[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-changelog] [xen staging-4.12] x86/ctxt-switch: Document and improve GDT handling



commit e5122c6c4a413e5d185b25094a15b4ee6d5daf4b
Author:     Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
AuthorDate: Fri Jul 19 16:14:08 2019 +0200
Commit:     Jan Beulich <jbeulich@xxxxxxxx>
CommitDate: Fri Jul 19 16:14:08 2019 +0200

    x86/ctxt-switch: Document and improve GDT handling
    
    Calling virt_to_mfn() in the context switch path is a lot
    of wasted cycles for a result which is constant after boot.
    
    Begin by documenting how Xen handles the GDTs across context switch.
    
    The loop in write_full_gdt_ptes() is unnecessary, because
    NR_RESERVED_GDT_PAGES is 1.  Dropping it makes the code substantially
    more clear, and with it dropped, write_full_gdt_ptes() becomes more
    obviously a poor name, so rename it to update_xen_slot_in_full_gdt().
    
    Furthermore, load_full_gdt() is completely independent of the current
    CPU, and load_default_gdt() only needs the current CPU's regular
    GDT.  (This is a change in behaviour, as previously it may have used the
    compat GDT, but either will do.)
    
    Add two extra per-cpu variables which cache the L1e for the regular and 
compat
    GDT, calculated in cpu_smpboot_alloc()/trap_init() as appropriate, so
    update_xen_slot_in_full_gdt() doesn't need to waste time performing the same
    calculation on every context switch.
    
    One performance scenario of Jüergen's (time to build the hypervisor on
    an 8 CPU system, with two single-vCPU MiniOS VMs constantly interrupting
    dom0 with events) shows the following, average over 5 measurements:
    
                elapsed  user   system
      Unpatched  66.51  232.93  109.21
      Patched    57.00  225.47  105.47
    
    which is a substantial improvement.
    
    Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
    Reviewed-by: Juergen Gross <jgross@xxxxxxxx>
    Tested-by: Juergen Gross <jgross@xxxxxxxx>
    Reviewed-by: Jan Beulich <jbeulich@xxxxxxxx>
    master commit: 564d261687c071c5a77fa81d693b7ca3d8e83b48
    master date: 2019-07-10 09:40:25 -0500
---
 xen/arch/x86/domain.c      | 52 ++++++++++++++++++++++++++++++----------------
 xen/arch/x86/smpboot.c     |  4 ++++
 xen/arch/x86/traps.c       | 10 +++++++++
 xen/include/asm-x86/desc.h |  2 ++
 4 files changed, 50 insertions(+), 18 deletions(-)

diff --git a/xen/arch/x86/domain.c b/xen/arch/x86/domain.c
index e07f2b37c7..2585327834 100644
--- a/xen/arch/x86/domain.c
+++ b/xen/arch/x86/domain.c
@@ -1608,23 +1608,42 @@ static void _update_runstate_area(struct vcpu *v)
         v->arch.pv.need_update_runstate_area = 1;
 }
 
+/*
+ * Overview of Xen's GDTs.
+ *
+ * Xen maintains per-CPU compat and regular GDTs which are both a single page
+ * in size.  Some content is specific to each CPU (the TSS, the per-CPU marker
+ * for #DF handling, and optionally the LDT).  The compat and regular GDTs
+ * differ by the layout and content of the guest accessible selectors.
+ *
+ * The Xen selectors live from 0xe000 (slot 14 of 16), and need to always
+ * appear in this position for interrupt/exception handling to work.
+ *
+ * A PV guest may specify GDT frames of their own (slots 0 to 13).  Room for a
+ * full GDT exists in the per-domain mappings.
+ *
+ * To schedule a PV vcpu, we point slot 14 of the guest's full GDT at the
+ * current CPU's compat or regular (as appropriate) GDT frame.  This is so
+ * that the per-CPU parts still work correctly after switching pagetables and
+ * loading the guests full GDT into GDTR.
+ *
+ * To schedule Idle or HVM vcpus, we load a GDT base address which causes the
+ * regular per-CPU GDT frame to appear with selectors at the appropriate
+ * offset.
+ */
 static inline bool need_full_gdt(const struct domain *d)
 {
     return is_pv_domain(d) && !is_idle_domain(d);
 }
 
-static void write_full_gdt_ptes(seg_desc_t *gdt, const struct vcpu *v)
+static void update_xen_slot_in_full_gdt(const struct vcpu *v, unsigned int cpu)
 {
-    unsigned long mfn = virt_to_mfn(gdt);
-    l1_pgentry_t *pl1e = pv_gdt_ptes(v);
-    unsigned int i;
-
-    for ( i = 0; i < NR_RESERVED_GDT_PAGES; i++ )
-        l1e_write(pl1e + FIRST_RESERVED_GDT_PAGE + i,
-                  l1e_from_pfn(mfn + i, __PAGE_HYPERVISOR_RW));
+    l1e_write(pv_gdt_ptes(v) + FIRST_RESERVED_GDT_PAGE,
+              !is_pv_32bit_vcpu(v) ? per_cpu(gdt_table_l1e, cpu)
+                                   : per_cpu(compat_gdt_table_l1e, cpu));
 }
 
-static void load_full_gdt(const struct vcpu *v, unsigned int cpu)
+static void load_full_gdt(const struct vcpu *v)
 {
     struct desc_ptr gdt_desc = {
         .limit = LAST_RESERVED_GDT_BYTE,
@@ -1634,11 +1653,12 @@ static void load_full_gdt(const struct vcpu *v, 
unsigned int cpu)
     lgdt(&gdt_desc);
 }
 
-static void load_default_gdt(const seg_desc_t *gdt, unsigned int cpu)
+static void load_default_gdt(unsigned int cpu)
 {
     struct desc_ptr gdt_desc = {
         .limit = LAST_RESERVED_GDT_BYTE,
-        .base  = (unsigned long)(gdt - FIRST_RESERVED_GDT_ENTRY),
+        .base  = (unsigned long)(per_cpu(gdt_table, cpu) -
+                                 FIRST_RESERVED_GDT_ENTRY),
     };
 
     lgdt(&gdt_desc);
@@ -1651,7 +1671,6 @@ static void __context_switch(void)
     struct vcpu          *p = per_cpu(curr_vcpu, cpu);
     struct vcpu          *n = current;
     struct domain        *pd = p->domain, *nd = n->domain;
-    seg_desc_t           *gdt;
 
     ASSERT(p != n);
     ASSERT(!vcpu_cpu_dirty(n));
@@ -1691,15 +1710,12 @@ static void __context_switch(void)
 
     psr_ctxt_switch_to(nd);
 
-    gdt = !is_pv_32bit_domain(nd) ? per_cpu(gdt_table, cpu) :
-                                    per_cpu(compat_gdt_table, cpu);
-
     if ( need_full_gdt(nd) )
-        write_full_gdt_ptes(gdt, n);
+        update_xen_slot_in_full_gdt(n, cpu);
 
     if ( need_full_gdt(pd) &&
          ((p->vcpu_id != n->vcpu_id) || !need_full_gdt(nd)) )
-        load_default_gdt(gdt, cpu);
+        load_default_gdt(cpu);
 
     write_ptbase(n);
 
@@ -1712,7 +1728,7 @@ static void __context_switch(void)
 
     if ( need_full_gdt(nd) &&
          ((p->vcpu_id != n->vcpu_id) || !need_full_gdt(pd)) )
-        load_full_gdt(n, cpu);
+        load_full_gdt(n);
 
     if ( pd != nd )
         cpumask_clear_cpu(cpu, pd->dirty_cpumask);
diff --git a/xen/arch/x86/smpboot.c b/xen/arch/x86/smpboot.c
index 7d1226d7bc..331cd14177 100644
--- a/xen/arch/x86/smpboot.c
+++ b/xen/arch/x86/smpboot.c
@@ -990,6 +990,8 @@ static int cpu_smpboot_alloc(unsigned int cpu)
     if ( gdt == NULL )
         goto out;
     per_cpu(gdt_table, cpu) = gdt;
+    per_cpu(gdt_table_l1e, cpu) =
+        l1e_from_pfn(virt_to_mfn(gdt), __PAGE_HYPERVISOR_RW);
     memcpy(gdt, boot_cpu_gdt_table, NR_RESERVED_GDT_PAGES * PAGE_SIZE);
     BUILD_BUG_ON(NR_CPUS > 0x10000);
     gdt[PER_CPU_GDT_ENTRY - FIRST_RESERVED_GDT_ENTRY].a = cpu;
@@ -997,6 +999,8 @@ static int cpu_smpboot_alloc(unsigned int cpu)
     per_cpu(compat_gdt_table, cpu) = gdt = alloc_xenheap_pages(order, 
memflags);
     if ( gdt == NULL )
         goto out;
+    per_cpu(compat_gdt_table_l1e, cpu) =
+        l1e_from_pfn(virt_to_mfn(gdt), __PAGE_HYPERVISOR_RW);
     memcpy(gdt, boot_cpu_compat_gdt_table, NR_RESERVED_GDT_PAGES * PAGE_SIZE);
     gdt[PER_CPU_GDT_ENTRY - FIRST_RESERVED_GDT_ENTRY].a = cpu;
 
diff --git a/xen/arch/x86/traps.c b/xen/arch/x86/traps.c
index 05ddc39bfe..d313d88e77 100644
--- a/xen/arch/x86/traps.c
+++ b/xen/arch/x86/traps.c
@@ -97,7 +97,9 @@ DEFINE_PER_CPU(uint64_t, efer);
 static DEFINE_PER_CPU(unsigned long, last_extable_addr);
 
 DEFINE_PER_CPU_READ_MOSTLY(seg_desc_t *, gdt_table);
+DEFINE_PER_CPU_READ_MOSTLY(l1_pgentry_t, gdt_table_l1e);
 DEFINE_PER_CPU_READ_MOSTLY(seg_desc_t *, compat_gdt_table);
+DEFINE_PER_CPU_READ_MOSTLY(l1_pgentry_t, compat_gdt_table_l1e);
 
 /* Master table, used by CPU0. */
 idt_entry_t __section(".bss.page_aligned") __aligned(PAGE_SIZE)
@@ -2056,6 +2058,14 @@ void __init trap_init(void)
         }
     }
 
+    /* Cache {,compat_}gdt_table_l1e now that physically relocation is done. */
+    this_cpu(gdt_table_l1e) =
+        l1e_from_pfn(virt_to_mfn(boot_cpu_gdt_table),
+                     __PAGE_HYPERVISOR_RW);
+    this_cpu(compat_gdt_table_l1e) =
+        l1e_from_pfn(virt_to_mfn(boot_cpu_compat_gdt_table),
+                     __PAGE_HYPERVISOR_RW);
+
     percpu_traps_init();
 
     cpu_init();
diff --git a/xen/include/asm-x86/desc.h b/xen/include/asm-x86/desc.h
index 85e83bcefb..e565727dc0 100644
--- a/xen/include/asm-x86/desc.h
+++ b/xen/include/asm-x86/desc.h
@@ -206,8 +206,10 @@ struct __packed desc_ptr {
 
 extern seg_desc_t boot_cpu_gdt_table[];
 DECLARE_PER_CPU(seg_desc_t *, gdt_table);
+DECLARE_PER_CPU(l1_pgentry_t, gdt_table_l1e);
 extern seg_desc_t boot_cpu_compat_gdt_table[];
 DECLARE_PER_CPU(seg_desc_t *, compat_gdt_table);
+DECLARE_PER_CPU(l1_pgentry_t, compat_gdt_table_l1e);
 
 extern void load_TR(void);
 
--
generated by git-patchbot for /home/xen/git/xen.git#staging-4.12

_______________________________________________
Xen-changelog mailing list
Xen-changelog@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/xen-changelog

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.