| Message ID | 1561162778-12669-1-git-send-email-megha.dey@linux.intel.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Introduce dynamic allocation/freeing of MSI-X vectors | expand |
Hi Megha, On Fri, Jun 21, 2019 at 05:19:32PM -0700, Megha Dey wrote: > Currently, MSI-X vector enabling and allocation for a PCIe device is > static i.e. a device driver gets only one chance to enable a specific > number of MSI-X vectors, usually during device probe. Also, in many > cases, drivers usually reserve more than required number of vectors > anticipating their use, which unnecessarily blocks resources that > could have been made available to other devices. Lastly, there is no > way for drivers to reserve more vectors, if the MSI-x has already been > enabled for that device. > > Hence, a dynamic MSI-X kernel infrastructure can benefit drivers by > deferring MSI-X allocation to post probe phase, where actual demand > information is available. > > This patchset enables the dynamic allocation/de-allocation of MSI-X > vectors by introducing 2 new APIs: > pci_alloc_irq_vectors_dyn() and pci_free_irq_vectors_grp(): > > We have had requests from some of the NIC/RDMA users who have lots of > interrupt resources and would like to allocate them on demand, > instead of using an all or none approach. > > The APIs are fairly well tested (multiple allocations/deallocations), > but we have no early adopters yet. Hence, sending this series as an > RFC for review and comments. > > The patches are based out of Linux 5.2-rc5. > > Megha Dey (6): > PCI/MSI: New structures/macros for dynamic MSI-X allocation > PCI/MSI: Dynamic allocation of MSI-X vectors by group > x86: Introduce the dynamic teardown function > PCI/MSI: Introduce new structure to manage MSI-x entries s/MSI-x/MSI-X/ If "entries" here means the same as "vectors" above, please use the same word. > PCI/MSI: Free MSI-X resources by group Is "resources" the same as "vectors"? > Documentation: PCI/MSI: Document dynamic MSI-X infrastructure When you post a v2 after addressing Thomas' comments, please make these subject lines imperative sentences beginning with a descriptive verb. You can run "git log --oneline drivers/pci" to see the style. If you're adding a specific interface or structure, mention it by name in the subject if it's practical. The "x86" line needs a little more context; I assume it should include "IRQ", "MSI-X", or something. > Documentation/PCI/MSI-HOWTO.txt | 38 +++++ > arch/x86/include/asm/x86_init.h | 1 + > arch/x86/kernel/x86_init.c | 6 + > drivers/pci/msi.c | 363 +++++++++++++++++++++++++++++++++++++--- > drivers/pci/probe.c | 9 + > include/linux/device.h | 3 + > include/linux/msi.h | 13 ++ > include/linux/pci.h | 61 +++++++ > kernel/irq/msi.c | 34 +++- > 9 files changed, 497 insertions(+), 31 deletions(-) > > -- > 2.7.4 >
On Thu, 2019-08-01 at 19:24 -0500, Bjorn Helgaas wrote: > Hi Megha, > > On Fri, Jun 21, 2019 at 05:19:32PM -0700, Megha Dey wrote: > > > > Currently, MSI-X vector enabling and allocation for a PCIe device > > is > > static i.e. a device driver gets only one chance to enable a > > specific > > number of MSI-X vectors, usually during device probe. Also, in many > > cases, drivers usually reserve more than required number of vectors > > anticipating their use, which unnecessarily blocks resources that > > could have been made available to other devices. Lastly, there is > > no > > way for drivers to reserve more vectors, if the MSI-x has already > > been > > enabled for that device. > > > > Hence, a dynamic MSI-X kernel infrastructure can benefit drivers by > > deferring MSI-X allocation to post probe phase, where actual demand > > information is available. > > > > This patchset enables the dynamic allocation/de-allocation of MSI-X > > vectors by introducing 2 new APIs: > > pci_alloc_irq_vectors_dyn() and pci_free_irq_vectors_grp(): > > > > We have had requests from some of the NIC/RDMA users who have lots > > of > > interrupt resources and would like to allocate them on demand, > > instead of using an all or none approach. > > > > The APIs are fairly well tested (multiple > > allocations/deallocations), > > but we have no early adopters yet. Hence, sending this series as an > > RFC for review and comments. > > > > The patches are based out of Linux 5.2-rc5. > > > > Megha Dey (6): > > PCI/MSI: New structures/macros for dynamic MSI-X allocation > > PCI/MSI: Dynamic allocation of MSI-X vectors by group > > x86: Introduce the dynamic teardown function > > PCI/MSI: Introduce new structure to manage MSI-x entries > s/MSI-x/MSI-X/ > If "entries" here means the same as "vectors" above, please use the > same word. > Hi Bjorn, Well, here entries basically mean the msi_desc entries. I thought MSI vectors are used for each address/data pair, hence used the term entries. Will update it to vectors to ensure uniformity. > > > > PCI/MSI: Free MSI-X resources by group > Is "resources" the same as "vectors"? > Yes, will update this in V2. > > > > Documentation: PCI/MSI: Document dynamic MSI-X infrastructure > When you post a v2 after addressing Thomas' comments, please make > these subject lines imperative sentences beginning with a descriptive > verb. You can run "git log --oneline drivers/pci" to see the style. Sure, I will update the subject lines in V2. > If you're adding a specific interface or structure, mention it by > name > in the subject if it's practical. The "x86" line needs a little more > context; I assume it should include "IRQ", "MSI-X", or something. > True, will change this in V2. > > > > Documentation/PCI/MSI-HOWTO.txt | 38 +++++ > > arch/x86/include/asm/x86_init.h | 1 + > > arch/x86/kernel/x86_init.c | 6 + > > drivers/pci/msi.c | 363 > > +++++++++++++++++++++++++++++++++++++--- > > drivers/pci/probe.c | 9 + > > include/linux/device.h | 3 + > > include/linux/msi.h | 13 ++ > > include/linux/pci.h | 61 +++++++ > > kernel/irq/msi.c | 34 +++- > > 9 files changed, 497 insertions(+), 31 deletions(-) > >