* Re: [PATCH v2 2/2] powerpc/powernv/npu: Don't explicitly flush nmmu tlb
2017-09-06 1:49 ` [PATCH v2 2/2] powerpc/powernv/npu: Don't explicitly flush nmmu tlb Alistair Popple
@ 2017-09-06 6:58 ` Frederic Barrat
2017-09-08 10:23 ` kbuild test robot
1 sibling, 0 replies; 6+ messages in thread
From: Frederic Barrat @ 2017-09-06 6:58 UTC (permalink / raw)
To: Alistair Popple, linuxppc-dev; +Cc: arbab, mpe, andrew.donnellan
Le 06/09/2017 à 03:49, Alistair Popple a écrit :
> The nest mmu required an explicit flush as a tlbi would not flush it in the
> same way as the core. However an alternate firmware fix exists which should
> eliminate the need for this flush, so instead add a device-tree property
> (ibm,nmmu-flush) on the NVLink2 PHB to enable it only if required.
>
> Signed-off-by: Alistair Popple <alistair@popple.id.au>
> ---
FWIW,
Reviewed-by: Frederic Barrat <fbarrat@linux.vnet.ibm.com>
> Changes for v2:
> - Use mm_context_add_copro()/mm_context_remove_copro() instead
> of inc_mm_active_cpus()/dec_mm_active_cpus()
>
> arch/powerpc/platforms/powernv/npu-dma.c | 28 +++++++++++++++++++++++-----
> arch/powerpc/platforms/powernv/pci.h | 3 +++
> 2 files changed, 26 insertions(+), 5 deletions(-)
>
> diff --git a/arch/powerpc/platforms/powernv/npu-dma.c b/arch/powerpc/platforms/powernv/npu-dma.c
> index 2fff9a65..f6cbc1a 100644
> --- a/arch/powerpc/platforms/powernv/npu-dma.c
> +++ b/arch/powerpc/platforms/powernv/npu-dma.c
> @@ -395,6 +395,7 @@ struct npu_context {
> struct pci_dev *npdev[NV_MAX_NPUS][NV_MAX_LINKS];
> struct mmu_notifier mn;
> struct kref kref;
> + bool nmmu_flush;
>
> /* Callback to stop translation requests on a given GPU */
> struct npu_context *(*release_cb)(struct npu_context *, void *);
> @@ -545,11 +546,13 @@ static void mmio_invalidate(struct npu_context *npu_context, int va,
> struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS];
> unsigned long pid = npu_context->mm->context.id;
>
> - /*
> - * Unfortunately the nest mmu does not support flushing specific
> - * addresses so we have to flush the whole mm.
> - */
> - flush_all_mm(npu_context->mm);
> + if (npu_context->nmmu_flush)
> + /*
> + * Unfortunately the nest mmu does not support flushing specific
> + * addresses so we have to flush the whole mm once before
> + * shooting down the GPU translation.
> + */
> + flush_all_mm(npu_context->mm);
>
> /*
> * Loop over all the NPUs this process is active on and launch
> @@ -722,6 +725,16 @@ struct npu_context *pnv_npu2_init_context(struct pci_dev *gpdev,
> return ERR_PTR(-ENODEV);
> npu_context->npdev[npu->index][nvlink_index] = npdev;
>
> + if (!nphb->npu.nmmu_flush) {
> + /*
> + * If we're not explicitly flushing ourselves we need to mark
> + * the thread for global flushes
> + */
> + npu_context->nmmu_flush = false;
> + mm_context_add_copro(mm);
> + } else
> + npu_context->nmmu_flush = true;
> +
> return npu_context;
> }
> EXPORT_SYMBOL(pnv_npu2_init_context);
> @@ -731,6 +744,9 @@ static void pnv_npu2_release_context(struct kref *kref)
> struct npu_context *npu_context =
> container_of(kref, struct npu_context, kref);
>
> + if (!npu_context->nmmu_flush)
> + mm_context_remove_copro(npu_context->mm);
> +
> npu_context->mm->context.npu_context = NULL;
> mmu_notifier_unregister(&npu_context->mn,
> npu_context->mm);
> @@ -819,6 +835,8 @@ int pnv_npu2_init(struct pnv_phb *phb)
> static int npu_index;
> uint64_t rc = 0;
>
> + phb->npu.nmmu_flush =
> + of_property_read_bool(phb->hose->dn, "ibm,nmmu-flush");
> for_each_child_of_node(phb->hose->dn, dn) {
> gpdev = pnv_pci_get_gpu_dev(get_pci_dev(dn));
> if (gpdev) {
> diff --git a/arch/powerpc/platforms/powernv/pci.h b/arch/powerpc/platforms/powernv/pci.h
> index a95273c..22025c6 100644
> --- a/arch/powerpc/platforms/powernv/pci.h
> +++ b/arch/powerpc/platforms/powernv/pci.h
> @@ -187,6 +187,9 @@ struct pnv_phb {
>
> /* Bitmask for MMIO register usage */
> unsigned long mmio_atsd_usage;
> +
> + /* Do we need to explicitly flush the nest mmu? */
> + bool nmmu_flush;
> } npu;
>
> #ifdef CONFIG_CXL_BASE
>
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH v2 2/2] powerpc/powernv/npu: Don't explicitly flush nmmu tlb
2017-09-06 1:49 ` [PATCH v2 2/2] powerpc/powernv/npu: Don't explicitly flush nmmu tlb Alistair Popple
2017-09-06 6:58 ` Frederic Barrat
@ 2017-09-08 10:23 ` kbuild test robot
1 sibling, 0 replies; 6+ messages in thread
From: kbuild test robot @ 2017-09-08 10:23 UTC (permalink / raw)
To: Alistair Popple
Cc: kbuild-all, linuxppc-dev, andrew.donnellan, arbab, fbarrat,
Alistair Popple
[-- Attachment #1: Type: text/plain, Size: 10056 bytes --]
Hi Alistair,
[auto build test ERROR on powerpc/next]
[also build test ERROR on next-20170907]
[cannot apply to v4.13]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system]
url: https://github.com/0day-ci/linux/commits/Alistair-Popple/powerpc-npu-Use-flush_all_mm-instead-of-flush_tlb_mm/20170908-080828
base: https://git.kernel.org/pub/scm/linux/kernel/git/powerpc/linux.git next
config: powerpc-allmodconfig (attached as .config)
compiler: powerpc64-linux-gnu-gcc (Debian 6.1.1-9) 6.1.1 20160705
reproduce:
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
chmod +x ~/bin/make.cross
# save the attached .config to linux build tree
make.cross ARCH=powerpc
All errors (new ones prefixed by >>):
arch/powerpc/platforms/powernv/npu-dma.c: In function 'mmio_invalidate':
arch/powerpc/platforms/powernv/npu-dma.c:555:3: error: implicit declaration of function 'flush_all_mm' [-Werror=implicit-function-declaration]
flush_all_mm(npu_context->mm);
^~~~~~~~~~~~
arch/powerpc/platforms/powernv/npu-dma.c: In function 'pnv_npu2_init_context':
>> arch/powerpc/platforms/powernv/npu-dma.c:744:3: error: implicit declaration of function 'mm_context_add_copro' [-Werror=implicit-function-declaration]
mm_context_add_copro(mm);
^~~~~~~~~~~~~~~~~~~~
arch/powerpc/platforms/powernv/npu-dma.c: In function 'pnv_npu2_release_context':
>> arch/powerpc/platforms/powernv/npu-dma.c:758:3: error: implicit declaration of function 'mm_context_remove_copro' [-Werror=implicit-function-declaration]
mm_context_remove_copro(npu_context->mm);
^~~~~~~~~~~~~~~~~~~~~~~
cc1: some warnings being treated as errors
vim +/mm_context_add_copro +744 arch/powerpc/platforms/powernv/npu-dma.c
534
535 /*
536 * Invalidate either a single address or an entire PID depending on
537 * the value of va.
538 */
539 static void mmio_invalidate(struct npu_context *npu_context, int va,
540 unsigned long address, bool flush)
541 {
542 int i, j;
543 struct npu *npu;
544 struct pnv_phb *nphb;
545 struct pci_dev *npdev;
546 struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS];
547 unsigned long pid = npu_context->mm->context.id;
548
549 if (npu_context->nmmu_flush)
550 /*
551 * Unfortunately the nest mmu does not support flushing specific
552 * addresses so we have to flush the whole mm once before
553 * shooting down the GPU translation.
554 */
> 555 flush_all_mm(npu_context->mm);
556
557 /*
558 * Loop over all the NPUs this process is active on and launch
559 * an invalidate.
560 */
561 for (i = 0; i <= max_npu2_index; i++) {
562 mmio_atsd_reg[i].reg = -1;
563 for (j = 0; j < NV_MAX_LINKS; j++) {
564 npdev = npu_context->npdev[i][j];
565 if (!npdev)
566 continue;
567
568 nphb = pci_bus_to_host(npdev->bus)->private_data;
569 npu = &nphb->npu;
570 mmio_atsd_reg[i].npu = npu;
571
572 if (va)
573 mmio_atsd_reg[i].reg =
574 mmio_invalidate_va(npu, address, pid,
575 flush);
576 else
577 mmio_atsd_reg[i].reg =
578 mmio_invalidate_pid(npu, pid, flush);
579
580 /*
581 * The NPU hardware forwards the shootdown to all GPUs
582 * so we only have to launch one shootdown per NPU.
583 */
584 break;
585 }
586 }
587
588 mmio_invalidate_wait(mmio_atsd_reg, flush);
589 if (flush)
590 /* Wait for the flush to complete */
591 mmio_invalidate_wait(mmio_atsd_reg, false);
592 }
593
594 static void pnv_npu2_mn_release(struct mmu_notifier *mn,
595 struct mm_struct *mm)
596 {
597 struct npu_context *npu_context = mn_to_npu_context(mn);
598
599 /* Call into device driver to stop requests to the NMMU */
600 if (npu_context->release_cb)
601 npu_context->release_cb(npu_context, npu_context->priv);
602
603 /*
604 * There should be no more translation requests for this PID, but we
605 * need to ensure any entries for it are removed from the TLB.
606 */
607 mmio_invalidate(npu_context, 0, 0, true);
608 }
609
610 static void pnv_npu2_mn_change_pte(struct mmu_notifier *mn,
611 struct mm_struct *mm,
612 unsigned long address,
613 pte_t pte)
614 {
615 struct npu_context *npu_context = mn_to_npu_context(mn);
616
617 mmio_invalidate(npu_context, 1, address, true);
618 }
619
620 static void pnv_npu2_mn_invalidate_page(struct mmu_notifier *mn,
621 struct mm_struct *mm,
622 unsigned long address)
623 {
624 struct npu_context *npu_context = mn_to_npu_context(mn);
625
626 mmio_invalidate(npu_context, 1, address, true);
627 }
628
629 static void pnv_npu2_mn_invalidate_range(struct mmu_notifier *mn,
630 struct mm_struct *mm,
631 unsigned long start, unsigned long end)
632 {
633 struct npu_context *npu_context = mn_to_npu_context(mn);
634 unsigned long address;
635
636 for (address = start; address < end; address += PAGE_SIZE)
637 mmio_invalidate(npu_context, 1, address, false);
638
639 /* Do the flush only on the final addess == end */
640 mmio_invalidate(npu_context, 1, address, true);
641 }
642
643 static const struct mmu_notifier_ops nv_nmmu_notifier_ops = {
644 .release = pnv_npu2_mn_release,
645 .change_pte = pnv_npu2_mn_change_pte,
646 .invalidate_page = pnv_npu2_mn_invalidate_page,
647 .invalidate_range = pnv_npu2_mn_invalidate_range,
648 };
649
650 /*
651 * Call into OPAL to setup the nmmu context for the current task in
652 * the NPU. This must be called to setup the context tables before the
653 * GPU issues ATRs. pdev should be a pointed to PCIe GPU device.
654 *
655 * A release callback should be registered to allow a device driver to
656 * be notified that it should not launch any new translation requests
657 * as the final TLB invalidate is about to occur.
658 *
659 * Returns an error if there no contexts are currently available or a
660 * npu_context which should be passed to pnv_npu2_handle_fault().
661 *
662 * mmap_sem must be held in write mode.
663 */
664 struct npu_context *pnv_npu2_init_context(struct pci_dev *gpdev,
665 unsigned long flags,
666 struct npu_context *(*cb)(struct npu_context *, void *),
667 void *priv)
668 {
669 int rc;
670 u32 nvlink_index;
671 struct device_node *nvlink_dn;
672 struct mm_struct *mm = current->mm;
673 struct pnv_phb *nphb;
674 struct npu *npu;
675 struct npu_context *npu_context;
676
677 /*
678 * At present we don't support GPUs connected to multiple NPUs and I'm
679 * not sure the hardware does either.
680 */
681 struct pci_dev *npdev = pnv_pci_get_npu_dev(gpdev, 0);
682
683 if (!firmware_has_feature(FW_FEATURE_OPAL))
684 return ERR_PTR(-ENODEV);
685
686 if (!npdev)
687 /* No nvlink associated with this GPU device */
688 return ERR_PTR(-ENODEV);
689
690 if (!mm || mm->context.id == 0) {
691 /*
692 * Kernel thread contexts are not supported and context id 0 is
693 * reserved on the GPU.
694 */
695 return ERR_PTR(-EINVAL);
696 }
697
698 nphb = pci_bus_to_host(npdev->bus)->private_data;
699 npu = &nphb->npu;
700
701 /*
702 * Setup the NPU context table for a particular GPU. These need to be
703 * per-GPU as we need the tables to filter ATSDs when there are no
704 * active contexts on a particular GPU.
705 */
706 rc = opal_npu_init_context(nphb->opal_id, mm->context.id, flags,
707 PCI_DEVID(gpdev->bus->number, gpdev->devfn));
708 if (rc < 0)
709 return ERR_PTR(-ENOSPC);
710
711 /*
712 * We store the npu pci device so we can more easily get at the
713 * associated npus.
714 */
715 npu_context = mm->context.npu_context;
716 if (!npu_context) {
717 npu_context = kzalloc(sizeof(struct npu_context), GFP_KERNEL);
718 if (!npu_context)
719 return ERR_PTR(-ENOMEM);
720
721 mm->context.npu_context = npu_context;
722 npu_context->mm = mm;
723 npu_context->mn.ops = &nv_nmmu_notifier_ops;
724 __mmu_notifier_register(&npu_context->mn, mm);
725 kref_init(&npu_context->kref);
726 } else {
727 kref_get(&npu_context->kref);
728 }
729
730 npu_context->release_cb = cb;
731 npu_context->priv = priv;
732 nvlink_dn = of_parse_phandle(npdev->dev.of_node, "ibm,nvlink", 0);
733 if (WARN_ON(of_property_read_u32(nvlink_dn, "ibm,npu-link-index",
734 &nvlink_index)))
735 return ERR_PTR(-ENODEV);
736 npu_context->npdev[npu->index][nvlink_index] = npdev;
737
738 if (!nphb->npu.nmmu_flush) {
739 /*
740 * If we're not explicitly flushing ourselves we need to mark
741 * the thread for global flushes
742 */
743 npu_context->nmmu_flush = false;
> 744 mm_context_add_copro(mm);
745 } else
746 npu_context->nmmu_flush = true;
747
748 return npu_context;
749 }
750 EXPORT_SYMBOL(pnv_npu2_init_context);
751
752 static void pnv_npu2_release_context(struct kref *kref)
753 {
754 struct npu_context *npu_context =
755 container_of(kref, struct npu_context, kref);
756
757 if (!npu_context->nmmu_flush)
> 758 mm_context_remove_copro(npu_context->mm);
759
760 npu_context->mm->context.npu_context = NULL;
761 mmu_notifier_unregister(&npu_context->mn,
762 npu_context->mm);
763
764 kfree(npu_context);
765 }
766
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
[-- Attachment #2: .config.gz --]
[-- Type: application/gzip, Size: 54017 bytes --]
^ permalink raw reply [flat|nested] 6+ messages in thread