Re: [PATCH V4: 3/3] pci: Provide Multiple Error Received supporton AER

From: Andrew Patterson
Date: Mon Jun 15 2009 - 00:01:26 EST


On Mon, 2009-06-15 at 09:47 +0800, Zhang, Yanmin wrote:
> On Fri, 2009-06-12 at 22:16 +0000, Andrew Patterson wrote:
> > On Fri, 2009-06-12 at 11:08 +0800, Zhang, Yanmin wrote:
> > > ïïïïïWhen a root port receive the same errors more than once before kernel
> > > process them, the Multiple Error Messages Received flags are set by
> > > hardware. Because root port could only save one kind of correctable
> > > error source id and another uncorrectable error source id at the same
> > > time, so the second message sender id is lost if the 2 messages are
> > > sent from 2 different devices. Below patch searches all devices under
> > > the root port when multiple messages are received.
> > >
> > > ïïïïSigned-off-by: ïZhang Yanmin <yanmin.zhang@xxxxxxxxxxxxxxx>
> > >
> > > ---
> > >
> > > diff -Nraup linux-2.6_next_aernoid/drivers/pci/pcie/aer/aerdrv_core.c linux-2.6_next_aermultierror/drivers/pci/pcie/aer/aerdrv_core.c
> > > --- linux-2.6_next_aernoid/drivers/pci/pcie/aer/aerdrv_core.c 2009-06-12 05:39:24.000000000 +0800
> > > +++ linux-2.6_next_aermultierror/drivers/pci/pcie/aer/aerdrv_core.c 2009-06-12 05:45:15.000000000 +0800
> > > @@ -145,13 +145,22 @@ static void set_downstream_devices_error
> > > pci_walk_bus(dev->subordinate, set_device_error_reporting, &enable);
> > > }
> > >
> > > +static int add_error_device(struct aer_err_info *e_info, struct pci_dev *dev)
> > > +{
> > > + if (e_info->error_dev_num < AER_MAX_MULTI_ERR_DEVICES) {
> > > + e_info->dev[e_info->error_dev_num ++] = dev;
> >
> > checkpatch reports:
> > ERROR: space prohibited before that '++' (ctx:WxB)
> > #46: FILE: drivers/pci/pcie/aer/aerdrv_core.c:151:
> > + e_info->dev[e_info->error_dev_num ++] = dev;
> Thanks. I will change it and use checkpatch to check it.
>
> >
> > Personally I would prefer:
> > e_info->dev[e_info->error_dev_num] = dev;
> > e_info->error_dev_num++;
> >
> > > + return 1;
> > > + } else
> > > + return 0;
> > > +}
> > > +
> >
> >
> > This function is now doing more than just comparing device ID's.
> > Perhaps you could rename it or put call add_error_device after
> > compare_device_id in find_device_iter?
> I will move the call ïadd_error_device to find_device_iter.
>
> > > static int compare_device_id(struct pci_dev *dev, struct aer_err_info *e_info)
> > > {
> > > if (e_info->id == ((dev->bus->number << 8) | dev->devfn)) {
> > > /*
> > > * Device ID match
> > > */
> > > - e_info->dev = dev;
> > > + add_error_device(e_info, dev);
> > > return 1;
> > > }
> > >
> > > @@ -166,20 +175,38 @@ static int find_device_iter(struct pci_d
> > > u32 status;
> > > u32 mask;
> > > u16 reg16;
> > > + int result;
> > > struct aer_err_info *e_info = (struct aer_err_info *)data;
> > >
> > > /*
> > > * When bus id is equal to 0, it might be a bad id
> > > * reported by root port.
> > > */
> > > - if (!nosourceid && (PCI_BUS(e_info->id) != 0))
> > > - return compare_device_id(dev, e_info);
> > > + if (!nosourceid && (PCI_BUS(e_info->id) != 0)) {
> > > + result = compare_device_id(dev, e_info);
> > > + /*
> > > + * If there is no multiple error, we stop
> > > + * or continue based on the id comparing.
> > > + */
> > > + if (!(e_info->flags & AER_MULTI_ERROR_VALID_FLAG))
> > > + return result;
> > > +
> > > + /*
> > > + * If there are multiple errors and id does match,
> > > + * We need continue to search other devices under
> > > + * the root port. Return 0 means that.
> > > + */
> > > + if (result)
> > > + return 0;
> > > + }
> > >
> > > /*
> > > - * Next is to check when bus id is equal to 0 or
> > > - * nosourceid==y. Some ports might lose the bus
> > > - * id of error source id. We check AER status
> > > - * registers to find the initial reporter.
> > > + * When either
> > > + * 1) nosourceid==y;
> > > + * 2) bus id is equal to 0. Some ports might lose the bus
> > > + * id of error source id;
> > > + * 3) There are multiple errors and prior id comparing fails;
> > > + * We check AER status registers to find the initial reporter.
> > > */
> > > if (atomic_read(&dev->enable_cnt) == 0)
> > > return 0;
> > > @@ -208,8 +235,8 @@ static int find_device_iter(struct pci_d
> > > pos + PCI_ERR_COR_MASK,
> > > &mask);
> > > if (status & ERR_CORRECTABLE_ERROR_MASK & ~mask) {
> > > - e_info->dev = dev;
> > > - return 1;
> > > + add_error_device(e_info, dev);
> > > + goto added;
> > > }
> > > } else {
> > > pci_read_config_dword(dev,
> > > @@ -219,12 +246,18 @@ static int find_device_iter(struct pci_d
> > > pos + PCI_ERR_UNCOR_MASK,
> > > &mask);
> > > if (status & ERR_UNCORRECTABLE_ERROR_MASK & ~mask) {
> > > - e_info->dev = dev;
> > > - return 1;
> > > + add_error_device(e_info, dev);
> > > + goto added;
> > > }
> > > }
> > >
> > > return 0;
> > > +
> > > +added:
> > > + if (e_info->flags & AER_MULTI_ERROR_VALID_FLAG) {
> > > + return 0;
> > > + } else
> > > + return 1;
> >
> > checkpatch reports:
> > WARNING: braces {} are not necessary for any arm of this statement
> > #133: FILE: drivers/pci/pcie/aer/aerdrv_core.c:257:
> I will change it.
>
> > + if (e_info->flags & AER_MULTI_ERROR_VALID_FLAG) {
> > [...]
> > + } else
> > [...]
> >
> >
> > > }
> > >
> > > /**
> > > @@ -705,6 +738,30 @@ static int get_device_error_info(struct
> > > return AER_SUCCESS;
> > > }
> > >
> > > +static inline void aer_process_err_devices(struct pcie_device *p_device,
> > > + struct aer_err_info *e_info)
> > > +{
> > > + int i;
> > > +
> > > + if (e_info->dev[0] == NULL) {
> > Minor not. Can we use
> Yes. We can.
>
> > if (!e_info->dev[0]) {
> >
> > > + printk(KERN_DEBUG "%s->can't find device of ID%04x\n",
> > > + __func__, e_info->id);
> >
> > I suspect we don't want to embed the function name here, and use
> > dev_printk.
> Ok.
>
> >
> > > + }
> > > +
> > > + for (i = 0; i < e_info->error_dev_num; i ++) {
> > checkpatch reports:
> > ERROR: space prohibited before that '++' (ctx:WxB)
> > #154: FILE: drivers/pci/pcie/aer/aerdrv_core.c:751:
> > + for (i = 0; i < e_info->error_dev_num; i +
> I will change it.
>
> >
> >
> > > + if (e_info->dev[i] == NULL)
> > again if (!e_info->dev[i])
> Will do.
>
> >
> > You could also put this check in the for loop.
> I planed to, but one guy helped me test it within a guest OS on XEN and
> reported a weired oops of guest OS. She said useing ïe_info->error_dev_num
> could avoid the oops.

I think something like:

for (i = 0; i < e_info->error_dev_num && e_info->dev[i]; i++)

is functionally equivalent.

>
> >
> >
> > > + break;
> > > +
> > > + if (get_device_error_info(e_info->dev[i], e_info) ==
> > > + AER_SUCCESS) {
> > > + aer_print_error(e_info->dev[i], e_info);
> > > + handle_error_source(p_device,
> > > + e_info->dev[i],
> > > + e_info);
> > > + }
> > > + }
> > > +}
> > > +
> > > /**
> > > * aer_isr_one_error - consume an error detected by root port
> > > * @p_device: pointer to error root port service device
> > > @@ -747,18 +804,7 @@ static void aer_isr_one_error(struct pci
> > > e_info->flags |= AER_MULTI_ERROR_VALID_FLAG;
> > >
> > > find_source_device(p_device->port, e_info);
> > > - if (e_info->dev == NULL) {
> > > - printk(KERN_DEBUG "%s->can't find device of ID%04x\n",
> > > - __func__, e_info->id);
> > > - continue;
> > > - }
> > > - if (get_device_error_info(e_info->dev, e_info) ==
> > > - AER_SUCCESS) {
> > > - aer_print_error(e_info->dev, e_info);
> > > - handle_error_source(p_device,
> > > - e_info->dev,
> > > - e_info);
> > > - }
> > > + aer_process_err_devices(p_device, e_info);
> > > }
> > >
> > > kfree(e_info);
> > > diff -Nraup linux-2.6_next_aernoid/drivers/pci/pcie/aer/aerdrv.h linux-2.6_next_aermultierror/drivers/pci/pcie/aer/aerdrv.h
> > > --- linux-2.6_next_aernoid/drivers/pci/pcie/aer/aerdrv.h 2009-06-12 05:39:24.000000000 +0800
> > > +++ linux-2.6_next_aermultierror/drivers/pci/pcie/aer/aerdrv.h 2009-06-12 05:45:15.000000000 +0800
> > > @@ -57,8 +57,10 @@ struct header_log_regs {
> > > unsigned int dw3;
> > > };
> > >
> > > +#define AER_MAX_MULTI_ERR_DEVICES 5
> > Is this number arbitrary or in the spec somewhere?
> It's arbitrary and not spec.

I suspected so.

> The startpoint is it's very rare that there are more
> than 5 devices under the same root port reporting errors at the same time.

Agreed.

> It's hard
> to say number 5 is the best. I just don't want the array is big.

I don't have a problem with that decision. But you might add a comment
saying so, e.g.,

#define AER_MAX_MULTI_ERR_DEVICES 5 /* Not likely to have more */

>
> >
> >
> > > struct aer_err_info {
> > > - struct pci_dev *dev;
> > > + struct pci_dev *dev[AER_MAX_MULTI_ERR_DEVICES];
> > > + int error_dev_num;
> > > u16 id;
> > > int severity; /* 0:NONFATAL | 1:FATAL | 2:COR */
> > > int flags;
> > >
>
>
--
Andrew Patterson
Hewlett-Packard

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/