From patchwork Wed Jul 1 22:53:48 2020 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Ralph Campbell X-Patchwork-Id: 11637453 X-Patchwork-Delegate: jgg@ziepe.ca Return-Path: Received: from mail.kernel.org (pdx-korg-mail-1.web.codeaurora.org [172.30.200.123]) by pdx-korg-patchwork-2.web.codeaurora.org (Postfix) with ESMTP id D8E2D13B4 for ; Wed, 1 Jul 2020 22:54:12 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id B259F20853 for ; Wed, 1 Jul 2020 22:54:12 +0000 (UTC) Authentication-Results: mail.kernel.org; dkim=pass (2048-bit key) header.d=nvidia.com header.i=@nvidia.com header.b="FRI4jtZl" Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1726631AbgGAWyL (ORCPT ); Wed, 1 Jul 2020 18:54:11 -0400 Received: from hqnvemgate24.nvidia.com ([216.228.121.143]:11147 "EHLO hqnvemgate24.nvidia.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1726586AbgGAWyL (ORCPT ); Wed, 1 Jul 2020 18:54:11 -0400 Received: from hqpgpgate101.nvidia.com (Not Verified[216.228.121.13]) by hqnvemgate24.nvidia.com (using TLS: TLSv1.2, DES-CBC3-SHA) id ; Wed, 01 Jul 2020 15:52:31 -0700 Received: from hqmail.nvidia.com ([172.20.161.6]) by hqpgpgate101.nvidia.com (PGP Universal service); Wed, 01 Jul 2020 15:54:11 -0700 X-PGP-Universal: processed; by hqpgpgate101.nvidia.com on Wed, 01 Jul 2020 15:54:11 -0700 Received: from HQMAIL101.nvidia.com (172.20.187.10) by HQMAIL105.nvidia.com (172.20.187.12) with Microsoft SMTP Server (TLS) id 15.0.1473.3; Wed, 1 Jul 2020 22:54:04 +0000 Received: from rnnvemgw01.nvidia.com (10.128.109.123) by HQMAIL101.nvidia.com (172.20.187.10) with Microsoft SMTP Server (TLS) id 15.0.1473.3 via Frontend Transport; Wed, 1 Jul 2020 22:54:04 +0000 Received: from rcampbell-dev.nvidia.com (Not Verified[10.110.48.66]) by rnnvemgw01.nvidia.com with Trustwave SEG (v7,5,8,10121) id ; Wed, 01 Jul 2020 15:54:04 -0700 From: Ralph Campbell To: , , , , CC: Jerome Glisse , John Hubbard , Christoph Hellwig , Jason Gunthorpe , "Andrew Morton" , Shuah Khan , "Ben Skeggs" , Ralph Campbell Subject: [PATCH v3 1/5] nouveau/hmm: fault one page at a time Date: Wed, 1 Jul 2020 15:53:48 -0700 Message-ID: <20200701225352.9649-2-rcampbell@nvidia.com> X-Mailer: git-send-email 2.20.1 In-Reply-To: <20200701225352.9649-1-rcampbell@nvidia.com> References: <20200701225352.9649-1-rcampbell@nvidia.com> MIME-Version: 1.0 X-NVConfidentiality: public DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=nvidia.com; s=n1; t=1593643951; bh=2WyHYZcAAd5j4j8YMKEb9Xnebvu/ReeLnPA7IJpaPHY=; h=X-PGP-Universal:From:To:CC:Subject:Date:Message-ID:X-Mailer: In-Reply-To:References:MIME-Version:X-NVConfidentiality: Content-Transfer-Encoding:Content-Type; b=FRI4jtZlIOtx6sdst9uuAiIkV5JRUspzQUhLl9rNHNt/y6cphOfL8DGqjh67NfJ/T 1cvaZbJd0uhA2TviV4kMKd/XXbJkkX/pqqRSI0y2Q/tSpZIjPSy5A+BmlDWw5gwT9d AZR5mu1OO3xYXqY/Bas+afFG8V4jyrkqrIMB2pIF6zRvL9+K6EvQILcM4Vl78VKb82 hd4iYLOa+5KeGdtjbC8DnWaW5iT0x2WbxJPItpRUPAjPcGt9l0pl0WHC4pt0n4Z1e7 +9sKiSBCv4Y9nBxJSBhK7Iwrx5PHNGjtDrX7FC84CXFpXj2tQlD96ZLwWRjgQMFny1 kWR9t84Kq0LDg== Sender: linux-rdma-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-rdma@vger.kernel.org The SVM page fault handler groups faults into a range of contiguous virtual addresses and requests hmm_range_fault() to populate and return the page frame number of system memory mapped by the CPU. In preparation for supporting large pages to be mapped by the GPU, process faults one page at a time. In addition, use the hmm_range default_flags to fix a corner case where the input hmm_pfns array is not reinitialized after hmm_range_fault() returns -EBUSY and must be called again. Signed-off-by: Ralph Campbell --- drivers/gpu/drm/nouveau/nouveau_svm.c | 199 +++++++++----------------- 1 file changed, 66 insertions(+), 133 deletions(-) diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c index ba9f9359c30e..665dede69bd1 100644 --- a/drivers/gpu/drm/nouveau/nouveau_svm.c +++ b/drivers/gpu/drm/nouveau/nouveau_svm.c @@ -516,7 +516,7 @@ static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = { static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm, struct hmm_range *range, u64 *ioctl_addr) { - unsigned long i, npages; + struct page *page; /* * The ioctl_addr prepared here is passed through nvif_object_ioctl() @@ -525,42 +525,38 @@ static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm, * This is all just encoding the internal hmm representation into a * different nouveau internal representation. */ - npages = (range->end - range->start) >> PAGE_SHIFT; - for (i = 0; i < npages; ++i) { - struct page *page; - - if (!(range->hmm_pfns[i] & HMM_PFN_VALID)) { - ioctl_addr[i] = 0; - continue; - } - - page = hmm_pfn_to_page(range->hmm_pfns[i]); - if (is_device_private_page(page)) - ioctl_addr[i] = nouveau_dmem_page_addr(page) | - NVIF_VMM_PFNMAP_V0_V | - NVIF_VMM_PFNMAP_V0_VRAM; - else - ioctl_addr[i] = page_to_phys(page) | - NVIF_VMM_PFNMAP_V0_V | - NVIF_VMM_PFNMAP_V0_HOST; - if (range->hmm_pfns[i] & HMM_PFN_WRITE) - ioctl_addr[i] |= NVIF_VMM_PFNMAP_V0_W; + if (!(range->hmm_pfns[0] & HMM_PFN_VALID)) { + ioctl_addr[0] = 0; + return; } + + page = hmm_pfn_to_page(range->hmm_pfns[0]); + if (is_device_private_page(page)) + ioctl_addr[0] = nouveau_dmem_page_addr(page) | + NVIF_VMM_PFNMAP_V0_V | + NVIF_VMM_PFNMAP_V0_VRAM; + else + ioctl_addr[0] = page_to_phys(page) | + NVIF_VMM_PFNMAP_V0_V | + NVIF_VMM_PFNMAP_V0_HOST; + if (range->hmm_pfns[0] & HMM_PFN_WRITE) + ioctl_addr[0] |= NVIF_VMM_PFNMAP_V0_W; } static int nouveau_range_fault(struct nouveau_svmm *svmm, struct nouveau_drm *drm, void *data, u32 size, - unsigned long hmm_pfns[], u64 *ioctl_addr, + u64 *ioctl_addr, unsigned long hmm_flags, struct svm_notifier *notifier) { unsigned long timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT); /* Have HMM fault pages within the fault window to the GPU. */ + unsigned long hmm_pfns[1]; struct hmm_range range = { .notifier = ¬ifier->notifier, .start = notifier->notifier.interval_tree.start, .end = notifier->notifier.interval_tree.last + 1, - .pfn_flags_mask = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE, + .default_flags = hmm_flags, .hmm_pfns = hmm_pfns, }; struct mm_struct *mm = notifier->notifier.mm; @@ -575,11 +571,6 @@ static int nouveau_range_fault(struct nouveau_svmm *svmm, ret = hmm_range_fault(&range); mmap_read_unlock(mm); if (ret) { - /* - * FIXME: the input PFN_REQ flags are destroyed on - * -EBUSY, we need to regenerate them, also for the - * other continue below - */ if (ret == -EBUSY) continue; return ret; @@ -614,17 +605,12 @@ nouveau_svm_fault(struct nvif_notify *notify) struct nvif_object *device = &svm->drm->client.device.object; struct nouveau_svmm *svmm; struct { - struct { - struct nvif_ioctl_v0 i; - struct nvif_ioctl_mthd_v0 m; - struct nvif_vmm_pfnmap_v0 p; - } i; - u64 phys[16]; + struct nouveau_pfnmap_args i; + u64 phys[1]; } args; - unsigned long hmm_pfns[ARRAY_SIZE(args.phys)]; - struct vm_area_struct *vma; + unsigned long hmm_flags; u64 inst, start, limit; - int fi, fn, pi, fill; + int fi, fn; int replay = 0, ret; /* Parse available fault buffer entries into a cache, and update @@ -691,66 +677,53 @@ nouveau_svm_fault(struct nvif_notify *notify) * window into a single update. */ start = buffer->fault[fi]->addr; - limit = start + (ARRAY_SIZE(args.phys) << PAGE_SHIFT); + limit = start + PAGE_SIZE; if (start < svmm->unmanaged.limit) limit = min_t(u64, limit, svmm->unmanaged.start); - SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit); - mm = svmm->notifier.mm; - if (!mmget_not_zero(mm)) { - nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]); - continue; - } - - /* Intersect fault window with the CPU VMA, cancelling - * the fault if the address is invalid. + /* + * Prepare the GPU-side update of all pages within the + * fault window, determining required pages and access + * permissions based on pending faults. */ - mmap_read_lock(mm); - vma = find_vma_intersection(mm, start, limit); - if (!vma) { - SVMM_ERR(svmm, "wndw %016llx-%016llx", start, limit); - mmap_read_unlock(mm); - mmput(mm); - nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]); - continue; + args.i.p.addr = start; + args.i.p.page = PAGE_SHIFT; + args.i.p.size = PAGE_SIZE; + /* + * Determine required permissions based on GPU fault + * access flags. + * XXX: atomic? + */ + switch (buffer->fault[fi]->access) { + case 0: /* READ. */ + hmm_flags = HMM_PFN_REQ_FAULT; + break; + case 3: /* PREFETCH. */ + hmm_flags = 0; + break; + default: + hmm_flags = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE; + break; } - start = max_t(u64, start, vma->vm_start); - limit = min_t(u64, limit, vma->vm_end); - mmap_read_unlock(mm); - SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit); - if (buffer->fault[fi]->addr != start) { - SVMM_ERR(svmm, "addr %016llx", buffer->fault[fi]->addr); - mmput(mm); + mm = svmm->notifier.mm; + if (!mmget_not_zero(mm)) { nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]); continue; } - /* Prepare the GPU-side update of all pages within the - * fault window, determining required pages and access - * permissions based on pending faults. - */ - args.i.p.page = PAGE_SHIFT; - args.i.p.addr = start; - for (fn = fi, pi = 0;;) { - /* Determine required permissions based on GPU fault - * access flags. - *XXX: atomic? - */ - switch (buffer->fault[fn]->access) { - case 0: /* READ. */ - hmm_pfns[pi++] = HMM_PFN_REQ_FAULT; - break; - case 3: /* PREFETCH. */ - hmm_pfns[pi++] = 0; - break; - default: - hmm_pfns[pi++] = HMM_PFN_REQ_FAULT | - HMM_PFN_REQ_WRITE; - break; - } - args.i.p.size = pi << PAGE_SHIFT; + notifier.svmm = svmm; + ret = mmu_interval_notifier_insert(¬ifier.notifier, mm, + args.i.p.addr, args.i.p.size, + &nouveau_svm_mni_ops); + if (!ret) { + ret = nouveau_range_fault(svmm, svm->drm, &args, + sizeof(args), args.phys, hmm_flags, ¬ifier); + mmu_interval_notifier_remove(¬ifier.notifier); + } + mmput(mm); + for (fn = fi; ++fn < buffer->fault_nr; ) { /* It's okay to skip over duplicate addresses from the * same SVMM as faults are ordered by access type such * that only the first one needs to be handled. @@ -758,61 +731,21 @@ nouveau_svm_fault(struct nvif_notify *notify) * ie. WRITE faults appear first, thus any handling of * pending READ faults will already be satisfied. */ - while (++fn < buffer->fault_nr && - buffer->fault[fn]->svmm == svmm && - buffer->fault[fn ]->addr == - buffer->fault[fn - 1]->addr); - - /* If the next fault is outside the window, or all GPU - * faults have been dealt with, we're done here. - */ - if (fn >= buffer->fault_nr || - buffer->fault[fn]->svmm != svmm || + if (buffer->fault[fn]->svmm != svmm || buffer->fault[fn]->addr >= limit) break; - - /* Fill in the gap between this fault and the next. */ - fill = (buffer->fault[fn ]->addr - - buffer->fault[fn - 1]->addr) >> PAGE_SHIFT; - while (--fill) - hmm_pfns[pi++] = 0; } - SVMM_DBG(svmm, "wndw %016llx-%016llx covering %d fault(s)", - args.i.p.addr, - args.i.p.addr + args.i.p.size, fn - fi); - - notifier.svmm = svmm; - ret = mmu_interval_notifier_insert(¬ifier.notifier, - svmm->notifier.mm, - args.i.p.addr, args.i.p.size, - &nouveau_svm_mni_ops); - if (!ret) { - ret = nouveau_range_fault( - svmm, svm->drm, &args, - sizeof(args.i) + pi * sizeof(args.phys[0]), - hmm_pfns, args.phys, ¬ifier); - mmu_interval_notifier_remove(¬ifier.notifier); - } - mmput(mm); + /* If handling failed completely, cancel all faults. */ + if (ret) { + while (fi < fn) { + struct nouveau_svm_fault *fault = + buffer->fault[fi++]; - /* Cancel any faults in the window whose pages didn't manage - * to keep their valid bit, or stay writeable when required. - * - * If handling failed completely, cancel all faults. - */ - while (fi < fn) { - struct nouveau_svm_fault *fault = buffer->fault[fi++]; - pi = (fault->addr - args.i.p.addr) >> PAGE_SHIFT; - if (ret || - !(args.phys[pi] & NVIF_VMM_PFNMAP_V0_V) || - (!(args.phys[pi] & NVIF_VMM_PFNMAP_V0_W) && - fault->access != 0 && fault->access != 3)) { nouveau_svm_fault_cancel_fault(svm, fault); - continue; } + } else replay++; - } } /* Issue fault replay to the GPU. */