[SCSI] Revert "sg: use rwsem to solve race during exclusive open"
authorJames Bottomley <JBottomley@Parallels.com>
Fri, 25 Oct 2013 09:27:02 +0000 (10:27 +0100)
committerJames Bottomley <JBottomley@Parallels.com>
Fri, 25 Oct 2013 09:59:54 +0000 (10:59 +0100)
This reverts commit 15b06f9a02406e5460001db6d5af5c738cd3d4e7.

This is one of four patches that was causing this bug

[  205.372823] ================================================
[  205.372901] [ BUG: lock held when returning to user space! ]
[  205.372979] 3.12.0-rc6-hw-debug-pagealloc+ #67 Not tainted
[  205.373055] ------------------------------------------------
[  205.373132] megarc.bin/5283 is leaving the kernel with locks still held!
[  205.373212] 1 lock held by megarc.bin/5283:
[  205.373285]  #0:  (&sdp->o_sem){.+.+..}, at: [<ffffffff8161e650>] sg_open+0x3a0/0x4d0

Cc: Vaughan Cao <vaughan.cao@oracle.com>
Acked-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
drivers/scsi/sg.c

index 4efa9b5884b73a7543fa87b0ca6fef60d28f8efe..df5e961484e108312f6f01765882dd4319d584f7 100644 (file)
@@ -170,11 +170,11 @@ typedef struct sg_fd {            /* holds the state of a file descriptor */
 
 typedef struct sg_device { /* holds the state of each scsi generic device */
        struct scsi_device *device;
+       wait_queue_head_t o_excl_wait;  /* queue open() when O_EXCL in use */
        int sg_tablesize;       /* adapter's max scatter-gather table size */
        u32 index;              /* device index number */
        /* sfds is protected by sg_index_lock */
        struct list_head sfds;
-       struct rw_semaphore o_sem;      /* exclude open should hold this rwsem */
        volatile char detached; /* 0->attached, 1->detached pending removal */
        /* exclude protected by sg_open_exclusive_lock */
        char exclude;           /* opened for exclusive access */
@@ -265,6 +265,7 @@ sg_open(struct inode *inode, struct file *filp)
        struct request_queue *q;
        Sg_device *sdp;
        Sg_fd *sfp;
+       int res;
        int retval;
 
        nonseekable_open(inode, filp);
@@ -293,35 +294,35 @@ sg_open(struct inode *inode, struct file *filp)
                goto error_out;
        }
 
-       if ((flags & O_EXCL) && (O_RDONLY == (flags & O_ACCMODE))) {
-               retval = -EPERM; /* Can't lock it with read only access */
-               goto error_out;
-       }
-       if (flags & O_NONBLOCK) {
-               if (flags & O_EXCL) {
-                       if (!down_write_trylock(&sdp->o_sem)) {
-                               retval = -EBUSY;
-                               goto error_out;
-                       }
-               } else {
-                       if (!down_read_trylock(&sdp->o_sem)) {
-                               retval = -EBUSY;
-                               goto error_out;
-                       }
+       if (flags & O_EXCL) {
+               if (O_RDONLY == (flags & O_ACCMODE)) {
+                       retval = -EPERM; /* Can't lock it with read only access */
+                       goto error_out;
+               }
+               if (!sfds_list_empty(sdp) && (flags & O_NONBLOCK)) {
+                       retval = -EBUSY;
+                       goto error_out;
+               }
+               res = wait_event_interruptible(sdp->o_excl_wait,
+                                          ((!sfds_list_empty(sdp) || get_exclude(sdp)) ? 0 : set_exclude(sdp, 1)));
+               if (res) {
+                       retval = res;   /* -ERESTARTSYS because signal hit process */
+                       goto error_out;
+               }
+       } else if (get_exclude(sdp)) {  /* some other fd has an exclusive lock on dev */
+               if (flags & O_NONBLOCK) {
+                       retval = -EBUSY;
+                       goto error_out;
+               }
+               res = wait_event_interruptible(sdp->o_excl_wait, !get_exclude(sdp));
+               if (res) {
+                       retval = res;   /* -ERESTARTSYS because signal hit process */
+                       goto error_out;
                }
-       } else {
-               if (flags & O_EXCL)
-                       down_write(&sdp->o_sem);
-               else
-                       down_read(&sdp->o_sem);
        }
-       /* Since write lock is held, no need to check sfd_list */
-       if (flags & O_EXCL)
-               set_exclude(sdp, 1);
-
        if (sdp->detached) {
                retval = -ENODEV;
-               goto sem_out;
+               goto error_out;
        }
        if (sfds_list_empty(sdp)) {     /* no existing opens on this device */
                sdp->sgdebug = 0;
@@ -330,18 +331,17 @@ sg_open(struct inode *inode, struct file *filp)
        }
        if ((sfp = sg_add_sfp(sdp, dev)))
                filp->private_data = sfp;
-               /* retval is already provably zero at this point because of the
-                * check after retval = scsi_autopm_get_device(sdp->device))
-                */
        else {
-               retval = -ENOMEM;
-sem_out:
                if (flags & O_EXCL) {
                        set_exclude(sdp, 0);    /* undo if error */
-                       up_write(&sdp->o_sem);
-               } else
-                       up_read(&sdp->o_sem);
+                       wake_up_interruptible(&sdp->o_excl_wait);
+               }
+               retval = -ENOMEM;
+               goto error_out;
+       }
+       retval = 0;
 error_out:
+       if (retval) {
                scsi_autopm_put_device(sdp->device);
 sdp_put:
                scsi_device_put(sdp->device);
@@ -358,18 +358,13 @@ sg_release(struct inode *inode, struct file *filp)
 {
        Sg_device *sdp;
        Sg_fd *sfp;
-       int excl;
 
        if ((!(sfp = (Sg_fd *) filp->private_data)) || (!(sdp = sfp->parentdp)))
                return -ENXIO;
        SCSI_LOG_TIMEOUT(3, printk("sg_release: %s\n", sdp->disk->disk_name));
 
-       excl = get_exclude(sdp);
        set_exclude(sdp, 0);
-       if (excl)
-               up_write(&sdp->o_sem);
-       else
-               up_read(&sdp->o_sem);
+       wake_up_interruptible(&sdp->o_excl_wait);
 
        scsi_autopm_put_device(sdp->device);
        kref_put(&sfp->f_ref, sg_remove_sfp);
@@ -1421,7 +1416,7 @@ static Sg_device *sg_alloc(struct gendisk *disk, struct scsi_device *scsidp)
        sdp->disk = disk;
        sdp->device = scsidp;
        INIT_LIST_HEAD(&sdp->sfds);
-       init_rwsem(&sdp->o_sem);
+       init_waitqueue_head(&sdp->o_excl_wait);
        sdp->sg_tablesize = queue_max_segments(q);
        sdp->index = k;
        kref_init(&sdp->d_ref);
@@ -2132,11 +2127,13 @@ static void sg_remove_sfp_usercontext(struct work_struct *work)
 static void sg_remove_sfp(struct kref *kref)
 {
        struct sg_fd *sfp = container_of(kref, struct sg_fd, f_ref);
+       struct sg_device *sdp = sfp->parentdp;
        unsigned long iflags;
 
        write_lock_irqsave(&sg_index_lock, iflags);
        list_del(&sfp->sfd_siblings);
        write_unlock_irqrestore(&sg_index_lock, iflags);
+       wake_up_interruptible(&sdp->o_excl_wait);
 
        INIT_WORK(&sfp->ew.work, sg_remove_sfp_usercontext);
        schedule_work(&sfp->ew.work);