mm: dmapool: add/remove sysfs file outside of the pool lock lock
authorSebastian Andrzej Siewior <bigeasy@linutronix.de>
Thu, 9 Oct 2014 22:28:50 +0000 (15:28 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 10 Oct 2014 02:25:59 +0000 (22:25 -0400)
cat /sys/.../pools followed by removal the device leads to:

|======================================================
|[ INFO: possible circular locking dependency detected ]
|3.17.0-rc4+ #1498 Not tainted
|-------------------------------------------------------
|rmmod/2505 is trying to acquire lock:
| (s_active#28){++++.+}, at: [<c017f754>] kernfs_remove_by_name_ns+0x3c/0x88
|
|but task is already holding lock:
| (pools_lock){+.+.+.}, at: [<c011494c>] dma_pool_destroy+0x18/0x17c
|
|which lock already depends on the new lock.
|the existing dependency chain (in reverse order) is:
|
|-> #1 (pools_lock){+.+.+.}:
|   [<c0114ae8>] show_pools+0x30/0xf8
|   [<c0313210>] dev_attr_show+0x1c/0x48
|   [<c0180e84>] sysfs_kf_seq_show+0x88/0x10c
|   [<c017f960>] kernfs_seq_show+0x24/0x28
|   [<c013efc4>] seq_read+0x1b8/0x480
|   [<c011e820>] vfs_read+0x8c/0x148
|   [<c011ea10>] SyS_read+0x40/0x8c
|   [<c000e960>] ret_fast_syscall+0x0/0x48
|
|-> #0 (s_active#28){++++.+}:
|   [<c017e9ac>] __kernfs_remove+0x258/0x2ec
|   [<c017f754>] kernfs_remove_by_name_ns+0x3c/0x88
|   [<c0114a7c>] dma_pool_destroy+0x148/0x17c
|   [<c03ad288>] hcd_buffer_destroy+0x20/0x34
|   [<c03a4780>] usb_remove_hcd+0x110/0x1a4

The problem is the lock order of pools_lock and kernfs_mutex in
dma_pool_destroy() vs show_pools() call path.

This patch breaks out the creation of the sysfs file outside of the
pools_lock mutex.  The newly added pools_reg_lock ensures that there is no
race of create vs destroy code path in terms whether or not the sysfs file
has to be deleted (and was it deleted before we try to create a new one)
and what to do if device_create_file() failed.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/dmapool.c

index ba8019b063e18ecb14da4cf3a64dfa671586f8e0..2372ed5a33d3dc6efe19f678686f0829c6011a29 100644 (file)
@@ -62,6 +62,7 @@ struct dma_page {             /* cacheable header for 'allocation' bytes */
 };
 
 static DEFINE_MUTEX(pools_lock);
+static DEFINE_MUTEX(pools_reg_lock);
 
 static ssize_t
 show_pools(struct device *dev, struct device_attribute *attr, char *buf)
@@ -132,6 +133,7 @@ struct dma_pool *dma_pool_create(const char *name, struct device *dev,
 {
        struct dma_pool *retval;
        size_t allocation;
+       bool empty = false;
 
        if (align == 0) {
                align = 1;
@@ -172,15 +174,34 @@ struct dma_pool *dma_pool_create(const char *name, struct device *dev,
 
        INIT_LIST_HEAD(&retval->pools);
 
+       /*
+        * pools_lock ensures that the ->dma_pools list does not get corrupted.
+        * pools_reg_lock ensures that there is not a race between
+        * dma_pool_create() and dma_pool_destroy() or within dma_pool_create()
+        * when the first invocation of dma_pool_create() failed on
+        * device_create_file() and the second assumes that it has been done (I
+        * know it is a short window).
+        */
+       mutex_lock(&pools_reg_lock);
        mutex_lock(&pools_lock);
-       if (list_empty(&dev->dma_pools) &&
-           device_create_file(dev, &dev_attr_pools)) {
-               kfree(retval);
-               retval = NULL;
-       } else
-               list_add(&retval->pools, &dev->dma_pools);
+       if (list_empty(&dev->dma_pools))
+               empty = true;
+       list_add(&retval->pools, &dev->dma_pools);
        mutex_unlock(&pools_lock);
-
+       if (empty) {
+               int err;
+
+               err = device_create_file(dev, &dev_attr_pools);
+               if (err) {
+                       mutex_lock(&pools_lock);
+                       list_del(&retval->pools);
+                       mutex_unlock(&pools_lock);
+                       mutex_unlock(&pools_reg_lock);
+                       kfree(retval);
+                       return NULL;
+               }
+       }
+       mutex_unlock(&pools_reg_lock);
        return retval;
 }
 EXPORT_SYMBOL(dma_pool_create);
@@ -251,11 +272,17 @@ static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
  */
 void dma_pool_destroy(struct dma_pool *pool)
 {
+       bool empty = false;
+
+       mutex_lock(&pools_reg_lock);
        mutex_lock(&pools_lock);
        list_del(&pool->pools);
        if (pool->dev && list_empty(&pool->dev->dma_pools))
-               device_remove_file(pool->dev, &dev_attr_pools);
+               empty = true;
        mutex_unlock(&pools_lock);
+       if (empty)
+               device_remove_file(pool->dev, &dev_attr_pools);
+       mutex_unlock(&pools_reg_lock);
 
        while (!list_empty(&pool->page_list)) {
                struct dma_page *page;