powerpc/spufs: Use state_mutex for switch_log locking, and prevent multiple openers

Currently, we use ctx->mapping_lock and ctx->switch_log->lock for the
context switch log. The mapping lock only prevents concurrent open()s,
so we require the switch_lock->lock for reads.

Since writes to the switch log buffer occur on context switches, we're
better off synchronising with the state_mutex, which is held during a
switch. Since we're serialised througout the buffer reads and writes,
we can use the state mutex to protect open and release too, and
can now kfree() the log buffer on release. This allows us to perform
the switch log notify without taking any extra locks.

Because the buffer is only present while the file is open, we can use
it to prevent multiple simultaneous openers.

Signed-off-by: Jeremy Kerr <jk@ozlabs.org>

diff --git a/arch/powerpc/platforms/cell/spufs/file.c b/arch/powerpc/platforms/cell/spufs/file.c
index 010a51f59796b7a9068103bb6598843f9195fff4..adb5abb9af5d71f0cb7cc6eeb319f400ec9a0eb4 100644 (file)
--- a/arch/powerpc/platforms/cell/spufs/file.c
+++ b/arch/powerpc/platforms/cell/spufs/file.c
@@ -2426,38 +2426,48 @@ static inline int spufs_switch_log_avail(struct spu_context *ctx)
 static int spufs_switch_log_open(struct inode *inode, struct file *file)
 {
        struct spu_context *ctx = SPUFS_I(inode)->i_ctx;
+       int rc;
+
+       rc = spu_acquire(ctx);
+       if (rc)
+               return rc;
 
-       /*
-        * We (ab-)use the mapping_lock here because it serves the similar
-        * purpose for synchronizing open/close elsewhere.  Maybe it should
-        * be renamed eventually.
-        */
-       mutex_lock(&ctx->mapping_lock);
        if (ctx->switch_log) {
-               spin_lock(&ctx->switch_log->lock);
-               ctx->switch_log->head = 0;
-               ctx->switch_log->tail = 0;
-               spin_unlock(&ctx->switch_log->lock);
-       } else {
-               /*
-                * We allocate the switch log data structures on first open.
-                * They will never be free because we assume a context will
-                * be traced until it goes away.
-                */
-               ctx->switch_log = kzalloc(sizeof(struct switch_log) +
-                       SWITCH_LOG_BUFSIZE * sizeof(struct switch_log_entry),
-                       GFP_KERNEL);
-               if (!ctx->switch_log)
-                       goto out;
-               spin_lock_init(&ctx->switch_log->lock);
-               init_waitqueue_head(&ctx->switch_log->wait);
+               rc = -EBUSY;
+               goto out;
        }
-       mutex_unlock(&ctx->mapping_lock);
+
+       ctx->switch_log = kzalloc(sizeof(struct switch_log) +
+               SWITCH_LOG_BUFSIZE * sizeof(struct switch_log_entry),
+               GFP_KERNEL);
+
+       if (!ctx->switch_log) {
+               rc = -ENOMEM;
+               goto out;
+       }
+
+       init_waitqueue_head(&ctx->switch_log->wait);
+       rc = 0;
+
+out:
+       spu_release(ctx);
+       return rc;
+}
+
+static int spufs_switch_log_release(struct inode *inode, struct file *file)
+{
+       struct spu_context *ctx = SPUFS_I(inode)->i_ctx;
+       int rc;
+
+       rc = spu_acquire(ctx);
+       if (rc)
+               return rc;
+
+       kfree(ctx->switch_log);
+       ctx->switch_log = NULL;
+       spu_release(ctx);
 
        return 0;
- out:
-       mutex_unlock(&ctx->mapping_lock);
-       return -ENOMEM;
 }
 
 static int switch_log_sprint(struct spu_context *ctx, char *tbuf, int n)
@@ -2485,42 +2495,46 @@ static ssize_t spufs_switch_log_read(struct file *file, char __user *buf,
        if (!buf || len < 0)
                return -EINVAL;
 
+       error = spu_acquire(ctx);
+       if (error)
+               return error;
+
        while (cnt < len) {
                char tbuf[128];
                int width;
 
                if (file->f_flags & O_NONBLOCK) {
-                       if (spufs_switch_log_used(ctx) <= 0)
-                               return cnt ? cnt : -EAGAIN;
+                       if (spufs_switch_log_used(ctx) == 0) {
+                               error = -EAGAIN;
+                               break;
+                       }
                } else {
-                       /* Wait for data in buffer */
-                       error = wait_event_interruptible(ctx->switch_log->wait,
+                       /* spufs_wait will drop the mutex and re-acquire,
+                        * but since we're in read(), the file cannot be
+                        * _released (and so ctx->switch_log is stable).
+                        */
+                       error = spufs_wait(ctx->switch_log->wait,
                                        spufs_switch_log_used(ctx) > 0);
+
+                       /* On error, spufs_wait returns without the
+                        * state mutex held */
                        if (error)
-                               break;
+                               return error;
                }
 
-               spin_lock(&ctx->switch_log->lock);
-               if (ctx->switch_log->head == ctx->switch_log->tail) {
-                       /* multiple readers race? */
-                       spin_unlock(&ctx->switch_log->lock);
+               /* We may have had entries read from underneath us while we
+                * dropped the mutex in spufs_wait, so re-check */
+               if (ctx->switch_log->head == ctx->switch_log->tail)
                        continue;
-               }
 
                width = switch_log_sprint(ctx, tbuf, sizeof(tbuf));
-               if (width < len) {
+               if (width < len)
                        ctx->switch_log->tail =
                                (ctx->switch_log->tail + 1) %
                                 SWITCH_LOG_BUFSIZE;
-               }
-
-               spin_unlock(&ctx->switch_log->lock);
-
-               /*
-                * If the record is greater than space available return
-                * partial buffer (so far)
-                */
-               if (width >= len)
+               else
+                       /* If the record is greater than space available return
+                        * partial buffer (so far) */
                        break;
 
                error = copy_to_user(buf + cnt, tbuf, width);
@@ -2529,6 +2543,8 @@ static ssize_t spufs_switch_log_read(struct file *file, char __user *buf,
                cnt += width;
        }
 
+       spu_release(ctx);
+
        return cnt == 0 ? error : cnt;
 }
 
@@ -2537,29 +2553,41 @@ static unsigned int spufs_switch_log_poll(struct file *file, poll_table *wait)
        struct inode *inode = file->f_path.dentry->d_inode;
        struct spu_context *ctx = SPUFS_I(inode)->i_ctx;
        unsigned int mask = 0;
+       int rc;
 
        poll_wait(file, &ctx->switch_log->wait, wait);
 
+       rc = spu_acquire(ctx);
+       if (rc)
+               return rc;
+
        if (spufs_switch_log_used(ctx) > 0)
                mask |= POLLIN;
 
+       spu_release(ctx);
+
        return mask;
 }
 
 static const struct file_operations spufs_switch_log_fops = {
-       .owner  = THIS_MODULE,
-       .open   = spufs_switch_log_open,
-       .read   = spufs_switch_log_read,
-       .poll   = spufs_switch_log_poll,
+       .owner          = THIS_MODULE,
+       .open           = spufs_switch_log_open,
+       .read           = spufs_switch_log_read,
+       .poll           = spufs_switch_log_poll,
+       .release        = spufs_switch_log_release,
 };
 
+/**
+ * Log a context switch event to a switch log reader.
+ *
+ * Must be called with ctx->state_mutex held.
+ */
 void spu_switch_log_notify(struct spu *spu, struct spu_context *ctx,
                u32 type, u32 val)
 {
        if (!ctx->switch_log)
                return;
 
-       spin_lock(&ctx->switch_log->lock);
        if (spufs_switch_log_avail(ctx) > 1) {
                struct switch_log_entry *p;
 
@@ -2573,7 +2601,6 @@ void spu_switch_log_notify(struct spu *spu, struct spu_context *ctx,
                ctx->switch_log->head =
                        (ctx->switch_log->head + 1) % SWITCH_LOG_BUFSIZE;
        }
-       spin_unlock(&ctx->switch_log->lock);
 
        wake_up(&ctx->switch_log->wait);
 }

diff --git a/arch/powerpc/platforms/cell/spufs/run.c b/arch/powerpc/platforms/cell/spufs/run.c
index c9bb7cfd3dca8069434a9071860a6b99cff19022..c58bd36b0c5b1045c170c7135a2e30f884f204fa 100644 (file)
--- a/arch/powerpc/platforms/cell/spufs/run.c
+++ b/arch/powerpc/platforms/cell/spufs/run.c
@@ -249,6 +249,7 @@ static int spu_run_fini(struct spu_context *ctx, u32 *npc,
 
        spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
        clear_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags);
+       spu_switch_log_notify(NULL, ctx, SWITCH_LOG_EXIT, *status);
        spu_release(ctx);
 
        if (signal_pending(current))
@@ -417,8 +418,6 @@ long spufs_run_spu(struct spu_context *ctx, u32 *npc, u32 *event)
        ret = spu_run_fini(ctx, npc, &status);
        spu_yield(ctx);
 
-       spu_switch_log_notify(NULL, ctx, SWITCH_LOG_EXIT, status);
-
        if ((status & SPU_STATUS_STOPPED_BY_STOP) &&
            (((status >> SPU_STOP_STATUS_SHIFT) & 0x3f00) == 0x2100))
                ctx->stats.libassist++;

diff --git a/arch/powerpc/platforms/cell/spufs/spufs.h b/arch/powerpc/platforms/cell/spufs/spufs.h
index 8ae8ef9dfc22dd5a73bfba0899ddffa7d375b693..15c62d3ca129f55d6ebfe56fe7b93fcb8939f42c 100644 (file)
--- a/arch/powerpc/platforms/cell/spufs/spufs.h
+++ b/arch/powerpc/platforms/cell/spufs/spufs.h
@@ -65,7 +65,6 @@ enum {
 };
 
 struct switch_log {
-       spinlock_t              lock;
        wait_queue_head_t       wait;
        unsigned long           head;
        unsigned long           tail;