From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from BL2PR02CU003.outbound.protection.outlook.com (mail-eastusazon11011035.outbound.protection.outlook.com [52.101.52.35]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.subspace.kernel.org (Postfix) with ESMTPS id 544542DE6E3 for ; Mon, 20 Apr 2026 16:05:54 +0000 (UTC) Authentication-Results: smtp.subspace.kernel.org; arc=fail smtp.client-ip=52.101.52.35 ARC-Seal:i=2; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1776701156; cv=fail; b=ASUZ8RL3KyRphpgGk3CIT6Y2iEqMa8ruIN0b9gaPfUOb0Un5zf+oSl4djygFcAdRmLHBqrfmcGM+fnXtckrOED4vC75lwQ8yQbwe0ZbpKK2vj0zSKXRGLqG1RLK9P7teUC1mam+6oPnoyvcJPXQMcDFdjX8Gix6gSSV9dQ6d/Ck= ARC-Message-Signature:i=2; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1776701156; c=relaxed/simple; bh=h7j76n5uSwgy8l7m36xibdqWCfvRDBQNp7KJFyGGbGM=; h=Message-ID:Date:MIME-Version:Subject:To:CC:References:From: In-Reply-To:Content-Type; b=V/PSnq3Su/8OhHfCnuTJQR/cpHVR87wwhUeuVnSwb1OHvV3+G9m9JIBhdiPuex1i2nVj7f91mZY5zNzjpBWX+iVADKk8UUb2jsyWolQbfxd4rraZPy6V145Z/5VoCIUZ3AnvsE4v1K9hjlEot7TLoAZLYdqx4sSxFP9z+fWQNz4= ARC-Authentication-Results:i=2; smtp.subspace.kernel.org; dmarc=pass (p=quarantine dis=none) header.from=amd.com; spf=fail smtp.mailfrom=amd.com; dkim=pass (1024-bit key) header.d=amd.com header.i=@amd.com header.b=ot/hrhDr; arc=fail smtp.client-ip=52.101.52.35 Authentication-Results: smtp.subspace.kernel.org; dmarc=pass (p=quarantine dis=none) header.from=amd.com Authentication-Results: smtp.subspace.kernel.org; spf=fail smtp.mailfrom=amd.com Authentication-Results: smtp.subspace.kernel.org; dkim=pass (1024-bit key) header.d=amd.com header.i=@amd.com header.b="ot/hrhDr" ARC-Seal: i=1; a=rsa-sha256; s=arcselector10001; d=microsoft.com; cv=none; b=W1GfIVHxNTB8nUe/KRMbPrND+2onO+tm+ZzD3JGprqxWMnDtwT7pD5kFHl55WIk7NPTDDm3uIH73rxvqhcd26r7piOqAp7SwoxVCd3kBK4BQghoHHNawup4Vr2q3f3rYUzxWY9xHijgaDh5QR1h+mx4jMkSsUN9noYPexOPVYqzJRroVk+wwg8KOEoKp2mQ3RKoK8KMHlFO4dd6Z60erw0XvgxTgWJVKli1A5xN5eIuPzy4gVkpi+zD/nSgLQ3uURTGK8Zt/zQ8t1v0NLdGesAe/nCA82nLSM1SmZc6bfGfE3gxmpAtOYMqCA4lB1OnCXjJmRM74Zag2TB47I5+5KQ== ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=microsoft.com; s=arcselector10001; h=From:Date:Subject:Message-ID:Content-Type:MIME-Version:X-MS-Exchange-AntiSpam-MessageData-ChunkCount:X-MS-Exchange-AntiSpam-MessageData-0:X-MS-Exchange-AntiSpam-MessageData-1; bh=o51zM9jktn1FMWaQeMyLV73lDnXdXfaYzqo7z3cwpm8=; b=tLuVTYF+hz4m2om1kwW1KGI1QpIbZo/SnjDrziXmWkmqP1i78ca1BOgWq53RxKKXs4HPD8hpkAPC1vayJ3GhGs3JK7zeXAc6XeCraO5/pQ3iaffrhXaxXjjUO7qD8MHhsr1eVLq3VhbeXGNQkdwckuN08otHNOG7l2e2aIG1YQErxzkqQ0MRpx0X9GllwMxIHcHD3J7MPc9jE2pap8kfu7H9tvxY8QfYzYTErjcB6xJvkgiHy8Ssxg++KtaSVjE8WaK9G8eyOpKZusCB/rUNQOKF06xCQz7ed4FlX4a81L1tNgi+tknEM30SuGX4QUyWYew9zXBkS17uUlarX1vrtQ== ARC-Authentication-Results: i=1; mx.microsoft.com 1; spf=pass (sender ip is 165.204.84.17) smtp.rcpttodomain=kernel.org smtp.mailfrom=amd.com; dmarc=pass (p=quarantine sp=quarantine pct=100) action=none header.from=amd.com; dkim=none (message not signed); arc=none (0) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=amd.com; s=selector1; h=From:Date:Subject:Message-ID:Content-Type:MIME-Version:X-MS-Exchange-SenderADCheck; bh=o51zM9jktn1FMWaQeMyLV73lDnXdXfaYzqo7z3cwpm8=; b=ot/hrhDrQTanQMm/Zl27si07Ca7cy93YGy9HYSFA8AwFZ1g7T60Isu6KxSMFtdCEIEzAN/3iKxeot3exkA28NQmjuljp7O0x/4XahYxocFyvMB+D33kyNegxIrf7uJioesABT33krThCucFp7VB8VpagGID8neo+1wC/0m2cUgE= Received: from SN7PR04CA0182.namprd04.prod.outlook.com (2603:10b6:806:126::7) by MN0PR12MB5812.namprd12.prod.outlook.com (2603:10b6:208:378::11) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id 15.20.9846.15; Mon, 20 Apr 2026 16:05:51 +0000 Received: from SN1PEPF00026368.namprd02.prod.outlook.com (2603:10b6:806:126:cafe::30) by SN7PR04CA0182.outlook.office365.com (2603:10b6:806:126::7) with Microsoft SMTP Server (version=TLS1_3, cipher=TLS_AES_256_GCM_SHA384) id 15.20.9791.48 via Frontend Transport; Mon, 20 Apr 2026 16:05:51 +0000 X-MS-Exchange-Authentication-Results: spf=pass (sender IP is 165.204.84.17) smtp.mailfrom=amd.com; dkim=none (message not signed) header.d=none;dmarc=pass action=none header.from=amd.com; Received-SPF: Pass (protection.outlook.com: domain of amd.com designates 165.204.84.17 as permitted sender) receiver=protection.outlook.com; client-ip=165.204.84.17; helo=satlexmb08.amd.com; pr=C Received: from satlexmb08.amd.com (165.204.84.17) by SN1PEPF00026368.mail.protection.outlook.com (10.167.241.133) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id 15.20.9791.48 via Frontend Transport; Mon, 20 Apr 2026 16:05:51 +0000 Received: from Satlexmb09.amd.com (10.181.42.218) by satlexmb08.amd.com (10.181.42.217) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id 15.2.2562.17; Mon, 20 Apr 2026 11:05:51 -0500 Received: from satlexmb08.amd.com (10.181.42.217) by satlexmb09.amd.com (10.181.42.218) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id 15.2.2562.17; Mon, 20 Apr 2026 09:05:50 -0700 Received: from [172.19.71.207] (10.180.168.240) by satlexmb08.amd.com (10.181.42.217) with Microsoft SMTP Server id 15.2.2562.17 via Frontend Transport; Mon, 20 Apr 2026 11:05:50 -0500 Message-ID: Date: Mon, 20 Apr 2026 09:05:50 -0700 Precedence: bulk X-Mailing-List: linux-kernel@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101 Thunderbird/91.11.0 Subject: Re: [PATCH V1] accel/amdxdna: Add carveout memory support for non-IOMMU systems Content-Language: en-US To: Mario Limonciello , , , , CC: Max Zhen , , References: <20260417210621.1173841-1-lizhi.hou@amd.com> From: Lizhi Hou In-Reply-To: Content-Type: text/plain; charset="UTF-8"; format=flowed Content-Transfer-Encoding: 8bit X-EOPAttributedMessage: 0 X-MS-PublicTrafficType: Email X-MS-TrafficTypeDiagnostic: SN1PEPF00026368:EE_|MN0PR12MB5812:EE_ X-MS-Office365-Filtering-Correlation-Id: ec7aac79-a4fd-4d51-012f-08de9ef6b1c6 X-MS-Exchange-SenderADCheck: 1 X-MS-Exchange-AntiSpam-Relay: 0 X-Microsoft-Antispam: BCL:0;ARA:13230040|36860700016|376014|1800799024|82310400026|56012099003|22082099003|18002099003; X-Microsoft-Antispam-Message-Info: Ez0YW3etcqhyoK1HjGYOsAywkP6Bx7GtKYOkq9/cR9qJEeGwQ7R5cSqIbOukLiLmjO0Sy0bNmIOnbjbW0A+bSwuCY3mv0toLzUgK3Ouj2AOEo8Zicr6lB9GpsZ9bpw/KrA736XaxM6KLu1tFwR7CnJjlaFas2l4gYAAym7ftWdA4vD0pyn96HhII/LB0NxlgKE1MaE+Snb+A+/kKIUYWXXsZFmGRjBrakAyYuzLvY5G9xWE3GcVT+jXNjsUn+2lJE1Q8A9S9iG2hT6o7VEHM+EV1sxat7oTy91TgfC/XCOHnUUrQR/d+Uz8nUsZozSJSd7tRJo0SjcFYZPQh04z7NUHAUp0XRMnP3M6FTkIrFBE368aLkR1G3RDNkq+kq+RBbTc5GhuQgAp4/O/TYNW5uCL56vrVxHX7c1M8UCVFZpfGu9q4g5sLkP1oIKx5y3USaB5smxmol6enxIKC/I5KDcktU9QLTccYq1pdQBtbfs3YNhVO3BiNiy9sRcEF+aVotdkXocwqW8c0hJoFqzOGXGarNnX/zN2tOdNCDEZYxxrzVuZIdzFVbIO1ETqRTwIBr8tz22Bscxk7ULdUS1TblaLZUmC50Wxm4m4CkU+WBwxvuj8kIbHFhE/rWgJhMfzJdpKj8Jg3g/uw97oHhFcClQjmpMb5ki89Lt7jdYZj0bRT+as3UYQZ8lcjYR2Pato3ktToNGDFP/Cjfhbs6ddUMUcc1BDdSgCSZDO1OJGwOVl2qZTWQmOSp6bC8m4pOTrDSJTHWkjUn9RPUTurXh5vYA== X-Forefront-Antispam-Report: CIP:165.204.84.17;CTRY:US;LANG:en;SCL:1;SRV:;IPV:NLI;SFV:NSPM;H:satlexmb08.amd.com;PTR:InfoDomainNonexistent;CAT:NONE;SFS:(13230040)(36860700016)(376014)(1800799024)(82310400026)(56012099003)(22082099003)(18002099003);DIR:OUT;SFP:1101; X-MS-Exchange-AntiSpam-MessageData-ChunkCount: 1 X-MS-Exchange-AntiSpam-MessageData-0: KFHcMNn9Ei1mi7b3TA5tOV5SBsoPQiX/SCQvtk4M1H+WxUeSbpSVgT9QdO8xC5Vg0svfkNaCcZnDHEg5IcR4dTqYW8v1Z8Ae1bIX3OrT4vMwqQEd0XTLDKyv22nOZbsQC4yiKuln+ep1TVApIv8xsHVM6Ezq/V7jzxVz/IQdYtqRwqySTCbwUMEaF6IzC2vZR1J20ySASRKJAluIBoUjbQY2DFHN0lfaUcHxqcVmMTho3peYc1YSclojYwl8UEj/YP9Pk1HGwpiTF90ZTvjpKv5QJwY8TpxFWegKJRKIsBus12IZ0RXchExrzF3HPiJRy04UeZpWesoArCKwxUOPDGs5A5GI5akq3xTwTYVUZZ25e760s4ddffqjjBhZ+8PMBFeWGleB8Y5vikfo3F9Yt2CtqVXrAAOIAcv1fL9PR247ioGAOaciFn9WYbzQueHR X-OriginatorOrg: amd.com X-MS-Exchange-CrossTenant-OriginalArrivalTime: 20 Apr 2026 16:05:51.4210 (UTC) X-MS-Exchange-CrossTenant-Network-Message-Id: ec7aac79-a4fd-4d51-012f-08de9ef6b1c6 X-MS-Exchange-CrossTenant-Id: 3dd8961f-e488-4e60-8e11-a82d994e183d X-MS-Exchange-CrossTenant-OriginalAttributedTenantConnectingIp: TenantId=3dd8961f-e488-4e60-8e11-a82d994e183d;Ip=[165.204.84.17];Helo=[satlexmb08.amd.com] X-MS-Exchange-CrossTenant-AuthSource: SN1PEPF00026368.namprd02.prod.outlook.com X-MS-Exchange-CrossTenant-AuthAs: Anonymous X-MS-Exchange-CrossTenant-FromEntityHeader: HybridOnPrem X-MS-Exchange-Transport-CrossTenantHeadersStamped: MN0PR12MB5812 On 4/19/26 11:13, Mario Limonciello wrote: > > > On 4/17/26 16:06, Lizhi Hou wrote: >> From: Max Zhen >> >> Add support for allocating buffers from reserved carveout memory when >> IOMMU is not available. This is useful during debugging or bring-up. >> >> In this configuration, the device uses physical addresses and does >> not support scatter-gather lists, requiring physically contiguous >> buffers. >> >> Implement carveout-backed allocation and integrate it into buffer >> management to support operation in physical address mode. > > Running with IOMMU disabled is unlikely to be a common production > scenario.  At first I was thinking this is OK, but considering Greg's > strong comments about driver specific module parameters I /wonder/ if > this should be gated behind a kconfig option to be used at bringup? > > The Kconfig option could be something like XDNA_CARVEOUT and default > to 0.  Then it can be set to any value necessary for debugging. > > You could go a step further and only let XDNA_CARVEOUT do the > reservation when IOMMU wasn't found. Thanks for your comments. It would be more useful to change the carveout setting without recompile the driver. Because this is a debug function, so we will implement debugfs to config the carveout address and size. Lizhi > >> >> Signed-off-by: Max Zhen >> Signed-off-by: Lizhi Hou >> --- >>   drivers/accel/amdxdna/Makefile          |   1 + >>   drivers/accel/amdxdna/amdxdna_cbuf.c    | 249 ++++++++++++++++++++++++ >>   drivers/accel/amdxdna/amdxdna_cbuf.h    |  16 ++ >>   drivers/accel/amdxdna/amdxdna_gem.c     |  95 +++++++-- >>   drivers/accel/amdxdna/amdxdna_iommu.c   |  77 +++++--- >>   drivers/accel/amdxdna/amdxdna_pci_drv.c |  91 ++++++--- >>   drivers/accel/amdxdna/amdxdna_pci_drv.h |   4 +- >>   7 files changed, 454 insertions(+), 79 deletions(-) >>   create mode 100644 drivers/accel/amdxdna/amdxdna_cbuf.c >>   create mode 100644 drivers/accel/amdxdna/amdxdna_cbuf.h >> >> diff --git a/drivers/accel/amdxdna/Makefile >> b/drivers/accel/amdxdna/Makefile >> index 79369e497540..a055aea36971 100644 >> --- a/drivers/accel/amdxdna/Makefile >> +++ b/drivers/accel/amdxdna/Makefile >> @@ -12,6 +12,7 @@ amdxdna-y := \ >>       aie2_solver.o \ >>       aie4_message.o \ >>       aie4_pci.o \ >> +    amdxdna_cbuf.o \ >>       amdxdna_ctx.o \ >>       amdxdna_gem.o \ >>       amdxdna_iommu.o \ >> diff --git a/drivers/accel/amdxdna/amdxdna_cbuf.c >> b/drivers/accel/amdxdna/amdxdna_cbuf.c >> new file mode 100644 >> index 000000000000..4a556199a461 >> --- /dev/null >> +++ b/drivers/accel/amdxdna/amdxdna_cbuf.c >> @@ -0,0 +1,249 @@ >> +// SPDX-License-Identifier: GPL-2.0 >> +/* >> + * Copyright (C) 2026, Advanced Micro Devices, Inc. >> + */ >> + >> +#include >> +#include >> + >> +#include "amdxdna_cbuf.h" >> +#include "amdxdna_pci_drv.h" >> + >> +/* >> + * This is a platform debug/bringup feature. >> + * >> + * Carveout memory is a chunk of memory which is physically >> contiguous and >> + * is reserved during early boot time. There is only one chunk of >> such memory >> + * per system. Once available, all BOs accessible from device should be >> + * allocated from this memory. >> + */ >> +u64 carveout_addr; >> +module_param(carveout_addr, ullong, 0400); >> +MODULE_PARM_DESC(carveout_addr, "Physical memory address for >> reserved memory chunk"); >> + >> +u64 carveout_size; >> +module_param(carveout_size, ullong, 0400); >> +MODULE_PARM_DESC(carveout_size, "Physical memory size for reserved >> memory chunk"); >> + >> +struct amdxdna_carveout { >> +    struct drm_mm    mm; >> +    struct mutex    lock; /* protect mm */ >> +} carveout; >> + >> +bool amdxdna_use_carveout(void) >> +{ >> +    return !!carveout_size; >> +} >> + >> +void amdxdna_carveout_init(void) >> +{ >> +    if (!amdxdna_use_carveout()) >> +        return; >> +    mutex_init(&carveout.lock); >> +    drm_mm_init(&carveout.mm, carveout_addr, carveout_size); >> +    pr_info("Use carveout mem, addr=0x%llx, size=0x%llx\n", >> carveout_addr, carveout_size); >> +} >> + >> +void amdxdna_carveout_fini(void) >> +{ >> +    if (!amdxdna_use_carveout()) >> +        return; >> +    drm_mm_takedown(&carveout.mm); >> +    mutex_destroy(&carveout.lock); >> +} >> + >> +struct amdxdna_cbuf_priv { >> +    struct drm_mm_node node; >> +}; >> + >> +static struct sg_table *amdxdna_cbuf_map(struct dma_buf_attachment >> *attach, >> +                     enum dma_data_direction direction) >> +{ >> +    struct amdxdna_cbuf_priv *cbuf = attach->dmabuf->priv; >> +    struct device *dev = attach->dev; >> +    struct scatterlist *sgl, *sg; >> +    int ret, n_entries, i; >> +    struct sg_table *sgt; >> +    dma_addr_t dma_addr; >> +    size_t dma_size; >> +    size_t max_seg; >> + >> +    sgt = kzalloc_obj(*sgt); >> +    if (!sgt) >> +        return ERR_PTR(-ENOMEM); >> + >> +    max_seg = min_t(size_t, UINT_MAX, dma_max_mapping_size(dev)); >> +    n_entries = (cbuf->node.size + max_seg - 1) / max_seg; >> +    sgl = kzalloc_objs(*sg, n_entries); >> +    if (!sgl) { >> +        ret = -ENOMEM; >> +        goto free_sgt; >> +    } >> +    sg_init_table(sgl, n_entries); >> +    sgt->orig_nents = n_entries; >> +    sgt->nents = n_entries; >> +    sgt->sgl = sgl; >> + >> +    dma_size = cbuf->node.size; >> +    dma_addr = dma_map_resource(dev, cbuf->node.start, dma_size, >> +                    direction, DMA_ATTR_SKIP_CPU_SYNC); >> +    ret = dma_mapping_error(dev, dma_addr); >> +    if (ret) { >> +        pr_err("Failed to dma_map_resource carveout dma buf, ret >> %d\n", ret); >> +        goto free_sgl; >> +    } >> + >> +    for_each_sgtable_dma_sg(sgt, sg, i) { >> +        size_t len = min_t(size_t, max_seg, dma_size); >> + >> +        sg_dma_address(sg) = dma_addr; >> +        sg_dma_len(sg) = len; >> +        dma_addr += len; >> +        dma_size -= len; >> +    } >> + >> +    return sgt; >> + >> +free_sgl: >> +    kfree(sgl); >> +free_sgt: >> +    kfree(sgt); >> +    return ERR_PTR(ret); >> +} >> + >> +static void amdxdna_cbuf_unmap(struct dma_buf_attachment *attach, >> +                   struct sg_table *sgt, >> +                   enum dma_data_direction direction) >> +{ >> +    dma_unmap_resource(attach->dev, sg_dma_address(sgt->sgl), >> +               drm_prime_get_contiguous_size(sgt), direction, >> +               DMA_ATTR_SKIP_CPU_SYNC); >> +    sg_free_table(sgt); >> +    kfree(sgt); >> +} >> + >> +static void amdxdna_cbuf_release(struct dma_buf *dbuf) >> +{ >> +    struct amdxdna_cbuf_priv *cbuf = dbuf->priv; >> + >> +    mutex_lock(&carveout.lock); >> +    drm_mm_remove_node(&cbuf->node); >> +    mutex_unlock(&carveout.lock); >> + >> +    kfree(cbuf); >> +} >> + >> +static vm_fault_t amdxdna_cbuf_vm_fault(struct vm_fault *vmf) >> +{ >> +    struct vm_area_struct *vma = vmf->vma; >> +    struct amdxdna_cbuf_priv *cbuf; >> +    unsigned long pfn; >> +    pgoff_t pgoff; >> + >> +    cbuf = vma->vm_private_data; >> +    pgoff = (vmf->address - vma->vm_start) >> PAGE_SHIFT; >> +    pfn = (cbuf->node.start >> PAGE_SHIFT) + pgoff; >> + >> +    return vmf_insert_pfn(vma, vmf->address, pfn); >> +} >> + >> +static const struct vm_operations_struct amdxdna_cbuf_vm_ops = { >> +    .fault = amdxdna_cbuf_vm_fault, >> +}; >> + >> +static int amdxdna_cbuf_mmap(struct dma_buf *dbuf, struct >> vm_area_struct *vma) >> +{ >> +    struct amdxdna_cbuf_priv *cbuf = dbuf->priv; >> + >> +    vma->vm_ops = &amdxdna_cbuf_vm_ops; >> +    vma->vm_private_data = cbuf; >> +    vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP); >> + >> +    return 0; >> +} >> + >> +static int amdxdna_cbuf_vmap(struct dma_buf *dbuf, struct iosys_map >> *map) >> +{ >> +    struct amdxdna_cbuf_priv *cbuf = dbuf->priv; >> +    void *kva; >> + >> +    kva = memremap(cbuf->node.start, cbuf->node.size, MEMREMAP_WB); >> +    if (!kva) { >> +        pr_err("Failed to vmap carveout dma buf\n"); >> +        return -ENOMEM; >> +    } >> + >> +    iosys_map_set_vaddr(map, kva); >> +    return 0; >> +} >> + >> +static void amdxdna_cbuf_vunmap(struct dma_buf *dbuf, struct >> iosys_map *map) >> +{ >> +    memunmap(map->vaddr); >> +} >> + >> +static const struct dma_buf_ops amdxdna_cbuf_dmabuf_ops = { >> +    .map_dma_buf = amdxdna_cbuf_map, >> +    .unmap_dma_buf = amdxdna_cbuf_unmap, >> +    .release = amdxdna_cbuf_release, >> +    .mmap = amdxdna_cbuf_mmap, >> +    .vmap = amdxdna_cbuf_vmap, >> +    .vunmap = amdxdna_cbuf_vunmap, >> +}; >> + >> +static int amdxdna_cbuf_clear(struct dma_buf *dbuf) >> +{ >> +    struct iosys_map vmap = IOSYS_MAP_INIT_VADDR(NULL); >> + >> +    dma_buf_vmap(dbuf, &vmap); >> +    if (!vmap.vaddr) >> +        return -EFAULT; >> + >> +    memset(vmap.vaddr, 0, dbuf->size); >> +    dma_buf_vunmap(dbuf, &vmap); >> + >> +    return 0; >> +} >> + >> +struct dma_buf *amdxdna_get_cbuf(struct drm_device *dev, size_t >> size, u64 alignment) >> +{ >> +    DEFINE_DMA_BUF_EXPORT_INFO(exp_info); >> +    struct amdxdna_cbuf_priv *cbuf; >> +    struct dma_buf *dbuf; >> +    int ret; >> + >> +    cbuf = kzalloc_obj(*cbuf); >> +    if (!cbuf) >> +        return ERR_PTR(-ENOMEM); >> + >> +    mutex_lock(&carveout.lock); >> +    ret = drm_mm_insert_node_generic(&carveout.mm, &cbuf->node, size, >> +                     alignment, 0, DRM_MM_INSERT_BEST); >> +    mutex_unlock(&carveout.lock); >> +    if (ret) >> +        goto free_cbuf; >> + >> +    exp_info.size = size; >> +    exp_info.ops = &amdxdna_cbuf_dmabuf_ops; >> +    exp_info.priv = cbuf; >> +    exp_info.flags = O_RDWR; >> +    dbuf = dma_buf_export(&exp_info); >> +    if (IS_ERR(dbuf)) { >> +        ret = PTR_ERR(dbuf); >> +        goto remove_node; >> +    } >> + >> +    ret = amdxdna_cbuf_clear(dbuf); >> +    if (ret) { >> +        dma_buf_put(dbuf); >> +        goto out; >> +    } >> +    return dbuf; >> + >> +remove_node: >> +    drm_mm_remove_node(&cbuf->node); >> +free_cbuf: >> +    kfree(cbuf); >> +out: >> +    return ERR_PTR(ret); >> +} >> diff --git a/drivers/accel/amdxdna/amdxdna_cbuf.h >> b/drivers/accel/amdxdna/amdxdna_cbuf.h >> new file mode 100644 >> index 000000000000..15e189ce779e >> --- /dev/null >> +++ b/drivers/accel/amdxdna/amdxdna_cbuf.h >> @@ -0,0 +1,16 @@ >> +/* SPDX-License-Identifier: GPL-2.0 */ >> +/* >> + * Copyright (C) 2026, Advanced Micro Devices, Inc. >> + */ >> +#ifndef _AMDXDNA_CBUF_H_ >> +#define _AMDXDNA_CBUF_H_ >> + >> +#include >> +#include >> + >> +bool amdxdna_use_carveout(void); >> +void amdxdna_carveout_init(void); >> +void amdxdna_carveout_fini(void); >> +struct dma_buf *amdxdna_get_cbuf(struct drm_device *dev, size_t >> size, u64 alignment); >> + >> +#endif >> diff --git a/drivers/accel/amdxdna/amdxdna_gem.c >> b/drivers/accel/amdxdna/amdxdna_gem.c >> index 238ee244d4a6..905514ec183c 100644 >> --- a/drivers/accel/amdxdna/amdxdna_gem.c >> +++ b/drivers/accel/amdxdna/amdxdna_gem.c >> @@ -16,6 +16,7 @@ >>   #include >>   #include >>   +#include "amdxdna_cbuf.h" >>   #include "amdxdna_ctx.h" >>   #include "amdxdna_gem.h" >>   #include "amdxdna_pci_drv.h" >> @@ -516,10 +517,6 @@ static void amdxdna_imported_obj_free(struct >> amdxdna_gem_obj *abo) >>   static inline bool >>   amdxdna_gem_skip_bo_usage(struct amdxdna_gem_obj *abo) >>   { >> -    /* Do not count imported BOs since the buffer is not allocated >> by us. */ >> -    if (is_import_bo(abo)) >> -        return true; >> - >>       /* Already counted as part of HEAP BO */ >>       if (abo->type == AMDXDNA_BO_DEV) >>           return true; >> @@ -571,9 +568,7 @@ static void amdxdna_gem_obj_free(struct >> drm_gem_object *gobj) >>       if (abo->type == AMDXDNA_BO_DEV_HEAP) >>           drm_mm_takedown(&abo->mm); >>   -    if (amdxdna_iova_on(xdna)) >> -        amdxdna_iommu_unmap_bo(xdna, abo); >> - >> +    amdxdna_dma_unmap_bo(xdna, abo); >>       amdxdna_gem_vunmap(abo); >>       mutex_destroy(&abo->lock); >>   @@ -591,18 +586,20 @@ static int amdxdna_gem_obj_open(struct >> drm_gem_object *gobj, struct drm_file *fi >>         guard(mutex)(&abo->lock); >>       abo->open_ref++; >> +    if (abo->open_ref > 1) >> +        return 0; >>   -    if (abo->open_ref == 1) { >> -        /* Attached to the client when first opened by it. */ >> -        abo->client = filp->driver_priv; >> -        amdxdna_gem_add_bo_usage(abo); >> -    } >> -    if (amdxdna_iova_on(xdna)) { >> -        ret = amdxdna_iommu_map_bo(xdna, abo); >> +    /* Attached to the client when first opened by it. */ >> +    abo->client = filp->driver_priv; >> + >> +    /* No need to set up dma addr mapping in PASID mode. */ >> +    if (!amdxdna_pasid_on(abo->client)) { >> +        ret = amdxdna_dma_map_bo(xdna, abo); >>           if (ret) >>               return ret; >>       } >>   +    amdxdna_gem_add_bo_usage(abo); >>       return 0; >>   } >>   @@ -620,6 +617,39 @@ static void amdxdna_gem_obj_close(struct >> drm_gem_object *gobj, struct drm_file * >>       } >>   } >>   +static int amdxdna_gem_obj_vmap(struct drm_gem_object *obj, struct >> iosys_map *map) >> +{ >> +    struct amdxdna_gem_obj *abo = to_xdna_obj(obj); >> +    int ret; >> + >> +    iosys_map_clear(map); >> + >> +    dma_resv_assert_held(obj->resv); >> + >> +    if (is_import_bo(abo)) >> +        ret = dma_buf_vmap(abo->dma_buf, map); >> +    else >> +        ret = drm_gem_shmem_object_vmap(obj, map); >> +    if (ret) >> +        return ret; >> +    if (!map->vaddr) >> +        return -ENOMEM; >> + >> +    return 0; >> +} >> + >> +static void amdxdna_gem_obj_vunmap(struct drm_gem_object *obj, >> struct iosys_map *map) >> +{ >> +    struct amdxdna_gem_obj *abo = to_xdna_obj(obj); >> + >> +    dma_resv_assert_held(obj->resv); >> + >> +    if (is_import_bo(abo)) >> +        dma_buf_vunmap(abo->dma_buf, map); >> +    else >> +        drm_gem_shmem_object_vunmap(obj, map); >> +} >> + >>   static int amdxdna_gem_dev_obj_vmap(struct drm_gem_object *obj, >> struct iosys_map *map) >>   { >>       struct amdxdna_gem_obj *abo = to_xdna_obj(obj); >> @@ -645,8 +675,8 @@ static const struct drm_gem_object_funcs >> amdxdna_gem_shmem_funcs = { >>       .pin = drm_gem_shmem_object_pin, >>       .unpin = drm_gem_shmem_object_unpin, >>       .get_sg_table = drm_gem_shmem_object_get_sg_table, >> -    .vmap = drm_gem_shmem_object_vmap, >> -    .vunmap = drm_gem_shmem_object_vunmap, >> +    .vmap = amdxdna_gem_obj_vmap, >> +    .vunmap = amdxdna_gem_obj_vunmap, >>       .mmap = amdxdna_gem_obj_mmap, >>       .vm_ops = &drm_gem_shmem_vm_ops, >>       .export = amdxdna_gem_prime_export, >> @@ -714,6 +744,36 @@ amdxdna_gem_create_ubuf_object(struct drm_device >> *dev, struct amdxdna_drm_create >>       return to_xdna_obj(gobj); >>   } >>   +static struct amdxdna_gem_obj * >> +amdxdna_gem_create_cbuf_object(struct drm_device *dev, struct >> amdxdna_drm_create_bo *args) >> +{ >> +    struct amdxdna_dev *xdna = to_xdna_dev(dev); >> +    size_t size = PAGE_ALIGN(args->size); >> +    struct drm_gem_object *gobj; >> +    struct amdxdna_gem_obj *ret; >> +    struct dma_buf *dma_buf; >> +    u64 align; >> + >> +    if (!size) { >> +        XDNA_ERR(xdna, "Invalid BO size 0x%llx", args->size); >> +        return ERR_PTR(-EINVAL); >> +    } >> + >> +    align = (args->type == AMDXDNA_BO_DEV_HEAP) ? >> xdna->dev_info->dev_mem_size : 0; >> +    dma_buf = amdxdna_get_cbuf(dev, size, align); >> +    if (IS_ERR(dma_buf)) >> +        return ERR_CAST(dma_buf); >> + >> +    gobj = amdxdna_gem_prime_import(dev, dma_buf); >> +    if (IS_ERR(gobj)) >> +        ret = ERR_CAST(gobj); >> +    else >> +        ret = to_xdna_obj(gobj); >> + >> +    dma_buf_put(dma_buf); >> +    return ret; >> +} >> + >>   struct drm_gem_object * >>   amdxdna_gem_prime_import(struct drm_device *dev, struct dma_buf >> *dma_buf) >>   { >> @@ -769,6 +829,8 @@ amdxdna_drm_create_share_bo(struct drm_device *dev, >>         if (args->vaddr) >>           abo = amdxdna_gem_create_ubuf_object(dev, args); >> +    else if (amdxdna_use_carveout()) >> +        abo = amdxdna_gem_create_cbuf_object(dev, args); >>       else >>           abo = amdxdna_gem_create_shmem_object(dev, args); >>       if (IS_ERR(abo)) >> @@ -884,7 +946,6 @@ int amdxdna_drm_create_bo_ioctl(struct drm_device >> *dev, void *data, struct drm_f >>            args->type, args->vaddr, args->size, args->flags); >>       switch (args->type) { >>       case AMDXDNA_BO_CMD: >> -        fallthrough; >>       case AMDXDNA_BO_SHARE: >>           abo = amdxdna_drm_create_share_bo(dev, args, filp); >>           break; >> diff --git a/drivers/accel/amdxdna/amdxdna_iommu.c >> b/drivers/accel/amdxdna/amdxdna_iommu.c >> index 5a9f06183487..eff00131d0f8 100644 >> --- a/drivers/accel/amdxdna/amdxdna_iommu.c >> +++ b/drivers/accel/amdxdna/amdxdna_iommu.c >> @@ -35,14 +35,15 @@ static struct iova >> *amdxdna_iommu_alloc_iova(struct amdxdna_dev *xdna, >>       return iova; >>   } >>   -int amdxdna_iommu_map_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo) >> +int amdxdna_dma_map_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo) >>   { >> +    unsigned long contig_sz; >>       struct sg_table *sgt; >>       dma_addr_t dma_addr; >>       struct iova *iova; >>       ssize_t size; >>   -    if (abo->type != AMDXDNA_BO_DEV_HEAP && abo->type != >> AMDXDNA_BO_SHMEM) >> +    if (abo->type != AMDXDNA_BO_DEV_HEAP && abo->type != >> AMDXDNA_BO_SHARE) >>           return 0; >>         sgt = drm_gem_shmem_get_pages_sgt(&abo->base); >> @@ -51,47 +52,63 @@ int amdxdna_iommu_map_bo(struct amdxdna_dev >> *xdna, struct amdxdna_gem_obj *abo) >>           return PTR_ERR(sgt); >>       } >>   -    if (!sgt->orig_nents || !sg_page(sgt->sgl)) { >> -        XDNA_ERR(xdna, "sgl is zero length or not page backed"); >> +    if (!sgt->orig_nents) { >> +        XDNA_ERR(xdna, "sgl is zero length"); >>           return -EOPNOTSUPP; >>       } >>   -    iova = amdxdna_iommu_alloc_iova(xdna, abo->mem.size, &dma_addr, >> -                    (abo->type == AMDXDNA_BO_DEV_HEAP)); >> -    if (IS_ERR(iova)) { >> -        XDNA_ERR(xdna, "Alloc iova failed, ret %ld", PTR_ERR(iova)); >> -        return PTR_ERR(iova); >> +    if (amdxdna_iova_on(xdna)) { >> +        if (!sg_page(sgt->sgl)) { >> +            XDNA_ERR(xdna, "sgl is not page backed"); >> +            return -EOPNOTSUPP; >> +        } >> + >> +        iova = amdxdna_iommu_alloc_iova(xdna, abo->mem.size, &dma_addr, >> +                        (abo->type == AMDXDNA_BO_DEV_HEAP)); >> +        if (IS_ERR(iova)) { >> +            XDNA_ERR(xdna, "Alloc iova failed, ret %ld", >> PTR_ERR(iova)); >> +            return PTR_ERR(iova); >> +        } >> + >> +        size = iommu_map_sgtable(xdna->domain, dma_addr, sgt, >> +                     IOMMU_READ | IOMMU_WRITE); >> +        if (size < 0) { >> +            XDNA_ERR(xdna, "iommu_map_sgtable failed: %zd", size); >> +            __free_iova(&xdna->iovad, iova); >> +            return size; >> +        } >> +        if (size < abo->mem.size) { >> +            iommu_unmap(xdna->domain, dma_addr, size); >> +            __free_iova(&xdna->iovad, iova); >> +            return -ENXIO; >> +        } >> +        abo->mem.dma_addr = dma_addr; >> +    } else { >> +        /* Device doesn't support scatter/gather list, fail >> non-contiguous mapping. */ >> +        contig_sz = drm_prime_get_contiguous_size(sgt); >> +        if (contig_sz < abo->mem.size) { >> +            XDNA_ERR(xdna, >> +                 "noncontiguous dma addr, contig size:%ld, expected >> size:%ld", >> +                 contig_sz, abo->mem.size); >> +            return -EINVAL; >> +        } >> +        abo->mem.dma_addr = sg_dma_address(sgt->sgl); >>       } >> - >> -    size = iommu_map_sgtable(xdna->domain, dma_addr, sgt, >> -                 IOMMU_READ | IOMMU_WRITE); >> -    if (size < 0) { >> -        XDNA_ERR(xdna, "iommu_map_sgtable failed: %zd", size); >> -        __free_iova(&xdna->iovad, iova); >> -        return size; >> -    } >> - >> -    if (size < abo->mem.size) { >> -        iommu_unmap(xdna->domain, dma_addr, size); >> -        __free_iova(&xdna->iovad, iova); >> -        return -ENXIO; >> -    } >> - >> -    abo->mem.dma_addr = dma_addr; >> - >>       return 0; >>   } >>   -void amdxdna_iommu_unmap_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo) >> +void amdxdna_dma_unmap_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo) >>   { >>       size_t size; >>         if (abo->mem.dma_addr == AMDXDNA_INVALID_ADDR) >>           return; >>   -    size = iova_align(&xdna->iovad, abo->mem.size); >> -    iommu_unmap(xdna->domain, abo->mem.dma_addr, size); >> -    free_iova(&xdna->iovad, iova_pfn(&xdna->iovad, abo->mem.dma_addr)); >> +    if (amdxdna_iova_on(xdna)) { >> +        size = iova_align(&xdna->iovad, abo->mem.size); >> +        iommu_unmap(xdna->domain, abo->mem.dma_addr, size); >> +        free_iova(&xdna->iovad, iova_pfn(&xdna->iovad, >> abo->mem.dma_addr)); >> +    } >>       abo->mem.dma_addr = AMDXDNA_INVALID_ADDR; >>   } >>   diff --git a/drivers/accel/amdxdna/amdxdna_pci_drv.c >> b/drivers/accel/amdxdna/amdxdna_pci_drv.c >> index 21eddfc538d0..b8c5dbc12489 100644 >> --- a/drivers/accel/amdxdna/amdxdna_pci_drv.c >> +++ b/drivers/accel/amdxdna/amdxdna_pci_drv.c >> @@ -14,6 +14,7 @@ >>   #include >>   #include >>   +#include "amdxdna_cbuf.h" >>   #include "amdxdna_ctx.h" >>   #include "amdxdna_gem.h" >>   #include "amdxdna_pci_drv.h" >> @@ -67,11 +68,40 @@ static const struct amdxdna_device_id >> amdxdna_ids[] = { >>       {0} >>   }; >>   +static int amdxdna_sva_init(struct amdxdna_client *client) >> +{ >> +    struct amdxdna_dev *xdna = client->xdna; >> + >> +    client->sva = iommu_sva_bind_device(xdna->ddev.dev, client->mm); >> +    if (IS_ERR(client->sva)) { >> +        XDNA_ERR(xdna, "SVA bind device failed, ret %ld", >> PTR_ERR(client->sva)); >> +        return PTR_ERR(client->sva); >> +    } >> + >> +    client->pasid = iommu_sva_get_pasid(client->sva); >> +    if (client->pasid == IOMMU_PASID_INVALID) { >> +        iommu_sva_unbind_device(client->sva); >> +        XDNA_ERR(xdna, "SVA get pasid failed"); >> +        return -ENODEV; >> +    } >> + >> +    return 0; >> +} >> + >> +static void amdxdna_sva_fini(struct amdxdna_client *client) >> +{ >> +    if (IS_ERR_OR_NULL(client->sva)) >> +        return; >> + >> +    iommu_sva_unbind_device(client->sva); >> +    client->sva = NULL; >> +    client->pasid = IOMMU_PASID_INVALID; >> +} >> + >>   static int amdxdna_drm_open(struct drm_device *ddev, struct >> drm_file *filp) >>   { >>       struct amdxdna_dev *xdna = to_xdna_dev(ddev); >>       struct amdxdna_client *client; >> -    int ret; >>         client = kzalloc_obj(*client); >>       if (!client) >> @@ -80,22 +110,13 @@ static int amdxdna_drm_open(struct drm_device >> *ddev, struct drm_file *filp) >>       client->pid = pid_nr(rcu_access_pointer(filp->pid)); >>       client->xdna = xdna; >>       client->pasid = IOMMU_PASID_INVALID; >> +    client->mm = current->mm; >>         if (!amdxdna_iova_on(xdna)) { >> -        client->sva = iommu_sva_bind_device(xdna->ddev.dev, >> current->mm); >> -        if (IS_ERR(client->sva)) { >> -            ret = PTR_ERR(client->sva); >> -            XDNA_ERR(xdna, "SVA bind device failed, ret %d", ret); >> -            goto failed; >> -        } >> -        client->pasid = iommu_sva_get_pasid(client->sva); >> -        if (client->pasid == IOMMU_PASID_INVALID) { >> -            XDNA_ERR(xdna, "SVA get pasid failed"); >> -            ret = -ENODEV; >> -            goto unbind_sva; >> -        } >> +        /* No need to fail open since user may use pa + carveout >> later. */ >> +        if (amdxdna_sva_init(client)) >> +            XDNA_WARN(xdna, "PASID not available for pid %d", >> client->pid); >>       } >> -    client->mm = current->mm; >>       mmgrab(client->mm); >>       init_srcu_struct(&client->hwctx_srcu); >>       xa_init_flags(&client->hwctx_xa, XA_FLAGS_ALLOC); >> @@ -110,14 +131,6 @@ static int amdxdna_drm_open(struct drm_device >> *ddev, struct drm_file *filp) >>         XDNA_DBG(xdna, "pid %d opened", client->pid); >>       return 0; >> - >> -unbind_sva: >> -    if (!IS_ERR_OR_NULL(client->sva)) >> -        iommu_sva_unbind_device(client->sva); >> -failed: >> -    kfree(client); >> - >> -    return ret; >>   } >>     static void amdxdna_client_cleanup(struct amdxdna_client *client) >> @@ -131,11 +144,8 @@ static void amdxdna_client_cleanup(struct >> amdxdna_client *client) >>           drm_gem_object_put(to_gobj(client->dev_heap)); >>         mutex_destroy(&client->mm_lock); >> - >> -    if (!IS_ERR_OR_NULL(client->sva)) >> -        iommu_sva_unbind_device(client->sva); >>       mmdrop(client->mm); >> - >> +    amdxdna_sva_fini(client); >>       kfree(client); >>   } >>   @@ -242,15 +252,17 @@ static void amdxdna_show_fdinfo(struct >> drm_printer *p, struct drm_file *filp) >>         /* >>        * Note for driver specific BO memory usage stat. >> -     * Total memory alloc = amdxdna-internal-alloc + >> amdxdna-external-alloc >> +     * Total memory in use = amdxdna-internal-alloc + >> amdxdna-external-alloc, which >> +     * includes both imported and created BOs. To avoid double >> counts, it includes >> +     * HEAP BO, but not DEV BO. DEV BO is counted by >> amdxdna-heap-alloc. >>        */ >>       drm_fdinfo_print_size(p, drv_name, "heap", "alloc", heap_usage); >>       drm_fdinfo_print_size(p, drv_name, "internal", "alloc", >> internal_usage); >>       drm_fdinfo_print_size(p, drv_name, "external", "alloc", >> external_usage); >>       /* >>        * Note for DRM standard BO memory stat. >> -     * drm-total-memory counts both DEV BO and HEAP BO >> -     * drm-shared-memory counts BO imported >> +     * drm-total-memory counts both DEV BO and HEAP BO. The DEV BO >> size is double counted. >> +     * drm-shared-memory counts BO shared with other processes/devices. >>        */ >>       drm_show_memory_stats(p, filp); >>   } >> @@ -420,7 +432,26 @@ static struct pci_driver amdxdna_pci_driver = { >>       .sriov_configure = amdxdna_sriov_configure, >>   }; >>   -module_pci_driver(amdxdna_pci_driver); >> +static int __init amdxdna_mod_init(void) >> +{ >> +    int ret; >> + >> +    amdxdna_carveout_init(); >> +    ret = pci_register_driver(&amdxdna_pci_driver); >> +    if (ret) >> +        amdxdna_carveout_fini(); >> + >> +    return ret; >> +} >> + >> +static void __exit amdxdna_mod_exit(void) >> +{ >> +    pci_unregister_driver(&amdxdna_pci_driver); >> +    amdxdna_carveout_fini(); >> +} >> + >> +module_init(amdxdna_mod_init); >> +module_exit(amdxdna_mod_exit); >>     MODULE_LICENSE("GPL"); >>   MODULE_IMPORT_NS("AMD_PMF"); >> diff --git a/drivers/accel/amdxdna/amdxdna_pci_drv.h >> b/drivers/accel/amdxdna/amdxdna_pci_drv.h >> index bdd0dc83f92e..07bd38281452 100644 >> --- a/drivers/accel/amdxdna/amdxdna_pci_drv.h >> +++ b/drivers/accel/amdxdna/amdxdna_pci_drv.h >> @@ -172,11 +172,11 @@ void amdxdna_sysfs_fini(struct amdxdna_dev *xdna); >>     int amdxdna_iommu_init(struct amdxdna_dev *xdna); >>   void amdxdna_iommu_fini(struct amdxdna_dev *xdna); >> -int amdxdna_iommu_map_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo); >> -void amdxdna_iommu_unmap_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo); >>   void *amdxdna_iommu_alloc(struct amdxdna_dev *xdna, size_t size, >> dma_addr_t *dma_addr); >>   void amdxdna_iommu_free(struct amdxdna_dev *xdna, size_t size, >>               void *cpu_addr, dma_addr_t dma_addr); >> +int amdxdna_dma_map_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo); >> +void amdxdna_dma_unmap_bo(struct amdxdna_dev *xdna, struct >> amdxdna_gem_obj *abo); >>     static inline bool amdxdna_iova_on(struct amdxdna_dev *xdna) >>   { >