Hi Bjorn, Hi Jan,
In an earlier version[0], I sought to apply the existing jailhouse special case
for isolated PCI functions to s390. As Bjorn noted in[1] there appears to be
some potential for cleaning things up and removing duplication though.
This series attempts to do this cleanup (Patches 1 and 2) followed by enabling
isolated PCI functions for s390 (Patches 3 and 4). If need be I can of course
split the cleanup off but for now I kept it as one as that's what I have
been testing.
Testing:
- On s390 with SR-IOV and a ConnectX NIC with PF 1 but not PF 0 passed throug
i.e. the isolated function case. Also of course with just VFs and an NVMe.
- On x86_64 on a desktop system where ARI is disabled and with an SR-IOV NIC
with non-contiguous VFs as well as the usual other PCI devices.
Thanks,
Niklas
Changes v4 -> v5:
- Remove unintended whitespace change in patch 1
Changes v3 -> v4:
- Use a do {} while loop in pci_scan_slot() as it is simpler (Bjorn)
- Explicitly check "fn == 0" as it is not a pointer or bool (Bjorn)
- Keep the "!dev" check in the ARI branch of next_fn() (Bjorn)
- Moved the "fn == 0 && !dev" condition out of next_fn() into pci_scan_slot().
This allows us to keep the "!dev" case in the ARI branch and means there are
no new conditions in next_fn() making it easier to verify that its behavior
is equivalent to the existing code.
- Guard the assignment of dev->multifunction with "fn > 0"
instead of "nr > 0". This matches the existing logic more closely and works
for the jailhouse case which unconditionally sets dev->multifunction for
"fn > 0". This also means fn == 0 is the single "first iteration" test.
- Remove some unneeded whitespace in patch 2
Changes v2 -> v3:
- Removed now unused nr_devs variable (kernel test robot)
Niklas Schnelle (4):
PCI: Clean up pci_scan_slot()
PCI: Move jailhouse's isolated function handling to pci_scan_slot()
PCI: Extend isolated function probing to s390
s390/pci: allow zPCI zbus without a function zero
arch/s390/pci/pci_bus.c | 82 ++++++++++----------------------------
drivers/pci/probe.c | 64 +++++++++++++----------------
include/linux/hypervisor.h | 8 ++++
3 files changed, 55 insertions(+), 99 deletions(-)
--
2.32.0
Currently the zPCI code block PCI bus creation and probing of a zPCI
zbus unless there is a PCI function with devfn 0. This is always the
case for the PCI functions with hidden RID but may keep PCI functions
from a multi-function PCI device with RID information invisible until
the function 0 becomes visible. Worse as a PCI bus is necessary to even
present a PCI hotplug slot even that remains invisible.
With the probing of these so called isolated PCI functions enabled for
s390 in common code this restriction is no longer necessary. On network
cards with multiple ports and a PF per port this also allows using each
port on its own while still providing the physical PCI topology
information in the devfn needed to associate VFs with their parent PF.
Signed-off-by: Niklas Schnelle <[email protected]>
---
arch/s390/pci/pci_bus.c | 82 ++++++++++-------------------------------
1 file changed, 20 insertions(+), 62 deletions(-)
diff --git a/arch/s390/pci/pci_bus.c b/arch/s390/pci/pci_bus.c
index 5d77acbd1c87..6a8da1b742ae 100644
--- a/arch/s390/pci/pci_bus.c
+++ b/arch/s390/pci/pci_bus.c
@@ -145,9 +145,6 @@ int zpci_bus_scan_bus(struct zpci_bus *zbus)
struct zpci_dev *zdev;
int devfn, rc, ret = 0;
- if (!zbus->function[0])
- return 0;
-
for (devfn = 0; devfn < ZPCI_FUNCTIONS_PER_BUS; devfn++) {
zdev = zbus->function[devfn];
if (zdev && zdev->state == ZPCI_FN_STATE_CONFIGURED) {
@@ -184,26 +181,26 @@ void zpci_bus_scan_busses(void)
/* zpci_bus_create_pci_bus - Create the PCI bus associated with this zbus
* @zbus: the zbus holding the zdevices
- * @f0: function 0 of the bus
+ * @fr: PCI root function that will determine the bus's domain, and bus speeed
* @ops: the pci operations
*
- * Function zero is taken as a parameter as this is used to determine the
- * domain, multifunction property and maximum bus speed of the entire bus.
+ * The PCI function @fr determines the domain (its UID), multifunction property
+ * and maximum bus speed of the entire bus.
*
* Return: 0 on success, an error code otherwise
*/
-static int zpci_bus_create_pci_bus(struct zpci_bus *zbus, struct zpci_dev *f0, struct pci_ops *ops)
+static int zpci_bus_create_pci_bus(struct zpci_bus *zbus, struct zpci_dev *fr, struct pci_ops *ops)
{
struct pci_bus *bus;
int domain;
- domain = zpci_alloc_domain((u16)f0->uid);
+ domain = zpci_alloc_domain((u16)fr->uid);
if (domain < 0)
return domain;
zbus->domain_nr = domain;
- zbus->multifunction = f0->rid_available;
- zbus->max_bus_speed = f0->max_bus_speed;
+ zbus->multifunction = fr->rid_available;
+ zbus->max_bus_speed = fr->max_bus_speed;
/*
* Note that the zbus->resources are taken over and zbus->resources
@@ -303,47 +300,6 @@ void pcibios_bus_add_device(struct pci_dev *pdev)
}
}
-/* zpci_bus_create_hotplug_slots - Add hotplug slot(s) for device added to bus
- * @zdev: the zPCI device that was newly added
- *
- * Add the hotplug slot(s) for the newly added PCI function. Normally this is
- * simply the slot for the function itself. If however we are adding the
- * function 0 on a zbus, it might be that we already registered functions on
- * that zbus but could not create their hotplug slots yet so add those now too.
- *
- * Return: 0 on success, an error code otherwise
- */
-static int zpci_bus_create_hotplug_slots(struct zpci_dev *zdev)
-{
- struct zpci_bus *zbus = zdev->zbus;
- int devfn, rc = 0;
-
- rc = zpci_init_slot(zdev);
- if (rc)
- return rc;
- zdev->has_hp_slot = 1;
-
- if (zdev->devfn == 0 && zbus->multifunction) {
- /* Now that function 0 is there we can finally create the
- * hotplug slots for those functions with devfn != 0 that have
- * been parked in zbus->function[] waiting for us to be able to
- * create the PCI bus.
- */
- for (devfn = 1; devfn < ZPCI_FUNCTIONS_PER_BUS; devfn++) {
- zdev = zbus->function[devfn];
- if (zdev && !zdev->has_hp_slot) {
- rc = zpci_init_slot(zdev);
- if (rc)
- return rc;
- zdev->has_hp_slot = 1;
- }
- }
-
- }
-
- return rc;
-}
-
static int zpci_bus_add_device(struct zpci_bus *zbus, struct zpci_dev *zdev)
{
int rc = -EINVAL;
@@ -352,21 +308,19 @@ static int zpci_bus_add_device(struct zpci_bus *zbus, struct zpci_dev *zdev)
pr_err("devfn %04x is already assigned\n", zdev->devfn);
return rc;
}
+
zdev->zbus = zbus;
zbus->function[zdev->devfn] = zdev;
zpci_nb_devices++;
- if (zbus->bus) {
- if (zbus->multifunction && !zdev->rid_available) {
- WARN_ONCE(1, "rid_available not set for multifunction\n");
- goto error;
- }
-
- zpci_bus_create_hotplug_slots(zdev);
- } else {
- /* Hotplug slot will be created once function 0 appears */
- zbus->multifunction = 1;
+ if (zbus->multifunction && !zdev->rid_available) {
+ WARN_ONCE(1, "rid_available not set for multifunction\n");
+ goto error;
}
+ rc = zpci_init_slot(zdev);
+ if (rc)
+ goto error;
+ zdev->has_hp_slot = 1;
return 0;
@@ -400,7 +354,11 @@ int zpci_bus_device_register(struct zpci_dev *zdev, struct pci_ops *ops)
return -ENOMEM;
}
- if (zdev->devfn == 0) {
+ if (!zbus->bus) {
+ /* The UID of the first PCI function registered with a zpci_bus
+ * is used as the domain number for that bus. Currently there
+ * is exactly one zpci_bus per domain.
+ */
rc = zpci_bus_create_pci_bus(zbus, zdev, ops);
if (rc)
goto error;
--
2.32.0
Like the jailhouse hypervisor s390's PCI architecture allows passing
isolated PCI functions to an OS instance. As of now this is was not
utilized even with multi-function support as the s390 PCI code makes
sure that only virtual PCI busses including a function with devfn 0 are
presented to the PCI subsystem. A subsequent change will remove this
restriction.
Allow probing such functions by replacing the existing check for
jailhouse_paravirt() with a new hypervisor_isolated_pci_functions()
helper.
Cc: Jan Kiszka <[email protected]>
Signed-off-by: Niklas Schnelle <[email protected]>
---
drivers/pci/probe.c | 2 +-
include/linux/hypervisor.h | 8 ++++++++
2 files changed, 9 insertions(+), 1 deletion(-)
diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index 3029edc68ff7..4e27cc929095 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -2663,7 +2663,7 @@ int pci_scan_slot(struct pci_bus *bus, int devfn)
* a hypervisor which passes through individual PCI
* functions.
*/
- if (!jailhouse_paravirt())
+ if (!hypervisor_isolated_pci_functions())
break;
}
fn = next_fn(bus, dev, fn);
diff --git a/include/linux/hypervisor.h b/include/linux/hypervisor.h
index fc08b433c856..33b1c0482aac 100644
--- a/include/linux/hypervisor.h
+++ b/include/linux/hypervisor.h
@@ -32,4 +32,12 @@ static inline bool jailhouse_paravirt(void)
#endif /* !CONFIG_X86 */
+static inline bool hypervisor_isolated_pci_functions(void)
+{
+ if (IS_ENABLED(CONFIG_S390))
+ return true;
+ else
+ return jailhouse_paravirt();
+}
+
#endif /* __LINUX_HYPEVISOR_H */
--
2.32.0
While determining the next PCI function is factored out of
pci_scan_slot() into next_fn() the former still handles the first
function as a special case. This duplicates the code from the scan loop.
Furthermore the non ARI branch of next_fn() is generally hard to
understand and especially the check for multifunction devices is hidden
in the handling of NULL devices for non-contiguous multifunction. It
also signals that no further functions need to be scanned by returning
0 via wraparound and this is a valid function number.
Improve upon this by transforming the conditions in next_fn() to be
easier to understand.
By changing next_fn() to return -ENODEV instead of 0 when there is no
next function we can then handle the initial function inside the loop
and deduplicate the shared handling. This also makes it more explicit
that only function 0 must exist.
No functional change is intended.
Cc: Jan Kiszka <[email protected]>
Signed-off-by: Niklas Schnelle <[email protected]>
---
drivers/pci/probe.c | 38 +++++++++++++++++++-------------------
1 file changed, 19 insertions(+), 19 deletions(-)
diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index 17a969942d37..78aa1bccab2f 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -2579,8 +2579,7 @@ struct pci_dev *pci_scan_single_device(struct pci_bus *bus, int devfn)
}
EXPORT_SYMBOL(pci_scan_single_device);
-static unsigned int next_fn(struct pci_bus *bus, struct pci_dev *dev,
- unsigned int fn)
+static int next_fn(struct pci_bus *bus, struct pci_dev *dev, int fn)
{
int pos;
u16 cap = 0;
@@ -2588,24 +2587,26 @@ static unsigned int next_fn(struct pci_bus *bus, struct pci_dev *dev,
if (pci_ari_enabled(bus)) {
if (!dev)
- return 0;
+ return -ENODEV;
pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_ARI);
if (!pos)
- return 0;
+ return -ENODEV;
pci_read_config_word(dev, pos + PCI_ARI_CAP, &cap);
next_fn = PCI_ARI_CAP_NFN(cap);
if (next_fn <= fn)
- return 0; /* protect against malformed list */
+ return -ENODEV; /* protect against malformed list */
return next_fn;
}
+ if (fn >= 7)
+ return -ENODEV;
- /* dev may be NULL for non-contiguous multifunction devices */
- if (!dev || dev->multifunction)
- return (fn + 1) % 8;
+ /* only multifunction devices may have more functions */
+ if (dev && !dev->multifunction)
+ return -ENODEV;
- return 0;
+ return fn + 1;
}
static int only_one_child(struct pci_bus *bus)
@@ -2643,26 +2644,25 @@ static int only_one_child(struct pci_bus *bus)
*/
int pci_scan_slot(struct pci_bus *bus, int devfn)
{
- unsigned int fn, nr = 0;
struct pci_dev *dev;
+ int fn = 0, nr = 0;
if (only_one_child(bus) && (devfn > 0))
return 0; /* Already scanned the entire slot */
- dev = pci_scan_single_device(bus, devfn);
- if (!dev)
- return 0;
- if (!pci_dev_is_added(dev))
- nr++;
-
- for (fn = next_fn(bus, dev, 0); fn > 0; fn = next_fn(bus, dev, fn)) {
+ do {
dev = pci_scan_single_device(bus, devfn + fn);
if (dev) {
if (!pci_dev_is_added(dev))
nr++;
- dev->multifunction = 1;
+ if (fn > 0)
+ dev->multifunction = 1;
+ } else if (fn == 0) {
+ /* function 0 is required */
+ break;
}
- }
+ fn = next_fn(bus, dev, fn);
+ } while (fn >= 0);
/* Only one slot has PCIe device */
if (bus->self && nr)
--
2.32.0
The special case of the jailhouse hypervisor passing through individual
PCI functions handles scanning for PCI functions even if function 0 does
not exist. Currently this is done with an extra loop duplicating the one
in pci_scan_slot(). By incorporating the check for jailhouse_paravirt()
into pci_scan_slot() we can instead do this as part of the normal
slot scan. Note that with the assignment of dev->multifunction gated by
fn > 0 we set dev->multifunction unconditionally for all functions if
function 0 is missing just as in the existing jailhouse loop.
The only functional change is that we now call
pcie_aspm_init_link_state() for these functions but this already
happened if function 0 was passed through and should not be a problem.
Cc: Jan Kiszka <[email protected]>
Link: https://lore.kernel.org/linux-pci/20220408224514.GA353445@bhelgaas/
Suggested-by: Bjorn Helgaas <[email protected]>
Signed-off-by: Niklas Schnelle <[email protected]>
---
drivers/pci/probe.c | 30 ++++++++++--------------------
1 file changed, 10 insertions(+), 20 deletions(-)
diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index 78aa1bccab2f..3029edc68ff7 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -2658,8 +2658,13 @@ int pci_scan_slot(struct pci_bus *bus, int devfn)
if (fn > 0)
dev->multifunction = 1;
} else if (fn == 0) {
- /* function 0 is required */
- break;
+ /*
+ * function 0 is required unless we are running on
+ * a hypervisor which passes through individual PCI
+ * functions.
+ */
+ if (!jailhouse_paravirt())
+ break;
}
fn = next_fn(bus, dev, fn);
} while (fn >= 0);
@@ -2858,29 +2863,14 @@ static unsigned int pci_scan_child_bus_extend(struct pci_bus *bus,
{
unsigned int used_buses, normal_bridges = 0, hotplug_bridges = 0;
unsigned int start = bus->busn_res.start;
- unsigned int devfn, fn, cmax, max = start;
+ unsigned int devfn, cmax, max = start;
struct pci_dev *dev;
- int nr_devs;
dev_dbg(&bus->dev, "scanning bus\n");
/* Go find them, Rover! */
- for (devfn = 0; devfn < 256; devfn += 8) {
- nr_devs = pci_scan_slot(bus, devfn);
-
- /*
- * The Jailhouse hypervisor may pass individual functions of a
- * multi-function device to a guest without passing function 0.
- * Look for them as well.
- */
- if (jailhouse_paravirt() && nr_devs == 0) {
- for (fn = 1; fn < 8; fn++) {
- dev = pci_scan_single_device(bus, devfn + fn);
- if (dev)
- dev->multifunction = 1;
- }
- }
- }
+ for (devfn = 0; devfn < 256; devfn += 8)
+ pci_scan_slot(bus, devfn);
/* Reserve buses for SR-IOV capability */
used_buses = pci_iov_bus_range(bus);
--
2.32.0
On Fri, 2022-05-13 at 09:07 -0500, Bjorn Helgaas wrote:
> On Thu, May 12, 2022 at 04:56:42PM +0200, Niklas Schnelle wrote:
> > On Thu, 2022-05-05 at 10:38 +0200, Niklas Schnelle wrote:
> > > While determining the next PCI function is factored out of
> > > pci_scan_slot() into next_fn() the former still handles the first
> > > function as a special case. This duplicates the code from the scan loop.
> > >
> > > Furthermore the non ARI branch of next_fn() is generally hard to
> > > understand and especially the check for multifunction devices is hidden
> > > in the handling of NULL devices for non-contiguous multifunction. It
> > > also signals that no further functions need to be scanned by returning
> > > 0 via wraparound and this is a valid function number.
> > >
> > > Improve upon this by transforming the conditions in next_fn() to be
> > > easier to understand.
> > >
> > > By changing next_fn() to return -ENODEV instead of 0 when there is no
> > > next function we can then handle the initial function inside the loop
> > > and deduplicate the shared handling. This also makes it more explicit
> > > that only function 0 must exist.
> > >
> > > No functional change is intended.
> > >
> > > Cc: Jan Kiszka <[email protected]>
> > > Signed-off-by: Niklas Schnelle <[email protected]>
> > > ---
> >
> > Friendly ping :-)
>
> Thanks and sorry for the delay. I'm off today for my daughter's
> wedding reception but will get back to it next week.
No worries, have a great day and congratulations!
> Just to expose
> some of my thought process (and not to request more work from you!)
> I've been wondering whether b1bd58e448f2 ("PCI: Consolidate
> "next-function" functions") is really causing us more trouble than
> it's worth. In some ways that makes the single next-function harder
> to read. But I guess the hypervisor special case is not exactly a
> "next-function" thing -- it's a "keep scanning even if there's no fn
> 0" thing.
>
> Bjorn
Yeah I do see your point. Let's discuss next week.
On Thu, May 12, 2022 at 04:56:42PM +0200, Niklas Schnelle wrote:
> On Thu, 2022-05-05 at 10:38 +0200, Niklas Schnelle wrote:
> > While determining the next PCI function is factored out of
> > pci_scan_slot() into next_fn() the former still handles the first
> > function as a special case. This duplicates the code from the scan loop.
> >
> > Furthermore the non ARI branch of next_fn() is generally hard to
> > understand and especially the check for multifunction devices is hidden
> > in the handling of NULL devices for non-contiguous multifunction. It
> > also signals that no further functions need to be scanned by returning
> > 0 via wraparound and this is a valid function number.
> >
> > Improve upon this by transforming the conditions in next_fn() to be
> > easier to understand.
> >
> > By changing next_fn() to return -ENODEV instead of 0 when there is no
> > next function we can then handle the initial function inside the loop
> > and deduplicate the shared handling. This also makes it more explicit
> > that only function 0 must exist.
> >
> > No functional change is intended.
> >
> > Cc: Jan Kiszka <[email protected]>
> > Signed-off-by: Niklas Schnelle <[email protected]>
> > ---
>
> Friendly ping :-)
Thanks and sorry for the delay. I'm off today for my daughter's
wedding reception but will get back to it next week. Just to expose
some of my thought process (and not to request more work from you!)
I've been wondering whether b1bd58e448f2 ("PCI: Consolidate
"next-function" functions") is really causing us more trouble than
it's worth. In some ways that makes the single next-function harder
to read. But I guess the hypervisor special case is not exactly a
"next-function" thing -- it's a "keep scanning even if there's no fn
0" thing.
Bjorn
On Thu, 2022-05-05 at 10:38 +0200, Niklas Schnelle wrote:
> While determining the next PCI function is factored out of
> pci_scan_slot() into next_fn() the former still handles the first
> function as a special case. This duplicates the code from the scan loop.
>
> Furthermore the non ARI branch of next_fn() is generally hard to
> understand and especially the check for multifunction devices is hidden
> in the handling of NULL devices for non-contiguous multifunction. It
> also signals that no further functions need to be scanned by returning
> 0 via wraparound and this is a valid function number.
>
> Improve upon this by transforming the conditions in next_fn() to be
> easier to understand.
>
> By changing next_fn() to return -ENODEV instead of 0 when there is no
> next function we can then handle the initial function inside the loop
> and deduplicate the shared handling. This also makes it more explicit
> that only function 0 must exist.
>
> No functional change is intended.
>
> Cc: Jan Kiszka <[email protected]>
> Signed-off-by: Niklas Schnelle <[email protected]>
> ---
Friendly ping :-)
On Fri, 2022-05-13 at 09:07 -0500, Bjorn Helgaas wrote:
> On Thu, May 12, 2022 at 04:56:42PM +0200, Niklas Schnelle wrote:
> > On Thu, 2022-05-05 at 10:38 +0200, Niklas Schnelle wrote:
> > > While determining the next PCI function is factored out of
> > > pci_scan_slot() into next_fn() the former still handles the first
> > > function as a special case. This duplicates the code from the scan loop.
> > >
> > > Furthermore the non ARI branch of next_fn() is generally hard to
> > > understand and especially the check for multifunction devices is hidden
> > > in the handling of NULL devices for non-contiguous multifunction. It
> > > also signals that no further functions need to be scanned by returning
> > > 0 via wraparound and this is a valid function number.
> > >
> > > Improve upon this by transforming the conditions in next_fn() to be
> > > easier to understand.
> > >
> > > By changing next_fn() to return -ENODEV instead of 0 when there is no
> > > next function we can then handle the initial function inside the loop
> > > and deduplicate the shared handling. This also makes it more explicit
> > > that only function 0 must exist.
> > >
> > > No functional change is intended.
> > >
> > > Cc: Jan Kiszka <[email protected]>
> > > Signed-off-by: Niklas Schnelle <[email protected]>
> > > ---
> >
> > Friendly ping :-)
>
> Thanks and sorry for the delay. I'm off today for my daughter's
> wedding reception but will get back to it next week. Just to expose
> some of my thought process (and not to request more work from you!)
> I've been wondering whether b1bd58e448f2 ("PCI: Consolidate
> "next-function" functions") is really causing us more trouble than
> it's worth. In some ways that makes the single next-function harder
> to read. But I guess the hypervisor special case is not exactly a
> "next-function" thing -- it's a "keep scanning even if there's no fn
> 0" thing.
>
> Bjorn
I've thought again about your comment. Personally what I like about
b1bd58e448f2 ("PCI: Consolidate "next-function" functions") is that it got rid of the next_fn function pointer complication. I agree though that on the other hand it removed a nice separation between the ARI and traditional cases. So I'm thinking maybe we should bring that part back. I think my patch as is makes it easier to see the equivalence to the existing code but then we could add a patch on top and turn it into the below, it's a bit more verbose but very easy to follow.
static int next_ari_fn(struct pci_bus *bus, struct pci_dev *dev, int fn)
{
…
}
static int next_trad_fn(struct pci_bus *bus, struct pci_dev *dev, int fn)
{
if (fn >= 7)
return -ENODEV;
/* only multifunction devices may have more functions */
if (dev && !dev->multifunction)
return -ENODEV;
return fn + 1;
}
static int next_fn(struct pci_bus *bus, struct pci_dev *dev, int fn)
{
if (pci_ari_enabled(bus)) {
return next_ari_fn(bus, dev, fn);
}
return next_trad_fn(bus, dev, fn);
}