lkml.org 
[lkml]   [2012]   [Nov]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/2] vfio powerpc: implemented IOMMU driver for VFIO
On 27/11/12 15:29, Alex Williamson wrote:
> On Tue, 2012-11-27 at 15:06 +1100, Alexey Kardashevskiy wrote:
>> On 27/11/12 05:20, Alex Williamson wrote:
>>> On Fri, 2012-11-23 at 20:03 +1100, Alexey Kardashevskiy wrote:
>>>> VFIO implements platform independent stuff such as
>>>> a PCI driver, BAR access (via read/write on a file descriptor
>>>> or direct mapping when possible) and IRQ signaling.
>>>>
>>>> The platform dependent part includes IOMMU initialization
>>>> and handling. This patch implements an IOMMU driver for VFIO
>>>> which does mapping/unmapping pages for the guest IO and
>>>> provides information about DMA window (required by a POWERPC
>>>> guest).
>>>>
>>>> The counterpart in QEMU is required to support this functionality.
>>>>
>>>> Cc: David Gibson <david@gibson.dropbear.id.au>
>>>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>>>> ---
>>>> drivers/vfio/Kconfig | 6 +
>>>> drivers/vfio/Makefile | 1 +
>>>> drivers/vfio/vfio_iommu_spapr_tce.c | 247 +++++++++++++++++++++++++++++++++++
>>>> include/linux/vfio.h | 20 +++
>>>> 4 files changed, 274 insertions(+)
>>>> create mode 100644 drivers/vfio/vfio_iommu_spapr_tce.c
>>>>
>>>> diff --git a/drivers/vfio/Kconfig b/drivers/vfio/Kconfig
>>>> index 7cd5dec..b464687 100644
>>>> --- a/drivers/vfio/Kconfig
>>>> +++ b/drivers/vfio/Kconfig
>>>> @@ -3,10 +3,16 @@ config VFIO_IOMMU_TYPE1
>>>> depends on VFIO
>>>> default n
>>>>
>>>> +config VFIO_IOMMU_SPAPR_TCE
>>>> + tristate
>>>> + depends on VFIO && SPAPR_TCE_IOMMU
>>>> + default n
>>>> +
>>>> menuconfig VFIO
>>>> tristate "VFIO Non-Privileged userspace driver framework"
>>>> depends on IOMMU_API
>>>> select VFIO_IOMMU_TYPE1 if X86
>>>> + select VFIO_IOMMU_SPAPR_TCE if PPC_POWERNV
>>>> help
>>>> VFIO provides a framework for secure userspace device drivers.
>>>> See Documentation/vfio.txt for more details.
>>>> diff --git a/drivers/vfio/Makefile b/drivers/vfio/Makefile
>>>> index 2398d4a..72bfabc 100644
>>>> --- a/drivers/vfio/Makefile
>>>> +++ b/drivers/vfio/Makefile
>>>> @@ -1,3 +1,4 @@
>>>> obj-$(CONFIG_VFIO) += vfio.o
>>>> obj-$(CONFIG_VFIO_IOMMU_TYPE1) += vfio_iommu_type1.o
>>>> +obj-$(CONFIG_VFIO_IOMMU_SPAPR_TCE) += vfio_iommu_spapr_tce.o
>>>> obj-$(CONFIG_VFIO_PCI) += pci/
>>>> diff --git a/drivers/vfio/vfio_iommu_spapr_tce.c b/drivers/vfio/vfio_iommu_spapr_tce.c
>>>> new file mode 100644
>>>> index 0000000..46a6298
>>>> --- /dev/null
>>>> +++ b/drivers/vfio/vfio_iommu_spapr_tce.c
>>>> @@ -0,0 +1,247 @@
>>>> +/*
>>>> + * VFIO: IOMMU DMA mapping support for TCE on POWER
>>>> + *
>>>> + * Copyright (C) 2012 IBM Corp. All rights reserved.
>>>> + * Author: Alexey Kardashevskiy <aik@ozlabs.ru>
>>>> + *
>>>> + * This program is free software; you can redistribute it and/or modify
>>>> + * it under the terms of the GNU General Public License version 2 as
>>>> + * published by the Free Software Foundation.
>>>> + *
>>>> + * Derived from original vfio_iommu_type1.c:
>>>> + * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
>>>> + * Author: Alex Williamson <alex.williamson@redhat.com>
>>>> + */
>>>> +
>>>> +#include <linux/module.h>
>>>> +#include <linux/pci.h>
>>>> +#include <linux/slab.h>
>>>> +#include <linux/uaccess.h>
>>>> +#include <linux/err.h>
>>>> +#include <linux/vfio.h>
>>>> +#include <asm/iommu.h>
>>>> +
>>>> +#define DRIVER_VERSION "0.1"
>>>> +#define DRIVER_AUTHOR "aik@ozlabs.ru"
>>>> +#define DRIVER_DESC "VFIO IOMMU SPAPR TCE"
>>>> +
>>>> +static void tce_iommu_detach_group(void *iommu_data,
>>>> + struct iommu_group *iommu_group);
>>>> +
>>>> +/*
>>>> + * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
>>>> + */
>>>> +
>>>> +/*
>>>> + * The container descriptor supports only a single group per container.
>>>> + * Required by the API as the container is not supplied with the IOMMU group
>>>> + * at the moment of initialization.
>>>> + */
>>>> +struct tce_container {
>>>> + struct mutex lock;
>>>> + struct iommu_table *tbl;
>>>> +};
>>>> +
>>>> +static void *tce_iommu_open(unsigned long arg)
>>>> +{
>>>> + struct tce_container *container;
>>>> +
>>>> + if (arg != VFIO_SPAPR_TCE_IOMMU) {
>>>> + printk(KERN_ERR "tce_vfio: Wrong IOMMU type\n");
>>>> + return ERR_PTR(-EINVAL);
>>>> + }
>>>> +
>>>> + container = kzalloc(sizeof(*container), GFP_KERNEL);
>>>> + if (!container)
>>>> + return ERR_PTR(-ENOMEM);
>>>> +
>>>> + mutex_init(&container->lock);
>>>> +
>>>> + return container;
>>>> +}
>>>> +
>>>> +static void tce_iommu_release(void *iommu_data)
>>>> +{
>>>> + struct tce_container *container = iommu_data;
>>>> +
>>>> + WARN_ON(container->tbl && !container->tbl->it_group);
>>>
>>> I think your patch ordering is backwards here. it_group isn't added
>>> until 2/2. I'd really like to see the arch/powerpc code approved and
>>> merged by the powerpc maintainer before we add the code that makes use
>>> of it into vfio. Otherwise we just get lots of churn if interfaces
>>> change or they disapprove of it altogether.
>>
>>
>> Makes sense, thanks.
>>
>>
>>>> + if (container->tbl && container->tbl->it_group)
>>>> + tce_iommu_detach_group(iommu_data, container->tbl->it_group);
>>>> +
>>>> + mutex_destroy(&container->lock);
>>>> +
>>>> + kfree(container);
>>>> +}
>>>> +
>>>> +static long tce_iommu_ioctl(void *iommu_data,
>>>> + unsigned int cmd, unsigned long arg)
>>>> +{
>>>> + struct tce_container *container = iommu_data;
>>>> + unsigned long minsz;
>>>> +
>>>> + switch (cmd) {
>>>> + case VFIO_CHECK_EXTENSION: {
>>>> + return (arg == VFIO_SPAPR_TCE_IOMMU) ? 1 : 0;
>>>> + }
>>>> + case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
>>>> + struct vfio_iommu_spapr_tce_info info;
>>>> + struct iommu_table *tbl = container->tbl;
>>>> +
>>>> + if (WARN_ON(!tbl))
>>>> + return -ENXIO;
>>>> +
>>>> + minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
>>>> + dma64_window_size);
>>>> +
>>>> + if (copy_from_user(&info, (void __user *)arg, minsz))
>>>> + return -EFAULT;
>>>> +
>>>> + if (info.argsz < minsz)
>>>> + return -EINVAL;
>>>> +
>>>> + info.dma32_window_start = tbl->it_offset << IOMMU_PAGE_SHIFT;
>>>> + info.dma32_window_size = tbl->it_size << IOMMU_PAGE_SHIFT;
>>>> + info.dma64_window_start = 0;
>>>> + info.dma64_window_size = 0;
>>>> + info.flags = 0;
>>>> +
>>>> + if (copy_to_user((void __user *)arg, &info, minsz))
>>>> + return -EFAULT;
>>>> +
>>>> + return 0;
>>>> + }
>>>> + case VFIO_IOMMU_MAP_DMA: {
>>>> + vfio_iommu_spapr_tce_dma_map param;
>>>> + struct iommu_table *tbl = container->tbl;
>>>> + enum dma_data_direction direction = DMA_NONE;
>>>> +
>>>> + if (WARN_ON(!tbl))
>>>> + return -ENXIO;
>>>> +
>>>> + minsz = offsetofend(vfio_iommu_spapr_tce_dma_map, size);
>>>> +
>>>> + if (copy_from_user(&param, (void __user *)arg, minsz))
>>>> + return -EFAULT;
>>>> +
>>>> + if (param.argsz < minsz)
>>>> + return -EINVAL;
>>>> +
>>>> + if ((param.flags & VFIO_DMA_MAP_FLAG_READ) &&
>>>> + (param.flags & VFIO_DMA_MAP_FLAG_WRITE)) {
>>>> + direction = DMA_BIDIRECTIONAL;
>>>> + } else if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
>>>> + direction = DMA_TO_DEVICE;
>>>> + } else if (param.flags & VFIO_DMA_MAP_FLAG_WRITE) {
>>>> + direction = DMA_FROM_DEVICE;
>>>> + }
>>>> +
>>>> + param.size += param.iova & ~IOMMU_PAGE_MASK;
>>>> + param.size = _ALIGN_UP(param.size, IOMMU_PAGE_SIZE);
>>>
>>> On x86 we force iova, vaddr, and size to all be aligned to the smallest
>>> page granularity of the iommu and return -EINVAL if it doesn't fit.
>>> What does it imply to the user if they're always aligned to work here?
>>> Won't this interface happily map overlapping entries with no indication
>>> to the user that the previous mapping is no longer valid?
>>> Maybe another reason why a combined unmap/map makes me nervous, we have
>>> to assume the user knows what they're doing.
>>
>>
>> I got used to guests which do know what they are doing so I am pretty calm :)
>> but ok, I'll move alignment to the QEMU, it makes sense.
>>
>>
>>>> +
>>>> + return iommu_put_tces(tbl, param.iova >> IOMMU_PAGE_SHIFT,
>>>> + param.vaddr & IOMMU_PAGE_MASK, direction,
>>>> + param.size >> IOMMU_PAGE_SHIFT);
>>>> + }
>>>> + case VFIO_IOMMU_UNMAP_DMA: {
>>>> + vfio_iommu_spapr_tce_dma_unmap param;
>>>> + struct iommu_table *tbl = container->tbl;
>>>> +
>>>> + if (WARN_ON(!tbl))
>>>> + return -ENXIO;
>>>> +
>>>> + minsz = offsetofend(vfio_iommu_spapr_tce_dma_unmap, size);
>>>> +
>>>> + if (copy_from_user(&param, (void __user *)arg, minsz))
>>>> + return -EFAULT;
>>>> +
>>>> + if (param.argsz < minsz)
>>>> + return -EINVAL;
>>>> +
>>>> + param.size += param.iova & ~IOMMU_PAGE_MASK;
>>>> + param.size = _ALIGN_UP(param.size, IOMMU_PAGE_SIZE);
>>>> +
>>>> + return iommu_put_tces(tbl, param.iova >> IOMMU_PAGE_SHIFT,
>>>> + 0, DMA_NONE, param.size >> IOMMU_PAGE_SHIFT);
>>>> + }
>>>> + default:
>>>> + printk(KERN_WARNING "tce_vfio: unexpected cmd %x\n", cmd);
>>>
>>> pr_warn
>>>
>>>> + }
>>>> +
>>>> + return -ENOTTY;
>>>> +}
>>>> +
>>>> +static int tce_iommu_attach_group(void *iommu_data,
>>>> + struct iommu_group *iommu_group)
>>>> +{
>>>> + struct tce_container *container = iommu_data;
>>>> + struct iommu_table *tbl = iommu_group_get_iommudata(iommu_group);
>>>> +
>>>> + BUG_ON(!tbl);
>>>> + mutex_lock(&container->lock);
>>>> + pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
>>>> + iommu_group_id(iommu_group), iommu_group);
>>>> + if (container->tbl) {
>>>> + printk(KERN_WARNING "tce_vfio: Only one group per IOMMU container is allowed, existing id=%d, attaching id=%d\n",
>>>
>>> pr_warn
>>>
>>>> + iommu_group_id(container->tbl->it_group),
>>>> + iommu_group_id(iommu_group));
>>>> + mutex_unlock(&container->lock);
>>>> + return -EBUSY;
>>>> + }
>>>> +
>>>> + container->tbl = tbl;
>>>
>>> Would it be too much paranoia to clear all the tce here as you do below
>>> on detach?
>>
>> Guess so. I do unmap on detach() and the guest calls put_tce(0) (i.e.
>> unmaps) the whole DMA window at the boot time.
>
> But that's just one user of this interface, we can't assume they'll all
> be so agreeable. If any tces were enabled here, a malicious user would
> have a window to host memory, right? Thanks,


But I still release pages on detach(), how can this code be not called on
the guest exit (normal or crashed)?



>
> Alex
>


--
Alexey


\
 
 \ /
  Last update: 2012-11-27 06:41    [W:0.082 / U:0.924 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site