Skip to contentSkip to conversation

eisa: fix eisa with PCI

PATCH In review Yinghai Lu 2 participants 18 messages Mar 28, 2013 – Apr 3, 2013 5 messages not in the archive PCI

lore
Share & tools
lore

Yinghai Lu

author
#0 lore raw

cover eisa: fix eisa with PCI

Looks like pci eisa bridge support is broken for a while.

one is root io resource reference, and other one is pnp related.

Please check if we can put them in v3.9.
eisa, PCI: Fix bus res reference eisa, PCI: init eisa early before pnp step in

Thanks

Yinghai Lu

author
#1 lore raw

PATCH 1/2 eisa, PCI: Fix bus res reference

Matthem found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

After looking at pci_eisa_init(), found it referring bus resource directly instead of pci_bus_resource_n().

After commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for pci root bus any more.

Fix it by using pci_bus_resource_n() and correct idx for root bus.

---
 drivers/eisa/pci_eisa.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
drivers/eisa/pci_eisa.c +7 -3
Index: linux-2.6/drivers/eisa/pci_eisa.c
===================================================================
@@ -22,7 +22,8 @@ static struct eisa_root_device pci_eisa_
 static int __init pci_eisa_init(struct pci_dev *pdev,
 				const struct pci_device_id *ent)
 {
-	int rc;
+	int rc, n = 0;
+	struct resource *bus_res;
 
 	if ((rc = pci_enable_device (pdev))) {
 		printk (KERN_ERR "pci_eisa : Could not enable device %s\n",
@@ -30,9 +31,12 @@ static int __init pci_eisa_init(struct p
 		return rc;
 	}
 
+	if (pci_is_root_bus(pdev->bus))
+		n = PCI_BRIDGE_RESOURCE_NUM;
+	bus_res = pci_bus_resource_n(pdev->bus, n);
 	pci_eisa_root.dev              = &pdev->dev;
-	pci_eisa_root.res	       = pdev->bus->resource[0];
-	pci_eisa_root.bus_base_addr    = pdev->bus->resource[0]->start;
+	pci_eisa_root.res	       = bus_res;
+	pci_eisa_root.bus_base_addr    = bus_res->start;
 	pci_eisa_root.slots	       = EISA_MAX_SLOTS;
 	pci_eisa_root.dma_mask         = pdev->dma_mask;
 	dev_set_drvdata(pci_eisa_root.dev, &pci_eisa_root);
maintainer
#2 lore raw

Re: [PATCH 1/2] eisa, PCI: Fix bus res reference

Quoted 5 lines · Yinghai Lu

Matthem found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

Quoted 8 lines

After looking at pci_eisa_init(), found it referring bus resource directly instead of pci_bus_resource_n().

After commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for pci root bus any more.

Fix it by using pci_bus_resource_n() and correct idx for root bus.

Please include URLs for the problem reports for these problems (bugzilla or mailing list discussion).

Quoted 4 lines

You are consistently using the wrong stable email address. It should be "[email protected]". I assume you see a bounce every time; at least, I get a bounce when I reply to your messages.

author
#3 lore raw

Re: [PATCH 1/2] eisa, PCI: Fix bus res reference

Quoted 9 lines · Bjorn Helgaas
Quoted 8 lines · Yinghai Lu

Matthem found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

After looking at pci_eisa_init(), found it referring bus resource directly instead of pci_bus_resource_n().

Quoted 8 lines
Quoted 5 lines

After commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for pci root bus any more.

Fix it by using pci_bus_resource_n() and correct idx for root bus.

Please include URLs for the problem reports for these problems (bugzilla or mailing list discussion).

Matthew sent private mail to me, and I sent him test patch to see if it fixes the problem. then I posted patches here after that.

Quoted 9 lines
Quoted 5 lines

You are consistently using the wrong stable email address. It should be "[email protected]". I assume you see a bounce every time; at least, I get a bounce when I reply to your messages.

I thought that is tag only, and stable maintainer will search that from Linus's tree.
but git send-mail will pick up Cc from the patch.

So [email protected] will never get fixed?

Yinghai

maintainer
#4 lore raw

Re: [PATCH 1/2] eisa, PCI: Fix bus res reference

Quoted 14 lines · Yinghai Lu
Quoted 13 lines · Bjorn Helgaas
Quoted 12 lines · Yinghai Lu

Matthem found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

After looking at pci_eisa_init(), found it referring bus resource directly instead of pci_bus_resource_n().

After commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for pci root bus any more.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Fix it by using pci_bus_resource_n() and correct idx for root bus.

Please include URLs for the problem reports for these problems (bugzilla or mailing list discussion).

Matthew sent private mail to me, and I sent him test patch to see if it fixes the problem. then I posted patches here after that.

Please add "cc: [email protected]" when you respond to private mail like that. As far as I'm concerned, if it doesn't appear in a public email archive, it didn't happen. I hate dealing with patches out of the blue like this that don't have any context. If it's on a mailing list, other people with the same problem have some hope of finding this solution.

Quoted 9 lines
Quoted 8 lines
Quoted 4 lines

You are consistently using the wrong stable email address. It should be "[email protected]". I assume you see a bounce every time; at least, I get a bounce when I reply to your messages.

Quoted 5 lines

I thought that is tag only, and stable maintainer will search that from Linus's tree.
but git send-mail will pick up Cc from the patch.

So [email protected] will never get fixed?

What's to fix? The doc (Documentation/stable_kernel_rules.txt) says to use "Cc: [email protected]" and using "Cc: [email protected]" causes bounces. Just use the documented address like everybody else and everything will be fine.

If you want "[email protected]" to work, first make the kernel.org change, then change Documentation/stable_kernel_rules.txt, *then* start using it in your patches.

Bjorn

author
#5 lore raw

Re: [PATCH 1/2] eisa, PCI: Fix bus res reference

Quoted 7 lines · Bjorn Helgaas
Quoted 2 lines

So [email protected] will never get fixed?

What's to fix? The doc (Documentation/stable_kernel_rules.txt) says to use "Cc: [email protected]" and using "Cc: [email protected]" causes bounces. Just use the documented address like everybody else and everything will be fine.

I don't think that I'm the only one is still using [email protected] in patch cc.
please check git log -p in recent commits in linus tree.

Quoted 4 lines

If you want "[email protected]" to work, first make the kernel.org change, then change Documentation/stable_kernel_rules.txt, *then* start using it in your patches.

[email protected] is not restored after last kernel.org get breached.

And I thought [email protected] should be just temporary one.

Yinghai

maintainer
#6 lore raw

Re: [PATCH 1/2] eisa, PCI: Fix bus res reference

In reply to Yinghai Lu
Quoted 38 lines · Yinghai Lu

Matthem found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

After looking at pci_eisa_init(), found it referring bus resource directly instead of pci_bus_resource_n().

After commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for pci root bus any more.

Fix it by using pci_bus_resource_n() and correct idx for root bus.

---
 drivers/eisa/pci_eisa.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

Index: linux-2.6/drivers/eisa/pci_eisa.c
===================================================================

drivers/eisa/pci_eisa.c +2 -1
@@ -22,7 +22,8 @@ static struct eisa_root_device pci_eisa_
 static int __init pci_eisa_init(struct pci_dev *pdev,
                                const struct pci_device_id *ent)
 {
-       int rc;
+       int rc, n = 0;
+       struct resource *bus_res;
 
        if ((rc = pci_enable_device (pdev))) {
                printk (KERN_ERR "pci_eisa : Could not enable device %s\n",

drivers/eisa/pci_eisa.c static int __init pci_eisa_init(struct p · line 36

Quoted 6 lines
drivers/eisa/pci_eisa.c +3 -0
@@ -30,9 +31,12 @@ static int __init pci_eisa_init(struct p
                return rc;
        }
 
+       if (pci_is_root_bus(pdev->bus))
+               n = PCI_BRIDGE_RESOURCE_NUM;
+       bus_res = pci_bus_resource_n(pdev->bus, n);

I haven't figured out why the pci_is_root_bus() test is here. Can you explain, maybe with a sample dmesg log and "lspci -vvx" output?

Doesn't the 82375 subtractively decode this PCI memory space, the same way a subtractive-decode P2P bridge does? Can we make this code look like the "if (dev->transparent)" block in pci_read_bridge_bases(), just breaking after the first resource we find, since eisa_root_register() can only deal with a single resource on the EISA bus?

Bjorn

Quoted 8 lines
drivers/eisa/pci_eisa.c +2 -2
        pci_eisa_root.dev              = &pdev->dev;
-       pci_eisa_root.res              = pdev->bus->resource[0];
-       pci_eisa_root.bus_base_addr    = pdev->bus->resource[0]->start;
+       pci_eisa_root.res              = bus_res;
+       pci_eisa_root.bus_base_addr    = bus_res->start;
        pci_eisa_root.slots            = EISA_MAX_SLOTS;
        pci_eisa_root.dma_mask         = pdev->dma_mask;
        dev_set_drvdata(pci_eisa_root.dev, &pci_eisa_root);
author
#7 lore raw

Re: [PATCH 1/2] eisa, PCI: Fix bus res reference

Quoted 42 lines · Bjorn Helgaas
Quoted 41 lines · Yinghai Lu

Matthem found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

After looking at pci_eisa_init(), found it referring bus resource directly instead of pci_bus_resource_n().

After commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for pci root bus any more.

Fix it by using pci_bus_resource_n() and correct idx for root bus.

---
 drivers/eisa/pci_eisa.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

Index: linux-2.6/drivers/eisa/pci_eisa.c
===================================================================

drivers/eisa/pci_eisa.c +2 -1
@@ -22,7 +22,8 @@ static struct eisa_root_device pci_eisa_
 static int __init pci_eisa_init(struct pci_dev *pdev,
                                const struct pci_device_id *ent)
 {
-       int rc;
+       int rc, n = 0;
+       struct resource *bus_res;
 
        if ((rc = pci_enable_device (pdev))) {
                printk (KERN_ERR "pci_eisa : Could not enable device %s\n",
@@ -30,9 +31,12 @@ static int __init pci_eisa_init(struct p
                return rc;
        }
 
Quoted 6 lines

drivers/eisa/pci_eisa.c static int __init pci_eisa_init(struct p · line 36

Quoted 3 lines
drivers/eisa/pci_eisa.c +3 -0
+       if (pci_is_root_bus(pdev->bus))
+               n = PCI_BRIDGE_RESOURCE_NUM;
+       bus_res = pci_bus_resource_n(pdev->bus, n);

I haven't figured out why the pci_is_root_bus() test is here. Can you explain, maybe with a sample dmesg log and "lspci -vvx" output?

Just forwarded two boot.log.

for root bus, pci_bus_resource_n[0, PCI_BRIDGE_RESOURCE_NUM) all equal to 0.

Quoted 7 lines

Doesn't the 82375 subtractively decode this PCI memory space, the same way a subtractive-decode P2P bridge does? Can we make this code look like the "if (dev->transparent)" block in pci_read_bridge_bases(), just breaking after the first resource we find, since eisa_root_register() can only deal with a single resource on the EISA bus?

they are using extra resource in pci_eisa_root.eisa_root_res to prevent multiple registering between pci_eisa and virtual platform.

and later eisa resource does get registered under system root ioport resource.

maintainer
#8 lore raw

Re: [PATCH 1/2] eisa, PCI: Fix bus res reference

Quoted 50 lines · Yinghai Lu
Quoted 42 lines · Bjorn Helgaas
Quoted 41 lines · Yinghai Lu

Matthem found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

After looking at pci_eisa_init(), found it referring bus resource directly instead of pci_bus_resource_n().

After commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for pci root bus any more.

Fix it by using pci_bus_resource_n() and correct idx for root bus.

---
 drivers/eisa/pci_eisa.c |   10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

Index: linux-2.6/drivers/eisa/pci_eisa.c
===================================================================

drivers/eisa/pci_eisa.c +2 -1
@@ -22,7 +22,8 @@ static struct eisa_root_device pci_eisa_
 static int __init pci_eisa_init(struct pci_dev *pdev,
                                const struct pci_device_id *ent)
 {
-       int rc;
+       int rc, n = 0;
+       struct resource *bus_res;
 
        if ((rc = pci_enable_device (pdev))) {
                printk (KERN_ERR "pci_eisa : Could not enable device %s\n",
@@ -30,9 +31,12 @@ static int __init pci_eisa_init(struct p
                return rc;
        }
 
Quoted 6 lines

drivers/eisa/pci_eisa.c static int __init pci_eisa_init(struct p · line 36

Quoted 3 lines
drivers/eisa/pci_eisa.c +3 -0
+       if (pci_is_root_bus(pdev->bus))
+               n = PCI_BRIDGE_RESOURCE_NUM;
+       bus_res = pci_bus_resource_n(pdev->bus, n);

I haven't figured out why the pci_is_root_bus() test is here. Can you explain, maybe with a sample dmesg log and "lspci -vvx" output?

Quoted 4 lines

Just forwarded two boot.log.

for root bus, pci_bus_resource_n[0, PCI_BRIDGE_RESOURCE_NUM) all equal to 0.

Thanks. The patch below is what I was thinking. That should work even in this scenario, I think, and it's nicer if we don't have to have this root bus internal knowledge here.

drivers/eisa/pci_eisa.c +23 -3
index cdae207..3e9dee9 100644
@@ -22,7 +22,8 @@ static struct eisa_root_device pci_eisa_root;
 static int __init pci_eisa_init(struct pci_dev *pdev,
 				const struct pci_device_id *ent)
 {
-	int rc;
+	int rc, i;
+	struct resource *res;
 
 	if ((rc = pci_enable_device (pdev))) {
 		printk (KERN_ERR "pci_eisa : Could not enable device %s\n",
@@ -30,9 +31,28 @@ static int __init pci_eisa_init(struct pci_dev *pdev,
 		return rc;
 	}
 
+	/*
+	 * The Intel 82375 PCI-EISA bridge is a subtractive-decode PCI
+	 * device, so the resources available on EISA are the same as those
+	 * available on the 82375 bus.  This works the same as a PCI-PCI
+	 * bridge in subtractive-decode mode (see pci_read_bridge_bases()).
+	 * We assume other PCI-EISA bridges are similar.
+	 *
+	 * eisa_root_register() can only deal with a single resource, so we
+	 * use the first valid one.
+	 */
+	pci_bus_for_each_resource(pdev->bus, res, i)
+		if (res)
+			break;
+
+	if (!res) {
+		dev_err(&pdev->dev, "No resources available\n");
+		return -1;
+	}
+
 	pci_eisa_root.dev              = &pdev->dev;
-	pci_eisa_root.res	       = pdev->bus->resource[0];
-	pci_eisa_root.bus_base_addr    = pdev->bus->resource[0]->start;
+	pci_eisa_root.res	       = res;
+	pci_eisa_root.bus_base_addr    = res->start;
 	pci_eisa_root.slots	       = EISA_MAX_SLOTS;
 	pci_eisa_root.dma_mask         = pdev->dma_mask;
 	dev_set_drvdata(pci_eisa_root.dev, &pci_eisa_root);
author
#9 lore raw

PATCH 2/2 eisa, PCI: init eisa early before pnp step in

In reply to Yinghai Lu

Mathhew reported kernels fail the pci_eisa probe and are later successful with the virtual_eisa_root_init force probe without slot0.

The reason for that is: pnp probing is early than pci_eisa_init get called as pci_eisa_init is called via pci_driver.

pnp 00:0f has 0xc80 - 0xc84 reserved.

[    9.700409] pnp 00:0f: [io  0x0c80-0x0c84]

so eisa_probe will fail from pci_eisa_init

				==>eisa_root_register
					==>eisa_probe path.

as force_probe is not set in pci_eisa_root, it will bail early when slot0 is not probed and initialized.

Try to use subsys_initcall_sync instead, and will keep following sequence:
pci_subsys_init
pci_eisa_init_early
pnpacpi_init/isapnp_init

After this patch eisa can be initialized properly, and pnp overlapping resource will not be reserved.

[   10.104434] system 00:0f: [io  0x0c80-0x0c84] could not be reserved
---
 drivers/eisa/pci_eisa.c |   41 ++++++++++++++++++++++-------------------
 1 file changed, 22 insertions(+), 19 deletions(-)
drivers/eisa/pci_eisa.c +22 -19
Index: linux-2.6/drivers/eisa/pci_eisa.c
===================================================================
@@ -19,8 +19,7 @@
 /* There is only *one* pci_eisa device per machine, right ? */
 static struct eisa_root_device pci_eisa_root;
 
-static int __init pci_eisa_init(struct pci_dev *pdev,
-				const struct pci_device_id *ent)
+static int __init pci_eisa_init(struct pci_dev *pdev)
 {
 	int rc, n = 0;
 	struct resource *bus_res;
@@ -49,22 +48,26 @@ static int __init pci_eisa_init(struct p
 	return 0;
 }
 
-static struct pci_device_id pci_eisa_pci_tbl[] = {
-	{ PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID,
-	  PCI_CLASS_BRIDGE_EISA << 8, 0xffff00, 0 },
-	{ 0, }
-};
-
-static struct pci_driver __refdata pci_eisa_driver = {
-	.name		= "pci_eisa",
-	.id_table	= pci_eisa_pci_tbl,
-	.probe		= pci_eisa_init,
-};
-
-static int __init pci_eisa_init_module (void)
+/*
+ * We have to call pci_eisa_init_early() before pnpacpi_init()/isapnp_init().
+ *   Otherwise pnp resource will get enabled early and could prevent eisa
+ *   to be initialized.
+ * Also need to make sure pci_eisa_init_early() is called after
+ * x86/pci_subsys_init().
+ * So need to use subsys_initcall_sync with it.
+ */
+static int __init pci_eisa_init_early(void)
 {
-	return pci_register_driver (&pci_eisa_driver);
-}
+	struct pci_dev *dev = NULL;
+	int ret;
+
+	for_each_pci_dev(dev)
+		if ((dev->class >> 8) == PCI_CLASS_BRIDGE_EISA) {
+			ret = pci_eisa_init(dev);
+			if (ret)
+				return ret;
+		}
 
-device_initcall(pci_eisa_init_module);
-MODULE_DEVICE_TABLE(pci, pci_eisa_pci_tbl);
+	return 0;
+}
+subsys_initcall_sync(pci_eisa_init_early);
maintainer
#10 lore raw

Re: [PATCH 2/2] eisa, PCI: init eisa early before pnp step in

[+cc Rafael, just FYI since it involves PNP resources]

Quoted 12 lines · Yinghai Lu

Mathhew reported kernels fail the pci_eisa probe and are later successful with the virtual_eisa_root_init force probe without slot0.

The reason for that is: pnp probing is early than pci_eisa_init get called as pci_eisa_init is called via pci_driver.

pnp 00:0f has 0xc80 - 0xc84 reserved.

[    9.700409] pnp 00:0f: [io  0x0c80-0x0c84]

so eisa_probe will fail from pci_eisa_init

                                ==>eisa_root_register
                                        ==>eisa_probe path.
Quoted 7 lines

as force_probe is not set in pci_eisa_root, it will bail early when slot0 is not probed and initialized.

Try to use subsys_initcall_sync instead, and will keep following sequence:
pci_subsys_init
pci_eisa_init_early
pnpacpi_init/isapnp_init

Is this a regression? This must have worked at one time, but it seems like we've had pnpacpi_init/isapnp_init/pnpbios_init before PCI drivers for quite a while.

Quoted 3 lines

After this patch eisa can be initialized properly, and pnp overlapping resource will not be reserved.

[   10.104434] system 00:0f: [io  0x0c80-0x0c84] could not be reserved

I think we should add logging in eisa_request_resources() similar to what drivers/pnp/system.c does so it's more clear what resources EISA is using.

Quoted 67 lines
---
 drivers/eisa/pci_eisa.c |   41 ++++++++++++++++++++++-------------------
 1 file changed, 22 insertions(+), 19 deletions(-)

Index: linux-2.6/drivers/eisa/pci_eisa.c
===================================================================

drivers/eisa/pci_eisa.c +22 -19
@@ -19,8 +19,7 @@
 /* There is only *one* pci_eisa device per machine, right ? */
 static struct eisa_root_device pci_eisa_root;
 
-static int __init pci_eisa_init(struct pci_dev *pdev,
-                               const struct pci_device_id *ent)
+static int __init pci_eisa_init(struct pci_dev *pdev)
 {
        int rc, n = 0;
        struct resource *bus_res;
@@ -49,22 +48,26 @@ static int __init pci_eisa_init(struct p
        return 0;
 }
 
-static struct pci_device_id pci_eisa_pci_tbl[] = {
-       { PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID,
-         PCI_CLASS_BRIDGE_EISA << 8, 0xffff00, 0 },
-       { 0, }
-};
-
-static struct pci_driver __refdata pci_eisa_driver = {
-       .name           = "pci_eisa",
-       .id_table       = pci_eisa_pci_tbl,
-       .probe          = pci_eisa_init,
-};
-
-static int __init pci_eisa_init_module (void)
+/*
+ * We have to call pci_eisa_init_early() before pnpacpi_init()/isapnp_init().
+ *   Otherwise pnp resource will get enabled early and could prevent eisa
+ *   to be initialized.
+ * Also need to make sure pci_eisa_init_early() is called after
+ * x86/pci_subsys_init().
+ * So need to use subsys_initcall_sync with it.
+ */
+static int __init pci_eisa_init_early(void)
 {
-       return pci_register_driver (&pci_eisa_driver);
-}
+       struct pci_dev *dev = NULL;
+       int ret;
+
+       for_each_pci_dev(dev)
+               if ((dev->class >> 8) == PCI_CLASS_BRIDGE_EISA) {
+                       ret = pci_eisa_init(dev);
+                       if (ret)
+                               return ret;
+               }
 
-device_initcall(pci_eisa_init_module);
-MODULE_DEVICE_TABLE(pci, pci_eisa_pci_tbl);
+       return 0;
+}
+subsys_initcall_sync(pci_eisa_init_early);
author
#11 lore raw

Re: [PATCH 2/2] eisa, PCI: init eisa early before pnp step in

Quoted 18 lines · Bjorn Helgaas

[+cc Rafael, just FYI since it involves PNP resources]

Quoted 15 lines · Yinghai Lu

Mathhew reported kernels fail the pci_eisa probe and are later successful with the virtual_eisa_root_init force probe without slot0.

The reason for that is: pnp probing is early than pci_eisa_init get called as pci_eisa_init is called via pci_driver.

pnp 00:0f has 0xc80 - 0xc84 reserved.

[    9.700409] pnp 00:0f: [io  0x0c80-0x0c84]

so eisa_probe will fail from pci_eisa_init

                                ==>eisa_root_register
                                        ==>eisa_probe path.

as force_probe is not set in pci_eisa_root, it will bail early when slot0 is not probed and initialized.

Quoted 8 lines
Quoted 4 lines

Try to use subsys_initcall_sync instead, and will keep following sequence:
pci_subsys_init
pci_eisa_init_early
pnpacpi_init/isapnp_init

Is this a regression? This must have worked at one time, but it seems like we've had pnpacpi_init/isapnp_init/pnpbios_init before PCI drivers for quite a while.

Yes.

Quoted 8 lines
Quoted 4 lines

After this patch eisa can be initialized properly, and pnp overlapping resource will not be reserved.

[   10.104434] system 00:0f: [io  0x0c80-0x0c84] could not be reserved

I think we should add logging in eisa_request_resources() similar to what drivers/pnp/system.c does so it's more clear what resources EISA is using.

Those eisa resource does get registered into io port resource tree.

Quoted 68 lines
Quoted 68 lines
---
 drivers/eisa/pci_eisa.c |   41 ++++++++++++++++++++++-------------------
 1 file changed, 22 insertions(+), 19 deletions(-)

Index: linux-2.6/drivers/eisa/pci_eisa.c
===================================================================

drivers/eisa/pci_eisa.c +22 -19
@@ -19,8 +19,7 @@
 /* There is only *one* pci_eisa device per machine, right ? */
 static struct eisa_root_device pci_eisa_root;
 
-static int __init pci_eisa_init(struct pci_dev *pdev,
-                               const struct pci_device_id *ent)
+static int __init pci_eisa_init(struct pci_dev *pdev)
 {
        int rc, n = 0;
        struct resource *bus_res;
@@ -49,22 +48,26 @@ static int __init pci_eisa_init(struct p
        return 0;
 }
 
-static struct pci_device_id pci_eisa_pci_tbl[] = {
-       { PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID,
-         PCI_CLASS_BRIDGE_EISA << 8, 0xffff00, 0 },
-       { 0, }
-};
-
-static struct pci_driver __refdata pci_eisa_driver = {
-       .name           = "pci_eisa",
-       .id_table       = pci_eisa_pci_tbl,
-       .probe          = pci_eisa_init,
-};
-
-static int __init pci_eisa_init_module (void)
+/*
+ * We have to call pci_eisa_init_early() before pnpacpi_init()/isapnp_init().
+ *   Otherwise pnp resource will get enabled early and could prevent eisa
+ *   to be initialized.
+ * Also need to make sure pci_eisa_init_early() is called after
+ * x86/pci_subsys_init().
+ * So need to use subsys_initcall_sync with it.
+ */
+static int __init pci_eisa_init_early(void)
 {
-       return pci_register_driver (&pci_eisa_driver);
-}
+       struct pci_dev *dev = NULL;
+       int ret;
+
+       for_each_pci_dev(dev)
+               if ((dev->class >> 8) == PCI_CLASS_BRIDGE_EISA) {
+                       ret = pci_eisa_init(dev);
+                       if (ret)
+                               return ret;
+               }
 
-device_initcall(pci_eisa_init_module);
-MODULE_DEVICE_TABLE(pci, pci_eisa_pci_tbl);
+       return 0;
+}
+subsys_initcall_sync(pci_eisa_init_early);
maintainer
#12 lore raw

Re: [PATCH 2/2] eisa, PCI: init eisa early before pnp step in

Quoted 24 lines · Yinghai Lu
Quoted 23 lines · Bjorn Helgaas

[+cc Rafael, just FYI since it involves PNP resources]

Quoted 19 lines · Yinghai Lu

Mathhew reported kernels fail the pci_eisa probe and are later successful with the virtual_eisa_root_init force probe without slot0.

The reason for that is: pnp probing is early than pci_eisa_init get called as pci_eisa_init is called via pci_driver.

pnp 00:0f has 0xc80 - 0xc84 reserved.

[    9.700409] pnp 00:0f: [io  0x0c80-0x0c84]

so eisa_probe will fail from pci_eisa_init

                                ==>eisa_root_register
                                        ==>eisa_probe path.

as force_probe is not set in pci_eisa_root, it will bail early when slot0 is not probed and initialized.

Try to use subsys_initcall_sync instead, and will keep following sequence:
pci_subsys_init
pci_eisa_init_early
pnpacpi_init/isapnp_init

Quoted 5 lines
Quoted 3 lines

Is this a regression? This must have worked at one time, but it seems like we've had pnpacpi_init/isapnp_init/pnpbios_init before PCI drivers for quite a while.

Yes.

Do you know when the regression occurred? If you do, I'll add that info to the "stable" tag.

Bjorn

author
#13 lore raw

Re: [PATCH 2/2] eisa, PCI: init eisa early before pnp step in

Quoted 8 lines · Bjorn Helgaas
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Is this a regression? This must have worked at one time, but it seems

like we've had pnpacpi_init/isapnp_init/pnpbios_init before PCI drivers for quite a while.

Yes.

Do you know when the regression occurred? If you do, I'll add that info to the "stable" tag.

No, I don't.

Looking at the current code, it should be quite before kernel was using git. as first drivers/Makefile in git already has pnp before eisa.

Do we need to dig before 2.6.12 and 2.4?

Yinghai

maintainer
#14 lore raw

Re: [PATCH 2/2] eisa, PCI: init eisa early before pnp step in

Quoted 10 lines · Yinghai Lu
Quoted 8 lines · Bjorn Helgaas
Quoted 5 lines
Quoted 3 lines
Quoted 1 line

Is this a regression? This must have worked at one time, but it seems

like we've had pnpacpi_init/isapnp_init/pnpbios_init before PCI drivers for quite a while.

Yes.

Do you know when the regression occurred? If you do, I'll add that info to the "stable" tag.

Quoted 6 lines

No, I don't.

Looking at the current code, it should be quite before kernel was using git. as first drivers/Makefile in git already has pnp before eisa.

Do we need to dig before 2.6.12 and 2.4?

No, I don't think we need to worry about fixing kernels that old.

But I do wonder whether we fully understand the cause. I know EISA is old and obsolete, but it's a little hard to believe that it's been broken since 2.6.12 and nobody noticed.

Bjorn

maintainer
#15 lore raw
In reply to Yinghai Lu

[+cc linux-pci]

Quoted 7 lines · Yinghai Lu

Looks like pci eisa bridge support is broken for a while.

one is root io resource reference, and other one is pnp related.

Please check if we can put them in v3.9.
eisa, PCI: Fix bus res reference eisa, PCI: init eisa early before pnp step in

Both of these patches are to drivers/eisa/pci_eisa.c. I don't see an EISA-specific maintainer, and I caused at least one of these problems, so I'll push these through my PCI tree unless somebody objects.

Bjorn

maintainer Tested-by Matthew Whitehead Tested-by Matthew Whitehead
#16 lore raw
Quoted 7 lines · Bjorn Helgaas

[+cc linux-pci]

Quoted 4 lines · Yinghai Lu

Looks like pci eisa bridge support is broken for a while.

one is root io resource reference, and other one is pnp related.

Quoted 7 lines
Quoted 3 lines

Please check if we can put them in v3.9.
eisa, PCI: Fix bus res reference eisa, PCI: init eisa early before pnp step in

Both of these patches are to drivers/eisa/pci_eisa.c. I don't see an EISA-specific maintainer, and I caused at least one of these problems, so I'll push these through my PCI tree unless somebody objects.

Hi Matthew,

Can you test either the attached patch (against v3.9-rc4) or the git
tree at http://git.kernel.org/cgit/linux/kernel/git/helgaas/pci.git/log/?h=pci/yinghai-eisa
and send me the complete dmesg log?

If all goes well, I'll ask Linus to pull these before v3.9.

Thanks!

Bjorn

commit 2cfda637e29ce9e3df31b59f64516b2e571cc985

Author: Yinghai Lu <[email protected]>
Date:   Mon Apr 1 11:48:59 2013 -0600

EISA/PCI: Fix bus res reference

Matthew found that 3.8.3 is having problems with an old (ancient) PCI-to-EISA bridge, the Intel 82375. It worked with the 3.2 kernel. He identified the 82375, but doesn't assign the struct resource *res pointer inside the struct eisa_root_device, and panics.

pci_eisa_init() was using bus->resource[] directly instead of pci_bus_resource_n(). The bus->resource[] array is a PCI-internal implementation detail, and after commit 45ca9e97 (PCI: add helpers for building PCI bus resource lists) and commit 0efd5aab (PCI: add struct pci_host_bridge_window with CPU/bus address offset), bus->resource[] is not used for PCI root buses any more.

The 82375 is a subtractive-decode PCI device, so handle it the same way we handle PCI-PCI bridges in subtractive-decode mode in pci_read_bridge_bases().

drivers/eisa/pci_eisa.c +25 -3
index cdae207..ef5c3ec 100644
@@ -22,7 +22,8 @@ static struct eisa_root_device pci_eisa_root;
 static int __init pci_eisa_init(struct pci_dev *pdev,
 				const struct pci_device_id *ent)
 {
-	int rc;
+	int rc, i;
+	struct resource *res, *bus_res = NULL;
 
 	if ((rc = pci_enable_device (pdev))) {
 		printk (KERN_ERR "pci_eisa : Could not enable device %s\n",
@@ -30,9 +31,30 @@ static int __init pci_eisa_init(struct pci_dev *pdev,
 		return rc;
 	}
 
+	/*
+	 * The Intel 82375 PCI-EISA bridge is a subtractive-decode PCI
+	 * device, so the resources available on EISA are the same as those
+	 * available on the 82375 bus.  This works the same as a PCI-PCI
+	 * bridge in subtractive-decode mode (see pci_read_bridge_bases()).
+	 * We assume other PCI-EISA bridges are similar.
+	 *
+	 * eisa_root_register() can only deal with a single io port resource,
+	*  so we use the first valid io port resource.
+	 */
+	pci_bus_for_each_resource(pdev->bus, res, i)
+		if (res && (res->flags & IORESOURCE_IO)) {
+			bus_res = res;
+			break;
+		}
+
+	if (!bus_res) {
+		dev_err(&pdev->dev, "No resources available\n");
+		return -1;
+	}
+
 	pci_eisa_root.dev              = &pdev->dev;
-	pci_eisa_root.res	       = pdev->bus->resource[0];
-	pci_eisa_root.bus_base_addr    = pdev->bus->resource[0]->start;
+	pci_eisa_root.res	       = bus_res;
+	pci_eisa_root.bus_base_addr    = bus_res->start;
 	pci_eisa_root.slots	       = EISA_MAX_SLOTS;
 	pci_eisa_root.dma_mask         = pdev->dma_mask;
 	dev_set_drvdata(pci_eisa_root.dev, &pci_eisa_root);

commit c5fb301ae83bec6892e54984e6ec765c47df8e10

Author: Yinghai Lu <[email protected]>
Date:   Wed Mar 27 21:28:05 2013 -0700

EISA/PCI: Init EISA early, before PNP

Matthew reported kernels fail the pci_eisa probe and are later successful with the virtual_eisa_root_init force probe without slot0.

The reason for that is: PNP probing is before pci_eisa_init gets called as pci_eisa_init is called via pci_driver.

pnp 00:0f has 0xc80 - 0xc84 reserved.
[ 9.700409] pnp 00:0f: [io 0x0c80-0x0c84]

so eisa_probe will fail from pci_eisa_init
==>eisa_root_register
==>eisa_probe path.
as force_probe is not set in pci_eisa_root, it will bail early when slot0 is not probed and initialized.

Try to use subsys_initcall_sync instead, and will keep following sequence:
pci_subsys_init
pci_eisa_init_early
pnpacpi_init/isapnp_init

After this patch EISA can be initialized properly, and PNP overlapping resource will not be reserved.
[ 10.104434] system 00:0f: [io 0x0c80-0x0c84] could not be reserved

drivers/eisa/pci_eisa.c +21 -18
index ef5c3ec..6c3fca9 100644
@@ -19,8 +19,7 @@
 /* There is only *one* pci_eisa device per machine, right ? */
 static struct eisa_root_device pci_eisa_root;
 
-static int __init pci_eisa_init(struct pci_dev *pdev,
-				const struct pci_device_id *ent)
+static int __init pci_eisa_init(struct pci_dev *pdev)
 {
 	int rc, i;
 	struct resource *res, *bus_res = NULL;
@@ -67,22 +66,26 @@ static int __init pci_eisa_init(struct pci_dev *pdev,
 	return 0;
 }
 
-static struct pci_device_id pci_eisa_pci_tbl[] = {
-	{ PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID,
-	  PCI_CLASS_BRIDGE_EISA << 8, 0xffff00, 0 },
-	{ 0, }
-};
+/*
+ * We have to call pci_eisa_init_early() before pnpacpi_init()/isapnp_init().
+ *   Otherwise pnp resource will get enabled early and could prevent eisa
+ *   to be initialized.
+ * Also need to make sure pci_eisa_init_early() is called after
+ * x86/pci_subsys_init().
+ * So need to use subsys_initcall_sync with it.
+ */
+static int __init pci_eisa_init_early(void)
+{
+	struct pci_dev *dev = NULL;
+	int ret;
 
-static struct pci_driver __refdata pci_eisa_driver = {
-	.name		= "pci_eisa",
-	.id_table	= pci_eisa_pci_tbl,
-	.probe		= pci_eisa_init,
-};
+	for_each_pci_dev(dev)
+		if ((dev->class >> 8) == PCI_CLASS_BRIDGE_EISA) {
+			ret = pci_eisa_init(dev);
+			if (ret)
+				return ret;
+		}
 
-static int __init pci_eisa_init_module (void)
-{
-	return pci_register_driver (&pci_eisa_driver);
+	return 0;
 }
-
-device_initcall(pci_eisa_init_module);
-MODULE_DEVICE_TABLE(pci, pci_eisa_pci_tbl);
+subsys_initcall_sync(pci_eisa_init_early);

commit d0e36df69863bfa054a4703406a737b0476e68e8

Author: Bjorn Helgaas <[email protected]>
Date:   Mon Apr 1 14:16:56 2013 -0600

EISA: Use dev_printk() when possible

Use dev_printk() when possible to make messages more useful.

drivers/eisa/eisa-bus.c +28 -35

113 lines not on this page · raw message

drivers/eisa/pci_eisa.c +2 -3
index 6c3fca9..a333bf3 100644
@@ -25,8 +25,7 @@ static int __init pci_eisa_init(struct pci_dev *pdev)
 	struct resource *res, *bus_res = NULL;
 
 	if ((rc = pci_enable_device (pdev))) {
-		printk (KERN_ERR "pci_eisa : Could not enable device %s\n",
-			pci_name(pdev));
+		dev_err(&pdev->dev, "Could not enable device\n");
 		return rc;
 	}
 
@@ -59,7 +58,7 @@ static int __init pci_eisa_init(struct pci_dev *pdev)
 	dev_set_drvdata(pci_eisa_root.dev, &pci_eisa_root);
 
 	if (eisa_root_register (&pci_eisa_root)) {
-		printk (KERN_ERR "pci_eisa : Could not register EISA root\n");
+		dev_err(&pdev->dev, "Could not register EISA root\n");
 		return -1;
 	}
 

commit 3a94edb7ba59d3ca30a3a54d8de628f1ccf004e4

Author: Bjorn Helgaas <[email protected]>
Date:   Mon Apr 1 14:40:05 2013 -0600

EISA: Log device resources in dmesg

Note the resources consumed by EISA devices in dmesg, similar to what we already do for PCI and PNP devices.

drivers/eisa/eisa-bus.c +1 -0
index 0b06df4..a6fbf98 100644
@@ -288,6 +288,7 @@ static int __init eisa_request_resources(struct eisa_root_device *root,
 			edev->res[i].flags = IORESOURCE_BUSY;
 		}
 
+		dev_printk(KERN_DEBUG, &edev->dev, "%pR\n", &edev->res[i]);
 		if (request_resource(root->res, &edev->res[i]))
 			goto failed;
 	}

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

maintainer
#22 lore raw
Quoted 1 line · Matthew Whitehead

It looks like it fixed /proc/ioports, but demsg is still (null). The boot log is attached.

Huh, I don't see the "(null)" in your patch4_boot.log. Here's what I see:

[    1.816154] pci 0000:00:07.0: Probing EISA bus 0
[    1.820819] eisa 00:00: [io  0x0c80-0x0c83]
[    1.824092] eisa 00:00: EISA: Mainboard ALR01A4 detected
[    1.833195] eisa 00:04: [io  0x4000-0x40ff]
[    1.836097] eisa 00:04: [io  0x4400-0x44ff]
[    1.840084] eisa 00:04: [io  0x4800-0x48ff]
[    1.844084] eisa 00:04: [io  0x4c00-0x4cff]
[    1.848082] eisa 00:04: EISA: slot 4: TCM5093 detected
[    1.852947] eisa 00:06: [io  0x6000-0x60ff]
[    1.856092] eisa 00:06: [io  0x6400-0x64ff]
[    1.860082] eisa 00:06: [io  0x6800-0x68ff]
[    1.864083] eisa 00:06: [io  0x6c00-0x6cff]
[    1.868083] eisa 00:06: EISA: slot 6: HWP1940 detected (forced enabled)
[    1.873097] eisa 00:07: [io  0x7000-0x70ff]
[    1.876097] eisa 00:07: [io  0x7400-0x74ff]
[    1.880083] eisa 00:07: [io  0x7800-0x78ff]
[    1.884082] eisa 00:07: [io  0x7c00-0x7cff]
[    1.888090] eisa 00:07: EISA: slot 7: TCM5920 detected
[    1.892702] pci 0000:00:07.0: EISA: Detected 3 cards

This now matches exactly what you have in /proc/ioports, which is good. Previously there were a bunch of these:

[    1.796731]  (null): [io  0x1000-0x10ff]
[    1.800077]  (null): [io  0x1400-0x14ff]

...

and I couldn't figure out where they came from. But those were an artifact of setting up the resources before checking to see whether there was a device there. When we discovered there was no device there, we discarded these resources, so it wasn't harmful, but dmesg was still ugly. Now that we check for a device first, we skip the resources for empty slots 1, 2, 3, 5, and 8, which is much nicer.

Thanks for all your testing!

I put the bugfixes (Yinghai's patches) in my "for-linus" branch for v3.9. The other printk fixes and such don't fix any actual bugs, so I'll put them in my "next" branch for merging into v3.10.

Bjorn