x86/ldt: Make modify_ldt synchronous
authorAndy Lutomirski <luto@kernel.org>
Thu, 30 Jul 2015 21:31:32 +0000 (14:31 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 21 Sep 2015 17:05:25 +0000 (10:05 -0700)
commit 37868fe113ff2ba814b3b4eb12df214df555f8dc upstream.

modify_ldt() has questionable locking and does not synchronize
threads.  Improve it: redesign the locking and synchronize all
threads' LDTs using an IPI on all modifications.

This will dramatically slow down modify_ldt in multithreaded
programs, but there shouldn't be any multithreaded programs that
care about modify_ldt's performance in the first place.

This fixes some fallout from the CVE-2015-5157 fixes.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Reviewed-by: Borislav Petkov <bp@suse.de>
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Jan Beulich <jbeulich@suse.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: security@kernel.org <security@kernel.org>
Cc: xen-devel <xen-devel@lists.xen.org>
Link: http://lkml.kernel.org/r/4c6978476782160600471bd865b318db34c7b628.1438291540.git.luto@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
arch/x86/include/asm/desc.h
arch/x86/include/asm/mmu.h
arch/x86/include/asm/mmu_context.h
arch/x86/kernel/cpu/common.c
arch/x86/kernel/cpu/perf_event.c
arch/x86/kernel/ldt.c
arch/x86/kernel/process_64.c
arch/x86/kernel/step.c
arch/x86/power/cpu.c

index a0bf89f..4e10d73 100644 (file)
@@ -280,21 +280,6 @@ static inline void clear_LDT(void)
        set_ldt(NULL, 0);
 }
 
-/*
- * load one particular LDT into the current CPU
- */
-static inline void load_LDT_nolock(mm_context_t *pc)
-{
-       set_ldt(pc->ldt, pc->size);
-}
-
-static inline void load_LDT(mm_context_t *pc)
-{
-       preempt_disable();
-       load_LDT_nolock(pc);
-       preempt_enable();
-}
-
 static inline unsigned long get_desc_base(const struct desc_struct *desc)
 {
        return (unsigned)(desc->base0 | ((desc->base1) << 16) | ((desc->base2) << 24));
index 09b9620..364d274 100644 (file)
@@ -9,8 +9,7 @@
  * we put the segment information here.
  */
 typedef struct {
-       void *ldt;
-       int size;
+       struct ldt_struct *ldt;
 
 #ifdef CONFIG_X86_64
        /* True if mm supports a task running in 32 bit compatibility mode. */
index e997f70..80d67dd 100644 (file)
@@ -34,6 +34,50 @@ static inline void load_mm_cr4(struct mm_struct *mm) {}
 #endif
 
 /*
+ * ldt_structs can be allocated, used, and freed, but they are never
+ * modified while live.
+ */
+struct ldt_struct {
+       /*
+        * Xen requires page-aligned LDTs with special permissions.  This is
+        * needed to prevent us from installing evil descriptors such as
+        * call gates.  On native, we could merge the ldt_struct and LDT
+        * allocations, but it's not worth trying to optimize.
+        */
+       struct desc_struct *entries;
+       int size;
+};
+
+static inline void load_mm_ldt(struct mm_struct *mm)
+{
+       struct ldt_struct *ldt;
+
+       /* lockless_dereference synchronizes with smp_store_release */
+       ldt = lockless_dereference(mm->context.ldt);
+
+       /*
+        * Any change to mm->context.ldt is followed by an IPI to all
+        * CPUs with the mm active.  The LDT will not be freed until
+        * after the IPI is handled by all such CPUs.  This means that,
+        * if the ldt_struct changes before we return, the values we see
+        * will be safe, and the new values will be loaded before we run
+        * any user code.
+        *
+        * NB: don't try to convert this to use RCU without extreme care.
+        * We would still need IRQs off, because we don't want to change
+        * the local LDT after an IPI loaded a newer value than the one
+        * that we can see.
+        */
+
+       if (unlikely(ldt))
+               set_ldt(ldt->entries, ldt->size);
+       else
+               clear_LDT();
+
+       DEBUG_LOCKS_WARN_ON(preemptible());
+}
+
+/*
  * Used for LDT copy/destruction.
  */
 int init_new_context(struct task_struct *tsk, struct mm_struct *mm);
@@ -78,12 +122,12 @@ static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
                 * was called and then modify_ldt changed
                 * prev->context.ldt but suppressed an IPI to this CPU.
                 * In this case, prev->context.ldt != NULL, because we
-                * never free an LDT while the mm still exists.  That
-                * means that next->context.ldt != prev->context.ldt,
-                * because mms never share an LDT.
+                * never set context.ldt to NULL while the mm still
+                * exists.  That means that next->context.ldt !=
+                * prev->context.ldt, because mms never share an LDT.
                 */
                if (unlikely(prev->context.ldt != next->context.ldt))
-                       load_LDT_nolock(&next->context);
+                       load_mm_ldt(next);
        }
 #ifdef CONFIG_SMP
          else {
@@ -106,7 +150,7 @@ static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next,
                        load_cr3(next->pgd);
                        trace_tlb_flush(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
                        load_mm_cr4(next);
-                       load_LDT_nolock(&next->context);
+                       load_mm_ldt(next);
                }
        }
 #endif
index a62cf04..205e0f3 100644 (file)
@@ -1434,7 +1434,7 @@ void cpu_init(void)
        load_sp0(t, &current->thread);
        set_tss_desc(cpu, t);
        load_TR_desc();
-       load_LDT(&init_mm.context);
+       load_mm_ldt(&init_mm);
 
        clear_all_debug_regs();
        dbg_restore_debug_regs();
@@ -1483,7 +1483,7 @@ void cpu_init(void)
        load_sp0(t, thread);
        set_tss_desc(cpu, t);
        load_TR_desc();
-       load_LDT(&init_mm.context);
+       load_mm_ldt(&init_mm);
 
        t->x86_tss.io_bitmap_base = offsetof(struct tss_struct, io_bitmap);
 
index aa4e3a7..4cc98a4 100644 (file)
@@ -2170,21 +2170,25 @@ static unsigned long get_segment_base(unsigned int segment)
        int idx = segment >> 3;
 
        if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
+               struct ldt_struct *ldt;
+
                if (idx > LDT_ENTRIES)
                        return 0;
 
-               if (idx > current->active_mm->context.size)
+               /* IRQs are off, so this synchronizes with smp_store_release */
+               ldt = lockless_dereference(current->active_mm->context.ldt);
+               if (!ldt || idx > ldt->size)
                        return 0;
 
-               desc = current->active_mm->context.ldt;
+               desc = &ldt->entries[idx];
        } else {
                if (idx > GDT_ENTRIES)
                        return 0;
 
-               desc = raw_cpu_ptr(gdt_page.gdt);
+               desc = raw_cpu_ptr(gdt_page.gdt) + idx;
        }
 
-       return get_desc_base(desc + idx);
+       return get_desc_base(desc);
 }
 
 #ifdef CONFIG_COMPAT
index c37886d..2bcc052 100644 (file)
@@ -12,6 +12,7 @@
 #include <linux/string.h>
 #include <linux/mm.h>
 #include <linux/smp.h>
+#include <linux/slab.h>
 #include <linux/vmalloc.h>
 #include <linux/uaccess.h>
 
 #include <asm/mmu_context.h>
 #include <asm/syscalls.h>
 
-#ifdef CONFIG_SMP
+/* context.lock is held for us, so we don't need any locking. */
 static void flush_ldt(void *current_mm)
 {
-       if (current->active_mm == current_mm)
-               load_LDT(&current->active_mm->context);
+       mm_context_t *pc;
+
+       if (current->active_mm != current_mm)
+               return;
+
+       pc = &current->active_mm->context;
+       set_ldt(pc->ldt->entries, pc->ldt->size);
 }
-#endif
 
-static int alloc_ldt(mm_context_t *pc, int mincount, int reload)
+/* The caller must call finalize_ldt_struct on the result. LDT starts zeroed. */
+static struct ldt_struct *alloc_ldt_struct(int size)
 {
-       void *oldldt, *newldt;
-       int oldsize;
-
-       if (mincount <= pc->size)
-               return 0;
-       oldsize = pc->size;
-       mincount = (mincount + (PAGE_SIZE / LDT_ENTRY_SIZE - 1)) &
-                       (~(PAGE_SIZE / LDT_ENTRY_SIZE - 1));
-       if (mincount * LDT_ENTRY_SIZE > PAGE_SIZE)
-               newldt = vmalloc(mincount * LDT_ENTRY_SIZE);
+       struct ldt_struct *new_ldt;
+       int alloc_size;
+
+       if (size > LDT_ENTRIES)
+               return NULL;
+
+       new_ldt = kmalloc(sizeof(struct ldt_struct), GFP_KERNEL);
+       if (!new_ldt)
+               return NULL;
+
+       BUILD_BUG_ON(LDT_ENTRY_SIZE != sizeof(struct desc_struct));
+       alloc_size = size * LDT_ENTRY_SIZE;
+
+       /*
+        * Xen is very picky: it requires a page-aligned LDT that has no
+        * trailing nonzero bytes in any page that contains LDT descriptors.
+        * Keep it simple: zero the whole allocation and never allocate less
+        * than PAGE_SIZE.
+        */
+       if (alloc_size > PAGE_SIZE)
+               new_ldt->entries = vzalloc(alloc_size);
        else
-               newldt = (void *)__get_free_page(GFP_KERNEL);
-
-       if (!newldt)
-               return -ENOMEM;
+               new_ldt->entries = kzalloc(PAGE_SIZE, GFP_KERNEL);
 
-       if (oldsize)
-               memcpy(newldt, pc->ldt, oldsize * LDT_ENTRY_SIZE);
-       oldldt = pc->ldt;
-       memset(newldt + oldsize * LDT_ENTRY_SIZE, 0,
-              (mincount - oldsize) * LDT_ENTRY_SIZE);
+       if (!new_ldt->entries) {
+               kfree(new_ldt);
+               return NULL;
+       }
 
-       paravirt_alloc_ldt(newldt, mincount);
+       new_ldt->size = size;
+       return new_ldt;
+}
 
-#ifdef CONFIG_X86_64
-       /* CHECKME: Do we really need this ? */
-       wmb();
-#endif
-       pc->ldt = newldt;
-       wmb();
-       pc->size = mincount;
-       wmb();
-
-       if (reload) {
-#ifdef CONFIG_SMP
-               preempt_disable();
-               load_LDT(pc);
-               if (!cpumask_equal(mm_cpumask(current->mm),
-                                  cpumask_of(smp_processor_id())))
-                       smp_call_function(flush_ldt, current->mm, 1);
-               preempt_enable();
-#else
-               load_LDT(pc);
-#endif
-       }
-       if (oldsize) {
-               paravirt_free_ldt(oldldt, oldsize);
-               if (oldsize * LDT_ENTRY_SIZE > PAGE_SIZE)
-                       vfree(oldldt);
-               else
-                       put_page(virt_to_page(oldldt));
-       }
-       return 0;
+/* After calling this, the LDT is immutable. */
+static void finalize_ldt_struct(struct ldt_struct *ldt)
+{
+       paravirt_alloc_ldt(ldt->entries, ldt->size);
 }
 
-static inline int copy_ldt(mm_context_t *new, mm_context_t *old)
+/* context.lock is held */
+static void install_ldt(struct mm_struct *current_mm,
+                       struct ldt_struct *ldt)
 {
-       int err = alloc_ldt(new, old->size, 0);
-       int i;
+       /* Synchronizes with lockless_dereference in load_mm_ldt. */
+       smp_store_release(&current_mm->context.ldt, ldt);
+
+       /* Activate the LDT for all CPUs using current_mm. */
+       on_each_cpu_mask(mm_cpumask(current_mm), flush_ldt, current_mm, true);
+}
 
-       if (err < 0)
-               return err;
+static void free_ldt_struct(struct ldt_struct *ldt)
+{
+       if (likely(!ldt))
+               return;
 
-       for (i = 0; i < old->size; i++)
-               write_ldt_entry(new->ldt, i, old->ldt + i * LDT_ENTRY_SIZE);
-       return 0;
+       paravirt_free_ldt(ldt->entries, ldt->size);
+       if (ldt->size * LDT_ENTRY_SIZE > PAGE_SIZE)
+               vfree(ldt->entries);
+       else
+               kfree(ldt->entries);
+       kfree(ldt);
 }
 
 /*
@@ -104,17 +105,37 @@ static inline int copy_ldt(mm_context_t *new, mm_context_t *old)
  */
 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 {
+       struct ldt_struct *new_ldt;
        struct mm_struct *old_mm;
        int retval = 0;
 
        mutex_init(&mm->context.lock);
-       mm->context.size = 0;
        old_mm = current->mm;
-       if (old_mm && old_mm->context.size > 0) {
-               mutex_lock(&old_mm->context.lock);
-               retval = copy_ldt(&mm->context, &old_mm->context);
-               mutex_unlock(&old_mm->context.lock);
+       if (!old_mm) {
+               mm->context.ldt = NULL;
+               return 0;
        }
+
+       mutex_lock(&old_mm->context.lock);
+       if (!old_mm->context.ldt) {
+               mm->context.ldt = NULL;
+               goto out_unlock;
+       }
+
+       new_ldt = alloc_ldt_struct(old_mm->context.ldt->size);
+       if (!new_ldt) {
+               retval = -ENOMEM;
+               goto out_unlock;
+       }
+
+       memcpy(new_ldt->entries, old_mm->context.ldt->entries,
+              new_ldt->size * LDT_ENTRY_SIZE);
+       finalize_ldt_struct(new_ldt);
+
+       mm->context.ldt = new_ldt;
+
+out_unlock:
+       mutex_unlock(&old_mm->context.lock);
        return retval;
 }
 
@@ -125,53 +146,47 @@ int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
  */
 void destroy_context(struct mm_struct *mm)
 {
-       if (mm->context.size) {
-#ifdef CONFIG_X86_32
-               /* CHECKME: Can this ever happen ? */
-               if (mm == current->active_mm)
-                       clear_LDT();
-#endif
-               paravirt_free_ldt(mm->context.ldt, mm->context.size);
-               if (mm->context.size * LDT_ENTRY_SIZE > PAGE_SIZE)
-                       vfree(mm->context.ldt);
-               else
-                       put_page(virt_to_page(mm->context.ldt));
-               mm->context.size = 0;
-       }
+       free_ldt_struct(mm->context.ldt);
+       mm->context.ldt = NULL;
 }
 
 static int read_ldt(void __user *ptr, unsigned long bytecount)
 {
-       int err;
+       int retval;
        unsigned long size;
        struct mm_struct *mm = current->mm;
 
-       if (!mm->context.size)
-               return 0;
+       mutex_lock(&mm->context.lock);
+
+       if (!mm->context.ldt) {
+               retval = 0;
+               goto out_unlock;
+       }
+
        if (bytecount > LDT_ENTRY_SIZE * LDT_ENTRIES)
                bytecount = LDT_ENTRY_SIZE * LDT_ENTRIES;
 
-       mutex_lock(&mm->context.lock);
-       size = mm->context.size * LDT_ENTRY_SIZE;
+       size = mm->context.ldt->size * LDT_ENTRY_SIZE;
        if (size > bytecount)
                size = bytecount;
 
-       err = 0;
-       if (copy_to_user(ptr, mm->context.ldt, size))
-               err = -EFAULT;
-       mutex_unlock(&mm->context.lock);
-       if (err < 0)
-               goto error_return;
+       if (copy_to_user(ptr, mm->context.ldt->entries, size)) {
+               retval = -EFAULT;
+               goto out_unlock;
+       }
+
        if (size != bytecount) {
-               /* zero-fill the rest */
-               if (clear_user(ptr + size, bytecount - size) != 0) {
-                       err = -EFAULT;
-                       goto error_return;
+               /* Zero-fill the rest and pretend we read bytecount bytes. */
+               if (clear_user(ptr + size, bytecount - size)) {
+                       retval = -EFAULT;
+                       goto out_unlock;
                }
        }
-       return bytecount;
-error_return:
-       return err;
+       retval = bytecount;
+
+out_unlock:
+       mutex_unlock(&mm->context.lock);
+       return retval;
 }
 
 static int read_default_ldt(void __user *ptr, unsigned long bytecount)
@@ -195,6 +210,8 @@ static int write_ldt(void __user *ptr, unsigned long bytecount, int oldmode)
        struct desc_struct ldt;
        int error;
        struct user_desc ldt_info;
+       int oldsize, newsize;
+       struct ldt_struct *new_ldt, *old_ldt;
 
        error = -EINVAL;
        if (bytecount != sizeof(ldt_info))
@@ -213,34 +230,39 @@ static int write_ldt(void __user *ptr, unsigned long bytecount, int oldmode)
                        goto out;
        }
 
-       mutex_lock(&mm->context.lock);
-       if (ldt_info.entry_number >= mm->context.size) {
-               error = alloc_ldt(&current->mm->context,
-                                 ldt_info.entry_number + 1, 1);
-               if (error < 0)
-                       goto out_unlock;
-       }
-
-       /* Allow LDTs to be cleared by the user. */
-       if (ldt_info.base_addr == 0 && ldt_info.limit == 0) {
-               if (oldmode || LDT_empty(&ldt_info)) {
-                       memset(&ldt, 0, sizeof(ldt));
-                       goto install;
+       if ((oldmode && !ldt_info.base_addr && !ldt_info.limit) ||
+           LDT_empty(&ldt_info)) {
+               /* The user wants to clear the entry. */
+               memset(&ldt, 0, sizeof(ldt));
+       } else {
+               if (!IS_ENABLED(CONFIG_X86_16BIT) && !ldt_info.seg_32bit) {
+                       error = -EINVAL;
+                       goto out;
                }
+
+               fill_ldt(&ldt, &ldt_info);
+               if (oldmode)
+                       ldt.avl = 0;
        }
 
-       if (!IS_ENABLED(CONFIG_X86_16BIT) && !ldt_info.seg_32bit) {
-               error = -EINVAL;
+       mutex_lock(&mm->context.lock);
+
+       old_ldt = mm->context.ldt;
+       oldsize = old_ldt ? old_ldt->size : 0;
+       newsize = max((int)(ldt_info.entry_number + 1), oldsize);
+
+       error = -ENOMEM;
+       new_ldt = alloc_ldt_struct(newsize);
+       if (!new_ldt)
                goto out_unlock;
-       }
 
-       fill_ldt(&ldt, &ldt_info);
-       if (oldmode)
-               ldt.avl = 0;
+       if (old_ldt)
+               memcpy(new_ldt->entries, old_ldt->entries, oldsize * LDT_ENTRY_SIZE);
+       new_ldt->entries[ldt_info.entry_number] = ldt;
+       finalize_ldt_struct(new_ldt);
 
-       /* Install the new entry ...  */
-install:
-       write_ldt_entry(mm->context.ldt, ldt_info.entry_number, &ldt);
+       install_ldt(mm, new_ldt);
+       free_ldt_struct(old_ldt);
        error = 0;
 
 out_unlock:
index ddfdbf7..5e0bf57 100644 (file)
@@ -122,11 +122,11 @@ void __show_regs(struct pt_regs *regs, int all)
 void release_thread(struct task_struct *dead_task)
 {
        if (dead_task->mm) {
-               if (dead_task->mm->context.size) {
+               if (dead_task->mm->context.ldt) {
                        pr_warn("WARNING: dead process %s still has LDT? <%p/%d>\n",
                                dead_task->comm,
                                dead_task->mm->context.ldt,
-                               dead_task->mm->context.size);
+                               dead_task->mm->context.ldt->size);
                        BUG();
                }
        }
index 9b4d51d..6273324 100644 (file)
@@ -5,6 +5,7 @@
 #include <linux/mm.h>
 #include <linux/ptrace.h>
 #include <asm/desc.h>
+#include <asm/mmu_context.h>
 
 unsigned long convert_ip_to_linear(struct task_struct *child, struct pt_regs *regs)
 {
@@ -30,10 +31,11 @@ unsigned long convert_ip_to_linear(struct task_struct *child, struct pt_regs *re
                seg &= ~7UL;
 
                mutex_lock(&child->mm->context.lock);
-               if (unlikely((seg >> 3) >= child->mm->context.size))
+               if (unlikely(!child->mm->context.ldt ||
+                            (seg >> 3) >= child->mm->context.ldt->size))
                        addr = -1L; /* bogus selector, access would fault */
                else {
-                       desc = child->mm->context.ldt + seg;
+                       desc = &child->mm->context.ldt->entries[seg];
                        base = get_desc_base(desc);
 
                        /* 16-bit code segment? */
index 757678f..bf93844 100644 (file)
@@ -23,6 +23,7 @@
 #include <asm/debugreg.h>
 #include <asm/fpu-internal.h> /* pcntxt_mask */
 #include <asm/cpu.h>
+#include <asm/mmu_context.h>
 
 #ifdef CONFIG_X86_32
 __visible unsigned long saved_context_ebx;
@@ -154,7 +155,7 @@ static void fix_processor_context(void)
        syscall_init();                         /* This sets MSR_*STAR and related */
 #endif
        load_TR_desc();                         /* This does ltr */
-       load_LDT(&current->active_mm->context); /* This does lldt */
+       load_mm_ldt(current->active_mm);        /* This does lldt */
 }
 
 /**