[Xen-devel] [PATCH v10 12/16] x86/microcode: Synchronize late microcode loading

[Chao Gao <chao.gao@xxxxxxxxx>]

This patch ports microcode improvement patches from linux kernel.

Before you read any further: the early loading method is still the
preferred one and you should always do that. The following patch is
improving the late loading mechanism for long running jobs and cloud use
cases.

Gather all cores and serialize the microcode update on them by doing it
one-by-one to make the late update process as reliable as possible and
avoid potential issues caused by the microcode update.

Signed-off-by: Chao Gao <chao.gao@xxxxxxxxx>
Tested-by: Chao Gao <chao.gao@xxxxxxxxx>
[linux commit: a5321aec6412b20b5ad15db2d6b916c05349dbff]
[linux commit: bb8c13d61a629276a162c1d2b1a20a815cbcfbb7]
Cc: Kevin Tian <kevin.tian@xxxxxxxxx>
Cc: Jun Nakajima <jun.nakajima@xxxxxxxxx>
Cc: Ashok Raj <ashok.raj@xxxxxxxxx>
Cc: Borislav Petkov <bp@xxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
Cc: Jan Beulich <jbeulich@xxxxxxxx>
---
Changes in v10:
 - introduce wait_for_state() and set_state() helper functions
 - make wait_for_condition() return bool and take const void *
 - disable/enable watchdog in control thread
 - rename "master" and "slave" thread to "primary" and "secondary"

Changes in v9:
 - log __buildin_return_address(0) when timeout
 - divide CPUs into three logical sets and they will call different
 functions during ucode loading. The 'control thread' is chosen to
 coordinate ucode loading on all CPUs. Since only control thread would
 set 'loading_state', we can get rid of 'cmpxchg' stuff in v8.
 - s/rep_nop/cpu_relax
 - each thread updates its revision number itself
 - add XENLOG_ERR prefix for each line of multi-line log messages

Changes in v8:
 - to support blocking #NMI handling during loading ucode
   * introduce a flag, 'loading_state', to mark the start or end of
     ucode loading.
   * use a bitmap for cpu callin since if cpu may stay in #NMI handling,
     there are two places for a cpu to call in. bitmap won't be counted
     twice.
   * don't wait for all CPUs callout, just wait for CPUs that perform the
     update. We have to do this because some threads may be stuck in NMI
     handling (where cannot reach the rendezvous).
 - emit a warning if the system stays in stop_machine context for more
 than 1s
 - comment that rdtsc is fine while loading an update
 - use cmpxchg() to avoid panic being called on multiple CPUs
 - Propagate revision number to other threads
 - refine comments and prompt messages

Changes in v7:
 - Check whether 'timeout' is 0 rather than "<=0" since it is unsigned int.
 - reword the comment above microcode_update_cpu() to clearly state that
 one thread per core should do the update.
---
 xen/arch/x86/microcode.c | 296 ++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 269 insertions(+), 27 deletions(-)

diff --git a/xen/arch/x86/microcode.c b/xen/arch/x86/microcode.c
index c2ea20f..049eda6 100644
--- a/xen/arch/x86/microcode.c
+++ b/xen/arch/x86/microcode.c
@@ -30,18 +30,52 @@
 #include <xen/smp.h>
 #include <xen/softirq.h>
 #include <xen/spinlock.h>
+#include <xen/stop_machine.h>
 #include <xen/tasklet.h>
 #include <xen/guest_access.h>
 #include <xen/earlycpio.h>
+#include <xen/watchdog.h>
 
+#include <asm/delay.h>
 #include <asm/msr.h>
 #include <asm/processor.h>
 #include <asm/setup.h>
 #include <asm/microcode.h>
 
+/*
+ * Before performing a late microcode update on any thread, we
+ * rendezvous all cpus in stop_machine context. The timeout for
+ * waiting for cpu rendezvous is 30ms. It is the timeout used by
+ * live patching
+ */
+#define MICROCODE_CALLIN_TIMEOUT_US 30000
+
+/*
+ * Timeout for each thread to complete update is set to 1s. It is a
+ * conservative choice considering all possible interference.
+ */
+#define MICROCODE_UPDATE_TIMEOUT_US 1000000
+
 static module_t __initdata ucode_mod;
 static signed int __initdata ucode_mod_idx;
 static bool_t __initdata ucode_mod_forced;
+static unsigned int nr_cores;
+
+/*
+ * These states help to coordinate CPUs during loading an update.
+ *
+ * The semantics of each state is as follow:
+ *  - LOADING_PREPARE: initial state of 'loading_state'.
+ *  - LOADING_CALLIN: CPUs are allowed to callin.
+ *  - LOADING_ENTER: all CPUs have called in. Initiate ucode loading.
+ *  - LOADING_EXIT: ucode loading is done or aborted.
+ */
+static enum {
+    LOADING_PREPARE,
+    LOADING_CALLIN,
+    LOADING_ENTER,
+    LOADING_EXIT,
+} loading_state;
 
 /*
  * If we scan the initramfs.cpio for the early microcode code
@@ -190,6 +224,16 @@ static DEFINE_SPINLOCK(microcode_mutex);
 DEFINE_PER_CPU(struct cpu_signature, cpu_sig);
 
 /*
+ * Count the CPUs that have entered, exited the rendezvous and succeeded in
+ * microcode update during late microcode update respectively.
+ *
+ * Note that a bitmap is used for callin to allow cpu to set a bit multiple
+ * times. It is required to do busy-loop in #NMI handling.
+ */
+static cpumask_t cpu_callin_map;
+static atomic_t cpu_out, cpu_updated;
+
+/*
  * Return a patch that covers current CPU. If there are multiple patches,
  * return the one with the highest revision number. Return error If no
  * patch is found and an error occurs during the parsing process. Otherwise
@@ -231,6 +275,34 @@ static bool microcode_update_cache(struct microcode_patch 
*patch)
     return true;
 }
 
+/* Wait for a condition to be met with a timeout (us). */
+static int wait_for_condition(bool (*func)(const void *data), void *data,
+                              unsigned int timeout)
+{
+    while ( !func(data) )
+    {
+        if ( !timeout-- )
+        {
+            printk("CPU%u: Timeout in %pS\n",
+                   smp_processor_id(), __builtin_return_address(0));
+            return -EBUSY;
+        }
+        udelay(1);
+    }
+
+    return 0;
+}
+
+static bool wait_cpu_callin(const void *nr)
+{
+    return cpumask_weight(&cpu_callin_map) >= (unsigned long)nr;
+}
+
+static bool wait_cpu_callout(const void *nr)
+{
+    return atomic_read(&cpu_out) >= (unsigned long)nr;
+}
+
 /*
  * Load a microcode update to current CPU.
  *
@@ -264,38 +336,158 @@ static int microcode_update_cpu(const struct 
microcode_patch *patch)
     return err;
 }
 
-static long do_microcode_update(void *patch)
+static bool wait_for_state(unsigned int state)
+{
+    while ( loading_state != state )
+    {
+        if ( state != LOADING_EXIT && loading_state == LOADING_EXIT )
+            return false;
+        cpu_relax();
+    }
+
+    return true;
+}
+
+static void set_state(unsigned int state)
+{
+    loading_state = state;
+    smp_wmb();
+}
+
+static int secondary_thread_fn(void)
+{
+    unsigned int primary = cpumask_first(this_cpu(cpu_sibling_mask));
+
+    if ( !wait_for_state(LOADING_CALLIN) )
+        return -EBUSY;
+
+    cpumask_set_cpu(smp_processor_id(), &cpu_callin_map);
+
+    if ( !wait_for_state(LOADING_EXIT) )
+        return -EBUSY;
+
+    /* Copy update revision from the primary thread. */
+    this_cpu(cpu_sig).rev = per_cpu(cpu_sig, primary).rev;
+
+    return 0;
+}
+
+static int primary_thread_fn(const struct microcode_patch *patch)
+{
+    int ret = 0;
+
+    if ( !wait_for_state(LOADING_CALLIN) )
+        return -EBUSY;
+
+    cpumask_set_cpu(smp_processor_id(), &cpu_callin_map);
+
+    if ( !wait_for_state(LOADING_ENTER) )
+        return -EBUSY;
+
+    ret = microcode_ops->apply_microcode(patch);
+    if ( !ret )
+        atomic_inc(&cpu_updated);
+    atomic_inc(&cpu_out);
+
+    return ret;
+}
+
+static int control_thread_fn(const struct microcode_patch *patch)
 {
-    unsigned int cpu;
-    int ret = microcode_update_cpu(patch);
+    unsigned int cpu = smp_processor_id(), done;
+    unsigned long tick;
+    int ret;
+
+    /*
+     * We intend to disable interrupt for long time, which may lead to
+     * watchdog timeout.
+     */
+    watchdog_disable();
 
-    /* Store the patch after a successful loading */
-    if ( !ret && patch )
+    /* Allow threads to call in */
+    set_state(LOADING_CALLIN);
+
+    cpumask_set_cpu(cpu, &cpu_callin_map);
+
+    /* Waiting for all threads calling in */
+    ret = wait_for_condition(wait_cpu_callin,
+                             (void *)(unsigned long)num_online_cpus(),
+                             MICROCODE_CALLIN_TIMEOUT_US);
+    if ( ret )
     {
-        spin_lock(&microcode_mutex);
-        microcode_update_cache(patch);
-        spin_unlock(&microcode_mutex);
-        patch = NULL;
+        set_state(LOADING_EXIT);
+        return ret;
     }
 
-    if ( microcode_ops->end_update_percpu )
-        microcode_ops->end_update_percpu();
+    /* Let primary threads load the given ucode update */
+    set_state(LOADING_ENTER);
 
+    ret = microcode_ops->apply_microcode(patch);
+    if ( !ret )
+        atomic_inc(&cpu_updated);
+    atomic_inc(&cpu_out);
+
+    tick = rdtsc_ordered();
+    /* Wait for primary threads finishing update */
+    done = atomic_read(&cpu_out);
+    while ( done != nr_cores )
+    {
+        /*
+         * During each timeout interval, at least a CPU is expected to
+         * finish its update. Otherwise, something goes wrong.
+         *
+         * Note that RDTSC (in wait_for_condition()) is safe for threads to
+         * execute while waiting for completion of loading an update.
+         */
+        if ( wait_for_condition(wait_cpu_callout,
+                                (void *)(unsigned long)(done + 1),
+                                MICROCODE_UPDATE_TIMEOUT_US) )
+            panic("Timeout when finished updating microcode (finished %u/%u)",
+                  done, nr_cores);
+
+        /* Print warning message once if long time is spent here */
+        if ( tick && rdtsc_ordered() - tick >= cpu_khz * 1000 )
+        {
+            printk(XENLOG_WARNING
+                   "WARNING: UPDATING MICROCODE HAS CONSUMED MORE THAN 1 
SECOND!\n");
+            tick = 0;
+        }
+        done = atomic_read(&cpu_out);
+    }
+
+    /* Mark loading is done to unblock other threads */
+    set_state(LOADING_EXIT);
+
+    watchdog_enable();
+
+    return ret;
+}
+
+static int do_microcode_update(void *patch)
+{
+    unsigned int cpu = smp_processor_id();
     /*
-     * Each thread tries to load ucode and only the first thread of a core
-     * would succeed. Ignore error other than -EIO.
+     * primary thread is the one with the lowest thread id among all siblings
+     * thread in a core or a compute unit. It is chosen to load a microcode
+     * update.
      */
-    if ( ret != -EIO )
-        ret = 0;
+    unsigned int primary = cpumask_first(this_cpu(cpu_sibling_mask));
+    int ret;
 
-    cpu = cpumask_next(smp_processor_id(), &cpu_online_map);
-    if ( cpu < nr_cpu_ids )
-        return continue_hypercall_on_cpu(cpu, do_microcode_update, patch) ?
-                                                                          : 
ret;
+    /*
+     * The control thread set state to coordinate ucode loading. Primary
+     * threads load the given ucode patch. Secondary threads just wait for
+     * the completion of the ucode loading process.
+     */
+    if ( cpu == cpumask_first(&cpu_online_map) )
+        ret = control_thread_fn(patch);
+    else if ( cpu == primary )
+        ret = primary_thread_fn(patch);
+    else
+        ret = secondary_thread_fn();
 
-    /* Free the patch if no CPU has loaded it successfully. */
-    if ( patch )
-        microcode_free_patch(patch);
+    if ( microcode_ops->end_update_percpu )
+        microcode_ops->end_update_percpu();
 
     return ret;
 }
@@ -304,6 +496,7 @@ int microcode_update(XEN_GUEST_HANDLE_PARAM(const_void) 
buf, unsigned long len)
 {
     int ret;
     void *buffer;
+    unsigned int cpu, updated;
     struct microcode_patch *patch;
 
     if ( len != (uint32_t)len )
@@ -322,18 +515,25 @@ int microcode_update(XEN_GUEST_HANDLE_PARAM(const_void) 
buf, unsigned long len)
         goto free;
     }
 
+    /* cpu_online_map must not change during update */
+    if ( !get_cpu_maps() )
+    {
+        ret = -EBUSY;
+        goto free;
+    }
+
     patch = parse_blob(buffer, len);
     if ( IS_ERR(patch) )
     {
         ret = PTR_ERR(patch);
         printk(XENLOG_WARNING "Parsing microcode blob error %d\n", ret);
-        goto free;
+        goto put;
     }
 
     if ( !patch )
     {
         ret = -ENOENT;
-        goto free;
+        goto put;
     }
 
     if ( microcode_ops->start_update )
@@ -342,13 +542,55 @@ int microcode_update(XEN_GUEST_HANDLE_PARAM(const_void) 
buf, unsigned long len)
         if ( ret != 0 )
         {
             microcode_free_patch(patch);
-            goto free;
+            goto put;
         }
     }
 
-    ret = continue_hypercall_on_cpu(cpumask_first(&cpu_online_map),
-                                    do_microcode_update, patch);
+    cpumask_clear(&cpu_callin_map);
+    atomic_set(&cpu_out, 0);
+    atomic_set(&cpu_updated, 0);
+    loading_state = LOADING_PREPARE;
+
+    /* Calculate the number of online CPU core */
+    nr_cores = 0;
+    for_each_online_cpu(cpu)
+        if ( cpu == cpumask_first(per_cpu(cpu_sibling_mask, cpu)) )
+            nr_cores++;
+
+    printk(XENLOG_INFO "%u cores are to update their microcode\n", nr_cores);
+
+    /*
+     * Late loading dance. Why the heavy-handed stop_machine effort?
+     *
+     * - HT siblings must be idle and not execute other code while the other
+     *   sibling is loading microcode in order to avoid any negative
+     *   interactions cause by the loading.
+     *
+     * - In addition, microcode update on the cores must be serialized until
+     *   this requirement can be relaxed in the future. Right now, this is
+     *   conservative and good.
+     */
+    ret = stop_machine_run(do_microcode_update, patch, NR_CPUS);
+
+    updated = atomic_read(&cpu_updated);
+    if ( updated > 0 )
+    {
+        spin_lock(&microcode_mutex);
+        microcode_update_cache(patch);
+        spin_unlock(&microcode_mutex);
+    }
+    else
+        microcode_free_patch(patch);
+
+    if ( updated && updated != nr_cores )
+        printk(XENLOG_ERR "ERROR: Updating microcode succeeded on %u cores and 
failed\n"
+               XENLOG_ERR "on other %u cores. A system with differing 
microcode\n"
+               XENLOG_ERR "revisions is considered unstable. Please reboot and 
do not\n"
+               XENLOG_ERR "load the microcode that triggers this warning!\n",
+               updated, nr_cores - updated);
 
+ put:
+    put_cpu_maps();
  free:
     xfree(buffer);
     return ret;
-- 
1.8.3.1


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel