From: Don Dutile <ddutile@redhat.com>
To: Ben Hutchings <bhutchings@solarflare.com>
Cc: linux-pci@vger.kernel.org, bhelgaas@google.com,
yuvalmin@broadcom.com, gregory.v.rose@intel.com,
yinghai@kernel.org, davem@davemloft.net
Subject: Re: [PATCH 3/8] PCI: sysfs per device SRIOV control and status
Date: Fri, 26 Oct 2012 11:07:05 -0400 [thread overview]
Message-ID: <508AA719.2060705@redhat.com> (raw)
In-Reply-To: <1351196234.2662.37.camel@bwh-desktop.uk.solarflarecom.com>
On 10/25/2012 04:17 PM, Ben Hutchings wrote:
> On Thu, 2012-10-25 at 14:38 -0400, Donald Dutile wrote:
> [...]
>> --- a/drivers/pci/pci-sysfs.c
>> +++ b/drivers/pci/pci-sysfs.c
>> @@ -14,6 +14,7 @@
>> *
>> */
>>
>> +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt /* has to precede printk.h */
>>
>> #include<linux/kernel.h>
>> #include<linux/sched.h>
>> @@ -66,7 +67,7 @@ static ssize_t broken_parity_status_store(struct device *dev,
>> struct pci_dev *pdev = to_pci_dev(dev);
>> unsigned long val;
>>
>> - if (strict_strtoul(buf, 0,&val)< 0)
>> + if (kstrtoul(buf, 0,&val)< 0)
>> return -EINVAL;
>>
>> pdev->broken_parity_status = !!val;
>> @@ -188,7 +189,7 @@ static ssize_t is_enabled_store(struct device *dev,
>> {
>> struct pci_dev *pdev = to_pci_dev(dev);
>> unsigned long val;
>> - ssize_t result = strict_strtoul(buf, 0,&val);
>> + ssize_t result = kstrtoul(buf, 0,&val);
>>
>> if (result< 0)
>> return result;
>> @@ -259,7 +260,7 @@ msi_bus_store(struct device *dev, struct device_attribute *attr,
>> struct pci_dev *pdev = to_pci_dev(dev);
>> unsigned long val;
>>
>> - if (strict_strtoul(buf, 0,&val)< 0)
>> + if (kstrtoul(buf, 0,&val)< 0)
>> return -EINVAL;
>>
>> /* bad things may happen if the no_msi flag is changed
>> @@ -292,7 +293,7 @@ static ssize_t bus_rescan_store(struct bus_type *bus, const char *buf,
>> unsigned long val;
>> struct pci_bus *b = NULL;
>>
>> - if (strict_strtoul(buf, 0,&val)< 0)
>> + if (kstrtoul(buf, 0,&val)< 0)
>> return -EINVAL;
>>
>> if (val) {
>> @@ -316,7 +317,7 @@ dev_rescan_store(struct device *dev, struct device_attribute *attr,
>> unsigned long val;
>> struct pci_dev *pdev = to_pci_dev(dev);
>>
>> - if (strict_strtoul(buf, 0,&val)< 0)
>> + if (kstrtoul(buf, 0,&val)< 0)
>> return -EINVAL;
>>
>> if (val) {
>> @@ -343,7 +344,7 @@ remove_store(struct device *dev, struct device_attribute *dummy,
>> int ret = 0;
>> unsigned long val;
>>
>> - if (strict_strtoul(buf, 0,&val)< 0)
>> + if (kstrtoul(buf, 0,&val)< 0)
>> return -EINVAL;
>>
>> /* An attribute cannot be unregistered by one of its own methods,
>> @@ -363,7 +364,7 @@ dev_bus_rescan_store(struct device *dev, struct device_attribute *attr,
>> unsigned long val;
>> struct pci_bus *bus = to_pci_bus(dev);
>>
>> - if (strict_strtoul(buf, 0,&val)< 0)
>> + if (kstrtoul(buf, 0,&val)< 0)
>> return -EINVAL;
>>
>> if (val) {
>> @@ -387,7 +388,7 @@ static ssize_t d3cold_allowed_store(struct device *dev,
>> struct pci_dev *pdev = to_pci_dev(dev);
>> unsigned long val;
>>
>> - if (strict_strtoul(buf, 0,&val)< 0)
>> + if (kstrtoul(buf, 0,&val)< 0)
>> return -EINVAL;
>>
>> pdev->d3cold_allowed = !!val;
>
> All this cleanup belongs in a separate patch.
Yes. Multiple people were getting anxious to see this V3, so I wanted
to get it out for review of its direction.
I did a checkpatch.pl run after I posted... ugh! what a mess!
you note a number of the issues below.
>
>> @@ -404,6 +405,114 @@ static ssize_t d3cold_allowed_show(struct device *dev,
>> }
>> #endif
>>
>> +#ifdef CONFIG_PCI_IOV
>> +static ssize_t sriov_totalvfs_show(struct device *dev,
>> + struct device_attribute *attr,
>> + char *buf)
>> +{
>> + struct pci_dev *pdev;
>> + u16 total;
>> +
>> + pdev = to_pci_dev(dev);
>> + total = pdev->sriov->total;
>> + return sprintf (buf, "%u\n", total);
>
> No space after sprintf, please.
>
yep, checkpatch...
>> +}
>> +
>> +
>> +static ssize_t sriov_numvfs_show(struct device *dev,
>> + struct device_attribute *attr,
>> + char *buf)
>> +{
>> + struct pci_dev *pdev;
>> +
>> + pdev = to_pci_dev(dev);
>> +
>> + return sprintf (buf, "%u\n", pdev->sriov->nr_virtfn);
>> +}
>> +
>> +/*
>> + * num_vfs> 0; number of vfs to enable
>> + * num_vfs = 0; disable all vfs
>> + *
>> + * Note: SRIOV spec doesn't allow partial VF
>> + * disable, so its all or none.
>> + */
>> +static ssize_t sriov_numvfs_store(struct device *dev,
>> + struct device_attribute *attr,
>> + const char *buf, size_t count)
>> +{
>> + struct pci_dev *pdev;
>> + int num_vfs_enabled = 0;
>> + int num_vfs;
>> + int ret = 0;
>> + u16 total;
>> +
>> + pdev = to_pci_dev(dev);
>> +
>> + /* Requested VFs to enable< totalvfs
>> + * and none enabled already
>> + */
>> + if (kstrtoint(buf, 0,&num_vfs)< 0)
>> + return -EINVAL;
>
> The above comment belongs with
> 'if ((num_vfs> 0)&& (num_vfs<= total))'.
>
yes, forgot to move it. thanks!
>> + /* is PF driver loaded w/callback */
>> + if (!pdev->driver || !pdev->driver->sriov_configure) {
>> + pr_info("Driver doesn't support sriov configuration via sysfs \n");
>> + return -ENOSYS;
>
> Use dev_err() to set the proper severity and provide context. Not sure
> whether this is the right error code.
>
well, i used pr_info() b/c that was what I was requested to do
when I did similar cleanup in the dmar.c file recently. I
can change to dev_*() formatting.
>> + }
>> +
>> + /* if enabling vf's ... */
>> + total = pdev->sriov->total;
>> + if ((num_vfs> 0)&& (num_vfs<= total)) {
>> + if (pdev->sriov->nr_virtfn == 0) { /* if not already enabled */
>> + num_vfs_enabled = pdev->driver->sriov_configure(pdev, num_vfs);
>> + if ((num_vfs_enabled>= 0)&& (num_vfs_enabled != num_vfs))
>> + pr_warn("%s: %04x:%02x:%02x.%d: Only %d VFs enabled \n",
>> + pdev->dev.driver->name, pci_domain_nr(pdev->bus),
>> + pdev->bus->number, PCI_SLOT(pdev->devfn),
>> + PCI_FUNC(pdev->devfn), num_vfs_enabled);
>
> Use dev_warn(), don't try to replicate it.
>
ditto.
>> + return count;
>> + } else {
>> + pr_warn("VFs already enabled. Disable before re-enabling\n");
>> + return -EINVAL;
>
> It would be helpful to make an 'enable' write succeed when
> num_vfs == pdev->sriov->nr_virtfn.
>
ah, functional feedback!
yes, setting numvfs == existing value could reply success,
but I had it this way since it could show a programming error
with a mgmt tool that is trying to do the enable when it
should not, or already had, and a successful return may be interpreted
as a correctness in the mgmt app's logic. .... I could go either way.
I'll wait for other feedback on this to see which way the
group wants to have it.
>> + }
>> + }
>> +
>> + /* disable vfs */
>> + if (num_vfs == 0) {
>> + if (pdev->sriov->nr_virtfn != 0) {
>> + ret = pdev->driver->sriov_configure(pdev, 0);
>> + return ret ? ret: count;
>> + } else {
>> + pr_err("Disable failed since no VFs enabled\n");
>> + return -EINVAL;
>
> This definitely should be treated as successful, not an error.
>
Same comment as above. it could be considered successful,
or it could be considered a rtn for programming logic error
of a mgmt app. again, I could go either way. ditto above.
>> + }
>> + }
>> +
>> + pr_err("Invalid value for number of VFs to enable: %d\n", num_vfs);
>> +
>> + return -EINVAL;
>> +}
>> +#else
>> +static ssize_t sriov_totalvfs_show(struct device *dev,
>> + struct device_attribute *attr,
>> + char *buf)
>> +{ return 0; }
>> +static ssize_t sriov_numvfs_show(struct device *dev,
>> + struct device_attribute *attr,
>> + char *buf)
>> +{ return 0; }
>
> Shouldn't these print "0\n"?
>
agree.
>> +static ssize_t sriov_numvfs_store(struct device *dev,
>> + struct device_attribute *attr,
>> + const char *buf, size_t count)
>> +{ return count; }
>> +#endif /* CONFIG_PCI_IOV */
>> +
>> +static struct device_attribute sriov_totalvfs_attr = __ATTR_RO(sriov_totalvfs);
>> +static struct device_attribute sriov_numvfs_attr =
>> + __ATTR(sriov_numvfs, (S_IRUGO|S_IWUSR|S_IWGRP),
>> + sriov_numvfs_show, sriov_numvfs_store);
>
> sriov_numvfs should be read-only if !CONFIG_PCI_IOV, rather than
> ignoring writes.
>
If !CONFIG_PCI_IOV, none of these attributes are visible,
so they won't exist.
>> struct device_attribute pci_dev_attrs[] = {
>> __ATTR_RO(resource),
>> __ATTR_RO(vendor),
>> @@ -1194,7 +1303,7 @@ static ssize_t reset_store(struct device *dev,
>> {
>> struct pci_dev *pdev = to_pci_dev(dev);
>> unsigned long val;
>> - ssize_t result = strict_strtoul(buf, 0,&val);
>> + ssize_t result = kstrtoul(buf, 0,&val);
>>
>> if (result< 0)
>> return result;
>> @@ -1408,8 +1517,14 @@ static struct attribute *pci_dev_dev_attrs[] = {
>> NULL,
>> };
>>
>> +static struct attribute *sriov_dev_attrs[] = {
>> +&sriov_totalvfs_attr.attr,
>> +&sriov_numvfs_attr.attr,
>
> These are wrongly indented with spaces.
>
checkpatch!
>> + NULL,
>> +};
>> +
>> static umode_t pci_dev_attrs_are_visible(struct kobject *kobj,
>> - struct attribute *a, int n)
>> + struct attribute *a, int n)
>> {
>> struct device *dev = container_of(kobj, struct device, kobj);
>> struct pci_dev *pdev = to_pci_dev(dev);
>> @@ -1421,13 +1536,33 @@ static umode_t pci_dev_attrs_are_visible(struct kobject *kobj,
>> return a->mode;
>> }
>>
>> +static umode_t sriov_attrs_are_visible(struct kobject *kobj,
>> + struct attribute *a, int n)
>> +{
>> + struct device *dev = container_of(kobj, struct device, kobj);
>> +
>> + if ((a ==&sriov_totalvfs_attr.attr) ||
>> + (a ==&sriov_numvfs_attr.attr)) {
>> + if (!dev_is_pf(dev))
>> + return 0;
> [...]
>
> An odd thing about dev_is_pf() is that if !CONFIG_PCI_IOV then it is
> false for all devices. Which means that the dummy attribute reader
> functions will actually be dead code. I think that either dev_is_pf()
> should be fixed to be accurate when !CONFIG_PCI_IOV, or the attributes
> should be completely removed if !CONFIG_PCI_IOV.
>
> Ben.
>
Since it's all in the same file, I could wrap the following code with
CONFIG_PCI_IOV:
static const struct attribute_group *pci_dev_attr_groups[] = {
&pci_dev_attr_group,
#ifdef CONFIG_PCI_IOV
&sriov_dev_attr_group,
#endif
NULL,
};
and then the dummy attribute function could be removed.
another good cleanup, thanks!
Thanks for the thorough review of the above!
Ok, my turn:
Any feedback on having the sysfs configure call pci_sriov_[enable/disable](),
as well as do the don't-disable if VFs are assigned to guests?
next prev parent reply other threads:[~2012-10-26 15:07 UTC|newest]
Thread overview: 17+ messages / expand[flat|nested] mbox.gz Atom feed top
2012-10-25 18:38 [RFC] SRIOV device enable and disable via sysfs Donald Dutile
2012-10-25 18:38 ` [PATCH 1/8] Yinghai's patch 1 of 2 Donald Dutile
2012-10-25 18:38 ` [PATCH 2/8] Yinghai's second patch for vga attr Donald Dutile
2012-10-25 18:38 ` [PATCH 3/8] PCI: sysfs per device SRIOV control and status Donald Dutile
2012-10-25 20:17 ` Ben Hutchings
2012-10-26 15:07 ` Don Dutile [this message]
2012-10-31 17:01 ` Rose, Gregory V
2012-10-31 17:36 ` Ben Hutchings
2012-10-31 18:18 ` Don Dutile
2012-10-31 18:25 ` Rose, Gregory V
2012-10-25 18:38 ` [PATCH 4/8] sriov: provide method to reduce the number of total VFs supported Donald Dutile
2012-10-25 20:24 ` Ben Hutchings
2012-10-26 15:11 ` Don Dutile
2012-10-25 18:38 ` [PATCH 5/8] ixgbe: refactor mailbox ops init Donald Dutile
2012-10-25 18:38 ` [PATCH 6/8] ixgbe: refactor SRIOV enable and disable for sysfs interface Donald Dutile
2012-10-25 18:38 ` [PATCH 7/8] ixgbe: sysfs sriov configuration callback support Donald Dutile
2012-10-25 18:38 ` [PATCH 8/8] ixgbe: change totalvfs to match support in driver Donald Dutile
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=508AA719.2060705@redhat.com \
--to=ddutile@redhat.com \
--cc=bhelgaas@google.com \
--cc=bhutchings@solarflare.com \
--cc=davem@davemloft.net \
--cc=gregory.v.rose@intel.com \
--cc=linux-pci@vger.kernel.org \
--cc=yinghai@kernel.org \
--cc=yuvalmin@broadcom.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).