Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / gpu / drm / msm / msm_rd.c
diff --git a/kernel/drivers/gpu/drm/msm/msm_rd.c b/kernel/drivers/gpu/drm/msm/msm_rd.c
new file mode 100644 (file)
index 0000000..9a78c48
--- /dev/null
@@ -0,0 +1,337 @@
+/*
+ * 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/>.
+ */
+
+/* For debugging crashes, userspace can:
+ *
+ *   tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
+ *
+ * To log the cmdstream in a format that is understood by freedreno/cffdump
+ * utility.  By comparing the last successfully completed fence #, to the
+ * cmdstream for the next fence, you can narrow down which process and submit
+ * caused the gpu crash/lockup.
+ *
+ * This bypasses drm_debugfs_create_files() mainly because we need to use
+ * our own fops for a bit more control.  In particular, we don't want to
+ * do anything if userspace doesn't have the debugfs file open.
+ */
+
+#ifdef CONFIG_DEBUG_FS
+
+#include <linux/kfifo.h>
+#include <linux/debugfs.h>
+#include <linux/circ_buf.h>
+#include <linux/wait.h>
+
+#include "msm_drv.h"
+#include "msm_gpu.h"
+#include "msm_gem.h"
+
+enum rd_sect_type {
+       RD_NONE,
+       RD_TEST,       /* ascii text */
+       RD_CMD,        /* ascii text */
+       RD_GPUADDR,    /* u32 gpuaddr, u32 size */
+       RD_CONTEXT,    /* raw dump */
+       RD_CMDSTREAM,  /* raw dump */
+       RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
+       RD_PARAM,      /* u32 param_type, u32 param_val, u32 bitlen */
+       RD_FLUSH,      /* empty, clear previous params */
+       RD_PROGRAM,    /* shader program, raw dump */
+       RD_VERT_SHADER,
+       RD_FRAG_SHADER,
+       RD_BUFFER_CONTENTS,
+       RD_GPU_ID,
+};
+
+#define BUF_SZ 512  /* should be power of 2 */
+
+/* space used: */
+#define circ_count(circ) \
+       (CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
+#define circ_count_to_end(circ) \
+       (CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
+/* space available: */
+#define circ_space(circ) \
+       (CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
+#define circ_space_to_end(circ) \
+       (CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
+
+struct msm_rd_state {
+       struct drm_device *dev;
+
+       bool open;
+
+       struct dentry *ent;
+       struct drm_info_node *node;
+
+       /* current submit to read out: */
+       struct msm_gem_submit *submit;
+
+       /* fifo access is synchronized on the producer side by
+        * struct_mutex held by submit code (otherwise we could
+        * end up w/ cmds logged in different order than they
+        * were executed).  And read_lock synchronizes the reads
+        */
+       struct mutex read_lock;
+
+       wait_queue_head_t fifo_event;
+       struct circ_buf fifo;
+
+       char buf[BUF_SZ];
+};
+
+static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
+{
+       struct circ_buf *fifo = &rd->fifo;
+       const char *ptr = buf;
+
+       while (sz > 0) {
+               char *fptr = &fifo->buf[fifo->head];
+               int n;
+
+               wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0);
+
+               n = min(sz, circ_space_to_end(&rd->fifo));
+               memcpy(fptr, ptr, n);
+
+               fifo->head = (fifo->head + n) & (BUF_SZ - 1);
+               sz  -= n;
+               ptr += n;
+
+               wake_up_all(&rd->fifo_event);
+       }
+}
+
+static void rd_write_section(struct msm_rd_state *rd,
+               enum rd_sect_type type, const void *buf, int sz)
+{
+       rd_write(rd, &type, 4);
+       rd_write(rd, &sz, 4);
+       rd_write(rd, buf, sz);
+}
+
+static ssize_t rd_read(struct file *file, char __user *buf,
+               size_t sz, loff_t *ppos)
+{
+       struct msm_rd_state *rd = file->private_data;
+       struct circ_buf *fifo = &rd->fifo;
+       const char *fptr = &fifo->buf[fifo->tail];
+       int n = 0, ret = 0;
+
+       mutex_lock(&rd->read_lock);
+
+       ret = wait_event_interruptible(rd->fifo_event,
+                       circ_count(&rd->fifo) > 0);
+       if (ret)
+               goto out;
+
+       n = min_t(int, sz, circ_count_to_end(&rd->fifo));
+       ret = copy_to_user(buf, fptr, n);
+       if (ret)
+               goto out;
+
+       fifo->tail = (fifo->tail + n) & (BUF_SZ - 1);
+       *ppos += n;
+
+       wake_up_all(&rd->fifo_event);
+
+out:
+       mutex_unlock(&rd->read_lock);
+       if (ret)
+               return ret;
+       return n;
+}
+
+static int rd_open(struct inode *inode, struct file *file)
+{
+       struct msm_rd_state *rd = inode->i_private;
+       struct drm_device *dev = rd->dev;
+       struct msm_drm_private *priv = dev->dev_private;
+       struct msm_gpu *gpu = priv->gpu;
+       uint64_t val;
+       uint32_t gpu_id;
+       int ret = 0;
+
+       mutex_lock(&dev->struct_mutex);
+
+       if (rd->open || !gpu) {
+               ret = -EBUSY;
+               goto out;
+       }
+
+       file->private_data = rd;
+       rd->open = true;
+
+       /* the parsing tools need to know gpu-id to know which
+        * register database to load.
+        */
+       gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
+       gpu_id = val;
+
+       rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
+
+out:
+       mutex_unlock(&dev->struct_mutex);
+       return ret;
+}
+
+static int rd_release(struct inode *inode, struct file *file)
+{
+       struct msm_rd_state *rd = inode->i_private;
+       rd->open = false;
+       return 0;
+}
+
+
+static const struct file_operations rd_debugfs_fops = {
+       .owner = THIS_MODULE,
+       .open = rd_open,
+       .read = rd_read,
+       .llseek = no_llseek,
+       .release = rd_release,
+};
+
+int msm_rd_debugfs_init(struct drm_minor *minor)
+{
+       struct msm_drm_private *priv = minor->dev->dev_private;
+       struct msm_rd_state *rd;
+
+       /* only create on first minor: */
+       if (priv->rd)
+               return 0;
+
+       rd = kzalloc(sizeof(*rd), GFP_KERNEL);
+       if (!rd)
+               return -ENOMEM;
+
+       rd->dev = minor->dev;
+       rd->fifo.buf = rd->buf;
+
+       mutex_init(&rd->read_lock);
+       priv->rd = rd;
+
+       init_waitqueue_head(&rd->fifo_event);
+
+       rd->node = kzalloc(sizeof(*rd->node), GFP_KERNEL);
+       if (!rd->node)
+               goto fail;
+
+       rd->ent = debugfs_create_file("rd", S_IFREG | S_IRUGO,
+                       minor->debugfs_root, rd, &rd_debugfs_fops);
+       if (!rd->ent) {
+               DRM_ERROR("Cannot create /sys/kernel/debug/dri/%s/rd\n",
+                               minor->debugfs_root->d_name.name);
+               goto fail;
+       }
+
+       rd->node->minor = minor;
+       rd->node->dent  = rd->ent;
+       rd->node->info_ent = NULL;
+
+       mutex_lock(&minor->debugfs_lock);
+       list_add(&rd->node->list, &minor->debugfs_list);
+       mutex_unlock(&minor->debugfs_lock);
+
+       return 0;
+
+fail:
+       msm_rd_debugfs_cleanup(minor);
+       return -1;
+}
+
+void msm_rd_debugfs_cleanup(struct drm_minor *minor)
+{
+       struct msm_drm_private *priv = minor->dev->dev_private;
+       struct msm_rd_state *rd = priv->rd;
+
+       if (!rd)
+               return;
+
+       priv->rd = NULL;
+
+       debugfs_remove(rd->ent);
+
+       if (rd->node) {
+               mutex_lock(&minor->debugfs_lock);
+               list_del(&rd->node->list);
+               mutex_unlock(&minor->debugfs_lock);
+               kfree(rd->node);
+       }
+
+       mutex_destroy(&rd->read_lock);
+
+       kfree(rd);
+}
+
+/* called under struct_mutex */
+void msm_rd_dump_submit(struct msm_gem_submit *submit)
+{
+       struct drm_device *dev = submit->dev;
+       struct msm_drm_private *priv = dev->dev_private;
+       struct msm_rd_state *rd = priv->rd;
+       char msg[128];
+       int i, n;
+
+       if (!rd->open)
+               return;
+
+       /* writing into fifo is serialized by caller, and
+        * rd->read_lock is used to serialize the reads
+        */
+       WARN_ON(!mutex_is_locked(&dev->struct_mutex));
+
+       n = snprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
+                       TASK_COMM_LEN, current->comm, task_pid_nr(current),
+                       submit->fence);
+
+       rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
+
+       /* could be nice to have an option (module-param?) to snapshot
+        * all the bo's associated with the submit.  Handy to see vtx
+        * buffers, etc.  For now just the cmdstream bo's is enough.
+        */
+
+       for (i = 0; i < submit->nr_cmds; i++) {
+               uint32_t idx  = submit->cmd[i].idx;
+               uint32_t iova = submit->cmd[i].iova;
+               uint32_t szd  = submit->cmd[i].size; /* in dwords */
+               struct msm_gem_object *obj = submit->bos[idx].obj;
+               const char *buf = msm_gem_vaddr_locked(&obj->base);
+
+               buf += iova - submit->bos[idx].iova;
+
+               rd_write_section(rd, RD_GPUADDR,
+                               (uint32_t[2]){ iova, szd * 4 }, 8);
+               rd_write_section(rd, RD_BUFFER_CONTENTS,
+                               buf, szd * 4);
+
+               switch (submit->cmd[i].type) {
+               case MSM_SUBMIT_CMD_IB_TARGET_BUF:
+                       /* ignore IB-targets, we've logged the buffer, the
+                        * parser tool will follow the IB based on the logged
+                        * buffer/gpuaddr, so nothing more to do.
+                        */
+                       break;
+               case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
+               case MSM_SUBMIT_CMD_BUF:
+                       rd_write_section(rd, RD_CMDSTREAM_ADDR,
+                                       (uint32_t[2]){ iova, szd }, 8);
+                       break;
+               }
+       }
+}
+#endif