[PATCH] zero-size E820 memory (kernel start-up failure)

[PATCH] zero-size E820 memory (kernel start-up failure)

[Alexander Achenbach]

[-- Attachment #1: Type: text/plain, Size: 1087 bytes --]

Hi everybody.

This patch fixes the

     Ok, booting the kernel.         [hangs]

problem with BIOSes that report E820 memory maps with zero-size entries.
Among the mainboards affected are

     EPOX 4G4A+ (presumably)
     EPOX 4BEA[-R]
     Chaintech 9EJL  .

In the past (kernels up to 2.4.20), these mainboards could only boot Linux
with explicit 'mem=' boot parameters.

Up to 2.4.20, a zero-size E820 memory region (eg 0xa0000 - 0xa0000)
reported by the BIOS causes 'sanitize_e820_map' to take the end of this
region as the start of another region, producing two overlapping regions
extending to the end of addressable memory. If the zero-size region is of
a higher type than 'usable RAM' (eg 'reserved'), this causes all memory
above the zero-size region to be marked unusable (leaving 640k as of the
above example -- the kernel fails to start at all).

The patch adds code to remove empty entries before sorting regions. It
was generated for a 2.4.20 kernel.

Best regards,
Alex

[ Please send answers to my Reply-To address.
   I'm not subscribed to the kernel mailing list. ]

[-- Attachment #2: nullmem.diff --]
[-- Type: text/plain, Size: 2365 bytes --]

diff -ur linux-2.4.20/arch/i386/kernel/setup.c linux-2.4.20-nullmem/arch/i386/kernel/setup.c
--- linux-2.4.20/arch/i386/kernel/setup.c	Fri Nov 29 00:53:09 2002
+++ linux-2.4.20-nullmem/arch/i386/kernel/setup.c	Mon Dec 16 14:52:21 2002
@@ -71,6 +71,9 @@
  *  CacheSize bug workaround updates for AMD, Intel & VIA Cyrix.
  *  Dave Jones <davej@suse.de>, September, October 2001.
  *
+ *  Provisions for empty E820 memory regions (reported by certain BIOSes).
+ *  Alex Achenbach <xela@slit.de>, December 2002.
+ *
  */
 
 /*
@@ -501,7 +504,7 @@
 	int chgidx, still_changing;
 	int overlap_entries;
 	int new_bios_entry;
-	int old_nr, new_nr;
+	int old_nr, new_nr, chg_nr;
 	int i;
 
 	/*
@@ -555,20 +558,24 @@
 	for (i=0; i < 2*old_nr; i++)
 		change_point[i] = &change_point_list[i];
 
-	/* record all known change-points (starting and ending addresses) */
+	/* record all known change-points (starting and ending addresses),
+	   omitting those that are for empty memory regions */
 	chgidx = 0;
 	for (i=0; i < old_nr; i++)	{
-		change_point[chgidx]->addr = biosmap[i].addr;
-		change_point[chgidx++]->pbios = &biosmap[i];
-		change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
-		change_point[chgidx++]->pbios = &biosmap[i];
+		if (biosmap[i].size != 0) {
+			change_point[chgidx]->addr = biosmap[i].addr;
+			change_point[chgidx++]->pbios = &biosmap[i];
+			change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
+			change_point[chgidx++]->pbios = &biosmap[i];
+		}
 	}
+	chg_nr = chgidx;    	/* true number of change-points */
 
 	/* sort change-point list by memory addresses (low -> high) */
 	still_changing = 1;
 	while (still_changing)	{
 		still_changing = 0;
-		for (i=1; i < 2*old_nr; i++)  {
+		for (i=1; i < chg_nr; i++)  {
 			/* if <current_addr> > <last_addr>, swap */
 			/* or, if current=<start_addr> & last=<end_addr>, swap */
 			if ((change_point[i]->addr < change_point[i-1]->addr) ||
@@ -591,7 +598,7 @@
 	last_type = 0;		 /* start with undefined memory type */
 	last_addr = 0;		 /* start with 0 as last starting address */
 	/* loop through change-points, determining affect on the new bios map */
-	for (chgidx=0; chgidx < 2*old_nr; chgidx++)
+	for (chgidx=0; chgidx < chg_nr; chgidx++)
 	{
 		/* keep track of all overlapping bios entries */
 		if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)

Re: [PATCH] zero-size E820 memory (kernel start-up failure)

[Dave Jones]

On Fri, Dec 20, 2002 at 12:02:57PM +0100, Alexander Achenbach wrote:

 > Up to 2.4.20, a zero-size E820 memory region (eg 0xa0000 - 0xa0000)
 > reported by the BIOS causes 'sanitize_e820_map' to take the end of this
 > region as the start of another region, producing two overlapping regions
 > extending to the end of addressable memory. If the zero-size region is of
 > a higher type than 'usable RAM' (eg 'reserved'), this causes all memory
 > above the zero-size region to be marked unusable (leaving 640k as of the
 > above example -- the kernel fails to start at all).
 > 
 > The patch adds code to remove empty entries before sorting regions. It
 > was generated for a 2.4.20 kernel.

Looks good to me. Care to do a similar patch for 2.5 ?

		Dave

-- 
| Dave Jones.        http://www.codemonkey.org.uk
| SuSE Labs

Re: [PATCH] zero-size E820 memory (kernel start-up failure)

[Alexander Achenbach]

[-- Attachment #1: Type: text/plain, Size: 336 bytes --]

Hi again.

Dave Jones wrote:
 >
 >  > The patch adds code to remove empty entries before sorting regions. It
 >  > was generated for a 2.4.20 kernel.
 >
 > Looks good to me. Care to do a similar patch for 2.5 ?

This is it, to be applied against 2.5.52.
I've tested the kernel on an EPOX 4BEA-R and it boots as it should.

Cheers,
Alex

[-- Attachment #2: nullmem-2.5.52.diff --]
[-- Type: text/plain, Size: 2338 bytes --]

diff -ru linux-2.5.52/arch/i386/kernel/setup.c linux-2.5.52-nullmem/arch/i386/kernel/setup.c
--- linux-2.5.52/arch/i386/kernel/setup.c	Mon Dec 16 03:07:56 2002
+++ linux-2.5.52-nullmem/arch/i386/kernel/setup.c	Fri Dec 20 15:48:35 2002
@@ -14,6 +14,9 @@
  * Moved CPU detection code to cpu/${cpu}.c
  *    Patrick Mochel <mochel@osdl.org>, March 2002
  *
+ *  Provisions for empty E820 memory regions (reported by certain BIOSes).
+ *  Alex Achenbach <xela@slit.de>, December 2002.
+ *
  */
 
 /*
@@ -275,7 +278,7 @@
 	int chgidx, still_changing;
 	int overlap_entries;
 	int new_bios_entry;
-	int old_nr, new_nr;
+	int old_nr, new_nr, chg_nr;
 	int i;
 
 	/*
@@ -329,20 +332,24 @@
 	for (i=0; i < 2*old_nr; i++)
 		change_point[i] = &change_point_list[i];
 
-	/* record all known change-points (starting and ending addresses) */
+	/* record all known change-points (starting and ending addresses),
+	   omitting those that are for empty memory regions */
 	chgidx = 0;
 	for (i=0; i < old_nr; i++)	{
-		change_point[chgidx]->addr = biosmap[i].addr;
-		change_point[chgidx++]->pbios = &biosmap[i];
-		change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
-		change_point[chgidx++]->pbios = &biosmap[i];
+		if (biosmap[i].size != 0) {
+			change_point[chgidx]->addr = biosmap[i].addr;
+			change_point[chgidx++]->pbios = &biosmap[i];
+			change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
+			change_point[chgidx++]->pbios = &biosmap[i];
+		}
 	}
+	chg_nr = chgidx;    	/* true number of change-points */
 
 	/* sort change-point list by memory addresses (low -> high) */
 	still_changing = 1;
 	while (still_changing)	{
 		still_changing = 0;
-		for (i=1; i < 2*old_nr; i++)  {
+		for (i=1; i < chg_nr; i++)  {
 			/* if <current_addr> > <last_addr>, swap */
 			/* or, if current=<start_addr> & last=<end_addr>, swap */
 			if ((change_point[i]->addr < change_point[i-1]->addr) ||
@@ -365,7 +372,7 @@
 	last_type = 0;		 /* start with undefined memory type */
 	last_addr = 0;		 /* start with 0 as last starting address */
 	/* loop through change-points, determining affect on the new bios map */
-	for (chgidx=0; chgidx < 2*old_nr; chgidx++)
+	for (chgidx=0; chgidx < chg_nr; chgidx++)
 	{
 		/* keep track of all overlapping bios entries */
 		if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)