| Message ID | 20180914161404.4685-2-s.miroshnichenko@yadro.com (mailing list archive) |
|---|---|
| State | New, archived |
| Delegated to: | Bjorn Helgaas |
| Headers | show |
| Series | PCI: Allow BAR movement during hotplug | expand |
Hi Sergey, On Fri, Sep 14, 2018 at 07:14:01PM +0300, Sergey Miroshnichenko wrote: > Introduce a new command line option "pci=pcie_movable_bars" that indicates > support of PCIe hotplug without prior reservation of memory regions by > BIOS/bootloader. > > If a new PCIe device has been hot-plugged between two active ones, which > have no (or not big enough) gap between their BARs, allocating new BARs > requires to move BARs of the following working devices: > > 1) dev 4 > | > v > .. | dev 3 | dev 3 | dev 5 | dev 7 | > .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > > 2) dev 4 > | > v > .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | > .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | > > 3) > > .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | > .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > > Not only BARs, but also bridge windows can be updated during a PCIe rescan, > threatening all memory transactions during this procedure, so the PCI > subsystem will instruct the drivers to pause by calling the reset_prepare() > and reset_done() callbacks. > > If a device may be affected by BAR movement, the BAR changes tracking must > be implemented in its driver. > > Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> > --- > .../admin-guide/kernel-parameters.txt | 6 +++ > drivers/pci/pci.c | 2 + > drivers/pci/probe.c | 43 +++++++++++++++++++ > include/linux/pci.h | 1 + > 4 files changed, 52 insertions(+) > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index 64a3bf54b974..f8132a709061 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -3311,6 +3311,12 @@ > bridges without forcing it upstream. Note: > this removes isolation between devices and > may put more devices in an IOMMU group. > + pcie_movable_bars Arrange a space at runtime for BARs of > + hotplugged PCIe devices - usable if bootloader > + doesn't reserve memory regions for them. Freeing > + a space may require moving BARs of active devices > + to higher addresses, so device drivers will be > + paused during rescan. > > pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power > Management. > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > index 1835f3a7aa8d..5f07a59b5924 100644 > --- a/drivers/pci/pci.c > +++ b/drivers/pci/pci.c > @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) > pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); > } else if (!strncmp(str, "disable_acs_redir=", 18)) { > disable_acs_redir_param = str + 18; > + } else if (!strncmp(str, "pcie_movable_bars", 17)) { > + pci_add_flags(PCI_MOVABLE_BARS); > } else { > printk(KERN_ERR "PCI: Unknown option `%s'\n", > str); > diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c > index 201f9e5ff55c..bdaafc48dc4c 100644 > --- a/drivers/pci/probe.c > +++ b/drivers/pci/probe.c > @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) > return max; > } > > +/* > + * Put all devices of the bus and its children to reset > + */ > +static void pci_bus_reset_prepare(struct pci_bus *bus) > +{ > + struct pci_dev *dev; > + > + list_for_each_entry(dev, &bus->devices, bus_list) { > + struct pci_bus *child = dev->subordinate; > + > + if (child) { > + pci_bus_reset_prepare(child); > + } else if (dev->driver && > + dev->driver->err_handler && > + dev->driver->err_handler->reset_prepare) { > + dev->driver->err_handler->reset_prepare(dev); > + } What about devices with drivers that don't have reset_prepare()? It looks like it will just reconfigure them anyway. Is that right? > + } > +} > + > +/* > + * Complete the reset of all devices for the bus and its children > + */ > +static void pci_bus_reset_done(struct pci_bus *bus) > +{ > + struct pci_dev *dev; > + > + list_for_each_entry(dev, &bus->devices, bus_list) { > + struct pci_bus *child = dev->subordinate; > + > + if (child) { > + pci_bus_reset_done(child); > + } else if (dev->driver && dev->driver->err_handler && > + dev->driver->err_handler->reset_done) { > + dev->driver->err_handler->reset_done(dev); > + } > + } > +} > + > /** > * pci_rescan_bus - Scan a PCI bus for devices > * @bus: PCI bus to scan > @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) > { > unsigned int max; > > + if (pci_has_flag(PCI_MOVABLE_BARS)) > + pci_bus_reset_prepare(bus); > max = pci_scan_child_bus(bus); > pci_assign_unassigned_bus_resources(bus); > + if (pci_has_flag(PCI_MOVABLE_BARS)) > + pci_bus_reset_done(bus); > pci_bus_add_devices(bus); > > return max; > diff --git a/include/linux/pci.h b/include/linux/pci.h > index 6925828f9f25..a8cb1a367c34 100644 > --- a/include/linux/pci.h > +++ b/include/linux/pci.h > @@ -847,6 +847,7 @@ enum { > PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ > PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ > PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ > + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ > }; > > /* These external functions are only available when PCI support is enabled */ > -- > 2.17.1 >
On Fri, Sep 14, 2018 at 9:21 AM Sergey Miroshnichenko <s.miroshnichenko@yadro.com> wrote: > > Introduce a new command line option "pci=pcie_movable_bars" that indicates > support of PCIe hotplug without prior reservation of memory regions by > BIOS/bootloader. > > If a new PCIe device has been hot-plugged between two active ones, which > have no (or not big enough) gap between their BARs, allocating new BARs > requires to move BARs of the following working devices: I think this is a great problem to solve. I have some questions (not objections, but trying to understand what are the limitations of the solution being proposed): * What about hot-pluggable root ports? Would this help (I'm guessing not)? * What about situations where the root port itself does not have enough resources for the new device being inserted. I'm guessing we are not going to expand root port allocation in those cases. But do we fail gracefully rejecting the hotplug by not assigning it resources, or do we manage to screw up the already working healthy devices (while attempting to move them)? * What about non-memory resources? E.g. cards may have pci bridges or switches on them, and they may need extra PCI bus numbers. Does this help that use case? > > 1) dev 4 > | > v > .. | dev 3 | dev 3 | dev 5 | dev 7 | > .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > > 2) dev 4 > | > v > .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | > .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | > > 3) > > .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | > .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > > Not only BARs, but also bridge windows can be updated during a PCIe rescan, > threatening all memory transactions during this procedure, so the PCI > subsystem will instruct the drivers to pause by calling the reset_prepare() > and reset_done() callbacks. > > If a device may be affected by BAR movement, the BAR changes tracking must > be implemented in its driver. This is quite a big thing right? I expect that this will break a lot of drivers because (a) they may not have reset_prepare() and reset_done() callbacks (I grepped in the sources and seems only 4 support it?). (b) Now, it is expected that in the reset_done() the drivers should re-read the BAR and remap any memory resources that may have changed. Note this may also include cleaning up any existing resource mappings that they had before they were paused. Not sure if this was the semantics of reset_done() already, or may be it was, I'm just not sure. Thanks, Rajat > > Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> > --- > .../admin-guide/kernel-parameters.txt | 6 +++ > drivers/pci/pci.c | 2 + > drivers/pci/probe.c | 43 +++++++++++++++++++ > include/linux/pci.h | 1 + > 4 files changed, 52 insertions(+) > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index 64a3bf54b974..f8132a709061 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -3311,6 +3311,12 @@ > bridges without forcing it upstream. Note: > this removes isolation between devices and > may put more devices in an IOMMU group. > + pcie_movable_bars Arrange a space at runtime for BARs of > + hotplugged PCIe devices - usable if bootloader > + doesn't reserve memory regions for them. Freeing > + a space may require moving BARs of active devices > + to higher addresses, so device drivers will be > + paused during rescan. > > pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power > Management. > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > index 1835f3a7aa8d..5f07a59b5924 100644 > --- a/drivers/pci/pci.c > +++ b/drivers/pci/pci.c > @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) > pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); > } else if (!strncmp(str, "disable_acs_redir=", 18)) { > disable_acs_redir_param = str + 18; > + } else if (!strncmp(str, "pcie_movable_bars", 17)) { > + pci_add_flags(PCI_MOVABLE_BARS); > } else { > printk(KERN_ERR "PCI: Unknown option `%s'\n", > str); > diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c > index 201f9e5ff55c..bdaafc48dc4c 100644 > --- a/drivers/pci/probe.c > +++ b/drivers/pci/probe.c > @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) > return max; > } > > +/* > + * Put all devices of the bus and its children to reset > + */ > +static void pci_bus_reset_prepare(struct pci_bus *bus) > +{ > + struct pci_dev *dev; > + > + list_for_each_entry(dev, &bus->devices, bus_list) { > + struct pci_bus *child = dev->subordinate; > + > + if (child) { > + pci_bus_reset_prepare(child); > + } else if (dev->driver && > + dev->driver->err_handler && > + dev->driver->err_handler->reset_prepare) { > + dev->driver->err_handler->reset_prepare(dev); > + } > + } > +} > + > +/* > + * Complete the reset of all devices for the bus and its children > + */ > +static void pci_bus_reset_done(struct pci_bus *bus) > +{ > + struct pci_dev *dev; > + > + list_for_each_entry(dev, &bus->devices, bus_list) { > + struct pci_bus *child = dev->subordinate; > + > + if (child) { > + pci_bus_reset_done(child); > + } else if (dev->driver && dev->driver->err_handler && > + dev->driver->err_handler->reset_done) { > + dev->driver->err_handler->reset_done(dev); > + } > + } > +} > + > /** > * pci_rescan_bus - Scan a PCI bus for devices > * @bus: PCI bus to scan > @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) > { > unsigned int max; > > + if (pci_has_flag(PCI_MOVABLE_BARS)) > + pci_bus_reset_prepare(bus); > max = pci_scan_child_bus(bus); > pci_assign_unassigned_bus_resources(bus); > + if (pci_has_flag(PCI_MOVABLE_BARS)) > + pci_bus_reset_done(bus); > pci_bus_add_devices(bus); > > return max; > diff --git a/include/linux/pci.h b/include/linux/pci.h > index 6925828f9f25..a8cb1a367c34 100644 > --- a/include/linux/pci.h > +++ b/include/linux/pci.h > @@ -847,6 +847,7 @@ enum { > PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ > PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ > PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ > + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ > }; > > /* These external functions are only available when PCI support is enabled */ > -- > 2.17.1 >
On Mon, Sep 17, 2018 at 12:00:22PM -0700, Rajat Jain wrote: > On Fri, Sep 14, 2018 at 9:21 AM Sergey Miroshnichenko <s.miroshnichenko@yadro.com> wrote: > > If a new PCIe device has been hot-plugged between two active ones, which > > have no (or not big enough) gap between their BARs, allocating new BARs > > requires to move BARs of the following working devices: > > * What about non-memory resources? E.g. cards may have pci bridges or > switches on them, and they may need extra PCI bus numbers. Does this > help that use case? FWIW, macOS has had a "PCI Pause" functionality since 2013, documented here: (the anchor is apparently overridden by Javascript, scroll down to "Supporting PCIe Pause") https://developer.apple.com/library/archive/documentation/HardwareDrivers/Conceptual/ThunderboltDevGuide/Basics02/Basics02.html#//apple_ref/doc/uid/TP40011138-CH4-SW14 In addition to memory resources, they also reallocate bus numbers and MSIs. Thanks, Lukas
On Mon, Sep 17, 2018 at 12:38 PM Lukas Wunner <lukas@wunner.de> wrote: > > On Mon, Sep 17, 2018 at 12:00:22PM -0700, Rajat Jain wrote: > > On Fri, Sep 14, 2018 at 9:21 AM Sergey Miroshnichenko <s.miroshnichenko@yadro.com> wrote: > > > If a new PCIe device has been hot-plugged between two active ones, which > > > have no (or not big enough) gap between their BARs, allocating new BARs > > > requires to move BARs of the following working devices: > > > > * What about non-memory resources? E.g. cards may have pci bridges or > > switches on them, and they may need extra PCI bus numbers. Does this > > help that use case? > > FWIW, macOS has had a "PCI Pause" functionality since 2013, documented here: > (the anchor is apparently overridden by Javascript, scroll down to > "Supporting PCIe Pause") > > https://developer.apple.com/library/archive/documentation/HardwareDrivers/Conceptual/ThunderboltDevGuide/Basics02/Basics02.html#//apple_ref/doc/uid/TP40011138-CH4-SW14 > > In addition to memory resources, they also reallocate bus numbers and MSIs. Thanks, that makes me wonder that would also mean that now the user space / UDEV may need to be notified (because some one may have taken a note of the PCI B:D:F). Also, the UIO drivers, ugh.... Thanks, Rajat > > Thanks, > > Lukas
Hello Sam, On 9/17/18 8:28 AM, Sam Bobroff wrote: > Hi Sergey, > > On Fri, Sep 14, 2018 at 07:14:01PM +0300, Sergey Miroshnichenko wrote: >> Introduce a new command line option "pci=pcie_movable_bars" that indicates >> support of PCIe hotplug without prior reservation of memory regions by >> BIOS/bootloader. >> >> If a new PCIe device has been hot-plugged between two active ones, which >> have no (or not big enough) gap between their BARs, allocating new BARs >> requires to move BARs of the following working devices: >> >> 1) dev 4 >> | >> v >> .. | dev 3 | dev 3 | dev 5 | dev 7 | >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >> >> 2) dev 4 >> | >> v >> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | >> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | >> >> 3) >> >> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >> >> Not only BARs, but also bridge windows can be updated during a PCIe rescan, >> threatening all memory transactions during this procedure, so the PCI >> subsystem will instruct the drivers to pause by calling the reset_prepare() >> and reset_done() callbacks. >> >> If a device may be affected by BAR movement, the BAR changes tracking must >> be implemented in its driver. >> >> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> >> --- >> .../admin-guide/kernel-parameters.txt | 6 +++ >> drivers/pci/pci.c | 2 + >> drivers/pci/probe.c | 43 +++++++++++++++++++ >> include/linux/pci.h | 1 + >> 4 files changed, 52 insertions(+) >> >> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt >> index 64a3bf54b974..f8132a709061 100644 >> --- a/Documentation/admin-guide/kernel-parameters.txt >> +++ b/Documentation/admin-guide/kernel-parameters.txt >> @@ -3311,6 +3311,12 @@ >> bridges without forcing it upstream. Note: >> this removes isolation between devices and >> may put more devices in an IOMMU group. >> + pcie_movable_bars Arrange a space at runtime for BARs of >> + hotplugged PCIe devices - usable if bootloader >> + doesn't reserve memory regions for them. Freeing >> + a space may require moving BARs of active devices >> + to higher addresses, so device drivers will be >> + paused during rescan. >> >> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power >> Management. >> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c >> index 1835f3a7aa8d..5f07a59b5924 100644 >> --- a/drivers/pci/pci.c >> +++ b/drivers/pci/pci.c >> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) >> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); >> } else if (!strncmp(str, "disable_acs_redir=", 18)) { >> disable_acs_redir_param = str + 18; >> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { >> + pci_add_flags(PCI_MOVABLE_BARS); >> } else { >> printk(KERN_ERR "PCI: Unknown option `%s'\n", >> str); >> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c >> index 201f9e5ff55c..bdaafc48dc4c 100644 >> --- a/drivers/pci/probe.c >> +++ b/drivers/pci/probe.c >> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) >> return max; >> } >> >> +/* >> + * Put all devices of the bus and its children to reset >> + */ >> +static void pci_bus_reset_prepare(struct pci_bus *bus) >> +{ >> + struct pci_dev *dev; >> + >> + list_for_each_entry(dev, &bus->devices, bus_list) { >> + struct pci_bus *child = dev->subordinate; >> + >> + if (child) { >> + pci_bus_reset_prepare(child); >> + } else if (dev->driver && >> + dev->driver->err_handler && >> + dev->driver->err_handler->reset_prepare) { >> + dev->driver->err_handler->reset_prepare(dev); >> + } > > What about devices with drivers that don't have reset_prepare()? It > looks like it will just reconfigure them anyway. Is that right? > It is possible that unprepared driver without these hooks will get BARs moved, I should put a warning message there. There three ways we can see to make this safe: - add the reset_prepare()/reset_done() hooks to *every* PCIe driver; - refuse BAR movement if at least one unprepared driver has been encountered during rescan; - reduce the number of drivers which can be affected to some controllable value and prepare them on demand. Applying the second proposal as a major restriction seems fairly reasonable, but for our particular setups and use-cases it is probably too stiff: - we've noticed that devices connected directly to the root bridge don't get moved BARs, and this covers our x86_64 servers: we only insert/remove devices into "second-level" and "lower" bridges there, but not root; - on PowerNV we have system devices (network interfaces, USB hub, etc.) grouped into dedicated domain, with all other domains ready for hotplug, and only these domains can be rescanned. With our scenarios currently reduced to these two, we can live with just a few drivers "prepared" for now: NVME and few Ethernet adapters, this gives us a possibility to use this feature before "converting" *all* the drivers, and even have the NVidia cards running on a closed proprietary driver. Should we make this behavior adjustable with something like "pcie_movable_bars=safe" and "pcie_movable_bars=always" ? Thanks! Best regards, Serge >> + } >> +} >> + >> +/* >> + * Complete the reset of all devices for the bus and its children >> + */ >> +static void pci_bus_reset_done(struct pci_bus *bus) >> +{ >> + struct pci_dev *dev; >> + >> + list_for_each_entry(dev, &bus->devices, bus_list) { >> + struct pci_bus *child = dev->subordinate; >> + >> + if (child) { >> + pci_bus_reset_done(child); >> + } else if (dev->driver && dev->driver->err_handler && >> + dev->driver->err_handler->reset_done) { >> + dev->driver->err_handler->reset_done(dev); >> + } >> + } >> +} >> + >> /** >> * pci_rescan_bus - Scan a PCI bus for devices >> * @bus: PCI bus to scan >> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) >> { >> unsigned int max; >> >> + if (pci_has_flag(PCI_MOVABLE_BARS)) >> + pci_bus_reset_prepare(bus); >> max = pci_scan_child_bus(bus); >> pci_assign_unassigned_bus_resources(bus); >> + if (pci_has_flag(PCI_MOVABLE_BARS)) >> + pci_bus_reset_done(bus); >> pci_bus_add_devices(bus); >> >> return max; >> diff --git a/include/linux/pci.h b/include/linux/pci.h >> index 6925828f9f25..a8cb1a367c34 100644 >> --- a/include/linux/pci.h >> +++ b/include/linux/pci.h >> @@ -847,6 +847,7 @@ enum { >> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ >> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ >> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ >> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ >> }; >> >> /* These external functions are only available when PCI support is enabled */ >> -- >> 2.17.1 >>
Hello Rajat, On 9/17/18 10:00 PM, Rajat Jain wrote: > On Fri, Sep 14, 2018 at 9:21 AM Sergey Miroshnichenko > <s.miroshnichenko@yadro.com> wrote: >> >> Introduce a new command line option "pci=pcie_movable_bars" that indicates >> support of PCIe hotplug without prior reservation of memory regions by >> BIOS/bootloader. >> >> If a new PCIe device has been hot-plugged between two active ones, which >> have no (or not big enough) gap between their BARs, allocating new BARs >> requires to move BARs of the following working devices: > > I think this is a great problem to solve. I have some questions (not > objections, but trying to understand what are the limitations of the > solution being proposed): > > * What about hot-pluggable root ports? Would this help (I'm guessing not)? > We don't have a hardware that is able to handle a hotplug directly into a root port (but we'll find) - there are intermediate bridges always, so I can't provide proofs right now. But after a thought experiment I can't see any obstacles, so just need to verify. > * What about situations where the root port itself does not have > enough resources for the new device being inserted. I'm guessing we > are not going to expand root port allocation in those cases. But do we > fail gracefully rejecting the hotplug by not assigning it resources, > or do we manage to screw up the already working healthy devices (while > attempting to move them)? > Thanks for a great point! The code currently is too simplistic to handle this case properly, and though it never happened to us yet, but I'm sure it can kick off a working device to use its resources for a hotplugged one. The struct resource doesn't have a "priority" field, so probably I need to check the status of the resource allocation procedure and retry it without some of new hotplugged devices in case of fail. > * What about non-memory resources? E.g. cards may have pci bridges or > switches on them, and they may need extra PCI bus numbers. Does this > help that use case? Actually, this patchset is a part of our bigger work on hotplugging bridges full of bridges full of NVME and GPU devices. And in case of success here I was planning to submit the last part (hotplugging bridges), which includes the movement of bus numbers - similar to BARs, but more awkward, with PCIe devices renaming and re-building the /sys/bus/pci and /proc/pci, otherwise the lspci tool gets mad. By device renaming I mean the "/dev/nvme0n1p1" and "eth0" will remain the same, but underlying "nvme 0022:04:00.0" and "tg3 0001:06:00.1" may change. In this regard I don't know which UDEV action is the best here - it's neither "add", "remove" nor "change". Also I've just realized that should take a closer look and check if MSIs are also needs to be reallocated. >> >> 1) dev 4 >> | >> v >> .. | dev 3 | dev 3 | dev 5 | dev 7 | >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >> >> 2) dev 4 >> | >> v >> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | >> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | >> >> 3) >> >> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >> >> Not only BARs, but also bridge windows can be updated during a PCIe rescan, >> threatening all memory transactions during this procedure, so the PCI >> subsystem will instruct the drivers to pause by calling the reset_prepare() >> and reset_done() callbacks. >> >> If a device may be affected by BAR movement, the BAR changes tracking must >> be implemented in its driver. > > This is quite a big thing right? I expect that this will break a lot > of drivers because > (a) they may not have reset_prepare() and reset_done() callbacks (I > grepped in the sources and seems only 4 support it?). > (b) Now, it is expected that in the reset_done() the drivers should > re-read the BAR and remap any memory resources that may have changed. > Note this may also include cleaning up any existing resource mappings > that they had before they were paused. Not sure if this was the > semantics of reset_done() already, or may be it was, I'm just not > sure. > Yes, this is the most painful part: all the drivers that may be affected with BARs getting moved, must be prepared in this way. Luckily, these operations have distinct patterns, just like power management callbacks like runtime_suspend()/runtime_resume() and friends. Just a few minutes earlier in this thread I've wrote to Sam why we think this approach still can be used even with a minimal subset of drivers being prepared: for now we've limited the configurations of hotplug usage to non-root bridges and to domains with similar devices. If this concept looks viable to the community, we'll be happy to share the patches for the nvme, ixgbe and tg3 drivers (and a bit later - for xhci). Right now these are the only drivers _we_ need on our ppc64 (PowerNV) and x86_64 servers. > Thanks, > > Rajat > Thanks for the review! Best regards, Serge >> >> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> >> --- >> .../admin-guide/kernel-parameters.txt | 6 +++ >> drivers/pci/pci.c | 2 + >> drivers/pci/probe.c | 43 +++++++++++++++++++ >> include/linux/pci.h | 1 + >> 4 files changed, 52 insertions(+) >> >> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt >> index 64a3bf54b974..f8132a709061 100644 >> --- a/Documentation/admin-guide/kernel-parameters.txt >> +++ b/Documentation/admin-guide/kernel-parameters.txt >> @@ -3311,6 +3311,12 @@ >> bridges without forcing it upstream. Note: >> this removes isolation between devices and >> may put more devices in an IOMMU group. >> + pcie_movable_bars Arrange a space at runtime for BARs of >> + hotplugged PCIe devices - usable if bootloader >> + doesn't reserve memory regions for them. Freeing >> + a space may require moving BARs of active devices >> + to higher addresses, so device drivers will be >> + paused during rescan. >> >> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power >> Management. >> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c >> index 1835f3a7aa8d..5f07a59b5924 100644 >> --- a/drivers/pci/pci.c >> +++ b/drivers/pci/pci.c >> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) >> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); >> } else if (!strncmp(str, "disable_acs_redir=", 18)) { >> disable_acs_redir_param = str + 18; >> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { >> + pci_add_flags(PCI_MOVABLE_BARS); >> } else { >> printk(KERN_ERR "PCI: Unknown option `%s'\n", >> str); >> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c >> index 201f9e5ff55c..bdaafc48dc4c 100644 >> --- a/drivers/pci/probe.c >> +++ b/drivers/pci/probe.c >> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) >> return max; >> } >> >> +/* >> + * Put all devices of the bus and its children to reset >> + */ >> +static void pci_bus_reset_prepare(struct pci_bus *bus) >> +{ >> + struct pci_dev *dev; >> + >> + list_for_each_entry(dev, &bus->devices, bus_list) { >> + struct pci_bus *child = dev->subordinate; >> + >> + if (child) { >> + pci_bus_reset_prepare(child); >> + } else if (dev->driver && >> + dev->driver->err_handler && >> + dev->driver->err_handler->reset_prepare) { >> + dev->driver->err_handler->reset_prepare(dev); >> + } >> + } >> +} >> + >> +/* >> + * Complete the reset of all devices for the bus and its children >> + */ >> +static void pci_bus_reset_done(struct pci_bus *bus) >> +{ >> + struct pci_dev *dev; >> + >> + list_for_each_entry(dev, &bus->devices, bus_list) { >> + struct pci_bus *child = dev->subordinate; >> + >> + if (child) { >> + pci_bus_reset_done(child); >> + } else if (dev->driver && dev->driver->err_handler && >> + dev->driver->err_handler->reset_done) { >> + dev->driver->err_handler->reset_done(dev); >> + } >> + } >> +} >> + >> /** >> * pci_rescan_bus - Scan a PCI bus for devices >> * @bus: PCI bus to scan >> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) >> { >> unsigned int max; >> >> + if (pci_has_flag(PCI_MOVABLE_BARS)) >> + pci_bus_reset_prepare(bus); >> max = pci_scan_child_bus(bus); >> pci_assign_unassigned_bus_resources(bus); >> + if (pci_has_flag(PCI_MOVABLE_BARS)) >> + pci_bus_reset_done(bus); >> pci_bus_add_devices(bus); >> >> return max; >> diff --git a/include/linux/pci.h b/include/linux/pci.h >> index 6925828f9f25..a8cb1a367c34 100644 >> --- a/include/linux/pci.h >> +++ b/include/linux/pci.h >> @@ -847,6 +847,7 @@ enum { >> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ >> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ >> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ >> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ >> }; >> >> /* These external functions are only available when PCI support is enabled */ >> -- >> 2.17.1 >>
[+cc Russell, Ben, Oliver, linuxppc-dev] On Mon, Sep 17, 2018 at 11:55:43PM +0300, Sergey Miroshnichenko wrote: > Hello Sam, > > On 9/17/18 8:28 AM, Sam Bobroff wrote: > > Hi Sergey, > > > > On Fri, Sep 14, 2018 at 07:14:01PM +0300, Sergey Miroshnichenko wrote: > >> Introduce a new command line option "pci=pcie_movable_bars" that indicates > >> support of PCIe hotplug without prior reservation of memory regions by > >> BIOS/bootloader. > >> > >> If a new PCIe device has been hot-plugged between two active ones, which > >> have no (or not big enough) gap between their BARs, allocating new BARs > >> requires to move BARs of the following working devices: > >> > >> 1) dev 4 > >> | > >> v > >> .. | dev 3 | dev 3 | dev 5 | dev 7 | > >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > >> > >> 2) dev 4 > >> | > >> v > >> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | > >> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | > >> > >> 3) > >> > >> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | > >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > >> > >> Not only BARs, but also bridge windows can be updated during a PCIe rescan, > >> threatening all memory transactions during this procedure, so the PCI > >> subsystem will instruct the drivers to pause by calling the reset_prepare() > >> and reset_done() callbacks. > >> > >> If a device may be affected by BAR movement, the BAR changes tracking must > >> be implemented in its driver. > >> > >> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> > >> --- > >> .../admin-guide/kernel-parameters.txt | 6 +++ > >> drivers/pci/pci.c | 2 + > >> drivers/pci/probe.c | 43 +++++++++++++++++++ > >> include/linux/pci.h | 1 + > >> 4 files changed, 52 insertions(+) > >> > >> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > >> index 64a3bf54b974..f8132a709061 100644 > >> --- a/Documentation/admin-guide/kernel-parameters.txt > >> +++ b/Documentation/admin-guide/kernel-parameters.txt > >> @@ -3311,6 +3311,12 @@ > >> bridges without forcing it upstream. Note: > >> this removes isolation between devices and > >> may put more devices in an IOMMU group. > >> + pcie_movable_bars Arrange a space at runtime for BARs of > >> + hotplugged PCIe devices - usable if bootloader > >> + doesn't reserve memory regions for them. Freeing > >> + a space may require moving BARs of active devices > >> + to higher addresses, so device drivers will be > >> + paused during rescan. > >> > >> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power > >> Management. > >> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > >> index 1835f3a7aa8d..5f07a59b5924 100644 > >> --- a/drivers/pci/pci.c > >> +++ b/drivers/pci/pci.c > >> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) > >> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); > >> } else if (!strncmp(str, "disable_acs_redir=", 18)) { > >> disable_acs_redir_param = str + 18; > >> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { > >> + pci_add_flags(PCI_MOVABLE_BARS); > >> } else { > >> printk(KERN_ERR "PCI: Unknown option `%s'\n", > >> str); > >> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c > >> index 201f9e5ff55c..bdaafc48dc4c 100644 > >> --- a/drivers/pci/probe.c > >> +++ b/drivers/pci/probe.c > >> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) > >> return max; > >> } > >> > >> +/* > >> + * Put all devices of the bus and its children to reset > >> + */ > >> +static void pci_bus_reset_prepare(struct pci_bus *bus) > >> +{ > >> + struct pci_dev *dev; > >> + > >> + list_for_each_entry(dev, &bus->devices, bus_list) { > >> + struct pci_bus *child = dev->subordinate; > >> + > >> + if (child) { > >> + pci_bus_reset_prepare(child); > >> + } else if (dev->driver && > >> + dev->driver->err_handler && > >> + dev->driver->err_handler->reset_prepare) { > >> + dev->driver->err_handler->reset_prepare(dev); > >> + } > > > > What about devices with drivers that don't have reset_prepare()? It > > looks like it will just reconfigure them anyway. Is that right? > > > > It is possible that unprepared driver without these hooks will get BARs > moved, I should put a warning message there. There three ways we can see > to make this safe: > - add the reset_prepare()/reset_done() hooks to *every* PCIe driver; > - refuse BAR movement if at least one unprepared driver has been > encountered during rescan; > - reduce the number of drivers which can be affected to some > controllable value and prepare them on demand. > > Applying the second proposal as a major restriction seems fairly > reasonable, but for our particular setups and use-cases it is probably > too stiff: > - we've noticed that devices connected directly to the root bridge > don't get moved BARs, and this covers our x86_64 servers: we only > insert/remove devices into "second-level" and "lower" bridges there, but > not root; > - on PowerNV we have system devices (network interfaces, USB hub, etc.) > grouped into dedicated domain, with all other domains ready for hotplug, > and only these domains can be rescanned. > > With our scenarios currently reduced to these two, we can live with just > a few drivers "prepared" for now: NVME and few Ethernet adapters, this > gives us a possibility to use this feature before "converting" *all* the > drivers, and even have the NVidia cards running on a closed proprietary > driver. > > Should we make this behavior adjustable with something like > "pcie_movable_bars=safe" and "pcie_movable_bars=always" ? I like the overall idea of this a lot. - Why do we need a command line parameter to enable this? Can't we do it unconditionally and automatically when it's possible? We could have a chicken switch to *disable* it in case this breaks something horribly, but I would like this functionality to be always available without a special option. - I'm not sure the existence of .reset_done() is evidence that a driver is prepared for its BARs to move. I don't see any documentation that refers to BAR movement, and I doubt it's been tested. But I only see 5 implementations in the tree, so it'd be easy to verify. - I think your second proposal above sounds right: we should regard any device whose driver lacks .reset_done() as immovable. We will likely be able to move some devices but not others. Implementing .reset_done() will increase flexibility but it shouldn't be a requirement for all drivers. > >> + } > >> +} > >> + > >> +/* > >> + * Complete the reset of all devices for the bus and its children > >> + */ > >> +static void pci_bus_reset_done(struct pci_bus *bus) > >> +{ > >> + struct pci_dev *dev; > >> + > >> + list_for_each_entry(dev, &bus->devices, bus_list) { > >> + struct pci_bus *child = dev->subordinate; > >> + > >> + if (child) { > >> + pci_bus_reset_done(child); > >> + } else if (dev->driver && dev->driver->err_handler && > >> + dev->driver->err_handler->reset_done) { > >> + dev->driver->err_handler->reset_done(dev); > >> + } > >> + } > >> +} > >> + > >> /** > >> * pci_rescan_bus - Scan a PCI bus for devices > >> * @bus: PCI bus to scan > >> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) > >> { > >> unsigned int max; > >> > >> + if (pci_has_flag(PCI_MOVABLE_BARS)) > >> + pci_bus_reset_prepare(bus); > >> max = pci_scan_child_bus(bus); > >> pci_assign_unassigned_bus_resources(bus); > >> + if (pci_has_flag(PCI_MOVABLE_BARS)) > >> + pci_bus_reset_done(bus); > >> pci_bus_add_devices(bus); > >> > >> return max; > >> diff --git a/include/linux/pci.h b/include/linux/pci.h > >> index 6925828f9f25..a8cb1a367c34 100644 > >> --- a/include/linux/pci.h > >> +++ b/include/linux/pci.h > >> @@ -847,6 +847,7 @@ enum { > >> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ > >> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ > >> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ > >> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ > >> }; > >> > >> /* These external functions are only available when PCI support is enabled */ > >> -- > >> 2.17.1 > >>
On Tue, Sep 18, 2018 at 6:55 AM, Sergey Miroshnichenko <s.miroshnichenko@yadro.com> wrote: > Hello Sam, > > On 9/17/18 8:28 AM, Sam Bobroff wrote: >> Hi Sergey, >> >> On Fri, Sep 14, 2018 at 07:14:01PM +0300, Sergey Miroshnichenko wrote: >>> Introduce a new command line option "pci=pcie_movable_bars" that indicates >>> support of PCIe hotplug without prior reservation of memory regions by >>> BIOS/bootloader. >>> >>> If a new PCIe device has been hot-plugged between two active ones, which >>> have no (or not big enough) gap between their BARs, allocating new BARs >>> requires to move BARs of the following working devices: >>> >>> 1) dev 4 >>> | >>> v >>> .. | dev 3 | dev 3 | dev 5 | dev 7 | >>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>> >>> 2) dev 4 >>> | >>> v >>> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | >>> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | >>> >>> 3) >>> >>> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | >>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>> >>> Not only BARs, but also bridge windows can be updated during a PCIe rescan, >>> threatening all memory transactions during this procedure, so the PCI >>> subsystem will instruct the drivers to pause by calling the reset_prepare() >>> and reset_done() callbacks. >>> >>> If a device may be affected by BAR movement, the BAR changes tracking must >>> be implemented in its driver. >>> >>> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> >>> --- >>> .../admin-guide/kernel-parameters.txt | 6 +++ >>> drivers/pci/pci.c | 2 + >>> drivers/pci/probe.c | 43 +++++++++++++++++++ >>> include/linux/pci.h | 1 + >>> 4 files changed, 52 insertions(+) >>> >>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt >>> index 64a3bf54b974..f8132a709061 100644 >>> --- a/Documentation/admin-guide/kernel-parameters.txt >>> +++ b/Documentation/admin-guide/kernel-parameters.txt >>> @@ -3311,6 +3311,12 @@ >>> bridges without forcing it upstream. Note: >>> this removes isolation between devices and >>> may put more devices in an IOMMU group. >>> + pcie_movable_bars Arrange a space at runtime for BARs of >>> + hotplugged PCIe devices - usable if bootloader >>> + doesn't reserve memory regions for them. Freeing >>> + a space may require moving BARs of active devices >>> + to higher addresses, so device drivers will be >>> + paused during rescan. >>> >>> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power >>> Management. >>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c >>> index 1835f3a7aa8d..5f07a59b5924 100644 >>> --- a/drivers/pci/pci.c >>> +++ b/drivers/pci/pci.c >>> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) >>> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); >>> } else if (!strncmp(str, "disable_acs_redir=", 18)) { >>> disable_acs_redir_param = str + 18; >>> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { >>> + pci_add_flags(PCI_MOVABLE_BARS); >>> } else { >>> printk(KERN_ERR "PCI: Unknown option `%s'\n", >>> str); >>> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c >>> index 201f9e5ff55c..bdaafc48dc4c 100644 >>> --- a/drivers/pci/probe.c >>> +++ b/drivers/pci/probe.c >>> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) >>> return max; >>> } >>> >>> +/* >>> + * Put all devices of the bus and its children to reset >>> + */ >>> +static void pci_bus_reset_prepare(struct pci_bus *bus) >>> +{ >>> + struct pci_dev *dev; >>> + >>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>> + struct pci_bus *child = dev->subordinate; >>> + >>> + if (child) { >>> + pci_bus_reset_prepare(child); >>> + } else if (dev->driver && >>> + dev->driver->err_handler && >>> + dev->driver->err_handler->reset_prepare) { >>> + dev->driver->err_handler->reset_prepare(dev); >>> + } >> >> What about devices with drivers that don't have reset_prepare()? It >> looks like it will just reconfigure them anyway. Is that right? > > It is possible that unprepared driver without these hooks will get BARs > moved, I should put a warning message there. There three ways we can see > to make this safe:> It might make more sense to allow drivers to opt-in to this rather than assuming any driver with the error handlers supports re-assigning BARs on the fly. As far as I know we restore the previous BAR values a part of the EEH recovery process when there's a driver bound, but I could be wrong. > - add the reset_prepare()/reset_done() hooks to *every* PCIe driver; > - refuse BAR movement if at least one unprepared driver has been > encountered during rescan; > - reduce the number of drivers which can be affected to some > controllable value and prepare them on demand. Is there any reason we couldn't just unbind the unaware drivers and re-bind them afterwards? That might be useful when hotplugging NVMe racks since you wouldn't want a switch management driver (or similar) to prevent re-assignments if they're required. > Applying the second proposal as a major restriction seems fairly > reasonable, but for our particular setups and use-cases it is probably > too stiff: > - we've noticed that devices connected directly to the root bridge > don't get moved BARs, and this covers our x86_64 servers: we only > insert/remove devices into "second-level" and "lower" bridges there, but > not root; > > - on PowerNV we have system devices (network interfaces, USB hub, etc.) > grouped into dedicated domain, with all other domains ready for hotplug, > and only these domains can be rescanned. Are you doing anything to enforce this or just relying on people not re-scanning the system device bus? > With our scenarios currently reduced to these two, we can live with just > a few drivers "prepared" for now: NVME and few Ethernet adapters, this > gives us a possibility to use this feature before "converting" *all* the > drivers, and even have the NVidia cards running on a closed proprietary > driver. > > Should we make this behavior adjustable with something like > "pcie_movable_bars=safe" and "pcie_movable_bars=always" ? > > Thanks! > > Best regards, > Serge > >>> + } >>> +} >>> + >>> +/* >>> + * Complete the reset of all devices for the bus and its children >>> + */ >>> +static void pci_bus_reset_done(struct pci_bus *bus) >>> +{ >>> + struct pci_dev *dev; >>> + >>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>> + struct pci_bus *child = dev->subordinate; >>> + >>> + if (child) { >>> + pci_bus_reset_done(child); >>> + } else if (dev->driver && dev->driver->err_handler && >>> + dev->driver->err_handler->reset_done) { >>> + dev->driver->err_handler->reset_done(dev); >>> + } >>> + } >>> +} >>> + >>> /** >>> * pci_rescan_bus - Scan a PCI bus for devices >>> * @bus: PCI bus to scan >>> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) >>> { >>> unsigned int max; >>> >>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>> + pci_bus_reset_prepare(bus); >>> max = pci_scan_child_bus(bus); >>> pci_assign_unassigned_bus_resources(bus); >>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>> + pci_bus_reset_done(bus); >>> pci_bus_add_devices(bus); >>> >>> return max; >>> diff --git a/include/linux/pci.h b/include/linux/pci.h >>> index 6925828f9f25..a8cb1a367c34 100644 >>> --- a/include/linux/pci.h >>> +++ b/include/linux/pci.h >>> @@ -847,6 +847,7 @@ enum { >>> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ >>> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ >>> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ >>> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ >>> }; >>> >>> /* These external functions are only available when PCI support is enabled */ >>> -- >>> 2.17.1 >>>
On 9/18/18 1:59 AM, Bjorn Helgaas wrote: > [+cc Russell, Ben, Oliver, linuxppc-dev] > > On Mon, Sep 17, 2018 at 11:55:43PM +0300, Sergey Miroshnichenko wrote: >> Hello Sam, >> >> On 9/17/18 8:28 AM, Sam Bobroff wrote: >>> Hi Sergey, >>> >>> On Fri, Sep 14, 2018 at 07:14:01PM +0300, Sergey Miroshnichenko wrote: >>>> Introduce a new command line option "pci=pcie_movable_bars" that indicates >>>> support of PCIe hotplug without prior reservation of memory regions by >>>> BIOS/bootloader. >>>> >>>> If a new PCIe device has been hot-plugged between two active ones, which >>>> have no (or not big enough) gap between their BARs, allocating new BARs >>>> requires to move BARs of the following working devices: >>>> >>>> 1) dev 4 >>>> | >>>> v >>>> .. | dev 3 | dev 3 | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>>> >>>> 2) dev 4 >>>> | >>>> v >>>> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | >>>> >>>> 3) >>>> >>>> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>>> >>>> Not only BARs, but also bridge windows can be updated during a PCIe rescan, >>>> threatening all memory transactions during this procedure, so the PCI >>>> subsystem will instruct the drivers to pause by calling the reset_prepare() >>>> and reset_done() callbacks. >>>> >>>> If a device may be affected by BAR movement, the BAR changes tracking must >>>> be implemented in its driver. >>>> >>>> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> >>>> --- >>>> .../admin-guide/kernel-parameters.txt | 6 +++ >>>> drivers/pci/pci.c | 2 + >>>> drivers/pci/probe.c | 43 +++++++++++++++++++ >>>> include/linux/pci.h | 1 + >>>> 4 files changed, 52 insertions(+) >>>> >>>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt >>>> index 64a3bf54b974..f8132a709061 100644 >>>> --- a/Documentation/admin-guide/kernel-parameters.txt >>>> +++ b/Documentation/admin-guide/kernel-parameters.txt >>>> @@ -3311,6 +3311,12 @@ >>>> bridges without forcing it upstream. Note: >>>> this removes isolation between devices and >>>> may put more devices in an IOMMU group. >>>> + pcie_movable_bars Arrange a space at runtime for BARs of >>>> + hotplugged PCIe devices - usable if bootloader >>>> + doesn't reserve memory regions for them. Freeing >>>> + a space may require moving BARs of active devices >>>> + to higher addresses, so device drivers will be >>>> + paused during rescan. >>>> >>>> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power >>>> Management. >>>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c >>>> index 1835f3a7aa8d..5f07a59b5924 100644 >>>> --- a/drivers/pci/pci.c >>>> +++ b/drivers/pci/pci.c >>>> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) >>>> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); >>>> } else if (!strncmp(str, "disable_acs_redir=", 18)) { >>>> disable_acs_redir_param = str + 18; >>>> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { >>>> + pci_add_flags(PCI_MOVABLE_BARS); >>>> } else { >>>> printk(KERN_ERR "PCI: Unknown option `%s'\n", >>>> str); >>>> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c >>>> index 201f9e5ff55c..bdaafc48dc4c 100644 >>>> --- a/drivers/pci/probe.c >>>> +++ b/drivers/pci/probe.c >>>> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) >>>> return max; >>>> } >>>> >>>> +/* >>>> + * Put all devices of the bus and its children to reset >>>> + */ >>>> +static void pci_bus_reset_prepare(struct pci_bus *bus) >>>> +{ >>>> + struct pci_dev *dev; >>>> + >>>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>>> + struct pci_bus *child = dev->subordinate; >>>> + >>>> + if (child) { >>>> + pci_bus_reset_prepare(child); >>>> + } else if (dev->driver && >>>> + dev->driver->err_handler && >>>> + dev->driver->err_handler->reset_prepare) { >>>> + dev->driver->err_handler->reset_prepare(dev); >>>> + } >>> >>> What about devices with drivers that don't have reset_prepare()? It >>> looks like it will just reconfigure them anyway. Is that right? >>> >> >> It is possible that unprepared driver without these hooks will get BARs >> moved, I should put a warning message there. There three ways we can see >> to make this safe: >> - add the reset_prepare()/reset_done() hooks to *every* PCIe driver; >> - refuse BAR movement if at least one unprepared driver has been >> encountered during rescan; >> - reduce the number of drivers which can be affected to some >> controllable value and prepare them on demand. >> >> Applying the second proposal as a major restriction seems fairly >> reasonable, but for our particular setups and use-cases it is probably >> too stiff: >> - we've noticed that devices connected directly to the root bridge >> don't get moved BARs, and this covers our x86_64 servers: we only >> insert/remove devices into "second-level" and "lower" bridges there, but >> not root; >> - on PowerNV we have system devices (network interfaces, USB hub, etc.) >> grouped into dedicated domain, with all other domains ready for hotplug, >> and only these domains can be rescanned. >> >> With our scenarios currently reduced to these two, we can live with just >> a few drivers "prepared" for now: NVME and few Ethernet adapters, this >> gives us a possibility to use this feature before "converting" *all* the >> drivers, and even have the NVidia cards running on a closed proprietary >> driver. >> >> Should we make this behavior adjustable with something like >> "pcie_movable_bars=safe" and "pcie_movable_bars=always" ? > > I like the overall idea of this a lot. > > - Why do we need a command line parameter to enable this? Can't we > do it unconditionally and automatically when it's possible? We > could have a chicken switch to *disable* it in case this breaks > something horribly, but I would like this functionality to be > always available without a special option. > After making this feature completely safe we could activate it with the existing option "pci=realloc". > - I'm not sure the existence of .reset_done() is evidence that a > driver is prepared for its BARs to move. I don't see any > documentation that refers to BAR movement, and I doubt it's been > tested. But I only see 5 implementations in the tree, so it'd be > easy to verify. > You are right, and we should clarify the description: - drivers which have the .reset_done() already - none of them are aware of movable BARs yet; - the rest of the drivers should both be able to pause and handle the changes in BARs. > - I think your second proposal above sounds right: we should regard > any device whose driver lacks .reset_done() as immovable. We will > likely be able to move some devices but not others. Implementing > .reset_done() will increase flexibility but it shouldn't be a > requirement for all drivers. > Thanks for the advice! This is more flexible and doesn't have any prerequisites. In this case the greater the "movable"/"immovable" ratio of the devices that was working before the hotplug event - the higher the probability to free some space for new BARs. But even a single "immovable" device at an undesirable place can block the re-arrangement, in this case all we can is just give up and print an error message. This patchset in its current form doesn't support marking a choosen BAR as immovable (just releasing all the resources of the root bridge and trying to sort and re-assign them back), so I'll have to implement that. Best regards, Serge >>>> + } >>>> +} >>>> + >>>> +/* >>>> + * Complete the reset of all devices for the bus and its children >>>> + */ >>>> +static void pci_bus_reset_done(struct pci_bus *bus) >>>> +{ >>>> + struct pci_dev *dev; >>>> + >>>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>>> + struct pci_bus *child = dev->subordinate; >>>> + >>>> + if (child) { >>>> + pci_bus_reset_done(child); >>>> + } else if (dev->driver && dev->driver->err_handler && >>>> + dev->driver->err_handler->reset_done) { >>>> + dev->driver->err_handler->reset_done(dev); >>>> + } >>>> + } >>>> +} >>>> + >>>> /** >>>> * pci_rescan_bus - Scan a PCI bus for devices >>>> * @bus: PCI bus to scan >>>> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) >>>> { >>>> unsigned int max; >>>> >>>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>>> + pci_bus_reset_prepare(bus); >>>> max = pci_scan_child_bus(bus); >>>> pci_assign_unassigned_bus_resources(bus); >>>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>>> + pci_bus_reset_done(bus); >>>> pci_bus_add_devices(bus); >>>> >>>> return max; >>>> diff --git a/include/linux/pci.h b/include/linux/pci.h >>>> index 6925828f9f25..a8cb1a367c34 100644 >>>> --- a/include/linux/pci.h >>>> +++ b/include/linux/pci.h >>>> @@ -847,6 +847,7 @@ enum { >>>> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ >>>> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ >>>> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ >>>> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ >>>> }; >>>> >>>> /* These external functions are only available when PCI support is enabled */ >>>> -- >>>> 2.17.1 >>>>
On 9/18/18 2:35 AM, Oliver wrote: > On Tue, Sep 18, 2018 at 6:55 AM, Sergey Miroshnichenko > <s.miroshnichenko@yadro.com> wrote: >> Hello Sam, >> >> On 9/17/18 8:28 AM, Sam Bobroff wrote: >>> Hi Sergey, >>> >>> On Fri, Sep 14, 2018 at 07:14:01PM +0300, Sergey Miroshnichenko wrote: >>>> Introduce a new command line option "pci=pcie_movable_bars" that indicates >>>> support of PCIe hotplug without prior reservation of memory regions by >>>> BIOS/bootloader. >>>> >>>> If a new PCIe device has been hot-plugged between two active ones, which >>>> have no (or not big enough) gap between their BARs, allocating new BARs >>>> requires to move BARs of the following working devices: >>>> >>>> 1) dev 4 >>>> | >>>> v >>>> .. | dev 3 | dev 3 | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>>> >>>> 2) dev 4 >>>> | >>>> v >>>> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | >>>> >>>> 3) >>>> >>>> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>>> >>>> Not only BARs, but also bridge windows can be updated during a PCIe rescan, >>>> threatening all memory transactions during this procedure, so the PCI >>>> subsystem will instruct the drivers to pause by calling the reset_prepare() >>>> and reset_done() callbacks. >>>> >>>> If a device may be affected by BAR movement, the BAR changes tracking must >>>> be implemented in its driver. >>>> >>>> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> >>>> --- >>>> .../admin-guide/kernel-parameters.txt | 6 +++ >>>> drivers/pci/pci.c | 2 + >>>> drivers/pci/probe.c | 43 +++++++++++++++++++ >>>> include/linux/pci.h | 1 + >>>> 4 files changed, 52 insertions(+) >>>> >>>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt >>>> index 64a3bf54b974..f8132a709061 100644 >>>> --- a/Documentation/admin-guide/kernel-parameters.txt >>>> +++ b/Documentation/admin-guide/kernel-parameters.txt >>>> @@ -3311,6 +3311,12 @@ >>>> bridges without forcing it upstream. Note: >>>> this removes isolation between devices and >>>> may put more devices in an IOMMU group. >>>> + pcie_movable_bars Arrange a space at runtime for BARs of >>>> + hotplugged PCIe devices - usable if bootloader >>>> + doesn't reserve memory regions for them. Freeing >>>> + a space may require moving BARs of active devices >>>> + to higher addresses, so device drivers will be >>>> + paused during rescan. >>>> >>>> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power >>>> Management. >>>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c >>>> index 1835f3a7aa8d..5f07a59b5924 100644 >>>> --- a/drivers/pci/pci.c >>>> +++ b/drivers/pci/pci.c >>>> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) >>>> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); >>>> } else if (!strncmp(str, "disable_acs_redir=", 18)) { >>>> disable_acs_redir_param = str + 18; >>>> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { >>>> + pci_add_flags(PCI_MOVABLE_BARS); >>>> } else { >>>> printk(KERN_ERR "PCI: Unknown option `%s'\n", >>>> str); >>>> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c >>>> index 201f9e5ff55c..bdaafc48dc4c 100644 >>>> --- a/drivers/pci/probe.c >>>> +++ b/drivers/pci/probe.c >>>> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) >>>> return max; >>>> } >>>> >>>> +/* >>>> + * Put all devices of the bus and its children to reset >>>> + */ >>>> +static void pci_bus_reset_prepare(struct pci_bus *bus) >>>> +{ >>>> + struct pci_dev *dev; >>>> + >>>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>>> + struct pci_bus *child = dev->subordinate; >>>> + >>>> + if (child) { >>>> + pci_bus_reset_prepare(child); >>>> + } else if (dev->driver && >>>> + dev->driver->err_handler && >>>> + dev->driver->err_handler->reset_prepare) { >>>> + dev->driver->err_handler->reset_prepare(dev); >>>> + } >>> >>> What about devices with drivers that don't have reset_prepare()? It >>> looks like it will just reconfigure them anyway. Is that right? >> >> It is possible that unprepared driver without these hooks will get BARs >> moved, I should put a warning message there. There three ways we can see >> to make this safe:> > > It might make more sense to allow drivers to opt-in to this rather > than assuming any driver with the error handlers supports re-assigning > BARs on the fly. As far as I know we restore the previous BAR values a > part of the EEH recovery process when there's a driver bound, but I > could be wrong. > If I add the new two callbacks rescan_prepare() and rescan_done() to the struct pci_driver, would they be a good indicator of driver's awareness of movable BARs? >> - add the reset_prepare()/reset_done() hooks to *every* PCIe driver; >> - refuse BAR movement if at least one unprepared driver has been >> encountered during rescan; >> - reduce the number of drivers which can be affected to some >> controllable value and prepare them on demand. > > Is there any reason we couldn't just unbind the unaware drivers and > re-bind them afterwards? That might be useful when hotplugging NVMe > racks since you wouldn't want a switch management driver (or similar) > to prevent re-assignments if they're required. > We would like the hotplug events to be as much transparent for the rest of the system as possible: it is ok if a device takes a pause for a few seconds, but it is not allowed to disappear, even if it will return soon. >> Applying the second proposal as a major restriction seems fairly >> reasonable, but for our particular setups and use-cases it is probably >> too stiff: >> - we've noticed that devices connected directly to the root bridge >> don't get moved BARs, and this covers our x86_64 servers: we only >> insert/remove devices into "second-level" and "lower" bridges there, but >> not root; >> >> - on PowerNV we have system devices (network interfaces, USB hub, etc.) >> grouped into dedicated domain, with all other domains ready for hotplug, >> and only these domains can be rescanned. > > Are you doing anything to enforce this or just relying on people not > re-scanning the system device bus? > It was the latter. But then Bjorn has suggested how we can avoid such prerequisites and assumptions: I'll change the implementation so the arrangement of a free space for BARs of new devices may fail because of memory fragmentation by "immovable" devices. In return, the whole operation of rescan becomes safe. Serge >> With our scenarios currently reduced to these two, we can live with just >> a few drivers "prepared" for now: NVME and few Ethernet adapters, this >> gives us a possibility to use this feature before "converting" *all* the >> drivers, and even have the NVidia cards running on a closed proprietary >> driver. >> >> Should we make this behavior adjustable with something like >> "pcie_movable_bars=safe" and "pcie_movable_bars=always" ? >> >> Thanks! >> >> Best regards, >> Serge >> >>>> + } >>>> +} >>>> + >>>> +/* >>>> + * Complete the reset of all devices for the bus and its children >>>> + */ >>>> +static void pci_bus_reset_done(struct pci_bus *bus) >>>> +{ >>>> + struct pci_dev *dev; >>>> + >>>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>>> + struct pci_bus *child = dev->subordinate; >>>> + >>>> + if (child) { >>>> + pci_bus_reset_done(child); >>>> + } else if (dev->driver && dev->driver->err_handler && >>>> + dev->driver->err_handler->reset_done) { >>>> + dev->driver->err_handler->reset_done(dev); >>>> + } >>>> + } >>>> +} >>>> + >>>> /** >>>> * pci_rescan_bus - Scan a PCI bus for devices >>>> * @bus: PCI bus to scan >>>> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) >>>> { >>>> unsigned int max; >>>> >>>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>>> + pci_bus_reset_prepare(bus); >>>> max = pci_scan_child_bus(bus); >>>> pci_assign_unassigned_bus_resources(bus); >>>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>>> + pci_bus_reset_done(bus); >>>> pci_bus_add_devices(bus); >>>> >>>> return max; >>>> diff --git a/include/linux/pci.h b/include/linux/pci.h >>>> index 6925828f9f25..a8cb1a367c34 100644 >>>> --- a/include/linux/pci.h >>>> +++ b/include/linux/pci.h >>>> @@ -847,6 +847,7 @@ enum { >>>> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ >>>> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ >>>> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ >>>> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ >>>> }; >>>> >>>> /* These external functions are only available when PCI support is enabled */ >>>> -- >>>> 2.17.1 >>>>
On Tue, Sep 18, 2018 at 05:01:48PM +0300, Sergey Miroshnichenko wrote: > On 9/18/18 1:59 AM, Bjorn Helgaas wrote: > > On Mon, Sep 17, 2018 at 11:55:43PM +0300, Sergey Miroshnichenko wrote: > >> On 9/17/18 8:28 AM, Sam Bobroff wrote: > >>> On Fri, Sep 14, 2018 at 07:14:01PM +0300, Sergey Miroshnichenko wrote: > >>>> Introduce a new command line option "pci=pcie_movable_bars" > >>>> that indicates support of PCIe hotplug without prior > >>>> reservation of memory regions by BIOS/bootloader. > >>> What about devices with drivers that don't have reset_prepare()? It > >>> looks like it will just reconfigure them anyway. Is that right? > >> > >> It is possible that unprepared driver without these hooks will get BARs > >> moved, I should put a warning message there. There three ways we can see > >> to make this safe: > >> - add the reset_prepare()/reset_done() hooks to *every* PCIe driver; > >> - refuse BAR movement if at least one unprepared driver has been > >> encountered during rescan; > >> - reduce the number of drivers which can be affected to some > >> controllable value and prepare them on demand. > >> > >> Applying the second proposal as a major restriction seems fairly > >> reasonable, but for our particular setups and use-cases it is probably > >> too stiff: > >> - we've noticed that devices connected directly to the root bridge > >> don't get moved BARs, and this covers our x86_64 servers: we only > >> insert/remove devices into "second-level" and "lower" bridges there, but > >> not root; > >> - on PowerNV we have system devices (network interfaces, USB hub, etc.) > >> grouped into dedicated domain, with all other domains ready for hotplug, > >> and only these domains can be rescanned. > >> > >> With our scenarios currently reduced to these two, we can live with just > >> a few drivers "prepared" for now: NVME and few Ethernet adapters, this > >> gives us a possibility to use this feature before "converting" *all* the > >> drivers, and even have the NVidia cards running on a closed proprietary > >> driver. > >> > >> Should we make this behavior adjustable with something like > >> "pcie_movable_bars=safe" and "pcie_movable_bars=always" ? > > > > I like the overall idea of this a lot. > > > > - Why do we need a command line parameter to enable this? Can't we > > do it unconditionally and automatically when it's possible? We > > could have a chicken switch to *disable* it in case this breaks > > something horribly, but I would like this functionality to be > > always available without a special option. > > After making this feature completely safe we could activate it with the > existing option "pci=realloc". That *sounds* good, but in practice it never happens that we decide a feature is completely safe and somebody makes it the default. If we're going to do this, I think we need to commit to making it work 100% of the time, with no option needed. > > - I'm not sure the existence of .reset_done() is evidence that a > > driver is prepared for its BARs to move. I don't see any > > documentation that refers to BAR movement, and I doubt it's been > > tested. But I only see 5 implementations in the tree, so it'd be > > easy to verify. > > You are right, and we should clarify the description: > - drivers which have the .reset_done() already - none of them are aware > of movable BARs yet; > - the rest of the drivers should both be able to pause and handle the > changes in BARs. This doesn't clarify it for me. If you want to update all existing .reset_done() methods so they deal with BAR changes, that would be fine with me. That would be done by preliminary patches in the series that adds the feature. > > - I think your second proposal above sounds right: we should regard > > any device whose driver lacks .reset_done() as immovable. We will > > likely be able to move some devices but not others. Implementing > > .reset_done() will increase flexibility but it shouldn't be a > > requirement for all drivers. > > Thanks for the advice! This is more flexible and doesn't have any > prerequisites. In this case the greater the "movable"/"immovable" ratio > of the devices that was working before the hotplug event - the higher > the probability to free some space for new BARs. But even a single > "immovable" device at an undesirable place can block the re-arrangement, > in this case all we can is just give up and print an error message. Right. There's nothing we can do about that except make the relevant drivers smarter. > This patchset in its current form doesn't support marking a choosen BAR > as immovable (just releasing all the resources of the root bridge and > trying to sort and re-assign them back), so I'll have to implement that. The current IORESOURCE_PCI_FIXED usage is for things that literally *cannot* be moved because there is no BAR at all (VGA or IDE legacy, enhanced allocation (see pci_ea_read()) or there's some platform quirk. It *might* make sense to also use it for devices where the *driver* isn't smart enough to deal with moving it, but I'm not sure. That would have to be done at driver probe time, I guess.
Hi Sergey, Sorry, I have one more basic question: The PCI express hierarchy is made up from PCI express switches and end points, so essentially there are only point to point links (except the switch internal logicl bus on which you cannot hotplug anything). The device being hotplugged shall be plugged onto a downstream port of a switch (or root port), so there shall be no other device on that bus (since each downstream port gets its own secondary bus number, and its own share of resources). So it seems to me, that in this case, IF the resources at the downstream port are not big enough, you will need to re-program the switch to get resources to this downstream port, by stealing from *another* downstream port. Is this correct understanding of the problem that you are trying to solve? Thanks, Rajat On Mon, Sep 17, 2018 at 2:25 PM Sergey Miroshnichenko <s.miroshnichenko@yadro.com> wrote: > > Hello Rajat, > > On 9/17/18 10:00 PM, Rajat Jain wrote: > > On Fri, Sep 14, 2018 at 9:21 AM Sergey Miroshnichenko > > <s.miroshnichenko@yadro.com> wrote: > >> > >> Introduce a new command line option "pci=pcie_movable_bars" that indicates > >> support of PCIe hotplug without prior reservation of memory regions by > >> BIOS/bootloader. > >> > >> If a new PCIe device has been hot-plugged between two active ones, which > >> have no (or not big enough) gap between their BARs, allocating new BARs > >> requires to move BARs of the following working devices: > > > > I think this is a great problem to solve. I have some questions (not > > objections, but trying to understand what are the limitations of the > > solution being proposed): > > > > * What about hot-pluggable root ports? Would this help (I'm guessing not)? > > > > We don't have a hardware that is able to handle a hotplug directly into > a root port (but we'll find) - there are intermediate bridges always, so > I can't provide proofs right now. But after a thought experiment I can't > see any obstacles, so just need to verify. > > > * What about situations where the root port itself does not have > > enough resources for the new device being inserted. I'm guessing we > > are not going to expand root port allocation in those cases. But do we > > fail gracefully rejecting the hotplug by not assigning it resources, > > or do we manage to screw up the already working healthy devices (while > > attempting to move them)? > > > > Thanks for a great point! The code currently is too simplistic to handle > this case properly, and though it never happened to us yet, but I'm sure > it can kick off a working device to use its resources for a hotplugged > one. The struct resource doesn't have a "priority" field, so probably I > need to check the status of the resource allocation procedure and retry > it without some of new hotplugged devices in case of fail. > > > * What about non-memory resources? E.g. cards may have pci bridges or > > switches on them, and they may need extra PCI bus numbers. Does this > > help that use case? > > Actually, this patchset is a part of our bigger work on hotplugging > bridges full of bridges full of NVME and GPU devices. And in case of > success here I was planning to submit the last part (hotplugging > bridges), which includes the movement of bus numbers - similar to BARs, > but more awkward, with PCIe devices renaming and re-building the > /sys/bus/pci and /proc/pci, otherwise the lspci tool gets mad. > > By device renaming I mean the "/dev/nvme0n1p1" and "eth0" will remain > the same, but underlying "nvme 0022:04:00.0" and "tg3 0001:06:00.1" may > change. In this regard I don't know which UDEV action is the best here - > it's neither "add", "remove" nor "change". > > > Also I've just realized that should take a closer look and check if MSIs > are also needs to be reallocated. > > >> > >> 1) dev 4 > >> | > >> v > >> .. | dev 3 | dev 3 | dev 5 | dev 7 | > >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > >> > >> 2) dev 4 > >> | > >> v > >> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | > >> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | > >> > >> 3) > >> > >> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | > >> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | > >> > >> Not only BARs, but also bridge windows can be updated during a PCIe rescan, > >> threatening all memory transactions during this procedure, so the PCI > >> subsystem will instruct the drivers to pause by calling the reset_prepare() > >> and reset_done() callbacks. > >> > >> If a device may be affected by BAR movement, the BAR changes tracking must > >> be implemented in its driver. > > > > This is quite a big thing right? I expect that this will break a lot > > of drivers because > > (a) they may not have reset_prepare() and reset_done() callbacks (I > > grepped in the sources and seems only 4 support it?). > > (b) Now, it is expected that in the reset_done() the drivers should > > re-read the BAR and remap any memory resources that may have changed. > > Note this may also include cleaning up any existing resource mappings > > that they had before they were paused. Not sure if this was the > > semantics of reset_done() already, or may be it was, I'm just not > > sure. > > > > Yes, this is the most painful part: all the drivers that may be affected > with BARs getting moved, must be prepared in this way. Luckily, these > operations have distinct patterns, just like power management callbacks > like runtime_suspend()/runtime_resume() and friends. > > Just a few minutes earlier in this thread I've wrote to Sam why we think > this approach still can be used even with a minimal subset of drivers > being prepared: for now we've limited the configurations of hotplug > usage to non-root bridges and to domains with similar devices. If this > concept looks viable to the community, we'll be happy to share the > patches for the nvme, ixgbe and tg3 drivers (and a bit later - for > xhci). Right now these are the only drivers _we_ need on our ppc64 > (PowerNV) and x86_64 servers. > > > Thanks, > > > > Rajat > > > > Thanks for the review! > > Best regards, > Serge > > >> > >> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> > >> --- > >> .../admin-guide/kernel-parameters.txt | 6 +++ > >> drivers/pci/pci.c | 2 + > >> drivers/pci/probe.c | 43 +++++++++++++++++++ > >> include/linux/pci.h | 1 + > >> 4 files changed, 52 insertions(+) > >> > >> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > >> index 64a3bf54b974..f8132a709061 100644 > >> --- a/Documentation/admin-guide/kernel-parameters.txt > >> +++ b/Documentation/admin-guide/kernel-parameters.txt > >> @@ -3311,6 +3311,12 @@ > >> bridges without forcing it upstream. Note: > >> this removes isolation between devices and > >> may put more devices in an IOMMU group. > >> + pcie_movable_bars Arrange a space at runtime for BARs of > >> + hotplugged PCIe devices - usable if bootloader > >> + doesn't reserve memory regions for them. Freeing > >> + a space may require moving BARs of active devices > >> + to higher addresses, so device drivers will be > >> + paused during rescan. > >> > >> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power > >> Management. > >> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c > >> index 1835f3a7aa8d..5f07a59b5924 100644 > >> --- a/drivers/pci/pci.c > >> +++ b/drivers/pci/pci.c > >> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) > >> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); > >> } else if (!strncmp(str, "disable_acs_redir=", 18)) { > >> disable_acs_redir_param = str + 18; > >> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { > >> + pci_add_flags(PCI_MOVABLE_BARS); > >> } else { > >> printk(KERN_ERR "PCI: Unknown option `%s'\n", > >> str); > >> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c > >> index 201f9e5ff55c..bdaafc48dc4c 100644 > >> --- a/drivers/pci/probe.c > >> +++ b/drivers/pci/probe.c > >> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) > >> return max; > >> } > >> > >> +/* > >> + * Put all devices of the bus and its children to reset > >> + */ > >> +static void pci_bus_reset_prepare(struct pci_bus *bus) > >> +{ > >> + struct pci_dev *dev; > >> + > >> + list_for_each_entry(dev, &bus->devices, bus_list) { > >> + struct pci_bus *child = dev->subordinate; > >> + > >> + if (child) { > >> + pci_bus_reset_prepare(child); > >> + } else if (dev->driver && > >> + dev->driver->err_handler && > >> + dev->driver->err_handler->reset_prepare) { > >> + dev->driver->err_handler->reset_prepare(dev); > >> + } > >> + } > >> +} > >> + > >> +/* > >> + * Complete the reset of all devices for the bus and its children > >> + */ > >> +static void pci_bus_reset_done(struct pci_bus *bus) > >> +{ > >> + struct pci_dev *dev; > >> + > >> + list_for_each_entry(dev, &bus->devices, bus_list) { > >> + struct pci_bus *child = dev->subordinate; > >> + > >> + if (child) { > >> + pci_bus_reset_done(child); > >> + } else if (dev->driver && dev->driver->err_handler && > >> + dev->driver->err_handler->reset_done) { > >> + dev->driver->err_handler->reset_done(dev); > >> + } > >> + } > >> +} > >> + > >> /** > >> * pci_rescan_bus - Scan a PCI bus for devices > >> * @bus: PCI bus to scan > >> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) > >> { > >> unsigned int max; > >> > >> + if (pci_has_flag(PCI_MOVABLE_BARS)) > >> + pci_bus_reset_prepare(bus); > >> max = pci_scan_child_bus(bus); > >> pci_assign_unassigned_bus_resources(bus); > >> + if (pci_has_flag(PCI_MOVABLE_BARS)) > >> + pci_bus_reset_done(bus); > >> pci_bus_add_devices(bus); > >> > >> return max; > >> diff --git a/include/linux/pci.h b/include/linux/pci.h > >> index 6925828f9f25..a8cb1a367c34 100644 > >> --- a/include/linux/pci.h > >> +++ b/include/linux/pci.h > >> @@ -847,6 +847,7 @@ enum { > >> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ > >> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ > >> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ > >> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ > >> }; > >> > >> /* These external functions are only available when PCI support is enabled */ > >> -- > >> 2.17.1 > >>
Hello Rajat, Sorry for the long absence! I've finally reworked the patchset heavily to address the proposals and issues (v2, sent today to linux-pci), so now I can give a better answer: On 9/19/18 12:22 AM, Rajat Jain wrote: > Hi Sergey, > > Sorry, I have one more basic question: The PCI express hierarchy is > made up from PCI express switches and end points, so essentially there > are only point to point links (except the switch internal logicl bus > on which you cannot hotplug anything). The device being hotplugged > shall be plugged onto a downstream port of a switch (or root port), so > there shall be no other device on that bus (since each downstream port > gets its own secondary bus number, and its own share of resources). So > it seems to me, that in this case, IF the resources at the downstream > port are not big enough, you will need to re-program the switch to get > resources to this downstream port, by stealing from *another* > downstream port. Is this correct understanding of the problem that you > are trying to solve? To some extent, yes. The patchset is intended to work with the "pci=realloc" command line argument, in this case the kernel will set bridge windows of zero size for empty buses. And for non-empty slots their bridge windows would be equal to to a sum of children BARs + alignments. After hotplugging a device into a downstream port of a switch, first the kernel needs to allocate a bridge window for this downstream port. Neighboring empty slots have nothing to steal from, and busy slots can not be shrunk (because all working devices must remain working after a hotplug event). So the switch's bridge window should be extended to fit the bridge window for the downstream port to fit the BARs of the hotplugged device. And here is the problem: switch's neighbors most probably reside very closely, and they can be movable (support movable BARs, or don't have connected drivers), immovable or movable in hard limits (depending on location of children immovable BARs). This patchset releases bridge windows and everything movable, marks the immovable BARs with IORESOURCE_PCI_FIXED, and calculates for every bridge its "fixed range" - the lowest and the highest addresses of children immovable BARs to determine how far away we can move these bridge windows. Then the re-assignment starts. If it fails (the root bridge window is not big enough to fit all the BARs or can't squeeze the new BARs between the immovable ones), the hotplugged devices are to be marked with the new flag PCI_DEV_IGNORE one by one, and re-assignment restarts. Then the kernel writes the recalculated BAR addresses and bridge windows into the ports and endpoints, and enables the MEM and IO flags if needed. Please ask if I haven't covered some aspects! Best regards, Serge > > Thanks, > > Rajat > > > > On Mon, Sep 17, 2018 at 2:25 PM Sergey Miroshnichenko > <s.miroshnichenko@yadro.com> wrote: >> >> Hello Rajat, >> >> On 9/17/18 10:00 PM, Rajat Jain wrote: >>> On Fri, Sep 14, 2018 at 9:21 AM Sergey Miroshnichenko >>> <s.miroshnichenko@yadro.com> wrote: >>>> >>>> Introduce a new command line option "pci=pcie_movable_bars" that indicates >>>> support of PCIe hotplug without prior reservation of memory regions by >>>> BIOS/bootloader. >>>> >>>> If a new PCIe device has been hot-plugged between two active ones, which >>>> have no (or not big enough) gap between their BARs, allocating new BARs >>>> requires to move BARs of the following working devices: >>> >>> I think this is a great problem to solve. I have some questions (not >>> objections, but trying to understand what are the limitations of the >>> solution being proposed): >>> >>> * What about hot-pluggable root ports? Would this help (I'm guessing not)? >>> >> >> We don't have a hardware that is able to handle a hotplug directly into >> a root port (but we'll find) - there are intermediate bridges always, so >> I can't provide proofs right now. But after a thought experiment I can't >> see any obstacles, so just need to verify. >> >>> * What about situations where the root port itself does not have >>> enough resources for the new device being inserted. I'm guessing we >>> are not going to expand root port allocation in those cases. But do we >>> fail gracefully rejecting the hotplug by not assigning it resources, >>> or do we manage to screw up the already working healthy devices (while >>> attempting to move them)? >>> >> >> Thanks for a great point! The code currently is too simplistic to handle >> this case properly, and though it never happened to us yet, but I'm sure >> it can kick off a working device to use its resources for a hotplugged >> one. The struct resource doesn't have a "priority" field, so probably I >> need to check the status of the resource allocation procedure and retry >> it without some of new hotplugged devices in case of fail. >> >>> * What about non-memory resources? E.g. cards may have pci bridges or >>> switches on them, and they may need extra PCI bus numbers. Does this >>> help that use case? >> >> Actually, this patchset is a part of our bigger work on hotplugging >> bridges full of bridges full of NVME and GPU devices. And in case of >> success here I was planning to submit the last part (hotplugging >> bridges), which includes the movement of bus numbers - similar to BARs, >> but more awkward, with PCIe devices renaming and re-building the >> /sys/bus/pci and /proc/pci, otherwise the lspci tool gets mad. >> >> By device renaming I mean the "/dev/nvme0n1p1" and "eth0" will remain >> the same, but underlying "nvme 0022:04:00.0" and "tg3 0001:06:00.1" may >> change. In this regard I don't know which UDEV action is the best here - >> it's neither "add", "remove" nor "change". >> >> >> Also I've just realized that should take a closer look and check if MSIs >> are also needs to be reallocated. >> >>>> >>>> 1) dev 4 >>>> | >>>> v >>>> .. | dev 3 | dev 3 | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>>> >>>> 2) dev 4 >>>> | >>>> v >>>> .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | >>>> >>>> 3) >>>> >>>> .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | >>>> .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | >>>> >>>> Not only BARs, but also bridge windows can be updated during a PCIe rescan, >>>> threatening all memory transactions during this procedure, so the PCI >>>> subsystem will instruct the drivers to pause by calling the reset_prepare() >>>> and reset_done() callbacks. >>>> >>>> If a device may be affected by BAR movement, the BAR changes tracking must >>>> be implemented in its driver. >>> >>> This is quite a big thing right? I expect that this will break a lot >>> of drivers because >>> (a) they may not have reset_prepare() and reset_done() callbacks (I >>> grepped in the sources and seems only 4 support it?). >>> (b) Now, it is expected that in the reset_done() the drivers should >>> re-read the BAR and remap any memory resources that may have changed. >>> Note this may also include cleaning up any existing resource mappings >>> that they had before they were paused. Not sure if this was the >>> semantics of reset_done() already, or may be it was, I'm just not >>> sure. >>> >> >> Yes, this is the most painful part: all the drivers that may be affected >> with BARs getting moved, must be prepared in this way. Luckily, these >> operations have distinct patterns, just like power management callbacks >> like runtime_suspend()/runtime_resume() and friends. >> >> Just a few minutes earlier in this thread I've wrote to Sam why we think >> this approach still can be used even with a minimal subset of drivers >> being prepared: for now we've limited the configurations of hotplug >> usage to non-root bridges and to domains with similar devices. If this >> concept looks viable to the community, we'll be happy to share the >> patches for the nvme, ixgbe and tg3 drivers (and a bit later - for >> xhci). Right now these are the only drivers _we_ need on our ppc64 >> (PowerNV) and x86_64 servers. >> >>> Thanks, >>> >>> Rajat >>> >> >> Thanks for the review! >> >> Best regards, >> Serge >> >>>> >>>> Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> >>>> --- >>>> .../admin-guide/kernel-parameters.txt | 6 +++ >>>> drivers/pci/pci.c | 2 + >>>> drivers/pci/probe.c | 43 +++++++++++++++++++ >>>> include/linux/pci.h | 1 + >>>> 4 files changed, 52 insertions(+) >>>> >>>> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt >>>> index 64a3bf54b974..f8132a709061 100644 >>>> --- a/Documentation/admin-guide/kernel-parameters.txt >>>> +++ b/Documentation/admin-guide/kernel-parameters.txt >>>> @@ -3311,6 +3311,12 @@ >>>> bridges without forcing it upstream. Note: >>>> this removes isolation between devices and >>>> may put more devices in an IOMMU group. >>>> + pcie_movable_bars Arrange a space at runtime for BARs of >>>> + hotplugged PCIe devices - usable if bootloader >>>> + doesn't reserve memory regions for them. Freeing >>>> + a space may require moving BARs of active devices >>>> + to higher addresses, so device drivers will be >>>> + paused during rescan. >>>> >>>> pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power >>>> Management. >>>> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c >>>> index 1835f3a7aa8d..5f07a59b5924 100644 >>>> --- a/drivers/pci/pci.c >>>> +++ b/drivers/pci/pci.c >>>> @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) >>>> pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); >>>> } else if (!strncmp(str, "disable_acs_redir=", 18)) { >>>> disable_acs_redir_param = str + 18; >>>> + } else if (!strncmp(str, "pcie_movable_bars", 17)) { >>>> + pci_add_flags(PCI_MOVABLE_BARS); >>>> } else { >>>> printk(KERN_ERR "PCI: Unknown option `%s'\n", >>>> str); >>>> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c >>>> index 201f9e5ff55c..bdaafc48dc4c 100644 >>>> --- a/drivers/pci/probe.c >>>> +++ b/drivers/pci/probe.c >>>> @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) >>>> return max; >>>> } >>>> >>>> +/* >>>> + * Put all devices of the bus and its children to reset >>>> + */ >>>> +static void pci_bus_reset_prepare(struct pci_bus *bus) >>>> +{ >>>> + struct pci_dev *dev; >>>> + >>>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>>> + struct pci_bus *child = dev->subordinate; >>>> + >>>> + if (child) { >>>> + pci_bus_reset_prepare(child); >>>> + } else if (dev->driver && >>>> + dev->driver->err_handler && >>>> + dev->driver->err_handler->reset_prepare) { >>>> + dev->driver->err_handler->reset_prepare(dev); >>>> + } >>>> + } >>>> +} >>>> + >>>> +/* >>>> + * Complete the reset of all devices for the bus and its children >>>> + */ >>>> +static void pci_bus_reset_done(struct pci_bus *bus) >>>> +{ >>>> + struct pci_dev *dev; >>>> + >>>> + list_for_each_entry(dev, &bus->devices, bus_list) { >>>> + struct pci_bus *child = dev->subordinate; >>>> + >>>> + if (child) { >>>> + pci_bus_reset_done(child); >>>> + } else if (dev->driver && dev->driver->err_handler && >>>> + dev->driver->err_handler->reset_done) { >>>> + dev->driver->err_handler->reset_done(dev); >>>> + } >>>> + } >>>> +} >>>> + >>>> /** >>>> * pci_rescan_bus - Scan a PCI bus for devices >>>> * @bus: PCI bus to scan >>>> @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) >>>> { >>>> unsigned int max; >>>> >>>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>>> + pci_bus_reset_prepare(bus); >>>> max = pci_scan_child_bus(bus); >>>> pci_assign_unassigned_bus_resources(bus); >>>> + if (pci_has_flag(PCI_MOVABLE_BARS)) >>>> + pci_bus_reset_done(bus); >>>> pci_bus_add_devices(bus); >>>> >>>> return max; >>>> diff --git a/include/linux/pci.h b/include/linux/pci.h >>>> index 6925828f9f25..a8cb1a367c34 100644 >>>> --- a/include/linux/pci.h >>>> +++ b/include/linux/pci.h >>>> @@ -847,6 +847,7 @@ enum { >>>> PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ >>>> PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ >>>> PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ >>>> + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ >>>> }; >>>> >>>> /* These external functions are only available when PCI support is enabled */ >>>> -- >>>> 2.17.1 >>>>
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt index 64a3bf54b974..f8132a709061 100644 --- a/Documentation/admin-guide/kernel-parameters.txt +++ b/Documentation/admin-guide/kernel-parameters.txt @@ -3311,6 +3311,12 @@ bridges without forcing it upstream. Note: this removes isolation between devices and may put more devices in an IOMMU group. + pcie_movable_bars Arrange a space at runtime for BARs of + hotplugged PCIe devices - usable if bootloader + doesn't reserve memory regions for them. Freeing + a space may require moving BARs of active devices + to higher addresses, so device drivers will be + paused during rescan. pcie_aspm= [PCIE] Forcibly enable or disable PCIe Active State Power Management. diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c index 1835f3a7aa8d..5f07a59b5924 100644 --- a/drivers/pci/pci.c +++ b/drivers/pci/pci.c @@ -6105,6 +6105,8 @@ static int __init pci_setup(char *str) pci_add_flags(PCI_SCAN_ALL_PCIE_DEVS); } else if (!strncmp(str, "disable_acs_redir=", 18)) { disable_acs_redir_param = str + 18; + } else if (!strncmp(str, "pcie_movable_bars", 17)) { + pci_add_flags(PCI_MOVABLE_BARS); } else { printk(KERN_ERR "PCI: Unknown option `%s'\n", str); diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c index 201f9e5ff55c..bdaafc48dc4c 100644 --- a/drivers/pci/probe.c +++ b/drivers/pci/probe.c @@ -3138,6 +3138,45 @@ unsigned int pci_rescan_bus_bridge_resize(struct pci_dev *bridge) return max; } +/* + * Put all devices of the bus and its children to reset + */ +static void pci_bus_reset_prepare(struct pci_bus *bus) +{ + struct pci_dev *dev; + + list_for_each_entry(dev, &bus->devices, bus_list) { + struct pci_bus *child = dev->subordinate; + + if (child) { + pci_bus_reset_prepare(child); + } else if (dev->driver && + dev->driver->err_handler && + dev->driver->err_handler->reset_prepare) { + dev->driver->err_handler->reset_prepare(dev); + } + } +} + +/* + * Complete the reset of all devices for the bus and its children + */ +static void pci_bus_reset_done(struct pci_bus *bus) +{ + struct pci_dev *dev; + + list_for_each_entry(dev, &bus->devices, bus_list) { + struct pci_bus *child = dev->subordinate; + + if (child) { + pci_bus_reset_done(child); + } else if (dev->driver && dev->driver->err_handler && + dev->driver->err_handler->reset_done) { + dev->driver->err_handler->reset_done(dev); + } + } +} + /** * pci_rescan_bus - Scan a PCI bus for devices * @bus: PCI bus to scan @@ -3151,8 +3190,12 @@ unsigned int pci_rescan_bus(struct pci_bus *bus) { unsigned int max; + if (pci_has_flag(PCI_MOVABLE_BARS)) + pci_bus_reset_prepare(bus); max = pci_scan_child_bus(bus); pci_assign_unassigned_bus_resources(bus); + if (pci_has_flag(PCI_MOVABLE_BARS)) + pci_bus_reset_done(bus); pci_bus_add_devices(bus); return max; diff --git a/include/linux/pci.h b/include/linux/pci.h index 6925828f9f25..a8cb1a367c34 100644 --- a/include/linux/pci.h +++ b/include/linux/pci.h @@ -847,6 +847,7 @@ enum { PCI_ENABLE_PROC_DOMAINS = 0x00000010, /* Enable domains in /proc */ PCI_COMPAT_DOMAIN_0 = 0x00000020, /* ... except domain 0 */ PCI_SCAN_ALL_PCIE_DEVS = 0x00000040, /* Scan all, not just dev 0 */ + PCI_MOVABLE_BARS = 0x00000080, /* Runtime BAR reassign after hotplug */ }; /* These external functions are only available when PCI support is enabled */
Introduce a new command line option "pci=pcie_movable_bars" that indicates support of PCIe hotplug without prior reservation of memory regions by BIOS/bootloader. If a new PCIe device has been hot-plugged between two active ones, which have no (or not big enough) gap between their BARs, allocating new BARs requires to move BARs of the following working devices: 1) dev 4 | v .. | dev 3 | dev 3 | dev 5 | dev 7 | .. | BAR 0 | BAR 1 | BAR 0 | BAR 0 | 2) dev 4 | v .. | dev 3 | dev 3 | --> --> | dev 5 | dev 7 | .. | BAR 0 | BAR 1 | --> --> | BAR 0 | BAR 0 | 3) .. | dev 3 | dev 3 | dev 4 | dev 4 | dev 5 | dev 7 | .. | BAR 0 | BAR 1 | BAR 0 | BAR 1 | BAR 0 | BAR 0 | Not only BARs, but also bridge windows can be updated during a PCIe rescan, threatening all memory transactions during this procedure, so the PCI subsystem will instruct the drivers to pause by calling the reset_prepare() and reset_done() callbacks. If a device may be affected by BAR movement, the BAR changes tracking must be implemented in its driver. Signed-off-by: Sergey Miroshnichenko <s.miroshnichenko@yadro.com> --- .../admin-guide/kernel-parameters.txt | 6 +++ drivers/pci/pci.c | 2 + drivers/pci/probe.c | 43 +++++++++++++++++++ include/linux/pci.h | 1 + 4 files changed, 52 insertions(+)