]> xenbits.xensource.com Git - xen.git/commitdiff
VT-d: introduce new fields in msi_desc to track binding with guest interrupt
authorFeng Wu <feng.wu@intel.com>
Fri, 7 Apr 2017 13:37:07 +0000 (15:37 +0200)
committerJan Beulich <jbeulich@suse.com>
Fri, 7 Apr 2017 13:37:07 +0000 (15:37 +0200)
msi_msg_to_remap_entry() is buggy when the live IRTE is in posted format. It
wrongly inherits the 'im' field meaning the IRTE is in posted format but
updates all the other fields to remapping format.

There are also two situations that lead to the above issue. One is some callers
really want to change the IRTE to remapped format. The other is some callers
only want to update msi message (e.g. set msi affinity) for they don't aware
that this msi is binded with a guest interrupt. We should suppress update
in the second situation. To distinguish them, straightforwardly, we can let
caller specify which format of IRTE they want update to. It isn't feasible for
making all callers be aware of the binding with guest interrupt will cause a
far more complicated change (including the interfaces exposed to IOAPIC and
MSI). Also some callings happen in interrupt context where we can't acquire
d->event_lock to read struct hvm_pirq_dpci.

This patch introduces two new fields in msi_desc to track binding with a guest
interrupt such that msi_msg_to_remap_entry() can get the binding and update
IRTE accordingly. After that change, pi_update_irte() can utilize
msi_msg_to_remap_entry() to update IRTE to posted format.

Signed-off-by: Feng Wu <feng.wu@intel.com>
Signed-off-by: Chao Gao <chao.gao@intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
xen/arch/x86/msi.c
xen/drivers/passthrough/io.c
xen/drivers/passthrough/vtd/intremap.c
xen/include/asm-x86/iommu.h
xen/include/asm-x86/msi.h

index a868007b1189b6618fe322ba2ba5896413b21d3f..3374cd459701a362c585c4ea349ca90dc405e351 100644 (file)
@@ -578,6 +578,7 @@ static struct msi_desc *alloc_msi_entry(unsigned int nr)
         entry[nr].dev = NULL;
         entry[nr].irq = -1;
         entry[nr].remap_index = -1;
+        entry[nr].pi_desc = NULL;
     }
 
     return entry;
index 4d1941374d6a2b716123398a9bbed91ba0199c94..3e0a10e2c66801f313ee7dcc9408102af3452f59 100644 (file)
@@ -415,7 +415,8 @@ int pt_irq_create_bind(
         if ( iommu_intpost )
         {
             if ( vcpu )
-                pi_update_irte(vcpu, info, pirq_dpci->gmsi.gvec);
+                pi_update_irte(&vcpu->arch.hvm_vmx.pi_desc, info,
+                               pirq_dpci->gmsi.gvec);
             else
                 dprintk(XENLOG_G_INFO,
                         "%pv: deliver interrupt in remapping mode,gvec:%02x\n",
@@ -619,6 +620,8 @@ int pt_irq_destroy_bind(
         else
             what = "bogus";
     }
+    else if ( pirq_dpci && pirq_dpci->gmsi.posted )
+        pi_update_irte(NULL, pirq, 0);
 
     if ( pirq_dpci && (pirq_dpci->flags & HVM_IRQ_DPCI_MAPPED) &&
          list_empty(&pirq_dpci->digl_list) )
index bfd468baf0f4fa7a77952da904540eb8cb22a9dc..6314cbffd9d6691f7569267ac0a5d82bded0e7d0 100644 (file)
@@ -551,12 +551,12 @@ static int msi_msg_to_remap_entry(
     struct iommu *iommu, struct pci_dev *pdev,
     struct msi_desc *msi_desc, struct msi_msg *msg)
 {
-    struct iremap_entry *iremap_entry = NULL, *iremap_entries;
-    struct iremap_entry new_ire;
+    struct iremap_entry *iremap_entry = NULL, *iremap_entries, new_ire = { };
     struct msi_msg_remap_entry *remap_rte;
     unsigned int index, i, nr = 1;
     unsigned long flags;
     struct ir_ctrl *ir_ctrl = iommu_ir_ctrl(iommu);
+    const struct pi_desc *pi_desc = msi_desc->pi_desc;
 
     if ( msi_desc->msi_attrib.type == PCI_CAP_ID_MSI )
         nr = msi_desc->msi.nvec;
@@ -595,33 +595,35 @@ static int msi_msg_to_remap_entry(
     GET_IREMAP_ENTRY(ir_ctrl->iremap_maddr, index,
                      iremap_entries, iremap_entry);
 
-    memcpy(&new_ire, iremap_entry, sizeof(struct iremap_entry));
-
-    /* Set interrupt remapping table entry */
-    new_ire.remap.fpd = 0;
-    new_ire.remap.dm = (msg->address_lo >> MSI_ADDR_DESTMODE_SHIFT) & 0x1;
-    new_ire.remap.tm = (msg->data >> MSI_DATA_TRIGGER_SHIFT) & 0x1;
-    new_ire.remap.dlm = (msg->data >> MSI_DATA_DELIVERY_MODE_SHIFT) & 0x1;
-    /* Hardware require RH = 1 for LPR delivery mode */
-    new_ire.remap.rh = (new_ire.remap.dlm == dest_LowestPrio);
-    new_ire.remap.avail = 0;
-    new_ire.remap.res_1 = 0;
-    new_ire.remap.vector = (msg->data >> MSI_DATA_VECTOR_SHIFT) &
-                            MSI_DATA_VECTOR_MASK;
-    new_ire.remap.res_2 = 0;
-    if ( x2apic_enabled )
-        new_ire.remap.dst = msg->dest32;
+    if ( !pi_desc )
+    {
+        new_ire.remap.dm = msg->address_lo >> MSI_ADDR_DESTMODE_SHIFT;
+        new_ire.remap.tm = msg->data >> MSI_DATA_TRIGGER_SHIFT;
+        new_ire.remap.dlm = msg->data >> MSI_DATA_DELIVERY_MODE_SHIFT;
+        /* Hardware requires RH = 1 for lowest priority delivery mode */
+        new_ire.remap.rh = (new_ire.remap.dlm == dest_LowestPrio);
+        new_ire.remap.vector = (msg->data >> MSI_DATA_VECTOR_SHIFT) &
+                                MSI_DATA_VECTOR_MASK;
+        if ( x2apic_enabled )
+            new_ire.remap.dst = msg->dest32;
+        else
+            new_ire.remap.dst =
+                MASK_EXTR(msg->address_lo, MSI_ADDR_DEST_ID_MASK) << 8;
+        new_ire.remap.p = 1;
+    }
     else
-        new_ire.remap.dst = ((msg->address_lo >> MSI_ADDR_DEST_ID_SHIFT)
-                             & 0xff) << 8;
+    {
+        new_ire.post.im = 1;
+        new_ire.post.vector = msi_desc->gvec;
+        new_ire.post.pda_l = virt_to_maddr(pi_desc) >> (32 - PDA_LOW_BIT);
+        new_ire.post.pda_h = virt_to_maddr(pi_desc) >> 32;
+        new_ire.post.p = 1;
+    }
 
     if ( pdev )
         set_msi_source_id(pdev, &new_ire);
     else
         set_hpet_source_id(msi_desc->hpet_id, &new_ire);
-    new_ire.remap.res_3 = 0;
-    new_ire.remap.res_4 = 0;
-    new_ire.remap.p = 1;    /* finally, set present bit */
 
     /* now construct new MSI/MSI-X rte entry */
     remap_rte = (struct msi_msg_remap_entry *)msg;
@@ -902,61 +904,16 @@ void iommu_disable_x2apic_IR(void)
         disable_qinval(drhd->iommu);
 }
 
-static void setup_posted_irte(
-    struct iremap_entry *new_ire, const struct iremap_entry *old_ire,
-    const struct pi_desc *pi_desc, const uint8_t gvec)
-{
-    memset(new_ire, 0, sizeof(*new_ire));
-
-    /*
-     * 'im' filed decides whether the irte is in posted format (with value 1)
-     * or remapped format (with value 0), if the old irte is in remapped format,
-     * we copy things from remapped part in 'struct iremap_entry', otherwise,
-     * we copy from posted part.
-     */
-    if ( !old_ire->remap.im )
-    {
-        new_ire->post.p = old_ire->remap.p;
-        new_ire->post.fpd = old_ire->remap.fpd;
-        new_ire->post.sid = old_ire->remap.sid;
-        new_ire->post.sq = old_ire->remap.sq;
-        new_ire->post.svt = old_ire->remap.svt;
-    }
-    else
-    {
-        new_ire->post.p = old_ire->post.p;
-        new_ire->post.fpd = old_ire->post.fpd;
-        new_ire->post.sid = old_ire->post.sid;
-        new_ire->post.sq = old_ire->post.sq;
-        new_ire->post.svt = old_ire->post.svt;
-        new_ire->post.urg = old_ire->post.urg;
-    }
-
-    new_ire->post.im = 1;
-    new_ire->post.vector = gvec;
-    new_ire->post.pda_l = virt_to_maddr(pi_desc) >> (32 - PDA_LOW_BIT);
-    new_ire->post.pda_h = virt_to_maddr(pi_desc) >> 32;
-}
-
 /*
  * This function is used to update the IRTE for posted-interrupt
  * when guest changes MSI/MSI-X information.
  */
-int pi_update_irte(const struct vcpu *v, const struct pirq *pirq,
+int pi_update_irte(const struct pi_desc *pi_desc, const struct pirq *pirq,
     const uint8_t gvec)
 {
     struct irq_desc *desc;
-    const struct msi_desc *msi_desc;
-    int remap_index;
-    int rc = 0;
-    const struct pci_dev *pci_dev;
-    const struct acpi_drhd_unit *drhd;
-    struct iommu *iommu;
-    struct ir_ctrl *ir_ctrl;
-    struct iremap_entry *iremap_entries = NULL, *p = NULL;
-    struct iremap_entry new_ire, old_ire;
-    const struct pi_desc *pi_desc = &v->arch.hvm_vmx.pi_desc;
-    __uint128_t ret;
+    struct msi_desc *msi_desc;
+    int rc;
 
     desc = pirq_spin_lock_irq_desc(pirq, NULL);
     if ( !desc )
@@ -968,59 +925,13 @@ int pi_update_irte(const struct vcpu *v, const struct pirq *pirq,
         rc = -ENODEV;
         goto unlock_out;
     }
-
-    pci_dev = msi_desc->dev;
-    if ( !pci_dev )
-    {
-        rc = -ENODEV;
-        goto unlock_out;
-    }
-
-    remap_index = msi_desc->remap_index;
+    msi_desc->pi_desc = pi_desc;
+    msi_desc->gvec = gvec;
 
     spin_unlock_irq(&desc->lock);
 
     ASSERT(pcidevs_locked());
-
-    /*
-     * FIXME: For performance reasons we should store the 'iommu' pointer in
-     * 'struct msi_desc' in some other place, so we don't need to waste
-     * time searching it here.
-     */
-    drhd = acpi_find_matched_drhd_unit(pci_dev);
-    if ( !drhd )
-        return -ENODEV;
-
-    iommu = drhd->iommu;
-    ir_ctrl = iommu_ir_ctrl(iommu);
-    if ( !ir_ctrl )
-        return -ENODEV;
-
-    spin_lock_irq(&ir_ctrl->iremap_lock);
-
-    GET_IREMAP_ENTRY(ir_ctrl->iremap_maddr, remap_index, iremap_entries, p);
-
-    old_ire = *p;
-
-    /* Setup/Update interrupt remapping table entry. */
-    setup_posted_irte(&new_ire, &old_ire, pi_desc, gvec);
-    ret = cmpxchg16b(p, &old_ire, &new_ire);
-
-    /*
-     * In the above, we use cmpxchg16 to atomically update the 128-bit IRTE,
-     * and the hardware cannot update the IRTE behind us, so the return value
-     * of cmpxchg16 should be the same as old_ire. This ASSERT validate it.
-     */
-    ASSERT(ret == old_ire.val);
-
-    iommu_flush_cache_entry(p, sizeof(*p));
-    iommu_flush_iec_index(iommu, 0, remap_index);
-
-    unmap_vtd_domain_page(iremap_entries);
-
-    spin_unlock_irq(&ir_ctrl->iremap_lock);
-
-    return 0;
+    return iommu_update_ire_from_msi(msi_desc, &msi_desc->msg);
 
  unlock_out:
     spin_unlock_irq(&desc->lock);
index 0008505699ce04fa3f9e39daea9ef9e8558f7636..043123333ee6c59bcc22394fede29a135dc34722 100644 (file)
@@ -18,6 +18,7 @@
 #include <xen/list.h>
 #include <xen/spinlock.h>
 #include <asm/processor.h>
+#include <asm/hvm/vmx/vmcs.h>
 
 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
 #define MAX_IOMMUS 32
@@ -91,7 +92,8 @@ bool_t iommu_supports_eim(void);
 int iommu_enable_x2apic_IR(void);
 void iommu_disable_x2apic_IR(void);
 
-int pi_update_irte(const struct vcpu *v, const struct pirq *pirq, const uint8_t gvec);
+int pi_update_irte(const struct pi_desc *pi_desc, const struct pirq *pirq,
+                   const uint8_t gvec);
 
 #endif /* !__ARCH_X86_IOMMU_H__ */
 /*
index 9c02945893ed3432e969fdfc224a0128f2e22a7f..bf243f8a3597eae98fc2fd1f107ec73023a6a4b7 100644 (file)
@@ -4,6 +4,7 @@
 #include <xen/cpumask.h>
 #include <xen/pci.h>
 #include <asm/byteorder.h>
+#include <asm/hvm/vmx/vmcs.h>
 
 /*
  * Constants for Intel APIC based MSI messages.
@@ -102,6 +103,9 @@ struct msi_desc {
                __u16   entry_nr;       /* specific enabled entry         */
        } msi_attrib;
 
+       uint8_t gvec;                   /* guest vector. valid when pi_desc isn't NULL */
+       const struct pi_desc *pi_desc;  /* pointer to posted descriptor */
+
        struct list_head list;
 
        union {