drivers/gpu/drm/msm/msm_rd.c

   1 /*
   2  * Copyright (C) 2013 Red Hat
   3  * Author: Rob Clark <robdclark@gmail.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 as published by
   7  * the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program.  If not, see <http://www.gnu.org/licenses/>.
  16  */
  17
  18 /* For debugging crashes, userspace can:
  19  *
  20  *   tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
  21  *
  22  * to log the cmdstream in a format that is understood by freedreno/cffdump
  23  * utility.  By comparing the last successfully completed fence #, to the
  24  * cmdstream for the next fence, you can narrow down which process and submit
  25  * caused the gpu crash/lockup.
  26  *
  27  * Additionally:
  28  *
  29  *   tail -f /sys/kernel/debug/dri/<minor>/hangrd > logfile.rd
  30  *
  31  * will capture just the cmdstream from submits which triggered a GPU hang.
  32  *
  33  * This bypasses drm_debugfs_create_files() mainly because we need to use
  34  * our own fops for a bit more control.  In particular, we don't want to
  35  * do anything if userspace doesn't have the debugfs file open.
  36  *
  37  * The module-param "rd_full", which defaults to false, enables snapshotting
  38  * all (non-written) buffers in the submit, rather than just cmdstream bo's.
  39  * This is useful to capture the contents of (for example) vbo's or textures,
  40  * or shader programs (if not emitted inline in cmdstream).
  41  */
  42
  43 #ifdef CONFIG_DEBUG_FS
  44
  45 #include <linux/kfifo.h>
  46 #include <linux/debugfs.h>
  47 #include <linux/circ_buf.h>
  48 #include <linux/wait.h>
  49
  50 #include "msm_drv.h"
  51 #include "msm_gpu.h"
  52 #include "msm_gem.h"
  53
  54 static bool rd_full = false;
  55 MODULE_PARM_DESC(rd_full, "If true, $debugfs/.../rd will snapshot all buffer contents");
  56 module_param_named(rd_full, rd_full, bool, 0600);
  57
  58 enum rd_sect_type {
  59         RD_NONE,
  60         RD_TEST,       /* ascii text */
  61         RD_CMD,        /* ascii text */
  62         RD_GPUADDR,    /* u32 gpuaddr, u32 size */
  63         RD_CONTEXT,    /* raw dump */
  64         RD_CMDSTREAM,  /* raw dump */
  65         RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
  66         RD_PARAM,      /* u32 param_type, u32 param_val, u32 bitlen */
  67         RD_FLUSH,      /* empty, clear previous params */
  68         RD_PROGRAM,    /* shader program, raw dump */
  69         RD_VERT_SHADER,
  70         RD_FRAG_SHADER,
  71         RD_BUFFER_CONTENTS,
  72         RD_GPU_ID,
  73 };
  74
  75 #define BUF_SZ 512  /* should be power of 2 */
  76
  77 /* space used: */
  78 #define circ_count(circ) \
  79         (CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
  80 #define circ_count_to_end(circ) \
  81         (CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
  82 /* space available: */
  83 #define circ_space(circ) \
  84         (CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
  85 #define circ_space_to_end(circ) \
  86         (CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
  87
  88 struct msm_rd_state {
  89         struct drm_device *dev;
  90
  91         bool open;
  92
  93         /* current submit to read out: */
  94         struct msm_gem_submit *submit;
  95
  96         /* fifo access is synchronized on the producer side by
  97          * struct_mutex held by submit code (otherwise we could
  98          * end up w/ cmds logged in different order than they
  99          * were executed).  And read_lock synchronizes the reads
 100          */
 101         struct mutex read_lock;
 102
 103         wait_queue_head_t fifo_event;
 104         struct circ_buf fifo;
 105
 106         char buf[BUF_SZ];
 107 };
 108
 109 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
 110 {
 111         struct circ_buf *fifo = &rd->fifo;
 112         const char *ptr = buf;
 113
 114         while (sz > 0) {
 115                 char *fptr = &fifo->buf[fifo->head];
 116                 int n;
 117
 118                 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
 119                 if (!rd->open)
 120                         return;
 121
 122                 /* Note that smp_load_acquire() is not strictly required
 123                  * as CIRC_SPACE_TO_END() does not access the tail more
 124                  * than once.
 125                  */
 126                 n = min(sz, circ_space_to_end(&rd->fifo));
 127                 memcpy(fptr, ptr, n);
 128
 129                 smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
 130                 sz  -= n;
 131                 ptr += n;
 132
 133                 wake_up_all(&rd->fifo_event);
 134         }
 135 }
 136
 137 static void rd_write_section(struct msm_rd_state *rd,
 138                 enum rd_sect_type type, const void *buf, int sz)
 139 {
 140         rd_write(rd, &type, 4);
 141         rd_write(rd, &sz, 4);
 142         rd_write(rd, buf, sz);
 143 }
 144
 145 static ssize_t rd_read(struct file *file, char __user *buf,
 146                 size_t sz, loff_t *ppos)
 147 {
 148         struct msm_rd_state *rd = file->private_data;
 149         struct circ_buf *fifo = &rd->fifo;
 150         const char *fptr = &fifo->buf[fifo->tail];
 151         int n = 0, ret = 0;
 152
 153         mutex_lock(&rd->read_lock);
 154
 155         ret = wait_event_interruptible(rd->fifo_event,
 156                         circ_count(&rd->fifo) > 0);
 157         if (ret)
 158                 goto out;
 159
 160         /* Note that smp_load_acquire() is not strictly required
 161          * as CIRC_CNT_TO_END() does not access the head more than
 162          * once.
 163          */
 164         n = min_t(int, sz, circ_count_to_end(&rd->fifo));
 165         if (copy_to_user(buf, fptr, n)) {
 166                 ret = -EFAULT;
 167                 goto out;
 168         }
 169
 170         smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
 171         *ppos += n;
 172
 173         wake_up_all(&rd->fifo_event);
 174
 175 out:
 176         mutex_unlock(&rd->read_lock);
 177         if (ret)
 178                 return ret;
 179         return n;
 180 }
 181
 182 static int rd_open(struct inode *inode, struct file *file)
 183 {
 184         struct msm_rd_state *rd = inode->i_private;
 185         struct drm_device *dev = rd->dev;
 186         struct msm_drm_private *priv = dev->dev_private;
 187         struct msm_gpu *gpu = priv->gpu;
 188         uint64_t val;
 189         uint32_t gpu_id;
 190         int ret = 0;
 191
 192         mutex_lock(&dev->struct_mutex);
 193
 194         if (rd->open || !gpu) {
 195                 ret = -EBUSY;
 196                 goto out;
 197         }
 198
 199         file->private_data = rd;
 200         rd->open = true;
 201
 202         /* the parsing tools need to know gpu-id to know which
 203          * register database to load.
 204          */
 205         gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
 206         gpu_id = val;
 207
 208         rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
 209
 210 out:
 211         mutex_unlock(&dev->struct_mutex);
 212         return ret;
 213 }
 214
 215 static int rd_release(struct inode *inode, struct file *file)
 216 {
 217         struct msm_rd_state *rd = inode->i_private;
 218
 219         rd->open = false;
 220         wake_up_all(&rd->fifo_event);
 221
 222         return 0;
 223 }
 224
 225
 226 static const struct file_operations rd_debugfs_fops = {
 227         .owner = THIS_MODULE,
 228         .open = rd_open,
 229         .read = rd_read,
 230         .llseek = no_llseek,
 231         .release = rd_release,
 232 };
 233
 234
 235 static void rd_cleanup(struct msm_rd_state *rd)
 236 {
 237         if (!rd)
 238                 return;
 239
 240         mutex_destroy(&rd->read_lock);
 241         kfree(rd);
 242 }
 243
 244 static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
 245 {
 246         struct msm_rd_state *rd;
 247         struct dentry *ent;
 248         int ret = 0;
 249
 250         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
 251         if (!rd)
 252                 return ERR_PTR(-ENOMEM);
 253
 254         rd->dev = minor->dev;
 255         rd->fifo.buf = rd->buf;
 256
 257         mutex_init(&rd->read_lock);
 258
 259         init_waitqueue_head(&rd->fifo_event);
 260
 261         ent = debugfs_create_file(name, S_IFREG | S_IRUGO,
 262                         minor->debugfs_root, rd, &rd_debugfs_fops);
 263         if (!ent) {
 264                 DRM_ERROR("Cannot create /sys/kernel/debug/dri/%pd/%s\n",
 265                                 minor->debugfs_root, name);
 266                 ret = -ENOMEM;
 267                 goto fail;
 268         }
 269
 270         return rd;
 271
 272 fail:
 273         rd_cleanup(rd);
 274         return ERR_PTR(ret);
 275 }
 276
 277 int msm_rd_debugfs_init(struct drm_minor *minor)
 278 {
 279         struct msm_drm_private *priv = minor->dev->dev_private;
 280         struct msm_rd_state *rd;
 281         int ret;
 282
 283         /* only create on first minor: */
 284         if (priv->rd)
 285                 return 0;
 286
 287         rd = rd_init(minor, "rd");
 288         if (IS_ERR(rd)) {
 289                 ret = PTR_ERR(rd);
 290                 goto fail;
 291         }
 292
 293         priv->rd = rd;
 294
 295         rd = rd_init(minor, "hangrd");
 296         if (IS_ERR(rd)) {
 297                 ret = PTR_ERR(rd);
 298                 goto fail;
 299         }
 300
 301         priv->hangrd = rd;
 302
 303         return 0;
 304
 305 fail:
 306         msm_rd_debugfs_cleanup(priv);
 307         return ret;
 308 }
 309
 310 void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
 311 {
 312         rd_cleanup(priv->rd);
 313         priv->rd = NULL;
 314
 315         rd_cleanup(priv->hangrd);
 316         priv->hangrd = NULL;
 317 }
 318
 319 static void snapshot_buf(struct msm_rd_state *rd,
 320                 struct msm_gem_submit *submit, int idx,
 321                 uint64_t iova, uint32_t size)
 322 {
 323         struct msm_gem_object *obj = submit->bos[idx].obj;
 324         unsigned offset = 0;
 325         const char *buf;
 326
 327         if (iova) {
 328                 offset = iova - submit->bos[idx].iova;
 329         } else {
 330                 iova = submit->bos[idx].iova;
 331                 size = obj->base.size;
 332         }
 333
 334         /*
 335          * Always write the GPUADDR header so can get a complete list of all the
 336          * buffers in the cmd
 337          */
 338         rd_write_section(rd, RD_GPUADDR,
 339                         (uint32_t[3]){ iova, size, iova >> 32 }, 12);
 340
 341         /* But only dump the contents of buffers marked READ */
 342         if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
 343                 return;
 344
 345         buf = msm_gem_get_vaddr_active(&obj->base);
 346         if (IS_ERR(buf))
 347                 return;
 348
 349         buf += offset;
 350
 351         rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
 352
 353         msm_gem_put_vaddr(&obj->base);
 354 }
 355
 356 static bool
 357 should_dump(struct msm_gem_submit *submit, int idx)
 358 {
 359         return rd_full || (submit->bos[idx].flags & MSM_SUBMIT_BO_DUMP);
 360 }
 361
 362 /* called under struct_mutex */
 363 void msm_rd_dump_submit(struct msm_rd_state *rd, struct msm_gem_submit *submit,
 364                 const char *fmt, ...)
 365 {
 366         struct drm_device *dev = submit->dev;
 367         struct task_struct *task;
 368         char msg[256];
 369         int i, n;
 370
 371         if (!rd->open)
 372                 return;
 373
 374         /* writing into fifo is serialized by caller, and
 375          * rd->read_lock is used to serialize the reads
 376          */
 377         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 378
 379         if (fmt) {
 380                 va_list args;
 381
 382                 va_start(args, fmt);
 383                 n = vscnprintf(msg, sizeof(msg), fmt, args);
 384                 va_end(args);
 385
 386                 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
 387         }
 388
 389         rcu_read_lock();
 390         task = pid_task(submit->pid, PIDTYPE_PID);
 391         if (task) {
 392                 n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
 393                                 TASK_COMM_LEN, task->comm,
 394                                 pid_nr(submit->pid), submit->seqno);
 395         } else {
 396                 n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
 397                                 pid_nr(submit->pid), submit->seqno);
 398         }
 399         rcu_read_unlock();
 400
 401         rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
 402
 403         for (i = 0; i < submit->nr_bos; i++)
 404                 if (should_dump(submit, i))
 405                         snapshot_buf(rd, submit, i, 0, 0);
 406
 407         for (i = 0; i < submit->nr_cmds; i++) {
 408                 uint64_t iova = submit->cmd[i].iova;
 409                 uint32_t szd  = submit->cmd[i].size; /* in dwords */
 410
 411                 /* snapshot cmdstream bo's (if we haven't already): */
 412                 if (!should_dump(submit, i)) {
 413                         snapshot_buf(rd, submit, submit->cmd[i].idx,
 414                                         submit->cmd[i].iova, szd * 4);
 415                 }
 416
 417                 switch (submit->cmd[i].type) {
 418                 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
 419                         /* ignore IB-targets, we've logged the buffer, the
 420                          * parser tool will follow the IB based on the logged
 421                          * buffer/gpuaddr, so nothing more to do.
 422                          */
 423                         break;
 424                 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
 425                 case MSM_SUBMIT_CMD_BUF:
 426                         rd_write_section(rd, RD_CMDSTREAM_ADDR,
 427                                 (uint32_t[3]){ iova, szd, iova >> 32 }, 12);
 428                         break;
 429                 }
 430         }
 431 }
 432 #endif