Skip to contentSkip to conversation

UPD: "do_IRQ: near stack overflow" when inserting CF disk

DISCUSSION Roger Larsson 3 participants 3 messages Dec 12, 2003 Pcmcia

lore
Share & tools
lore
#0 lore raw

OK, I have done some more testing....

How to avoid getting HTML messages from Outlook?
I have specified text - but I guess it automatically selects HTML when adding attachments...

please cc me, currently not subscribed.

/RogerL

---------- Forwarded Message ----------

Subject: "do_IRQ: near stack overflow" when inserting CF disk
Date: Friday 12 December 2003 11.17
From: Larsson Roger <[email protected]>
To: 'Kernel Mailing List' <[email protected]>, 'David Hinds' <[email protected]>
Cc: "'[email protected]'" <[email protected]>

I get these three printouts when inserting a compact flash disk (SuSE Linux 2.4.21-144-smp) - is the stack intentionally this fit? or is it an unexpected code path?

I have had other interrupt related problems, it usually locks up :-( This time I run with "nosmp noapic"

The ide-cs is now forced to use irq 3 (When running noapic others seems busy... with apic and using irq 6 or 10 computer has locked up...)

I will try irq 3 without one of the options to see if the computer still survives...

- new data -

Easier said than done - they require each other! But I noticed that reducing 'stackwarn' and 'stackdefer' to under the level warned about helps (might be luck)

Summary:
Some routines allocate to much stack: validate_cis?

	several subtypes in the union cisparse_t are quite
	big (>250 bytes)

This eats stack (to much?) or only close enough to trigger 'stackdefer' which might stress the drivers SMPness?

/RogerL

Signature · 3 lines
Roger Larsson
Skellefteå
Sweden
#1 lore raw
Quoted 7 lines · Roger Larsson

Summary:
Some routines allocate to much stack: validate_cis?

	several subtypes in the union cisparse_t are quite
	big (>250 bytes)

This eats stack (to much?) or only close enough to trigger 'stackdefer' which might stress the drivers SMPness?

Marcelo,

Please apply this patch. It reduces kernel stack usage for the PCMCIA CIS processing code.

-- Dave

drivers/pcmcia/cistpl.c renamed from linux-2.4.23/drivers/pcmcia/cistpl.c +58 -32
@@ -351,9 +351,12 @@
 
 int verify_cis_cache(socket_info_t *s)
 {
-    char buf[256], *caddr;
+    char *buf, *caddr;
     int i;
-    
+
+    buf = kmalloc(256, GFP_KERNEL);
+    if (buf == NULL)
+	return -1;
     caddr = s->cis_cache;
     for (i = 0; i < s->cis_used; i++) {
 #ifdef CONFIG_CARDBUS
@@ -368,6 +371,7 @@
 	    break;
 	caddr += s->cis_table[i].len;
     }
+    kfree(buf);
     return (i < s->cis_used);
 }
 
@@ -1410,20 +1414,31 @@
 
 int read_tuple(client_handle_t handle, cisdata_t code, void *parse)
 {
-    tuple_t tuple;
-    cisdata_t buf[255];
+    tuple_t *tuple;
+    cisdata_t *buf;
     int ret;
-    
-    tuple.DesiredTuple = code;
-    tuple.Attributes = TUPLE_RETURN_COMMON;
-    ret = pcmcia_get_first_tuple(handle, &tuple);
-    if (ret != CS_SUCCESS) return ret;
-    tuple.TupleData = buf;
-    tuple.TupleOffset = 0;
-    tuple.TupleDataMax = sizeof(buf);
-    ret = pcmcia_get_tuple_data(handle, &tuple);
-    if (ret != CS_SUCCESS) return ret;
-    ret = pcmcia_parse_tuple(handle, &tuple, parse);
+
+    buf = kmalloc(256, GFP_KERNEL);
+    if (buf == NULL)
+	return CS_OUT_OF_RESOURCE;
+    tuple = kmalloc(sizeof(*tuple), GFP_KERNEL);
+    if (tuple == NULL) {
+	kfree(buf);
+	return CS_OUT_OF_RESOURCE;
+    }
+    tuple->DesiredTuple = code;
+    tuple->Attributes = TUPLE_RETURN_COMMON;
+    ret = pcmcia_get_first_tuple(handle, tuple);
+    if (ret != CS_SUCCESS) goto done;
+    tuple->TupleData = buf;
+    tuple->TupleOffset = 0;
+    tuple->TupleDataMax = 255;
+    ret = pcmcia_get_tuple_data(handle, tuple);
+    if (ret != CS_SUCCESS) goto done;
+    ret = pcmcia_parse_tuple(handle, tuple, parse);
+done:
+    kfree(tuple);
+    kfree(buf);
     return ret;
 }
 
@@ -1439,50 +1454,61 @@
 
 int pcmcia_validate_cis(client_handle_t handle, cisinfo_t *info)
 {
-    tuple_t tuple;
-    cisparse_t p;
+    tuple_t *tuple;
+    cisparse_t *p;
     int ret, reserved, dev_ok = 0, ident_ok = 0;
 
     if (CHECK_HANDLE(handle))
 	return CS_BAD_HANDLE;
+    tuple = kmalloc(sizeof(*tuple), GFP_KERNEL);
+    if (tuple == NULL)
+	return CS_OUT_OF_RESOURCE;
+    p = kmalloc(sizeof(*p), GFP_KERNEL);
+    if (p == NULL) {
+	kfree(tuple);
+	return CS_OUT_OF_RESOURCE;
+    }
 
     info->Chains = reserved = 0;
-    tuple.DesiredTuple = RETURN_FIRST_TUPLE;
-    tuple.Attributes = TUPLE_RETURN_COMMON;
-    ret = pcmcia_get_first_tuple(handle, &tuple);
+    tuple->DesiredTuple = RETURN_FIRST_TUPLE;
+    tuple->Attributes = TUPLE_RETURN_COMMON;
+    ret = pcmcia_get_first_tuple(handle, tuple);
     if (ret != CS_SUCCESS)
-	return CS_SUCCESS;
+	goto done;
 
     /* First tuple should be DEVICE; we should really have either that
        or a CFTABLE_ENTRY of some sort */
-    if ((tuple.TupleCode == CISTPL_DEVICE) ||
-	(read_tuple(handle, CISTPL_CFTABLE_ENTRY, &p) == CS_SUCCESS) ||
-	(read_tuple(handle, CISTPL_CFTABLE_ENTRY_CB, &p) == CS_SUCCESS))
+    if ((tuple->TupleCode == CISTPL_DEVICE) ||
+	(read_tuple(handle, CISTPL_CFTABLE_ENTRY, p) == CS_SUCCESS) ||
+	(read_tuple(handle, CISTPL_CFTABLE_ENTRY_CB, p) == CS_SUCCESS))
 	dev_ok++;
 
     /* All cards should have a MANFID tuple, and/or a VERS_1 or VERS_2
        tuple, for card identification.  Certain old D-Link and Linksys
        cards have only a broken VERS_2 tuple; hence the bogus test. */
-    if ((read_tuple(handle, CISTPL_MANFID, &p) == CS_SUCCESS) ||
-	(read_tuple(handle, CISTPL_VERS_1, &p) == CS_SUCCESS) ||
-	(read_tuple(handle, CISTPL_VERS_2, &p) != CS_NO_MORE_ITEMS))
+    if ((read_tuple(handle, CISTPL_MANFID, p) == CS_SUCCESS) ||
+	(read_tuple(handle, CISTPL_VERS_1, p) == CS_SUCCESS) ||
+	(read_tuple(handle, CISTPL_VERS_2, p) != CS_NO_MORE_ITEMS))
 	ident_ok++;
 
     if (!dev_ok && !ident_ok)
-	return CS_SUCCESS;
+	goto done;
 
     for (info->Chains = 1; info->Chains < MAX_TUPLES; info->Chains++) {
-	ret = pcmcia_get_next_tuple(handle, &tuple);
+	ret = pcmcia_get_next_tuple(handle, tuple);
 	if (ret != CS_SUCCESS) break;
-	if (((tuple.TupleCode > 0x23) && (tuple.TupleCode < 0x40)) ||
-	    ((tuple.TupleCode > 0x47) && (tuple.TupleCode < 0x80)) ||
-	    ((tuple.TupleCode > 0x90) && (tuple.TupleCode < 0xff)))
+	if (((tuple->TupleCode > 0x23) && (tuple->TupleCode < 0x40)) ||
+	    ((tuple->TupleCode > 0x47) && (tuple->TupleCode < 0x80)) ||
+	    ((tuple->TupleCode > 0x90) && (tuple->TupleCode < 0xff)))
 	    reserved++;
     }
     if ((info->Chains == MAX_TUPLES) || (reserved > 5) ||
 	((!dev_ok || !ident_ok) && (info->Chains > 10)))
 	info->Chains = 0;
 
+done:
+    kfree(tuple);
+    kfree(p);
     return CS_SUCCESS;
 }
 
#2 lore raw
In reply to Roger Larsson
Quoted 21 lines · Roger Larsson

I get these three printouts when inserting a compact flash disk (SuSE Linux 2.4.21-144-smp) - is the stack intentionally this fit? or is it an unexpected code path?

I have had other interrupt related problems, it usually locks up :-( This time I run with "nosmp noapic"

The ide-cs is now forced to use irq 3 (When running noapic others seems busy... with apic and using irq 6 or 10 computer has locked up...)

I will try irq 3 without one of the options to see if the computer still survives...

- new data -

Easier said than done - they require each other! But I noticed that reducing 'stackwarn' and 'stackdefer' to under the level warned about helps (might be luck)

Quoted 6 lines

Summary:
Some routines allocate to much stack: validate_cis?

	several subtypes in the union cisparse_t are quite
	big (>250 bytes)

This eats stack (to much?) or only close enough to trigger 'stackdefer' which might stress the drivers SMPness?

Can you send me your .config? I'd like to run checkstack on it and see what other functions in your trace consume too much stack. David's patch shoundn't shave off more than 2k, so either my estimate is wrong or there are more hungry wolves in your list.

Jörn

Signature · 3 lines
Beware of bugs in the above code; I have only proved it correct, but
not tried it.
-- Donald Knuth