Skip to contentSkip to conversation

8xx: prevent a machine check in scc_init().

PATCH New Gary Jennejohn 3 participants 7 messages Sep 14, 2008 – Sep 26, 2008

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 8xx: prevent a machine check in scc_init().

---
 cpu/mpc8xx/scc.c |   17 +++++++++++++++++
 1 files changed, 17 insertions(+), 0 deletions(-)
cpu/mpc8xx/scc.c +17 -0
index 09a3db1..9ffeb11 100644
@@ -70,6 +70,9 @@ static int scc_recv(struct eth_device* dev);
 static int scc_init (struct eth_device* dev, bd_t * bd);
 static void scc_halt(struct eth_device* dev);
 
+/* avoid unnecessary reinitialization of the SCC */
+static int scc_init_completed;
+
 int scc_initialize(bd_t *bis)
 {
 	struct eth_device* dev;
@@ -193,6 +196,19 @@ static int scc_init (struct eth_device *dev, bd_t * bis)
 
 	volatile immap_t *immr = (immap_t *) CFG_IMMR;
 
+	/*
+	 * This routine is called again and again from eth_init(),
+	 * especially when CONFIG_NETCONSOLE is defined and
+	 * stdout=nc.
+	 * Avoid unneccessary flailing, otherwise we can get a panic here.
+	 */
+	if (scc_init_completed) {
+		immr->im_ioport.iop_pcso |= (PC_ENET_CLSN | PC_ENET_RENA);
+		immr->im_cpm.cp_scc[SCC_ENET].scc_gsmrl |=
+			(SCC_GSMRL_ENR | SCC_GSMRL_ENT);
+		return 1;
+	}
+
 #if defined(CONFIG_LWMON)
 	reset_phy();
 #endif
@@ -544,6 +560,7 @@ static int scc_init (struct eth_device *dev, bd_t * bis)
 	udelay (100000);	/* wait 100 ms */
 #endif
 
+	scc_init_completed = 1;
 	return 1;
 }
 
Signature · 9 lines
1.5.4.3

---
Gary Jennejohn
*********************************************************************
DENX Software Engineering GmbH,     MD: Wolfgang Denk & Detlev Zundel
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: +49-8142-66989-0 Fax: +49-8142-66989-80  Email: office at denx.de
*********************************************************************
#1 lore raw

PATCH 1/1 8xx: prevent a machine check in scc_init().

Dear Gary Jennejohn,

Quoted 2 lines · In message <[email protected]> you wrote:

Sorry, but I don't understand what you're doing here, or why.

Why would there be any machine checks in scc_init()? Such problems have never been repoorted for any systems. Actuallym this part of the code is one of the oldest in U-Boot - if there were machine check issues in any configruation, these should be known.

Quoted 16 lines
cpu/mpc8xx/scc.c +3 -0
@@ -70,6 +70,9 @@ static int scc_recv(struct eth_device* dev);
 static int scc_init (struct eth_device* dev, bd_t * bd);
 static void scc_halt(struct eth_device* dev);
 
+/* avoid unnecessary reinitialization of the SCC */
+static int scc_init_completed;
+
 int scc_initialize(bd_t *bis)
 {
 	struct eth_device* dev;
@@ -193,6 +196,19 @@ static int scc_init (struct eth_device *dev, bd_t * bis)
 
 	volatile immap_t *immr = (immap_t *) CFG_IMMR;
 

cpu/mpc8xx/scc.c static int scc_init (struct eth_device *dev, bd_t * bis) · line 208

Quoted 10 lines
cpu/mpc8xx/scc.c +10 -0
+	/*
+	 * This routine is called again and again from eth_init(),
+	 * especially when CONFIG_NETCONSOLE is defined and
+	 * stdout=nc.
+	 * Avoid unneccessary flailing, otherwise we can get a panic here.
+	 */
+	if (scc_init_completed) {
+		immr->im_ioport.iop_pcso |= (PC_ENET_CLSN | PC_ENET_RENA);
+		immr->im_cpm.cp_scc[SCC_ENET].scc_gsmrl |=
+			(SCC_GSMRL_ENR | SCC_GSMRL_ENT);

If the initialization has already been done, why would we repeat these two steps here?

And why exactle these two, and not the others that are performed for a regular init sequence?

cpu/mpc8xx/scc.c static int scc_init (struct eth_device *dev, bd_t * bis) · line 214

Quoted 6 lines
cpu/mpc8xx/scc.c +3 -0
+		return 1;
+	}
+
 #if defined(CONFIG_LWMON)
 	reset_phy();
 #endif

I'm not sure if your early return here is actually valid. As you can see, other actions follow - for example here on the LWMON baord the PHY reset; I suspect that your change might eventually work on your current system, but break many others.

Also, I'm not sure if you are4 aware that the eth_init() part may be called from a interface switching sequence, i. e. when cycling between several ethenret interfaces. These will have been shut down before by eth_halt() - skipping the init sequence here seems the wrong thing to me.

Maybe you could explain what problem you are actually truing to fix, and especially under which situations you observe machine checks.

Best regards,

Wolfgang Denk

Signature · 6 lines
DENX Software Engineering GmbH,     MD: Wolfgang Denk & Detlev Zundel
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: wd at denx.de
The reasonable man adapts himself to the world; the unreasonable  one
persists  in  trying  to  adapt  the  world to himself. Therefore all
progress depends on the unreasonable man."      - George Bernard Shaw
#2 lore raw

PATCH 1/1 8xx: prevent a machine check in scc_init().

Quoted 65 lines · On Sun, 14 Sep 2008 19:07:52 +0200 Wolfgang Denk <[email protected]> wrote:
Quoted 2 lines · In message <[email protected]> you wrote:

Sorry, but I don't understand what you're doing here, or why.

Why would there be any machine checks in scc_init()? Such problems have never been repoorted for any systems. Actuallym this part of the code is one of the oldest in U-Boot - if there were machine check issues in any configruation, these should be known.

Quoted 16 lines
cpu/mpc8xx/scc.c +3 -0
@@ -70,6 +70,9 @@ static int scc_recv(struct eth_device* dev);
 static int scc_init (struct eth_device* dev, bd_t * bd);
 static void scc_halt(struct eth_device* dev);
 
+/* avoid unnecessary reinitialization of the SCC */
+static int scc_init_completed;
+
 int scc_initialize(bd_t *bis)
 {
 	struct eth_device* dev;
@@ -193,6 +196,19 @@ static int scc_init (struct eth_device *dev, bd_t * bis)
 
 	volatile immap_t *immr = (immap_t *) CFG_IMMR;
 

cpu/mpc8xx/scc.c static int scc_init (struct eth_device *dev, bd_t * bis) · line 208

Quoted 10 lines
cpu/mpc8xx/scc.c +10 -0
+	/*
+	 * This routine is called again and again from eth_init(),
+	 * especially when CONFIG_NETCONSOLE is defined and
+	 * stdout=nc.
+	 * Avoid unneccessary flailing, otherwise we can get a panic here.
+	 */
+	if (scc_init_completed) {
+		immr->im_ioport.iop_pcso |= (PC_ENET_CLSN | PC_ENET_RENA);
+		immr->im_cpm.cp_scc[SCC_ENET].scc_gsmrl |=
+			(SCC_GSMRL_ENR | SCC_GSMRL_ENT);

If the initialization has already been done, why would we repeat these two steps here?

And why exactle these two, and not the others that are performed for a regular init sequence?

cpu/mpc8xx/scc.c static int scc_init (struct eth_device *dev, bd_t * bis) · line 214

Quoted 6 lines
cpu/mpc8xx/scc.c +3 -0
+		return 1;
+	}
+
 #if defined(CONFIG_LWMON)
 	reset_phy();
 #endif

I'm not sure if your early return here is actually valid. As you can see, other actions follow - for example here on the LWMON baord the PHY reset; I suspect that your change might eventually work on your current system, but break many others.

Also, I'm not sure if you are4 aware that the eth_init() part may be called from a interface switching sequence, i. e. when cycling between several ethenret interfaces. These will have been shut down before by eth_halt() - skipping the init sequence here seems the wrong thing to me.

Quoted 3 lines

Maybe you could explain what problem you are actually truing to fix, and especially under which situations you observe machine checks.

The first call to scc_init() executes all the code. Only subsequent calls are short-circuited.

OK, just think about it a little.

When nc is active as a console it is often invoked for every byte.

The network stack code calls eth_init() prior to sending the byte and eth_halt() after sending the byte.

scc_init() happily tries to reallocate dpram at every call when CFG_ALLOC_DPRAM is defined, as it was in my case. Eventually dpram is exhausted which evidently results in a machine check.

Also note that scc_halt() merely clears some bits in scc_gsmrl and iop_pcso. It really doesn't seem necessary to reinitialize every register at every call to scc_init().

Most network drivers behave this way, which works in the general use case, but not when nc is used.

I doubt that nc is used very often, otherwise this problem would have reared its ugly head a lot sooner. The general use case is also to merely invoke tftp to load the kernel and maybe fdt. This succeeds because it takes numerous (in this case about 14) calls to the init routine before the machine check happens.

---
Gary Jennejohn
*********************************************************************
DENX Software Engineering GmbH, MD: Wolfgang Denk & Detlev Zundel
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: +49-8142-66989-0 Fax: +49-8142-66989-80 Email: office at denx.de
*********************************************************************

#3 lore raw

PATCH 1/1 8xx: prevent a machine check in scc_init().

Dear Gary Jennejohn,

Quoted 3 lines · In message <[email protected]> you wrote:

The first call to scc_init() executes all the code. Only subsequent calls are short-circuited.

I am aware of that. But after the first call, eth_halt() might have been run, deactivating the interface; then for example interfaces might have been switched to another port, and finally switched back to the old port. In this case, a full initialization sequence will be necessary.

Quoted 1 line

OK, just think about it a little.

Indeed, please do.

Quoted 1 line

When nc is active as a console it is often invoked for every byte.

Ah! That's the origin of your concern. But I think you're trying to attack it at the wrong place. Your modification will most probably break some (or many) boards.

Quoted 6 lines

The network stack code calls eth_init() prior to sending the byte and eth_halt() after sending the byte.

scc_init() happily tries to reallocate dpram at every call when CFG_ALLOC_DPRAM is defined, as it was in my case. Eventually dpram is exhausted which evidently results in a machine check.

-> grep -l CONFIG_NETCONSOLE include/configs/* | xargs egrep CFG_ALLOC_DPRAM | wc -l 0
->

You are right. None of the boards that use netconsole so far (and that have been tested with that feature, obviously) is using CFG_ALLOC_DPRAM. So a fix is necessary not to re-allocate DPRAM here.

It would be great if you could submit a patch that fixes this problem.

Quoted 5 lines

I doubt that nc is used very often, otherwise this problem would have reared its ugly head a lot sooner. The general use case is also to merely invoke tftp to load the kernel and maybe fdt. This succeeds because it takes numerous (in this case about 14) calls to the init routine before the machine check happens.

In fact I think the problem never showed before because nodody used CFG_ALLOC_DPRAM with netconsole. I agree that this should be fixed.

Best regards,

Wolfgang Denk

Signature · 4 lines
DENX Software Engineering GmbH,     MD: Wolfgang Denk & Detlev Zundel
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: wd at denx.de
                  Nail here --X-- for new monitor.
#4 lore raw

PATCH v2 1/1 8xx: prevent a machine check in scc_init().

In reply to Gary Jennejohn

Without this change DPRAM can be exhausted when CFG_ALLOC_DPRAM is defined, which eventually leads to a machine check. This change assures that DPRAM is allocated only once in that case.

---
 cpu/mpc8xx/scc.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)
cpu/mpc8xx/scc.c +3 -1
index 09a3db1..d91289a 100644
@@ -216,7 +216,9 @@ static int scc_init (struct eth_device *dev, bd_t * bis)
 	txIdx = 0;
 
 #ifdef CFG_ALLOC_DPRAM
-	rtx = (RTXBD *) (immr->im_cpm.cp_dpmem +
+	/* Avoid exhausting DPRAM. */
+	if (rtx == NULL)
+		rtx = (RTXBD *) (immr->im_cpm.cp_dpmem +
 			 dpram_alloc_align (sizeof (RTXBD), 8));
 #else
 	rtx = (RTXBD *) (immr->im_cpm.cp_dpmem + CPM_SCC_BASE);
Signature · 9 lines
1.5.4.3

---
Gary Jennejohn
*********************************************************************
DENX Software Engineering GmbH,     MD: Wolfgang Denk & Detlev Zundel
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: +49-8142-66989-0 Fax: +49-8142-66989-80  Email: office at denx.de
*********************************************************************
#5 lore raw

PATCH v2 1/1 8xx: prevent a machine check in scc_init().

Dear Gary Jennejohn,

Quoted 9 lines · In message <[email protected]> you wrote:

Without this change DPRAM can be exhausted when CFG_ALLOC_DPRAM is defined, which eventually leads to a machine check. This change assures that DPRAM is allocated only once in that case.

---
 cpu/mpc8xx/scc.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)

Applied, with minor re-editing to make the code look similar to what we have in cpu/mpc8xx/fec.c

Thanks.

Bein, I hope it is OK with you that I pulled this directly.

Best regards,

Wolfgang Denk

Signature · 5 lines
DENX Software Engineering GmbH,     MD: Wolfgang Denk & Detlev Zundel
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: wd at denx.de
It would be illogical to kill without reason
	-- Spock, "Journey to Babel", stardate 3842.4
#6 lore raw

PATCH v2 1/1 8xx: prevent a machine check in scc_init().

Quoted 14 lines · Wolfgang Denk

Dear Gary Jennejohn,

Quoted 9 lines · In message <[email protected]> you wrote:

Without this change DPRAM can be exhausted when CFG_ALLOC_DPRAM is defined, which eventually leads to a machine check. This change assures that DPRAM is allocated only once in that case.

---
 cpu/mpc8xx/scc.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)
Quoted 9 lines

Applied, with minor re-editing to make the code look similar to what we have in cpu/mpc8xx/fec.c

Thanks.

Bein, I hope it is OK with you that I pulled this directly.

Totally fine. I've been off-grid for a while and am finally starting to catch up. Thanks for taking care of this.

Quoted 5 lines

Best regards,

Wolfgang Denk