[lkml]   [2020]   [Aug]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
Subject[PATCH v2 1/9] docs: Document IO Address Space ID (IOASID) APIs
IOASID is used to identify address spaces that can be targeted by device
DMA. It is a system-wide resource that is essential to its many users.
This document is an attempt to help developers from all vendors navigate
the APIs. At this time, ARM SMMU and Intel’s Scalable IO Virtualization
(SIOV) enabled platforms are the primary users of IOASID. Examples of
how SIOV components interact with IOASID APIs are provided in that many
APIs are driven by the requirements from SIOV.

Signed-off-by: Liu Yi L <>
Signed-off-by: Wu Hao <>
Signed-off-by: Jacob Pan <>
Documentation/ioasid.rst | 618 +++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 618 insertions(+)
create mode 100644 Documentation/ioasid.rst

diff --git a/Documentation/ioasid.rst b/Documentation/ioasid.rst
new file mode 100644
index 000000000000..b6a8cdc885ff
--- /dev/null
+++ b/Documentation/ioasid.rst
@@ -0,0 +1,618 @@
+.. ioasid:
+IO Address Space ID
+IOASID is a generic name for PCIe Process Address ID (PASID) or ARM
+SMMU sub-stream ID. An IOASID identifies an address space that DMA
+requests can target.
+The primary use cases for IOASID are Shared Virtual Address (SVA) and
+IO Virtual Address (IOVA). However, the requirements for IOASID
+management can vary among hardware architectures.
+This document covers the generic features supported by IOASID
+APIs. Vendor-specific use cases are also illustrated with Intel's VT-d
+based platforms as the first example.
+.. contents:: :local:
+PASID - Process Address Space ID
+IOASID - IO Address Space ID (generic term for PCIe PASID and
+sub-stream ID in SMMU)
+SVA/SVM - Shared Virtual Addressing/Memory
+ENQCMD - New Intel X86 ISA for efficient workqueue submission [1]
+DSA - Intel Data Streaming Accelerator [2]
+VDCM - Virtual device composition module [3]
+SIOV - Intel Scalable IO Virtualization
+Key Concepts
+An IOASID set is a group of IOASIDs allocated from the system-wide
+IOASID pool. An IOASID set is created and can be identified by a
+token of u64. Refer to IOASID set APIs for more details.
+IOASID set is particularly useful for guest SVA where each guest could
+have its own IOASID set for security and efficiency reasons.
+IOASID Set Private ID (SPID)
+SPIDs are introduced as IOASIDs within its set. Each SPID maps to a
+system-wide IOASID but the namespace of SPID is within its IOASID
+set. SPIDs can be used as guest IOASIDs where each guest could do
+IOASID allocation from its own pool and map them to host physical
+IOASIDs. SPIDs are particularly useful for supporting live migration
+where decoupling guest and host physical resources are necessary.
+For example, two VMs can both allocate guest PASID/SPID #101 but map to
+different host PASIDs #201 and #202 respectively as shown in the
+diagram below.
+ .------------------. .------------------.
+ | VM 1 | | VM 2 |
+ | | | |
+ |------------------| |------------------|
+ | GPASID/SPID 101 | | GPASID/SPID 101 |
+ '------------------' -------------------' Guest
+ __________|______________________|______________________
+ | | Host
+ v v
+ .------------------. .------------------.
+ | Host IOASID 201 | | Host IOASID 202 |
+ '------------------' '------------------'
+ | IOASID set 1 | | IOASID set 2 |
+ '------------------' '------------------'
+Guest PASID is treated as IOASID set private ID (SPID) within an
+IOASID set, mappings between guest and host IOASIDs are stored in the
+set for inquiry.
+To get the IOASID APIs, users must #include <linux/ioasid.h>. These APIs
+serve the following functionalities:
+ - IOASID allocation/Free
+ - Group management in the form of ioasid_set
+ - Private data storage and lookup
+ - Reference counting
+ - Event notification in case of state change
+IOASID Set Level APIs
+For use cases such as guest SVA it is necessary to manage IOASIDs at
+a group level. For example, VMs may allocate multiple IOASIDs for
+guest process address sharing (vSVA). It is imperative to enforce
+VM-IOASID ownership such that malicious guest cannot target DMA
+traffic outside its own IOASIDs, or free an active IOASID belong to
+another VM.
+ struct ioasid_set *ioasid_alloc_set(void *token, ioasid_t quota, u32 type)
+ int ioasid_adjust_set(struct ioasid_set *set, int quota);
+ void ioasid_set_get(struct ioasid_set *set)
+ void ioasid_set_put(struct ioasid_set *set)
+ void ioasid_set_get_locked(struct ioasid_set *set)
+ void ioasid_set_put_locked(struct ioasid_set *set)
+ int ioasid_set_for_each_ioasid(struct ioasid_set *sdata,
+ void (*fn)(ioasid_t id, void *data),
+ void *data)
+IOASID set concept is introduced to represent such IOASID groups. Each
+IOASID set is created with a token which can be one of the following
+ - IOASID_SET_TYPE_NULL (Arbitrary u64 value)
+ - IOASID_SET_TYPE_MM (Set token is a mm_struct)
+The explicit MM token type is useful when multiple users of an IOASID
+set under the same process need to communicate about their shared IOASIDs.
+E.g. An IOASID set created by VFIO for one guest can be associated
+with the KVM instance for the same guest since they share a common mm_struct.
+The IOASID set APIs serve the following purposes:
+ - Ownership/permission enforcement
+ - Take collective actions, e.g. free an entire set
+ - Event notifications within a set
+ - Look up a set based on token
+ - Quota enforcement
+Individual IOASID APIs
+Once an ioasid_set is created, IOASIDs can be allocated from the set.
+Within the IOASID set namespace, set private ID (SPID) is supported. In
+the VM use case, SPID can be used for storing guest PASID.
+ ioasid_t ioasid_alloc(struct ioasid_set *set, ioasid_t min, ioasid_t max,
+ void *private);
+ int ioasid_get(struct ioasid_set *set, ioasid_t ioasid);
+ void ioasid_put(struct ioasid_set *set, ioasid_t ioasid);
+ int ioasid_get_locked(struct ioasid_set *set, ioasid_t ioasid);
+ void ioasid_put_locked(struct ioasid_set *set, ioasid_t ioasid);
+ void *ioasid_find(struct ioasid_set *set, ioasid_t ioasid,
+ bool (*getter)(void *));
+ ioasid_t ioasid_find_by_spid(struct ioasid_set *set, ioasid_t spid)
+ int ioasid_attach_data(struct ioasid_set *set, ioasid_t ioasid,
+ void *data);
+ int ioasid_attach_spid(struct ioasid_set *set, ioasid_t ioasid,
+ ioasid_t ssid);
+An IOASID may have multiple users, each user may have hardware context
+associated with an IOASID. When the status of an IOASID changes,
+e.g. an IOASID is being freed, users need to be notified such that the
+associated hardware context can be cleared, flushed, and drained.
+ int ioasid_register_notifier(struct ioasid_set *set, struct
+ notifier_block *nb)
+ void ioasid_unregister_notifier(struct ioasid_set *set,
+ struct notifier_block *nb)
+ int ioasid_register_notifier_mm(struct mm_struct *mm, struct
+ notifier_block *nb)
+ void ioasid_unregister_notifier_mm(struct mm_struct *mm, struct
+ notifier_block *nb)
+ int ioasid_notify(ioasid_t ioasid, enum ioasid_notify_val cmd,
+ unsigned int flags)
+Notification events are pertinent to individual IOASIDs, they can be
+one of the following:
+ - FREE
+ - BIND
+Ordering is supported by IOASID notification priorities as the
+following (in ascending order):
+ enum ioasid_notifier_prios {
+ };
+The typical use case is when an IOASID is freed due to an exception, DMA
+source should be quiesced before tearing down other hardware contexts
+in the system. This will reduce the churn in handling faults. DMA work
+submission is performed by the CPU which is granted higher priority than
+There are two types of notifiers in IOASID core: system-wide and
+System-wide notifier is catering for users that need to handle all
+IOASIDs in the system. E.g. The IOMMU driver handles all IOASIDs.
+Per ioasid_set notifier can be used by VM specific components such as
+KVM. After all, each KVM instance only cares about IOASIDs within its
+own set.
+IOASID notifiers are atomic due to spinlocks used inside the IOASID
+core. For tasks cannot be completed in the notifier handler, async work
+can be submitted to complete the work later as long as there is no
+ordering requirement.
+Reference counting
+IOASID lifecycle management is based on reference counting. Users of
+IOASID intend to align lifecycle with the IOASID need to hold
+reference of the IOASID. IOASID will not be returned to the pool for
+allocation until all references are dropped. Calling ioasid_free()
+will mark the IOASID as FREE_PENDING if the IOASID has outstanding
+reference. ioasid_get() is not allowed once an IOASID is in the
+Event notifications are used to inform users of IOASID status change.
+IOASID_FREE event prompts users to drop their references after
+clearing its context.
+For example, on VT-d platform when an IOASID is freed, teardown
+actions are performed on KVM, device driver, and IOMMU driver.
+KVM shall register notifier block with::
+ static struct notifier_block pasid_nb_kvm = {
+ .notifier_call = pasid_status_change_kvm,
+ .priority = IOASID_PRIO_CPU,
+ };
+VDCM driver shall register notifier block with::
+ static struct notifier_block pasid_nb_vdcm = {
+ .notifier_call = pasid_status_change_vdcm,
+ .priority = IOASID_PRIO_DEVICE,
+ };
+In both cases, notifier blocks shall be registered on the IOASID set
+such that *only* events from the matching VM is received.
+If KVM attempts to register notifier block before the IOASID set is
+created for the MM token, the notifier block will be placed on a
+pending list inside IOASID core. Once the token matching IOASID set
+is created, IOASID will register the notifier block automatically.
+IOASID core does not replay events for the existing IOASIDs in the
+set. For IOASID set of MM type, notification blocks can be registered
+on empty sets only. This is to avoid lost events.
+IOMMU driver shall register notifier block on global chain::
+ static struct notifier_block pasid_nb_vtd = {
+ .notifier_call = pasid_status_change_vtd,
+ .priority = IOASID_PRIO_IOMMU,
+ };
+Custom allocator APIs
+ int ioasid_register_allocator(struct ioasid_allocator_ops *allocator);
+ void ioasid_unregister_allocator(struct ioasid_allocator_ops *allocator);
+Allocator Choices
+IOASIDs are allocated for both host and guest SVA/IOVA usage. However,
+allocators can be different. For example, on VT-d guest PASID
+allocation must be performed via a virtual command interface which is
+emulated by VMM.
+IOASID core has the notion of "custom allocator" such that guest can
+register virtual command allocator that precedes the default one.
+IOASIDs are limited system resources that default to 20 bits in
+size. Since each device has its own table, theoretically the namespace
+can be per device also. However, for security reasons sharing PASID
+tables among devices are not good for isolation. Therefore, IOASID
+namespace is system-wide.
+There are also other reasons to have this simpler system-wide
+namespace. Take VT-d as an example, VT-d supports shared workqueue
+and ENQCMD[1] where one IOASID could be used to submit work on
+multiple devices that are shared with other VMs. This requires IOASID
+to be system-wide. This is also the reason why guests must use an
+emulated virtual command interface to allocate IOASID from the host.
+Life cycle
+This section covers IOASID lifecycle management for both bare-metal
+and guest usages. In bare-metal SVA, MMU notifier is directly hooked
+up with IOMMU driver, therefore the process address space (MM)
+lifecycle is aligned with IOASID.
+However, guest MMU notifier is not available to host IOMMU driver,
+when guest MM terminates unexpectedly, the events have to go through
+VFIO and IOMMU UAPI to reach host IOMMU driver. There are also more
+parties involved in guest SVA, e.g. on Intel VT-d platform, IOASIDs
+are used by IOMMU driver, KVM, VDCM, and VFIO.
+Native IOASID Life Cycle (VT-d Example)
+The normal flow of native SVA code with Intel Data Streaming
+Accelerator(DSA) [2] as example:
+1. Host user opens accelerator FD, e.g. DSA driver, or uacce;
+2. DSA driver allocate WQ, do sva_bind_device();
+3. IOMMU driver calls ioasid_alloc(), then bind PASID with device,
+ mmu_notifier_get()
+4. DMA starts by DSA driver userspace
+5. DSA userspace close FD
+6. DSA/uacce kernel driver handles FD.close()
+7. DSA driver stops DMA
+8. DSA driver calls sva_unbind_device();
+9. IOMMU driver does unbind, clears PASID context in IOMMU, flush
+ TLBs. mmu_notifier_put() called.
+10. mmu_notifier.release() called, IOMMU SVA code calls ioasid_free()*
+11. The IOASID is returned to the pool, reclaimed.
+ * With ENQCMD, PASID used on VT-d is not released in mmu_notifier() but
+ mmdrop(). mmdrop comes after FD close. Should not matter.
+ If the user process dies unexpectedly, Step #10 may come before
+ Step #5, in between, all DMA faults discarded. PRQ responded with
+During the normal teardown, the following three steps would happen in
+1. Device driver stops DMA request
+2. IOMMU driver unbinds PASID and mm, flush all TLBs, drain in-flight
+ requests.
+3. IOASID freed
+Exception happens when process terminates *before* device driver stops
+DMA and call IOMMU driver to unbind. The flow of process exists are as
+ do_exit() {
+ exit_mm() {
+ mm_put();
+ exit_mmap() {
+ intel_invalidate_range() //mmu notifier
+ tlb_finish_mmu()
+ mmu_notifier_release(mm) {
+ intel_iommu_release() {
+ [2] intel_iommu_teardown_pasid();
+ intel_iommu_flush_tlbs();
+ }
+ // tlb_invalidate_range cb removed
+ }
+ unmap_vmas();
+ free_pgtables(); // IOMMU cannot walk PGT after this
+ };
+ }
+ exit_files(tsk) {
+ close_files() {
+ dsa_close();
+ [1] dsa_stop_dma();
+ intel_svm_unbind_pasid(); //nothing to do
+ }
+ }
+ }
+ mmdrop() /* some random time later, lazy mm user */ {
+ mm_free_pgd();
+ destroy_context(mm); {
+ [3] ioasid_free();
+ }
+ }
+As shown in the list above, step #2 could happen before
+#1. Unrecoverable(UR) faults could happen between #2 and #1.
+Also notice that TLB invalidation occurs at mmu_notifier
+invalidate_range callback as well as the release callback. The reason
+is that release callback will delete IOMMU driver from the notifier
+chain which may skip invalidate_range() calls during the exit path.
+To avoid unnecessary reporting of UR fault, IOMMU driver shall disable
+fault reporting after free and before unbind.
+Guest IOASID Life Cycle (VT-d Example)
+Guest IOASID life cycle starts with guest driver open(), this could be
+uacce or individual accelerator driver such as DSA. At FD open,
+sva_bind_device() is called which triggers a series of actions.
+The example below is an illustration of *normal* operations that
+involves *all* the SW components in VT-d. The flow can be simpler if
+no ENQCMD is supported.
+ ..................................................................
+ 1 ioasid_register_notifier/_mm()
+ 2 ioasid_alloc() 1
+ 3 bind_gpasid()
+ 4 iommu_bind()->ioasid_get() 2
+ 5 ioasid_notify(BIND)
+ 6 -> ioasid_get() 3
+ 7 -> vmcs_update_atomic()
+ 8 mdev_write(gpasid)
+ 9 hpasid=
+ 10 find_by_spid(gpasid) 4
+ 11 vdev_write(hpasid)
+ 12 -------- GUEST STARTS DMA --------------------------
+ 13 -------- GUEST STOPS DMA --------------------------
+ 14 mdev_clear(gpasid)
+ 15 vdev_clear(hpasid)
+ 16 ioasid_put() 3
+ 17 unbind_gpasid()
+ 18 iommu_ubind()
+ 19 ioasid_notify(UNBIND)
+ 20 -> vmcs_update_atomic()
+ 21 -> ioasid_put() 2
+ 22 ioasid_free() 1
+ 23 ioasid_put() 0
+ 24 Reclaimed
+ -------------- New Life Cycle Begin ----------------------------
+ 1 ioasid_alloc() -> 1
+ Note: IOASID Notification Events: FREE, BIND, UNBIND
+Exception cases arise when a guest crashes or a malicious guest
+attempts to cause disruption on the host system. The fault handling
+rules are:
+1. IOASID free must *always* succeed.
+2. An inactive period may be required before the freed IOASID is
+ reclaimed. During this period, consumers of IOASID perform cleanup.
+3. Malfunction is limited to the guest owned resources for all
+ programming errors.
+The primary source of exception is when the following are out of
+1. Start/Stop of DMA activity
+ (Guest device driver, mdev via VFIO)
+2. Setup/Teardown of IOMMU PASID context, IOTLB, DevTLB flushes
+ (Host IOMMU driver bind/unbind)
+3. Setup/Teardown of VMCS PASID translation table entries (KVM) in
+ case of ENQCMD
+4. Programming/Clearing host PASID in VDCM (Host VDCM driver)
+5. IOASID alloc/free (Host IOASID)
+VFIO is the *only* user-kernel interface, which is ultimately
+responsible for exception handlings.
+#1 is processed the same way as the assigned device today based on
+device file descriptors and events. There is no special handling.
+#3 is based on bind/unbind events emitted by #2.
+#4 is naturally aligned with IOASID life cycle in that an illegal
+guest PASID programming would fail in obtaining reference of the
+matching host IOASID.
+#5 is similar to #4. The fault will be reported to the user if PASID
+used in the ENQCMD is not set up in VMCS PASID translation table.
+Therefore, the remaining out of order problem is between #2 and
+#5. I.e. unbind vs. free. More specifically, free before unbind.
+IOASID notifier and refcounting are used to ensure order. Following
+a publisher-subscriber pattern where:
+- Publishers: VFIO & IOMMU
+- Subscribers: KVM, VDCM, IOMMU
+IOASID notifier is atomic which requires subscribers to do quick
+handling of the event in the atomic context. Workqueue can be used for
+any processing that requires thread context. IOASID reference must be
+acquired before receiving the FREE event. The reference must be
+dropped at the end of the processing in order to return the IOASID to
+the pool.
+Let's examine the IOASID life cycle again when free happens *before*
+unbind. This could be a result of misbehaving guests or crash. Assuming
+VFIO cannot enforce unbind->free order. Notice that the setup part up
+until step #12 is identical to the normal case, the flow below starts
+with step 13.
+ ..................................................................
+ 13 -------- GUEST STARTS DMA --------------------------
+ 14 -------- *GUEST MISBEHAVES!!!* ----------------
+ 15 ioasid_free()
+ 16 ioasid_notify(FREE)
+ 17 mark_ioasid_inactive[1]
+ 18 kvm_nb_handler(FREE)
+ 19 vmcs_update_atomic()
+ 20 ioasid_put_locked() -> 3
+ 21 vdcm_nb_handler(FREE)
+ 22 iomm_nb_handler(FREE)
+ 23 ioasid_free() returns[2] schedule_work() 2
+ 24 schedule_work() vdev_clear_wk(hpasid)
+ 25 teardown_pasid_wk()
+ 26 ioasid_put() -> 1
+ 27 ioasid_put() 0
+ 28 Reclaimed
+ 29 unbind_gpasid()
+ 30 iommu_unbind()->ioasid_find() Fails[3]
+ -------------- New Life Cycle Begin ----------------------------
+1. By marking IOASID inactive at step #17, no new references can be
+ held. ioasid_get/find() will return -ENOENT;
+2. After step #23, all events can go out of order. Shall not affect
+ the outcome.
+3. IOMMU driver fails to find private data for unbinding. If unbind is
+ called after the same IOASID is allocated for the same guest again,
+ this is a programming error. The damage is limited to the guest
+ itself since unbind performs permission checking based on the
+ IOASID set associated with the guest process.
+KVM PASID Translation Table Updates
+Per VM PASID translation table is maintained by KVM in order to
+support ENQCMD in the guest. The table contains host-guest PASID
+translations to be consumed by CPU ucode. The synchronization of the
+PASID states depends on VFIO/IOMMU driver, where IOCTL and atomic
+notifiers are used. KVM must register IOASID notifier per VM instance
+during launch time. The following events are handled:
+2. FREE
+1. Multiple devices can bind with the same PASID, this can be different PCI
+ devices or mdevs within the same PCI device. However, only the
+ *first* BIND and *last* UNBIND emit notifications.
+2. IOASID code is responsible for ensuring the correctness of H-G
+ PASID mapping. There is no need for KVM to validate the
+ notification data.
+3. When UNBIND happens *after* FREE, KVM will see error in
+ ioasid_get() even when the reclaim is not done. IOMMU driver will
+ also avoid sending UNBIND if the PASID is already FREE.
+4. When KVM terminates *before* FREE & UNBIND, references will be
+ dropped for all host PASIDs.
+VDCM PASID Programming
+VDCM composes virtual devices and exposes them to the guests. When
+the guest allocates a PASID then program it to the virtual device, VDCM
+intercepts the programming attempt then program the matching host
+PASID on to the hardware.
+Conversely, when a device is going away, VDCM must be informed such
+that PASID context on the hardware can be cleared. There could be
+multiple mdevs assigned to different guests in the same VDCM. Since
+the PASID table is shared at PCI device level, lazy clearing is not
+secure. A malicious guest can attack by using newly freed PASIDs that
+are allocated by another guest.
+By holding a reference of the PASID until VDCM cleans up the HW context,
+it is guaranteed that PASID life cycles do not cross within the same
 \ /
  Last update: 2020-08-22 06:28    [W:0.664 / U:0.024 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site