Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id 3C234C3ABC0 for ; Thu, 8 May 2025 13:25:21 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id CF8D310E095; Thu, 8 May 2025 13:25:20 +0000 (UTC) Authentication-Results: gabe.freedesktop.org; dkim=pass (2048-bit key; unprotected) header.d=intel.com header.i=@intel.com header.b="jnmnerfu"; dkim-atps=neutral Received: from mgamail.intel.com (mgamail.intel.com [192.198.163.13]) by gabe.freedesktop.org (Postfix) with ESMTPS id 512AC10E095 for ; Thu, 8 May 2025 13:25:15 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=intel.com; i=@intel.com; q=dns/txt; s=Intel; t=1746710715; x=1778246715; h=message-id:subject:from:to:cc:date:in-reply-to: references:content-transfer-encoding:mime-version; bh=REK1nwahW0CyM1EF/1l2jo2p6Vb5m954ezPdDAmN8Uo=; b=jnmnerfu9k/AM594E6yxu+up6Ni3IrvXKgUQYNzP8OhN1Lf/TUm3oRI2 bmCcekP7XUWvnl9kYS4lKzcpaSEqyURIp0uCJhsR2a2Pa/xmnG5f9n9yS 1zVa3j7aUDkinuY4t/ls2EBRUFO2kNRZm/Tm2FOFdyJSkumqihzE/V4jI FXKvimxpkqHjSwYguEK7dwS3WHa3GArjIwA3wTgkRGn2h3asNTa4hyDQF /fFW/KXh15lT1wHoLXup4UiWeakeZzbne2Qj2wvOh+EXeROt8YfVspOyZ IaY7sezfriq6BcHW7SsSHelodnsRyys6y7p3q/c6JcADiGTmcMsMvD4IY A==; X-CSE-ConnectionGUID: z0OOl5gISNey6mxDSzftxA== X-CSE-MsgGUID: yYcYzk3TRT+kcqMNwl3qPA== X-IronPort-AV: E=McAfee;i="6700,10204,11426"; a="51154620" X-IronPort-AV: E=Sophos;i="6.15,272,1739865600"; d="scan'208";a="51154620" Received: from orviesa002.jf.intel.com ([10.64.159.142]) by fmvoesa107.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 08 May 2025 06:25:14 -0700 X-CSE-ConnectionGUID: Ue11/2EBR5GKvc7dc3rbJw== X-CSE-MsgGUID: Ot3LYgGGTyeV8hHkRU+qwQ== X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="6.15,272,1739865600"; d="scan'208";a="167230457" Received: from uaeoff-desk2.amr.corp.intel.com (HELO [10.124.222.207]) ([10.124.222.207]) by orviesa002-auth.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 08 May 2025 06:25:11 -0700 Message-ID: Subject: Re: [PATCH v6 2/5] drm/xe: Strict migration policy for atomic SVM faults From: Thomas =?ISO-8859-1?Q?Hellstr=F6m?= To: Matthew Brost Cc: intel-xe@lists.freedesktop.org, himal.prasad.ghimiray@intel.com Date: Thu, 08 May 2025 15:25:08 +0200 In-Reply-To: References: <20250506230132.1188732-1-matthew.brost@intel.com> <20250506230132.1188732-3-matthew.brost@intel.com> <7a7385b2d2a81961963e4df4a34ad4c55bd49bfb.camel@linux.intel.com> Organization: Intel Sweden AB, Registration Number: 556189-6027 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: quoted-printable User-Agent: Evolution 3.54.3 (3.54.3-1.fc41) MIME-Version: 1.0 X-BeenThere: intel-xe@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Intel Xe graphics driver List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: intel-xe-bounces@lists.freedesktop.org Sender: "Intel-xe" Hi, Matthew. On Wed, 2025-05-07 at 08:41 -0700, Matthew Brost wrote: > On Wed, May 07, 2025 at 02:01:56PM +0200, Thomas Hellstr=C3=B6m wrote: > >=20 > > This is v7, right? (Says v6 in the subject). > >=20 >=20 > Yea, I somehow mismatched version in change log and patches sent. >=20 > Ok, getting a little bogged down on comments, let try fix them inline > in > this reply to make sure they look good before sending next rev. So since we're not adhering to the drm locking guidelines, I want to ensure that what we ask to PULL is at least properly documented. So what I've been asking here is two things. 1) If the intended usage of a function is advisory / opportunistic, please document that at the function header rather than as a comment within the function definition. 2) Document what a barrier / READ_ONCE / WRITE_ONCE pairs with. Ideally pointing to the specific function. In particular for 1) I've mentioned it a couple of times without it beeing addressed or commented on so that's why I'm bringing it up again in case it was missed. So let's check. >=20 > > On Tue, 2025-05-06 at 16:01 -0700, Matthew Brost wrote: > > > Mixing GPU and CPU atomics does not work unless a strict > > > migration > > > policy of GPU atomics must be device memory. Enforce a policy of > > > must > > > be > > > in VRAM with a retry loop of 3 attempts, if retry loop fails > > > abort > > > fault. > > >=20 > > > Removing always_migrate_to_vram modparam as we now have real > > > migration > > > policy. > > >=20 > > > v2: > > > =C2=A0- Only retry migration on atomics > > > =C2=A0- Drop alway migrate modparam > > > v3: > > > =C2=A0- Only set vram_only on DGFX (Himal) > > > =C2=A0- Bail on get_pages failure if vram_only and retry count > > > exceeded > > > (Himal) > > > =C2=A0- s/vram_only/devmem_only > > > =C2=A0- Update xe_svm_range_is_valid to accept devmem_only argument > > > v4: > > > =C2=A0- Fix logic bug get_pages failure > > > v5: > > > =C2=A0- Fix commit message (Himal) > > > =C2=A0- Mention removing always_migrate_to_vram in commit message > > > (Lucas) > > > =C2=A0- Fix xe_svm_range_is_valid to check for devmem pages > > > =C2=A0- Bail on devmem_only && !migrate_devmem (Thomas) > > > v6: > > > =C2=A0- Add READ_ONCE barriers for opportunistic checks (Thomas) > > > v7: > > > =C2=A0- Pair READ_ONCE with WRITE_ONCE (Thomas) > > >=20 > > > Fixes: 2f118c949160 ("drm/xe: Add SVM VRAM migration") > > > Cc: stable@vger.kernel.org > > > Signed-off-by: Himal Prasad Ghimiray > > > > > > Signed-off-by: Matthew Brost > > > Acked-by: Himal Prasad Ghimiray > > > --- > > > =C2=A0drivers/gpu/drm/drm_gpusvm.c=C2=A0=C2=A0 |=C2=A0 23 +++++-- > > > =C2=A0drivers/gpu/drm/xe/xe_module.c |=C2=A0=C2=A0 3 - > > > =C2=A0drivers/gpu/drm/xe/xe_module.h |=C2=A0=C2=A0 1 - > > > =C2=A0drivers/gpu/drm/xe/xe_pt.c=C2=A0=C2=A0=C2=A0=C2=A0 |=C2=A0 14 += +++- > > > =C2=A0drivers/gpu/drm/xe/xe_svm.c=C2=A0=C2=A0=C2=A0 | 109 +++++++++++= ++++++++++++++-- > > > ---- > > > -- > > > =C2=A0drivers/gpu/drm/xe/xe_svm.h=C2=A0=C2=A0=C2=A0 |=C2=A0=C2=A0 5 -= - > > > =C2=A0include/drm/drm_gpusvm.h=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0 |= =C2=A0 40 +++++++----- > > > =C2=A07 files changed, 138 insertions(+), 57 deletions(-) > > >=20 > > > diff --git a/drivers/gpu/drm/drm_gpusvm.c > > > b/drivers/gpu/drm/drm_gpusvm.c > > > index a58d03e6cac2..41f6616bcf76 100644 > > > --- a/drivers/gpu/drm/drm_gpusvm.c > > > +++ b/drivers/gpu/drm/drm_gpusvm.c > > > @@ -1118,6 +1118,10 @@ static void > > > __drm_gpusvm_range_unmap_pages(struct drm_gpusvm *gpusvm, > > > =C2=A0 lockdep_assert_held(&gpusvm->notifier_lock); > > > =C2=A0 > > > =C2=A0 if (range->flags.has_dma_mapping) { > > > + struct drm_gpusvm_range_flags flags =3D { > > > + .__flags =3D range->flags.__flags, > > > + }; > > > + > > > =C2=A0 for (i =3D 0, j =3D 0; i < npages; j++) { > > > =C2=A0 struct drm_pagemap_device_addr *addr =3D > > > &range->dma_addr[j]; > > > =C2=A0 > > > @@ -1131,8 +1135,12 @@ static void > > > __drm_gpusvm_range_unmap_pages(struct drm_gpusvm *gpusvm, > > > =C2=A0 =C2=A0=C2=A0=C2=A0 dev, > > > *addr); > > > =C2=A0 i +=3D 1 << addr->order; > > > =C2=A0 } > > > - range->flags.has_devmem_pages =3D false; > > > - range->flags.has_dma_mapping =3D false; > > > + > > > + /* WRITE_ONCE pairs with READ_ONCE for > > > opportunistic > > > checks */ >=20 > Is this one ok? Can't really quote driver side here... Yeah makes sense. >=20 > > > + flags.has_devmem_pages =3D false; > > > + flags.has_dma_mapping =3D false; > > > + WRITE_ONCE(range->flags.__flags, flags.__flags); > > > + > > > =C2=A0 range->dpagemap =3D NULL; > > > =C2=A0 } > > > =C2=A0} > > > @@ -1334,6 +1342,7 @@ int drm_gpusvm_range_get_pages(struct > > > drm_gpusvm *gpusvm, > > > =C2=A0 int err =3D 0; > > > =C2=A0 struct dev_pagemap *pagemap; > > > =C2=A0 struct drm_pagemap *dpagemap; > > > + struct drm_gpusvm_range_flags flags; > > > =C2=A0 > > > =C2=A0retry: > > > =C2=A0 hmm_range.notifier_seq =3D > > > mmu_interval_read_begin(notifier); > > > @@ -1378,7 +1387,8 @@ int drm_gpusvm_range_get_pages(struct > > > drm_gpusvm *gpusvm, > > > =C2=A0 */ > > > =C2=A0 drm_gpusvm_notifier_lock(gpusvm); > > > =C2=A0 > > > - if (range->flags.unmapped) { > > > + flags.__flags =3D range->flags.__flags; > > > + if (flags.unmapped) { > > > =C2=A0 drm_gpusvm_notifier_unlock(gpusvm); > > > =C2=A0 err =3D -EFAULT; > > > =C2=A0 goto err_free; > > > @@ -1474,14 +1484,17 @@ int drm_gpusvm_range_get_pages(struct > > > drm_gpusvm *gpusvm, > > > =C2=A0 } > > > =C2=A0 i +=3D 1 << order; > > > =C2=A0 num_dma_mapped =3D i; > > > - range->flags.has_dma_mapping =3D true; > > > + flags.has_dma_mapping =3D true; > > > =C2=A0 } > > > =C2=A0 > > > =C2=A0 if (zdd) { > > > - range->flags.has_devmem_pages =3D true; > > > + flags.has_devmem_pages =3D true; > > > =C2=A0 range->dpagemap =3D dpagemap; > > > =C2=A0 } > > > =C2=A0 > > > + /* WRITE_ONCE pairs with READ_ONCE for opportunistic > > > checks > > > */ >=20 > Same as above. Makes sense. >=20 > > > + WRITE_ONCE(range->flags.__flags, flags.__flags); > > > + > > > =C2=A0 drm_gpusvm_notifier_unlock(gpusvm); > > > =C2=A0 kvfree(pfns); > > > =C2=A0set_seqno: > > > diff --git a/drivers/gpu/drm/xe/xe_module.c > > > b/drivers/gpu/drm/xe/xe_module.c > > > index 05c7d0ae6d83..1c4dfafbcd0b 100644 > > > --- a/drivers/gpu/drm/xe/xe_module.c > > > +++ b/drivers/gpu/drm/xe/xe_module.c > > > @@ -33,9 +33,6 @@ struct xe_modparam xe_modparam =3D { > > > =C2=A0module_param_named(svm_notifier_size, > > > xe_modparam.svm_notifier_size, > > > uint, 0600); > > > =C2=A0MODULE_PARM_DESC(svm_notifier_size, "Set the svm notifier > > > size(in > > > MiB), must be power of 2"); > > > =C2=A0 > > > -module_param_named(always_migrate_to_vram, > > > xe_modparam.always_migrate_to_vram, bool, 0444); > > > -MODULE_PARM_DESC(always_migrate_to_vram, "Always migrate to VRAM > > > on > > > GPU fault"); > > > - > > > =C2=A0module_param_named_unsafe(force_execlist, > > > xe_modparam.force_execlist, bool, 0444); > > > =C2=A0MODULE_PARM_DESC(force_execlist, "Force Execlist submission"); > > > =C2=A0 > > > diff --git a/drivers/gpu/drm/xe/xe_module.h > > > b/drivers/gpu/drm/xe/xe_module.h > > > index 84339e509c80..5a3bfea8b7b4 100644 > > > --- a/drivers/gpu/drm/xe/xe_module.h > > > +++ b/drivers/gpu/drm/xe/xe_module.h > > > @@ -12,7 +12,6 @@ > > > =C2=A0struct xe_modparam { > > > =C2=A0 bool force_execlist; > > > =C2=A0 bool probe_display; > > > - bool always_migrate_to_vram; > > > =C2=A0 u32 force_vram_bar_size; > > > =C2=A0 int guc_log_level; > > > =C2=A0 char *guc_firmware_path; > > > diff --git a/drivers/gpu/drm/xe/xe_pt.c > > > b/drivers/gpu/drm/xe/xe_pt.c > > > index b42cf5d1b20c..b487e094178d 100644 > > > --- a/drivers/gpu/drm/xe/xe_pt.c > > > +++ b/drivers/gpu/drm/xe/xe_pt.c > > > @@ -2270,11 +2270,19 @@ static void op_commit(struct xe_vm *vm, > > > =C2=A0 } > > > =C2=A0 case DRM_GPUVA_OP_DRIVER: > > > =C2=A0 { > > > + /* WRITE_ONCE pairs with READ_ONCE opportunistic > > > checks */ > > > + >=20 > /* WRITE_ONCE pairs with READ_ONCE in xe_svm.c */ Fine. >=20 > > > =C2=A0 if (op->subop =3D=3D XE_VMA_SUBOP_MAP_RANGE) { > > > - op->map_range.range->tile_present |=3D > > > BIT(tile->id); > > > - op->map_range.range->tile_invalidated &=3D > > > ~BIT(tile->id); > > > + WRITE_ONCE(op->map_range.range- > > > > tile_present, > > > + =C2=A0=C2=A0 op->map_range.range- > > > >tile_present > > > >=20 > > > + =C2=A0=C2=A0 BIT(tile->id)); > > > + WRITE_ONCE(op->map_range.range- > > > > tile_invalidated, > > > + =C2=A0=C2=A0 op->map_range.range- > > > > tile_invalidated & > > > + =C2=A0=C2=A0 ~BIT(tile->id)); > > > =C2=A0 } else if (op->subop =3D=3D > > > XE_VMA_SUBOP_UNMAP_RANGE) { > > > - op->unmap_range.range->tile_present &=3D > > > ~BIT(tile->id); > > > + WRITE_ONCE(op->unmap_range.range- > > > > tile_present, > > > + =C2=A0=C2=A0 op->unmap_range.range- > > > > tile_present & > > > + =C2=A0=C2=A0 ~BIT(tile->id)); > > > =C2=A0 } > > > =C2=A0 break; > > > =C2=A0 } > > > diff --git a/drivers/gpu/drm/xe/xe_svm.c > > > b/drivers/gpu/drm/xe/xe_svm.c > > > index 890f6b2f40e9..8c0ef55cfd57 100644 > > > --- a/drivers/gpu/drm/xe/xe_svm.c > > > +++ b/drivers/gpu/drm/xe/xe_svm.c > > > @@ -16,8 +16,15 @@ > > > =C2=A0 > >=20 > > So what I meant in the review comments is something along > >=20 > > /* Advisory only check whether the range is currently backed by > > VRAM > > memory. */ > >=20 > > (As said we shouldn need to look at the function implementation to > > understand that it's opportunistic). > >=20 > > > =C2=A0static bool xe_svm_range_in_vram(struct xe_svm_range *range) > > > =C2=A0{ > > > - /* Not reliable without notifier lock */ > > > - return range->base.flags.has_devmem_pages; > > > + /* > > > + * Not reliable without notifier lock, opportunistic > > > only, > > > pairs with > > > + * WRITE_ONCE > >=20 > > So if we assume that this function should is never be called with > > the > > notifier lock held, we can leave that comment out, just > >=20 > > /* Pairs with WRITE_ONCE() in xe_pt.c */=20 > >=20 > >=20 > > And similar for the other places. > >=20 > > Thanks, > > Thomas > >=20 > > > + */ >=20 > /* > =C2=A0* Advisory only check whether the range is currently backed by VRAM > =C2=A0* memory, READ_ONCE pairs with WRITE_ONCE in drm_gpusvm.c > =C2=A0*/ >=20 Here I'd put the READ_ONCE part of the comment where it's actually used. > > > + struct drm_gpusvm_range_flags flags =3D { /* Pairs with WRITE_ONCE() in drm_gpusvm.c */ > > > + .__flags =3D READ_ONCE(range->base.flags.__flags), > > > + }; > > > + > > > + return flags.has_devmem_pages; > > > =C2=A0} > > > =C2=A0 > > > =C2=A0static bool xe_svm_range_has_vram_binding(struct xe_svm_range > > > *range) > > > @@ -650,9 +657,16 @@ void xe_svm_fini(struct xe_vm *vm) > > > =C2=A0} > > > =C2=A0 > > > =C2=A0static bool xe_svm_range_is_valid(struct xe_svm_range *range, > > > - =C2=A0 struct xe_tile *tile) > > > + =C2=A0 struct xe_tile *tile, > > > + =C2=A0 bool devmem_only) > > > =C2=A0{ > > > - return (range->tile_present & ~range->tile_invalidated) > > > & > > > BIT(tile->id); > > > + /* > > > + * Not reliable without notifier lock, opportunistic > > > only, > > > pairs with > > > + * WRITE_ONCE > > > + */ >=20 > /* > =C2=A0* Advisory only check whether the range currently has a valid > mapping, > =C2=A0* READ_ONCE pairs with WRITE_ONCE in xe_pt.c > =C2=A0*/ >=20 > > > + return ((READ_ONCE(range->tile_present) & > > > + ~READ_ONCE(range->tile_invalidated)) & > > > BIT(tile- > > > > id)) && > > > + (!devmem_only || xe_svm_range_in_vram(range)); > > > =C2=A0} > > > =C2=A0 > > > =C2=A0#if IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR) > > > @@ -726,6 +740,36 @@ static int xe_svm_alloc_vram(struct xe_vm > > > *vm,static bool supports_4K_migration(struct xe_device *xe) > > > +{ > > > + if (xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K) > > > + return false; > > > + > > > + return true; > > > +} > > > + > > > +static bool xe_svm_range_needs_migrate_to_vram(struct > > > xe_svm_range > > > *range, > > > + =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0 struct xe_vma > > > *vma) > > > +{ > > > + struct xe_vm *vm =3D range_to_vm(&range->base); > > > + u64 range_size =3D xe_svm_range_size(range); > > > + > > > + if (!range->base.flags.migrate_devmem) > > > + return false; > > > + > > > + /* Not reliable without notifier lock, opportunistic > > > only */ >=20 > Drop this one as xe_svm_range_in_vram has a comment. Yes, Thanks, Thomas >=20 > Matt >=20 > > > + if (xe_svm_range_in_vram(range)) { > > > + drm_dbg(&vm->xe->drm, "Range is already in > > > VRAM\n"); > > > + return false; > > > + } > > > + > > > + if (range_size <=3D SZ_64K && !supports_4K_migration(vm- > > > >xe)) > > > { > > > + drm_dbg(&vm->xe->drm, "Platform doesn't support > > > SZ_4K range migration\n"); > > > + return false; > > > + } > > > + > > > + return true; > > > +} > > > =C2=A0 > > > =C2=A0/** > > > =C2=A0 * xe_svm_handle_pagefault() - SVM handle page fault > > > @@ -750,12 +794,15 @@ int xe_svm_handle_pagefault(struct xe_vm > > > *vm, > > > struct xe_vma *vma, > > > =C2=A0 IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR), > > > =C2=A0 .check_pages_threshold =3D IS_DGFX(vm->xe) && > > > =C2=A0 IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR) > > > ? > > > SZ_64K : 0, > > > + .devmem_only =3D atomic && IS_DGFX(vm->xe) && > > > + IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR), > > > =C2=A0 }; > > > =C2=A0 struct xe_svm_range *range; > > > =C2=A0 struct drm_gpusvm_range *r; > > > =C2=A0 struct drm_exec exec; > > > =C2=A0 struct dma_fence *fence; > > > =C2=A0 struct xe_tile *tile =3D gt_to_tile(gt); > > > + int migrate_try_count =3D ctx.devmem_only ? 3 : 1; > > > =C2=A0 ktime_t end =3D 0; > > > =C2=A0 int err; > > > =C2=A0 > > > @@ -776,24 +823,30 @@ int xe_svm_handle_pagefault(struct xe_vm > > > *vm, > > > struct xe_vma *vma, > > > =C2=A0 if (IS_ERR(r)) > > > =C2=A0 return PTR_ERR(r); > > > =C2=A0 > > > + if (ctx.devmem_only && !r->flags.migrate_devmem) > > > + return -EACCES; > > > + > > > =C2=A0 range =3D to_xe_range(r); > > > - if (xe_svm_range_is_valid(range, tile)) > > > + if (xe_svm_range_is_valid(range, tile, ctx.devmem_only)) > > > =C2=A0 return 0; > > > =C2=A0 > > > =C2=A0 range_debug(range, "PAGE FAULT"); > > > =C2=A0 > > > - /* XXX: Add migration policy, for now migrate range once > > > */ > > > - if (!range->skip_migrate && range- > > > >base.flags.migrate_devmem > > > && > > > - =C2=A0=C2=A0=C2=A0 xe_svm_range_size(range) >=3D SZ_64K) { > > > - range->skip_migrate =3D true; > > > - > > > + if (--migrate_try_count >=3D 0 && > > > + =C2=A0=C2=A0=C2=A0 xe_svm_range_needs_migrate_to_vram(range, vma)) = { > > > =C2=A0 err =3D xe_svm_alloc_vram(vm, tile, range, &ctx); > > > =C2=A0 if (err) { > > > - drm_dbg(&vm->xe->drm, > > > - "VRAM allocation failed, falling > > > back to " > > > - "retrying fault, asid=3D%u, > > > errno=3D%pe\n", > > > - vm->usm.asid, ERR_PTR(err)); > > > - goto retry; > > > + if (migrate_try_count || > > > !ctx.devmem_only) { > > > + drm_dbg(&vm->xe->drm, > > > + "VRAM allocation failed, > > > falling back to retrying fault, asid=3D%u, errno=3D%pe\n", > > > + vm->usm.asid, > > > ERR_PTR(err)); > > > + goto retry; > > > + } else { > > > + drm_err(&vm->xe->drm, > > > + "VRAM allocation failed, > > > retry count exceeded, asid=3D%u, errno=3D%pe\n", > > > + vm->usm.asid, > > > ERR_PTR(err)); > > > + return err; > > > + } > > > =C2=A0 } > > > =C2=A0 } > > > =C2=A0 > > > @@ -801,15 +854,22 @@ int xe_svm_handle_pagefault(struct xe_vm > > > *vm, > > > struct xe_vma *vma, > > > =C2=A0 err =3D drm_gpusvm_range_get_pages(&vm->svm.gpusvm, r, > > > &ctx); > > > =C2=A0 /* Corner where CPU mappings have changed */ > > > =C2=A0 if (err =3D=3D -EOPNOTSUPP || err =3D=3D -EFAULT || err =3D=3D= - > > > EPERM) { > > > - if (err =3D=3D -EOPNOTSUPP) { > > > - range_debug(range, "PAGE FAULT - EVICT > > > PAGES"); > > > - drm_gpusvm_range_evict(&vm->svm.gpusvm, > > > &range->base); > > > + if (migrate_try_count > 0 || !ctx.devmem_only) { > > > + if (err =3D=3D -EOPNOTSUPP) { > > > + range_debug(range, "PAGE FAULT - > > > EVICT PAGES"); > > > + drm_gpusvm_range_evict(&vm- > > > > svm.gpusvm, > > > + =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0 &range- > > > > base); > > > + } > > > + drm_dbg(&vm->xe->drm, > > > + "Get pages failed, falling back > > > to > > > retrying, asid=3D%u, gpusvm=3D%p, errno=3D%pe\n", > > > + vm->usm.asid, &vm->svm.gpusvm, > > > ERR_PTR(err)); > > > + range_debug(range, "PAGE FAULT - RETRY > > > PAGES"); > > > + goto retry; > > > + } else { > > > + drm_err(&vm->xe->drm, > > > + "Get pages failed, retry count > > > exceeded, asid=3D%u, gpusvm=3D%p, errno=3D%pe\n", > > > + vm->usm.asid, &vm->svm.gpusvm, > > > ERR_PTR(err)); > > > =C2=A0 } > > > - drm_dbg(&vm->xe->drm, > > > - "Get pages failed, falling back to > > > retrying, > > > asid=3D%u, gpusvm=3D%p, errno=3D%pe\n", > > > - vm->usm.asid, &vm->svm.gpusvm, > > > ERR_PTR(err)); > > > - range_debug(range, "PAGE FAULT - RETRY PAGES"); > > > - goto retry; > > > =C2=A0 } > > > =C2=A0 if (err) { > > > =C2=A0 range_debug(range, "PAGE FAULT - FAIL PAGE > > > COLLECT"); > > > @@ -843,9 +903,6 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, > > > struct xe_vma *vma, > > > =C2=A0 } > > > =C2=A0 drm_exec_fini(&exec); > > > =C2=A0 > > > - if (xe_modparam.always_migrate_to_vram) > > > - range->skip_migrate =3D false; > > > - > > > =C2=A0 dma_fence_wait(fence, false); > > > =C2=A0 dma_fence_put(fence); > > > =C2=A0 > > > diff --git a/drivers/gpu/drm/xe/xe_svm.h > > > b/drivers/gpu/drm/xe/xe_svm.h > > > index 3d441eb1f7ea..0e1f376a7471 100644 > > > --- a/drivers/gpu/drm/xe/xe_svm.h > > > +++ b/drivers/gpu/drm/xe/xe_svm.h > > > @@ -39,11 +39,6 @@ struct xe_svm_range { > > > =C2=A0 * range. Protected by GPU SVM notifier lock. > > > =C2=A0 */ > > > =C2=A0 u8 tile_invalidated; > > > - /** > > > - * @skip_migrate: Skip migration to VRAM, protected by > > > GPU > > > fault handler > > > - * locking. > > > - */ > > > - u8 skip_migrate :1; > > > =C2=A0}; > > > =C2=A0 > > > =C2=A0/** > > > diff --git a/include/drm/drm_gpusvm.h b/include/drm/drm_gpusvm.h > > > index 9fd25fc880a4..653d48dbe1c1 100644 > > > --- a/include/drm/drm_gpusvm.h > > > +++ b/include/drm/drm_gpusvm.h > > > @@ -185,6 +185,31 @@ struct drm_gpusvm_notifier { > > > =C2=A0 } flags; > > > =C2=A0}; > > > =C2=A0 > > > +/** > > > + * struct drm_gpusvm_range_flags - Structure representing a GPU > > > SVM > > > range flags > > > + * > > > + * @migrate_devmem: Flag indicating whether the range can be > > > migrated to device memory > > > + * @unmapped: Flag indicating if the range has been unmapped > > > + * @partial_unmap: Flag indicating if the range has been > > > partially > > > unmapped > > > + * @has_devmem_pages: Flag indicating if the range has devmem > > > pages > > > + * @has_dma_mapping: Flag indicating if the range has a DMA > > > mapping > > > + * @__flags: Flags for range in u16 form (used for READ_ONCE) > > > + */ > > > +struct drm_gpusvm_range_flags { > > > + union { > > > + struct { > > > + /* All flags below must be set upon > > > creation > > > */ > > > + u16 migrate_devmem : 1; > > > + /* All flags below must be set / cleared > > > under notifier lock */ > > > + u16 unmapped : 1; > > > + u16 partial_unmap : 1; > > > + u16 has_devmem_pages : 1; > > > + u16 has_dma_mapping : 1; > > > + }; > > > + u16 __flags; > > > + }; > > > +}; > > > + > > > =C2=A0/** > > > =C2=A0 * struct drm_gpusvm_range - Structure representing a GPU SVM > > > range > > > =C2=A0 * > > > @@ -198,11 +223,6 @@ struct drm_gpusvm_notifier { > > > =C2=A0 * @dpagemap: The struct drm_pagemap of the device pages we're > > > dma- > > > mapping. > > > =C2=A0 *=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0= =C2=A0 Note this is assuming only one drm_pagemap per > > > range > > > is allowed. > > > =C2=A0 * @flags: Flags for range > > > - * @flags.migrate_devmem: Flag indicating whether the range can > > > be > > > migrated to device memory > > > - * @flags.unmapped: Flag indicating if the range has been > > > unmapped > > > - * @flags.partial_unmap: Flag indicating if the range has been > > > partially unmapped > > > - * @flags.has_devmem_pages: Flag indicating if the range has > > > devmem > > > pages > > > - * @flags.has_dma_mapping: Flag indicating if the range has a > > > DMA > > > mapping > > > =C2=A0 * > > > =C2=A0 * This structure represents a GPU SVM range used for tracking > > > memory ranges > > > =C2=A0 * mapped in a DRM device. > > > @@ -216,15 +236,7 @@ struct drm_gpusvm_range { > > > =C2=A0 unsigned long notifier_seq; > > > =C2=A0 struct drm_pagemap_device_addr *dma_addr; > > > =C2=A0 struct drm_pagemap *dpagemap; > > > - struct { > > > - /* All flags below must be set upon creation */ > > > - u16 migrate_devmem : 1; > > > - /* All flags below must be set / cleared under > > > notifier lock */ > > > - u16 unmapped : 1; > > > - u16 partial_unmap : 1; > > > - u16 has_devmem_pages : 1; > > > - u16 has_dma_mapping : 1; > > > - } flags; > > > + struct drm_gpusvm_range_flags flags; > > > =C2=A0}; > > > =C2=A0 > > > =C2=A0/** > >=20