[PATCH 5/5] iommu/vt-d: Add page request draining support
From: Lu Baolu
Date: Tue Mar 17 2020 - 03:05:49 EST
From: Jacob Pan <jacob.jun.pan@xxxxxxxxxxxxxxx>
When a PASID is stopped or terminated, there can be pending
PRQs (requests that haven't received responses) in remapping
hardware. This adds the interface to drain page requests and
call it when a PASID is terminated.
Signed-off-by: Jacob Pan <jacob.jun.pan@xxxxxxxxxxxxxxx>
Signed-off-by: Liu Yi L <yi.l.liu@xxxxxxxxx>
Signed-off-by: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx>
---
drivers/iommu/intel-svm.c | 156 +++++++++++++++++++++++++++++++++++-
include/linux/intel-iommu.h | 1 +
2 files changed, 155 insertions(+), 2 deletions(-)
diff --git a/drivers/iommu/intel-svm.c b/drivers/iommu/intel-svm.c
index 6ce96dd541a6..ab13e33bc6ff 100644
--- a/drivers/iommu/intel-svm.c
+++ b/drivers/iommu/intel-svm.c
@@ -22,9 +22,14 @@
#include "intel-pasid.h"
+static int intel_svm_process_prq(struct intel_iommu *iommu,
+ struct page_req_dsc *prq,
+ int head, int tail);
+static void intel_svm_drain_prq(struct device *dev, int pasid);
static irqreturn_t prq_event_thread(int irq, void *d);
-#define PRQ_ORDER 0
+#define PRQ_ORDER 0
+#define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x20)
int intel_svm_enable_prq(struct intel_iommu *iommu)
{
@@ -208,6 +213,7 @@ static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
rcu_read_lock();
list_for_each_entry_rcu(sdev, &svm->devs, list) {
intel_pasid_tear_down_entry(svm->iommu, sdev->dev, svm->pasid);
+ intel_svm_drain_prq(sdev->dev, svm->pasid);
intel_flush_svm_range_dev(svm, sdev, 0, -1, 0);
}
rcu_read_unlock();
@@ -226,6 +232,152 @@ static LIST_HEAD(global_svm_list);
list_for_each_entry((sdev), &(svm)->devs, list) \
if ((d) != (sdev)->dev) {} else
+/*
+ * When a PASID is stopped or terminated, there can be pending PRQs
+ * (requests have not received responses) in remapping hardware.
+ *
+ * There are at least below scenarios that PRQ drains are required:
+ * - unbind a PASID;
+ * - resume phase of the PASID suspend/resume cycle.
+ *
+ * Steps to be performed:
+ * - Disable PR interrupt;
+ * - Take a snapshot of the page request queue, record current head
+ * and tail, and mark PRQ entries with PASID to be dropped;
+ * - Mark queue empty, a.k.a. Head = Tail;
+ * - PRQ draining as described in 7.11 of the spec. Unnecessary to
+ * check queue full since queue was empty at the point of drain;
+ * - Tail could have been moved due to new PRQ written by HW;
+ * - Process snapshot copy of PR queue;
+ * - Process hardware PR queue, enable interrupt again.
+ *
+ * For an example, consider the following timeline going downward.
+ * VT-d HW VT-d Driver User(KMD, guest)
+ * --------------------------------------------------------
+ * [PR1.2.3]
+ * [PR1.1.3] <tail>
+ * [PR1.2.2]
+ * [PR1.2.1]
+ * [PR1.1.2]
+ * [PR1.1.1] <head>
+ * [IRQ] ->
+ *
+ *
+ * <- [unbind PASID 1]
+ * [delete pending PR]
+ * [drain PRQs]
+ *
+ * Decoder:
+ * - PR.PASID.GroupID.Index, e.g. PR.1.2.3 indicates the Page request
+ * with PASID = 1, GroupID = 2, 3rd request in the group.
+ * - LPIG: last page in group
+ * - PDP: private data present
+ * - KMD: kernel mode driver for native SVA
+ *
+ * Note:
+ * - Caller of unbind/suspend/resume PASID APIs must ensure no pending
+ * DMA activities prior to call.
+ */
+static void intel_svm_drain_prq(struct device *dev, int pasid)
+{
+ struct device_domain_info *info;
+ struct qi_desc desc[3] = { 0 };
+ struct dmar_domain *domain;
+ struct intel_iommu *iommu;
+ struct intel_svm *svm;
+ struct pci_dev *pdev;
+ unsigned long flags;
+ int head, tail;
+ u16 sid, did;
+ void *prqq;
+ int qdep;
+
+ info = get_domain_info(dev);
+ if (WARN_ON(!info || !dev_is_pci(dev)))
+ return;
+
+ iommu = info->iommu;
+ domain = info->domain;
+ pdev = to_pci_dev(dev);
+
+ rcu_read_lock();
+ svm = ioasid_find(NULL, pasid, NULL);
+ if (WARN_ON(!svm)) {
+ rcu_read_unlock();
+ return;
+ }
+ rcu_read_unlock();
+
+ spin_lock_irqsave(&iommu->lock, flags);
+ tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
+ head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
+
+ /*
+ * Make a copy of the PR queue then process it offline without
+ * blocking PRQ interrupts.
+ */
+ prqq = kmemdup(iommu->prq, PAGE_SIZE ^ PRQ_ORDER, GFP_ATOMIC);
+ if (!prqq) {
+ spin_unlock_irqrestore(&iommu->lock, flags);
+ return;
+ }
+ /*
+ * Make queue empty to allow further events and avoid the queue
+ * full condition while we drain the queue.
+ */
+ dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
+ spin_unlock_irqrestore(&iommu->lock, flags);
+
+ /*
+ * Process the copy of PRQ, drained PASID already marked to be
+ * dropped.
+ */
+ if (intel_svm_process_prq(iommu, prqq, head, tail)) {
+ kfree(prqq);
+ return;
+ }
+
+ /*
+ * Perform steps prescribed in VT-d spec CH7.11 to drain page
+ * request and responses.
+ */
+ sid = PCI_DEVID(info->bus, info->devfn);
+ did = domain->iommu_did[iommu->seq_id];
+ qdep = pci_ats_queue_depth(pdev);
+
+ desc[0].qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
+ QI_IWD_FENCE |
+ QI_IWD_TYPE;
+ desc[1].qw0 = QI_EIOTLB_PASID(pasid) |
+ QI_EIOTLB_DID(did) |
+ QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
+ QI_EIOTLB_TYPE;
+ desc[2].qw0 = QI_DEV_EIOTLB_PASID(pasid) |
+ QI_DEV_EIOTLB_SID(sid) |
+ QI_DEV_EIOTLB_QDEP(qdep) |
+ QI_DEIOTLB_TYPE |
+ QI_DEV_IOTLB_PFSID(info->pfsid);
+
+ qi_submit_sync(iommu, desc, 3, QI_OPT_WAIT_DRAIN);
+
+ /*
+ * If new requests come in while we processing the copy, we should
+ * process it now, otherwise the new request may be stuck until the
+ * next IRQ.
+ */
+ if (dmar_readq(iommu->reg + DMAR_PRS_REG) & DMA_PRS_PPR) {
+ tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
+ head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
+ intel_svm_process_prq(iommu, iommu->prq, head, tail);
+ }
+
+ /* Allow new pending PRQ to generate interrupts. */
+ writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
+
+ kfree(prqq);
+ return;
+}
+
/* Caller must hold pasid_mutex, mm reference */
static int intel_svm_bind_mm(struct device *dev, int flags, struct svm_dev_ops *ops,
struct mm_struct *mm, struct intel_svm_dev **sd)
@@ -439,6 +591,7 @@ int intel_svm_unbind_mm(struct device *dev, int pasid)
* large and has to be physically contiguous. So it's
* hard to be as defensive as we might like. */
intel_pasid_tear_down_entry(iommu, dev, svm->pasid);
+ intel_svm_drain_prq(dev, svm->pasid);
intel_flush_svm_range_dev(svm, sdev, 0, -1, 0);
kfree_rcu(sdev, rcu);
@@ -491,7 +644,6 @@ struct page_req_dsc {
u64 priv_data[2];
};
-#define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x20)
static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
{
diff --git a/include/linux/intel-iommu.h b/include/linux/intel-iommu.h
index 255b23f59a78..fcbc40d3ce18 100644
--- a/include/linux/intel-iommu.h
+++ b/include/linux/intel-iommu.h
@@ -324,6 +324,7 @@ enum {
#define QI_IWD_STATUS_DATA(d) (((u64)d) << 32)
#define QI_IWD_STATUS_WRITE (((u64)1) << 5)
+#define QI_IWD_FENCE (((u64)1) << 6)
#define QI_IWD_PRQ_DRAIN (((u64)1) << 7)
#define QI_IOTLB_DID(did) (((u64)did) << 16)
--
2.17.1