nvk_addr_high(tls_size));
    moche_push_data(moche, nvk_addr_low(tls_size) & ~0x7fff);
    moche_push_data(moche, 0x000000ff);
    moche_push_data(moche, nvk_sq_cmd(1, 0x02f0, 3));
    moche_push_data(moche, nvk_addr_high(tls_size));
    moche_push_data(moche, nvk_addr_low(tls_size) & ~0x7fff);
    moche_push_data(moche, 0x000000ff);
    moche_push_data(moche, nvk_sq_cmd(1, 0x077c, 1));
    moche_push_data(moche, 0xff000000);
    moche_push_data(moche, nvk_sq_cmd(1, 0x0214, 1));
    moche_push_data(moche, 0xfe000000);
    moche_push_data(moche, nvk_sq_cmd(1, 0x1608, 2));
    moche_push_data(moche, nvk_addr_high(bom->offset + 1024 * 4));
    moche_push_data(moche, nvk_addr_low(bom->offset + 1024 * 4));
    moche_push_data(moche, nvk_sq_cmd(1, 0x0310, 1));
    moche_push_data(moche, 0x00000400);
    moche_push_data(moche, nvk_ni_cmd(1, 0x0248, 64));
    for (int i = 63; i >= 0; --i) {
        moche_push_data(moche, 0x00038000 | i);
    }
    moche_push_data(moche, nvk_ni_cmd(1, 0x0110, 1));
    moche_push_data(moche, 0x00000000);
    moche_push_data(moche, nvk_sq_cmd(1, 0x2608, 1));
    moche_push_data(moche, 0x00000000);
    moche_push_data(moche, nvk_sq_cmd(1, 0x1698, 1));
    moche_push_data(moche, 0x00001000);
    moche_push_data(moche, nvk_sq_cmd(1, 0x021c, 1));
    moche_push_data(moche, 0x00001017);
    moche_push_data(moche, nvk_sq_cmd(1, 0x02b4, 1));
    moche_push_data(moche, (bom->offset + 128 * 4) >> 8);
    moche_push_data(moche, nvk_sq_cmd(1, 0x02bc, 1));
    moche_push_data(moche, 0x00000003);
    moche_push_data(moche, nvk_sq_cmd(1, 0x0110, 1));
    moche_push_data(moche, 0x00000000);

    moche_push_data(moche, nvk_sq_cmd(1, 0x0110, 1));
    moche_push_data(moche, 0x00000000);
    moche_push_data(moche, nvk_sq_cmd(1, 0x1b00, 4));
    moche_push_data(moche, nvk_addr_high(bom->offset));
    moche_push_data(moche, nvk_addr_low(bom->offset));
    moche_push_data(moche, 0xdeadcafe);
    moche_push_data(moche, 0x00000000);

    if ((r = moche_push_kick(moche))) {
        printf("EE: pushbuf kick (%d)\n", r);
        goto nouveau_push_kick;
    }

    for (i = 0; i < 10 && mptr[0] != 0xdeadcafe; ++i) {
        nouveau_bo_wait(bom, NOUVEAU_BO_RD, moche->nvclient);
    }
    if (mptr[0] == 0xdeadcafe) {
        printf("OK: compute test 0x%08x 0x%08x 0x%08x 0x%08x\n",
               mptr[0], rptr[0], rptr[1], rptr[nelem - 1]);
    } else {
        printf("EE: compute test 0x%08x\n", mptr[0]);
    }
    for (i = 0; i < nelem; ++i) {
        if (rptr[i] != i) {
            printf("EE: rptr[%d] != %d -> %d\n", i, i, rptr[i]);
            break;
        }
    }

nouveau_push_kick:
nouveau_push_space:
nouveau_bom_map:
    nouveau_bo_ref(NULL, &bom);
nouveau_bom_new:
    nouveau_bo_ref(NULL, &tls);
nouveau_tls_new:
    return r;
}

int main(int argc, char *argv[])
{
    struct nouveau_bo *bor = NULL;
    unsigned nelem = 512 * 1024;
    moche_t moche = {0};
    int r;

    if ((r = moche_init(&moche))) {
        return r;
    }

    printf("OK: fd %d ok chipset 0x%08x (press enter to run)\n",
           moche.fd, moche.nvdevice->chipset);
    getchar();

    if ((r = moche_pushbuf_test(&moche))) {
        return r;
    }

    if (0) {
        uint32_t *rptr;

        if ((r = nouveau_bo_new(moche.nvdevice, NOUVEAU_BO_GART,
                                0, 4 * nelem, NULL, &bor))) {
            printf("EE: failed to create nouveau bo (%d)\n", r);
            goto nouveau_bor_new;
        }
        if ((r = nouveau_bo_map(bor, NOUVEAU_BO_WR, moche.nvclient))) {
            printf("EE: failed to map nouveau bo (%d)\n", r);
            goto nouveau_bor_map;
        }
        rptr = bor->map;
        printf("bor offset 0x%lx\n", (unsigned long)bor->offset);

        if ((r = moche_compute_test(&moche, bor->offset, rptr, nelem, bor))) {
            return r;
        }
    } else {
        uint32_t *rptr;

        rptr = malloc(nelem * 4);
        printf("bor offset 0x%lx\n", (unsigned long)rptr);
        if ((r = moche_compute_test(&moche, (uint64_t)rptr, rptr, nelem, NULL))) {
            return r;
        }
    }

nouveau_bor_map:
    if (bor)
        nouveau_bo_ref(NULL, &bor);
nouveau_bor_new:
    moche_fini(&moche);
    return 0;
}