Skip to contentSkip to conversation

cio: kernel stack overflow.

S390 Martin Schwidefsky 5 participants 10 messages Aug 30, 2006 – Aug 31, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

From: Heiko Carstens <[email protected]>

[S390] cio: kernel stack overflow.

Use different kind of assignment to make sure gcc doesn't create code that creates temp variables on the stack, assigns values to it and copies the content of the whole temp variable to the destination. This reduces stack usage of e.g. ccwgroup_driver_register from 976 to 48 bytes instead.

---

 drivers/s390/cio/ccwgroup.c   |   14 +++++---------
 drivers/s390/cio/chsc.c       |    6 ++----
 drivers/s390/cio/device.c     |   19 +++++++------------
 drivers/s390/cio/device_fsm.c |   20 ++++++++------------
 4 files changed, 22 insertions(+), 37 deletions(-)
drivers/s390/cio/ccwgroup.c +5 -9
diff -urpN linux-2.6/drivers/s390/cio/ccwgroup.c linux-2.6-patched/drivers/s390/cio/ccwgroup.c
@@ -183,11 +183,9 @@ ccwgroup_create(struct device *root,
 
 	gdev->creator_id = creator_id;
 	gdev->count = argc;
-	gdev->dev = (struct device ) {
-		.bus = &ccwgroup_bus_type,
-		.parent = root,
-		.release = ccwgroup_release,
-	};
+	gdev->dev.bus = &ccwgroup_bus_type;
+	gdev->dev.parent = root;
+	gdev->dev.release = ccwgroup_release;
 
 	snprintf (gdev->dev.bus_id, BUS_ID_SIZE, "%s",
 			gdev->cdev[0]->dev.bus_id);
@@ -391,10 +389,8 @@ int
 ccwgroup_driver_register (struct ccwgroup_driver *cdriver)
 {
 	/* register our new driver with the core */
-	cdriver->driver = (struct device_driver) {
-		.bus = &ccwgroup_bus_type,
-		.name = cdriver->name,
-	};
+	cdriver->driver.bus = &ccwgroup_bus_type;
+	cdriver->driver.name = cdriver->name;
 
 	return driver_register(&cdriver->driver);
 }
drivers/s390/cio/chsc.c +2 -4
diff -urpN linux-2.6/drivers/s390/cio/chsc.c linux-2.6-patched/drivers/s390/cio/chsc.c
@@ -1391,10 +1391,8 @@ new_channel_path(int chpid)
 	/* fill in status, etc. */
 	chp->id = chpid;
 	chp->state = 1;
-	chp->dev = (struct device) {
-		.parent  = &css[0]->device,
-		.release = chp_release,
-	};
+	chp->dev.parent = &css[0]->device;
+	chp->dev.release = chp_release;
 	snprintf(chp->dev.bus_id, BUS_ID_SIZE, "chp0.%x", chpid);
 
 	/* Obtain channel path description and fill it in. */
drivers/s390/cio/device.c +7 -12
diff -urpN linux-2.6/drivers/s390/cio/device.c linux-2.6-patched/drivers/s390/cio/device.c
@@ -556,12 +556,11 @@ get_disc_ccwdev_by_devno(unsigned int de
 			 struct ccw_device *sibling)
 {
 	struct device *dev;
-	struct match_data data = {
-		.devno   = devno,
-		.ssid    = ssid,
-		.sibling = sibling,
-	};
+	struct match_data data;
 
+	data.devno = devno;
+	data.ssid = ssid;
+	data.sibling = sibling;
 	dev = bus_find_device(&ccw_bus_type, NULL, &data, match_devno);
 
 	return dev ? to_ccwdev(dev) : NULL;
@@ -835,10 +834,8 @@ io_subchannel_probe (struct subchannel *
 		return -ENOMEM;
 	}
 	atomic_set(&cdev->private->onoff, 0);
-	cdev->dev = (struct device) {
-		.parent = &sch->dev,
-		.release = ccw_device_release,
-	};
+	cdev->dev.parent = &sch->dev;
+	cdev->dev.release = ccw_device_release;
 	INIT_LIST_HEAD(&cdev->private->kick_work.entry);
 	/* Do first half of device_register. */
 	device_initialize(&cdev->dev);
@@ -977,9 +974,7 @@ ccw_device_console_enable (struct ccw_de
 	int rc;
 
 	/* Initialize the ccw_device structure. */
-	cdev->dev = (struct device) {
-		.parent = &sch->dev,
-	};
+	cdev->dev.parent= &sch->dev;
 	rc = io_subchannel_recog(cdev, sch);
 	if (rc)
 		return rc;
drivers/s390/cio/device_fsm.c +8 -12
diff -urpN linux-2.6/drivers/s390/cio/device_fsm.c linux-2.6-patched/drivers/s390/cio/device_fsm.c
@@ -267,12 +267,10 @@ ccw_device_recog_done(struct ccw_device 
 			notify = 1;
 		}
 		/* fill out sense information */
-		cdev->id = (struct ccw_device_id) {
-			.cu_type   = cdev->private->senseid.cu_type,
-			.cu_model  = cdev->private->senseid.cu_model,
-			.dev_type  = cdev->private->senseid.dev_type,
-			.dev_model = cdev->private->senseid.dev_model,
-		};
+		cdev->id.cu_type   = cdev->private->senseid.cu_type;
+		cdev->id.cu_model  = cdev->private->senseid.cu_model;
+		cdev->id.dev_type  = cdev->private->senseid.dev_type;
+		cdev->id.dev_model = cdev->private->senseid.dev_model;
 		if (notify) {
 			cdev->private->state = DEV_STATE_OFFLINE;
 			if (same_dev) {
@@ -566,12 +564,10 @@ ccw_device_verify_done(struct ccw_device
 		/* Deliver fake irb to device driver, if needed. */
 		if (cdev->private->flags.fake_irb) {
 			memset(&cdev->private->irb, 0, sizeof(struct irb));
-			cdev->private->irb.scsw = (struct scsw) {
-				.cc = 1,
-				.fctl = SCSW_FCTL_START_FUNC,
-				.actl = SCSW_ACTL_START_PEND,
-				.stctl = SCSW_STCTL_STATUS_PEND,
-			};
+			cdev->private->irb.scsw.cc = 1;
+			cdev->private->irb.scsw.fctl = SCSW_FCTL_START_FUNC;
+			cdev->private->irb.scsw.actl = SCSW_ACTL_START_PEND;
+			cdev->private->irb.scsw.stctl = SCSW_STCTL_STATUS_PEND;
 			cdev->private->flags.fake_irb = 0;
 			if (cdev->handler)
 				cdev->handler(cdev, cdev->private->intparm,
#1 lore raw

Have you checked that in all cases all fields of the struct have been overwritten? For instance, look at this:

drivers/s390/cio/chsc.c new_channel_path(int chpid) · line 1395

Quoted 6 lines · Martin Schwidefsky
drivers/s390/cio/chsc.c +2 -4
-	chp->dev = (struct device) {
-		.parent  = &css[0]->device,
-		.release = chp_release,
-	};
+	chp->dev.parent = &css[0]->device;
+	chp->dev.release = chp_release;

Doesn't this leave chp->dev.bus still holding whatever old value it had laying around before? Unless I'm missing something, it looks to me like this diff causes a change in the semantics of the code.

Perhaps it would be better to memset() the entire struct (chp->dev, in this case) to zero, before assigning to individual fields, so there is no possibility of old remnant data still being left laying around?

#2 lore raw
Quoted 11 lines · David Wagner

Have you checked that in all cases all fields of the struct have been overwritten? For instance, look at this:

drivers/s390/cio/chsc.c new_channel_path(int chpid) · line 1395

Quoted 6 lines · Martin Schwidefsky
drivers/s390/cio/chsc.c +2 -4
-      chp->dev = (struct device) {
-              .parent  = &css[0]->device,
-              .release = chp_release,
-      };
+      chp->dev.parent = &css[0]->device;
+      chp->dev.release = chp_release;
Quoted 7 lines

Doesn't this leave chp->dev.bus still holding whatever old value it had laying around before? Unless I'm missing something, it looks to me like this diff causes a change in the semantics of the code.

Perhaps it would be better to memset() the entire struct (chp->dev, in this case) to zero, before assigning to individual fields, so there is no possibility of old remnant data still being left laying around?

The structure is allocated with kzalloc().

Signature · 2 lines
blue skies,
 Martin
#3 lore raw
In reply to David Wagner
Quoted 11 lines · David Wagner

Have you checked that in all cases all fields of the struct have been overwritten? For instance, look at this:

drivers/s390/cio/chsc.c new_channel_path(int chpid) · line 1395

Quoted 6 lines · Martin Schwidefsky
drivers/s390/cio/chsc.c +2 -4
-      chp->dev = (struct device) {
-              .parent  = &css[0]->device,
-              .release = chp_release,
-      };
+      chp->dev.parent = &css[0]->device;
+      chp->dev.release = chp_release;
Quoted 2 lines

Doesn't this leave chp->dev.bus still holding whatever old value it had laying around before?

You're correct. While this eliminates the possibility of getting random values from the stack, it still leaves space for letting previous unwanted values unchanged.
Unless, of course, the structure in question is kcalloc()'d (which is not the case of gdev in the beginning of the patch - I haven't had the time to check the other cases). But we surely can't rely on that.

Quoted 2 lines

Unless I'm missing something, it looks to
me like this diff causes a change in the semantics of the code.

I can't see the semantic change.

Quoted 4 lines

Perhaps it would be better to memset() the entire struct (chp->dev, in this case) to zero, before assigning to individual fields, so there is no possibility of old remnant data still being left laying around?

Yes. Since the code can't be depend on the caller passing a zeroed structure, it's definitely more safe to memset to 0 (or use kcalloc(), instead of kmalloc(), when the allocation is the routine's own responsability).

Changing subjects a little bit: where's the stack overflow in the code? Either I'm too stupid to find it myself by looking at the patch or the e-mail is mistitled.

Cheers,

Julio Auto

#4 lore raw
Quoted 3 lines · Julio Auto

Unless, of course, the structure in question is kcalloc()'d (which is not the case of gdev in the beginning of the patch - I haven't had the time to check the other cases).

Sorry, actually gdev is kzalloc()'d (which works exactly kcalloc() without the check for n*size integer overflows). I was looking at a much older version of the code.

I still think, however, that memset()'ing to zero is still something to consider, in the cases where the structure is passed to the routine (as opposed to when it's _created by_ the routine). The code shouldn't rely on the awareness of future developers when adding other calls to these functions.

#5 lore raw
In reply to Julio Auto
Quoted 10 lines · Martin Schwidefsky
drivers/s390/cio/chsc.c +2 -4
-      chp->dev = (struct device) {
-              .parent  = &css[0]->device,
-              .release = chp_release,
-      };
+      chp->dev.parent = &css[0]->device;
+      chp->dev.release = chp_release;

On 8/30/06, David Wagner <[email protected]> wrote:
Unless I'm missing something, it looks to
me like this diff causes a change in the semantics of the code.

Quoted 1 line · Julio Auto

I can't see the semantic change.

Maybe "semantic change" isn't the right term. I'm not sure how gcc handles the case where some field (like .bus) is missing from the C99 initializer as the struct device is created -- but it seemed plausible to me that gcc might guarantee that all missing fields are automatically initialized to 0 when the struct is constructed. If this is the case, then the semantic change is that the original code guarantees to initialize chp->dev.bus to 0, whereas the proposed replacement code does not. I don't know whether gcc actually promises to initialize to 0 all unmentioned fields, or whether the C standard requires that, but you could imagine some compiler making that promise (it is a reasonable thing to do).

#6 lore raw
In reply to Martin Schwidefsky

Thanks for pointing out that in most cases there was immediately preceding code that zeroes out the whole struct using kzalloc() or memset(.., 0, ..). Sorry that I overlooked that; my mistake. That takes care of all but one of these. But in the interests of caution, let me ask about the following one:

drivers/s390/cio/device_fsm.c ccw_device_recog_done(struct ccw_device · line 273

Quoted 10 lines · Martin Schwidefsky
drivers/s390/cio/device_fsm.c +4 -6
-		cdev->id = (struct ccw_device_id) {
-			.cu_type   = cdev->private->senseid.cu_type,
-			.cu_model  = cdev->private->senseid.cu_model,
-			.dev_type  = cdev->private->senseid.dev_type,
-			.dev_model = cdev->private->senseid.dev_model,
-		};
+		cdev->id.cu_type   = cdev->private->senseid.cu_type;
+		cdev->id.cu_model  = cdev->private->senseid.cu_model;
+		cdev->id.dev_type  = cdev->private->senseid.dev_type;
+		cdev->id.dev_model = cdev->private->senseid.dev_model;

I don't see any obvious place that zeroes out cdev->id.
In particular, it looks like cdev->id.match_flags and .driver_info are never cleared (i.e., they retain whatever old garbage they had before). More importantly, if anyone ever adds any more fields to struct ccw_device_id, then they will also be retain old garbage values, which is a maintenance pitfall. Is this right, or did I miss something again?

#7 lore raw
Quoted 18 lines · David Wagner

Thanks for pointing out that in most cases there was immediately preceding code that zeroes out the whole struct using kzalloc() or memset(.., 0, ..). Sorry that I overlooked that; my mistake. That takes care of all but one of these. But in the interests of caution, let me ask about the following one:

drivers/s390/cio/device_fsm.c ccw_device_recog_done(struct ccw_device · line 273

Quoted 10 lines · Martin Schwidefsky
drivers/s390/cio/device_fsm.c +4 -6
-		cdev->id = (struct ccw_device_id) {
-			.cu_type   = cdev->private->senseid.cu_type,
-			.cu_model  = cdev->private->senseid.cu_model,
-			.dev_type  = cdev->private->senseid.dev_type,
-			.dev_model = cdev->private->senseid.dev_model,
-		};
+		cdev->id.cu_type   = cdev->private->senseid.cu_type;
+		cdev->id.cu_model  = cdev->private->senseid.cu_model;
+		cdev->id.dev_type  = cdev->private->senseid.dev_type;
+		cdev->id.dev_model = cdev->private->senseid.dev_model;
Quoted 7 lines

I don't see any obvious place that zeroes out cdev->id.
In particular, it looks like cdev->id.match_flags and .driver_info are never cleared (i.e., they retain whatever old garbage they had before). More importantly, if anyone ever adds any more fields to struct ccw_device_id, then they will also be retain old garbage values, which is a maintenance pitfall. Is this right, or did I miss something again?

You're right. Thanks for pointing this out! I will take care of it.

#9 lore raw
In reply to David Wagner

First of all, about your previous e-mail: you're correct, the members not explicitly initialized behave like when a _static_ object is not explicitly initialized (ie., zero'ing it), instead of behaving like an _automatic_ object not being explicitly initialized (which was the kind of behavior I was expecting). This is part of the C99 specification, indeed.

See section 6.7.8, constraints no. 10 and 19, for more info.

Quoted 7 lines · David Wagner

I don't see any obvious place that zeroes out cdev->id.
In particular, it looks like cdev->id.match_flags and .driver_info are never cleared (i.e., they retain whatever old garbage they had before). More importantly, if anyone ever adds any more fields to struct ccw_device_id, then they will also be retain old garbage values, which is a maintenance pitfall. Is this right, or did I miss something again?

Nicely pointed out, I hadn't thought about this possible maintenance issue. Looks like a nice place for a memset() to reside.

Cheers,

Julio Auto

#8 lore raw
In reply to Heiko Carstens
Quoted 19 lines · Heiko Carstens
Quoted 11 lines

drivers/s390/cio/device_fsm.c ccw_device_recog_done(struct ccw_device · line 273

Quoted 10 lines
drivers/s390/cio/device_fsm.c +4 -6
-            cdev->id = (struct ccw_device_id) {
-                    .cu_type   = cdev->private->senseid.cu_type,
-                    .cu_model  = cdev->private->senseid.cu_model,
-                    .dev_type  = cdev->private->senseid.dev_type,
-                    .dev_model = cdev->private->senseid.dev_model,
-            };
+            cdev->id.cu_type   = cdev->private->senseid.cu_type;
+            cdev->id.cu_model  = cdev->private->senseid.cu_model;
+            cdev->id.dev_type  = cdev->private->senseid.dev_type;
+            cdev->id.dev_model = cdev->private->senseid.dev_model;
Quoted 7 lines

I don't see any obvious place that zeroes out cdev->id.
In particular, it looks like cdev->id.match_flags and .driver_info are never cleared (i.e., they retain whatever old garbage they had before). More importantly, if anyone ever adds any more fields to struct ccw_device_id, then they will also be retain old garbage values, which is a maintenance pitfall. Is this right, or did I miss something again?

Quoted 1 line

You're right. Thanks for pointing this out! I will take care of it.

The ccw_device_id structure contains two more fields in addition to the field that are set up in ccw_device_recog_done, namely match_flags and driver_info. driver_info is set later in ccw_bus_match, so that is fine. match_flags of the device is never used, only the match_flags of the drivers version of the ccw_device_id is used. So the code is correct even without the memset. But your point about the maintenance pitfall is valid, we will add a memset after 2.6.18. I don't want to push yet another patch.

Signature · 8 lines
blue skies,
  Martin.

Martin Schwidefsky
Linux for zSeries Development & Services
IBM Deutschland Entwicklung GmbH

"Reality continues to ruin my life." - Calvin.