When using the write()/read() interface for submitting commands, the bsg
driver does not call blk_put_request() on a completed SCSI command until
userspace calls read() to get the command completion. Since scsi-mq
uses a fixed number of preallocated requests, this makes it possible for
userspace to exhaust the entire preallocated supply of requests, leading
to deadlock with the user process stuck in a permanent unkillable I/O
wait in bsg_write() -> ... -> blk_get_request() -> ... -> bt_get().
Note that this deadlock can happen only if scsi-mq is enabled. Prevent
the deadlock by calling blk_put_request() as soon as the SCSI command
completes instead of waiting for userspace to call read().
Cc: <[email protected]> # 3.17+
Signed-off-by: Tony Battersby <[email protected]>
---
For inclusion in kernel 3.20.
Note: I did a number of tests with this patch, but I do not have any
programs to test commands with bidirectional data transfer. I would
appreciate if someone could test that.
description of changes:
*) For ioctl(SG_IO), allocate a struct bsg_command on the stack instead
of allocating all the component fields on the stack. This helps
simplify the code.
*) Split blk_complete_sgv4_hdr_rq() into two functions:
bsg_complete_command_rq() and bsg_complete_command_usercontext().
*) Call bsg_complete_command_rq() from the asynchronous bsg_rq_end_io()
or after a synchronous call to blk_execute_rq().
*) The important change to fix the deadlock is that bsg_rq_end_io() now
calls __blk_put_request().
*) Replace calls to blk_complete_sgv4_hdr_rq() with calls to
bsg_complete_command_usercontext().
*) Use bc->err to pass around negative error values.
*) If rq->errors < 0, do not use it to set
device_status/transport_status/driver_status.
*) Check the return value of blk_rq_unmap_user().
*) Remove bc->rq as it is no longer needed.
I will send a separate patch to fix the same problem in the sg driver.
--- linux-3.19.0/block/bsg.c.orig 2015-02-08 21:54:22.000000000 -0500
+++ linux-3.19.0/block/bsg.c 2015-02-11 11:00:57.000000000 -0500
@@ -80,7 +80,6 @@ static struct kmem_cache *bsg_cmd_cachep
struct bsg_command {
struct bsg_device *bd;
struct list_head list;
- struct request *rq;
struct bio *bio;
struct bio *bidi_bio;
int err;
@@ -88,6 +87,10 @@ struct bsg_command {
char sense[SCSI_SENSE_BUFFERSIZE];
};
+static void bsg_complete_command_rq(struct bsg_command *bc,
+ struct request *rq,
+ bool holding_queue_lock);
+
static void bsg_free_command(struct bsg_command *bc)
{
struct bsg_device *bd = bc->bd;
@@ -346,6 +349,8 @@ static void bsg_rq_end_io(struct request
bc->hdr.duration = jiffies_to_msecs(jiffies - bc->hdr.duration);
+ bsg_complete_command_rq(bc, rq, true);
+
spin_lock_irqsave(&bd->lock, flags);
list_move_tail(&bc->list, &bd->done_list);
bd->done_cmds++;
@@ -366,7 +371,6 @@ static void bsg_add_command(struct bsg_d
/*
* add bc command to busy queue and submit rq for io
*/
- bc->rq = rq;
bc->bio = rq->bio;
if (rq->next_rq)
bc->bidi_bio = rq->next_rq->bio;
@@ -426,60 +430,100 @@ static struct bsg_command *bsg_get_done_
return bc;
}
-static int blk_complete_sgv4_hdr_rq(struct request *rq, struct sg_io_v4 *hdr,
- struct bio *bio, struct bio *bidi_bio)
+/*
+ * Post-process a completed request. This should be called immediately after
+ * the request completes so that its resources can be reused for other
+ * commands.
+ */
+static void bsg_complete_command_rq(struct bsg_command *bc,
+ struct request *rq,
+ bool holding_queue_lock)
{
- int ret = 0;
-
- dprintk("rq %p bio %p 0x%x\n", rq, bio, rq->errors);
+ dprintk("rq %p 0x%x\n", rq, rq->errors);
/*
* fill in all the output members
+ *
+ * If the request generated a negative error number, return it from
+ * the system call (e.g. read() or ioctl()); if it's just a protocol
+ * response (i.e. non negative), include it in the response hdr.
*/
- hdr->device_status = rq->errors & 0xff;
- hdr->transport_status = host_byte(rq->errors);
- hdr->driver_status = driver_byte(rq->errors);
- hdr->info = 0;
- if (hdr->device_status || hdr->transport_status || hdr->driver_status)
- hdr->info |= SG_INFO_CHECK;
- hdr->response_len = 0;
-
- if (rq->sense_len && hdr->response) {
- int len = min_t(unsigned int, hdr->max_response_len,
- rq->sense_len);
-
- ret = copy_to_user((void __user *)(unsigned long)hdr->response,
- rq->sense, len);
- if (!ret)
- hdr->response_len = len;
- else
- ret = -EFAULT;
- }
+ if (unlikely(rq->errors < 0)) {
+ bc->err = rq->errors;
+ bc->hdr.device_status = 0;
+ bc->hdr.transport_status = DID_ERROR;
+ bc->hdr.driver_status = DRIVER_ERROR;
+ } else {
+ bc->hdr.device_status = rq->errors & 0xff;
+ bc->hdr.transport_status = host_byte(rq->errors);
+ bc->hdr.driver_status = driver_byte(rq->errors);
+ }
+ bc->hdr.info = 0;
+ if (bc->hdr.device_status || bc->hdr.transport_status ||
+ bc->hdr.driver_status)
+ bc->hdr.info |= SG_INFO_CHECK;
+
+ bc->hdr.response_len =
+ (!bc->hdr.response) ? 0 : min_t(unsigned int,
+ bc->hdr.max_response_len,
+ rq->sense_len);
if (rq->next_rq) {
- hdr->dout_resid = rq->resid_len;
- hdr->din_resid = rq->next_rq->resid_len;
- blk_rq_unmap_user(bidi_bio);
- blk_put_request(rq->next_rq);
+ bc->hdr.dout_resid = rq->resid_len;
+ bc->hdr.din_resid = rq->next_rq->resid_len;
+ if (holding_queue_lock)
+ __blk_put_request(rq->q, rq->next_rq);
+ else
+ blk_put_request(rq->next_rq);
} else if (rq_data_dir(rq) == READ)
- hdr->din_resid = rq->resid_len;
+ bc->hdr.din_resid = rq->resid_len;
else
- hdr->dout_resid = rq->resid_len;
+ bc->hdr.dout_resid = rq->resid_len;
/*
- * If the request generated a negative error number, return it
- * (providing we aren't already returning an error); if it's
- * just a protocol response (i.e. non negative), that gets
- * processed above.
+ * Free the request as soon as it is complete so that its resources
+ * can be reused without waiting for userspace to read() the
+ * result. But keep the associated bios (if any) around until
+ * bsg_complete_command_usercontext() can be called from user context.
*/
- if (!ret && rq->errors < 0)
- ret = rq->errors;
-
- blk_rq_unmap_user(bio);
if (rq->cmd != rq->__cmd)
kfree(rq->cmd);
- blk_put_request(rq);
+ if (holding_queue_lock)
+ __blk_put_request(rq->q, rq);
+ else
+ blk_put_request(rq);
+}
- return ret;
+/*
+ * Post-process a completed request from user context.
+ */
+static int bsg_complete_command_usercontext(struct bsg_command *bc)
+{
+ int ret;
+
+ dprintk("bio %p\n", bc->bio);
+
+ if (bc->hdr.response_len) {
+ ret = copy_to_user(
+ (void __user *)(unsigned long)bc->hdr.response,
+ bc->sense, bc->hdr.response_len);
+ if (unlikely(ret)) {
+ bc->hdr.response_len = 0;
+ bc->err = -EFAULT;
+ }
+ }
+
+ if (bc->bidi_bio) {
+ ret = blk_rq_unmap_user(bc->bidi_bio);
+ if (unlikely(ret))
+ bc->err = ret;
+ }
+ if (bc->bio) {
+ ret = blk_rq_unmap_user(bc->bio);
+ if (unlikely(ret))
+ bc->err = ret;
+ }
+
+ return bc->err;
}
static int bsg_complete_all_commands(struct bsg_device *bd)
@@ -520,8 +564,7 @@ static int bsg_complete_all_commands(str
if (IS_ERR(bc))
break;
- tret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
- bc->bidi_bio);
+ tret = bsg_complete_command_usercontext(bc);
if (!ret)
ret = tret;
@@ -555,8 +598,7 @@ __bsg_read(char __user *buf, size_t coun
* after completing the request. so do that here,
* bsg_complete_work() cannot do that for us
*/
- ret = blk_complete_sgv4_hdr_rq(bc->rq, &bc->hdr, bc->bio,
- bc->bidi_bio);
+ ret = bsg_complete_command_usercontext(bc);
if (copy_to_user(buf, &bc->hdr, sizeof(bc->hdr)))
ret = -EFAULT;
@@ -926,28 +968,29 @@ static long bsg_ioctl(struct file *file,
return scsi_cmd_ioctl(bd->queue, NULL, file->f_mode, cmd, uarg);
}
case SG_IO: {
+ struct bsg_command bc;
struct request *rq;
- struct bio *bio, *bidi_bio = NULL;
- struct sg_io_v4 hdr;
int at_head;
- u8 sense[SCSI_SENSE_BUFFERSIZE];
- if (copy_from_user(&hdr, uarg, sizeof(hdr)))
+ if (copy_from_user(&bc.hdr, uarg, sizeof(bc.hdr)))
return -EFAULT;
- rq = bsg_map_hdr(bd, &hdr, file->f_mode & FMODE_WRITE, sense);
+ rq = bsg_map_hdr(bd, &bc.hdr, file->f_mode & FMODE_WRITE,
+ bc.sense);
if (IS_ERR(rq))
return PTR_ERR(rq);
- bio = rq->bio;
- if (rq->next_rq)
- bidi_bio = rq->next_rq->bio;
+ bc.bd = bd;
+ bc.bio = rq->bio;
+ bc.bidi_bio = (rq->next_rq) ? rq->next_rq->bio : NULL;
+ bc.err = 0;
- at_head = (0 == (hdr.flags & BSG_FLAG_Q_AT_TAIL));
+ at_head = (0 == (bc.hdr.flags & BSG_FLAG_Q_AT_TAIL));
blk_execute_rq(bd->queue, NULL, rq, at_head);
- ret = blk_complete_sgv4_hdr_rq(rq, &hdr, bio, bidi_bio);
+ bsg_complete_command_rq(&bc, rq, false);
+ ret = bsg_complete_command_usercontext(&bc);
- if (copy_to_user(uarg, &hdr, sizeof(hdr)))
+ if (copy_to_user(uarg, &bc.hdr, sizeof(bc.hdr)))
return -EFAULT;
return ret;
On 02/11/2015 11:31 AM, Tony Battersby wrote:
> Note: I did a number of tests with this patch, but I do not have any
> programs to test commands with bidirectional data transfer. I would
> appreciate if someone could test that.
Replying to myself, I was able to test bidirectional data transfer using
scsi_debug and sg_raw, and it worked just fine.
Tony