[Qemu-devel] [RFC v4 2/9] softmmu: Add new TLB_EXCL flag

[Alvise Rigo <a.rigo@virtualopensystems.com>]

Add a new TLB flag to force all the accesses made to a page to follow
the slow-path.

In the case we remove a TLB entry marked as EXCL, we unset the
corresponding exclusive bit in the bitmap.

Suggested-by: Jani Kokkonen <jani.kokkonen@huawei.com>
Suggested-by: Claudio Fontana <claudio.fontana@huawei.com>
Signed-off-by: Alvise Rigo <a.rigo@virtualopensystems.com>
---
 cputlb.c                |  39 ++++++++++++++++-
 include/exec/cpu-all.h  |   8 ++++
 include/exec/cpu-defs.h |  12 ++++++
 softmmu_template.h      | 112 ++++++++++++++++++++++++++++++++++++++----------
 4 files changed, 148 insertions(+), 23 deletions(-)

diff --git a/cputlb.c b/cputlb.c
index a506086..251eec8 100644
--- a/cputlb.c
+++ b/cputlb.c
@@ -299,6 +299,13 @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
     env->tlb_v_table[mmu_idx][vidx] = *te;
     env->iotlb_v[mmu_idx][vidx] = env->iotlb[mmu_idx][index];
 
+    if (unlikely(!(te->addr_write & TLB_MMIO) && (te->addr_write & TLB_EXCL))) {
+        /* We are removing an exclusive entry, set the page to dirty. */
+        hwaddr hw_addr = (env->iotlb[mmu_idx][index].addr & TARGET_PAGE_MASK) +
+                                          (te->addr_write & TARGET_PAGE_MASK);
+        cpu_physical_memory_set_excl_dirty(hw_addr, cpu->cpu_index);
+    }
+
     /* refill the tlb */
     env->iotlb[mmu_idx][index].addr = iotlb - vaddr;
     env->iotlb[mmu_idx][index].attrs = attrs;
@@ -324,7 +331,15 @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
                                                    + xlat)) {
             te->addr_write = address | TLB_NOTDIRTY;
         } else {
-            te->addr_write = address;
+            if (!(address & TLB_MMIO) &&
+                cpu_physical_memory_excl_atleast_one_clean(section->mr->ram_addr
+                                                           + xlat)) {
+                /* There is at least one vCPU that has flagged the address as
+                 * exclusive. */
+                te->addr_write = address | TLB_EXCL;
+            } else {
+                te->addr_write = address;
+            }
         }
     } else {
         te->addr_write = -1;
@@ -376,6 +391,28 @@ tb_page_addr_t get_page_addr_code(CPUArchState *env1, target_ulong addr)
     return qemu_ram_addr_from_host_nofail(p);
 }
 
+/* Atomic insn translation TLB support. */
+#define EXCLUSIVE_RESET_ADDR ULLONG_MAX
+/* For every vCPU compare the exclusive address and reset it in case of a
+ * match. Since only one vCPU is running at once, no lock has to be held to
+ * guard this operation. */
+static inline void lookup_and_reset_cpus_ll_addr(hwaddr addr, hwaddr size)
+{
+    CPUState *cpu;
+    CPUArchState *acpu;
+
+    CPU_FOREACH(cpu) {
+        acpu = (CPUArchState *)cpu->env_ptr;
+
+        if (acpu->excl_protected_range.begin != EXCLUSIVE_RESET_ADDR &&
+            ranges_overlap(acpu->excl_protected_range.begin,
+            acpu->excl_protected_range.end - acpu->excl_protected_range.begin,
+            addr, size)) {
+            acpu->excl_protected_range.begin = EXCLUSIVE_RESET_ADDR;
+        }
+    }
+}
+
 #define MMUSUFFIX _mmu
 
 #define SHIFT 0
diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h
index ea6a9a6..ad6afcb 100644
--- a/include/exec/cpu-all.h
+++ b/include/exec/cpu-all.h
@@ -320,6 +320,14 @@ extern RAMList ram_list;
 #define TLB_NOTDIRTY    (1 << 4)
 /* Set if TLB entry is an IO callback.  */
 #define TLB_MMIO        (1 << 5)
+/* Set if TLB entry references a page that requires exclusive access.  */
+#define TLB_EXCL        (1 << 6)
+
+/* Do not allow a TARGET_PAGE_MASK which covers one or more bits defined
+ * above. */
+#if TLB_EXCL >= TARGET_PAGE_SIZE
+#error TARGET_PAGE_MASK covering the low bits of the TLB virtual address
+#endif
 
 void dump_exec_info(FILE *f, fprintf_function cpu_fprintf);
 void dump_opcount_info(FILE *f, fprintf_function cpu_fprintf);
diff --git a/include/exec/cpu-defs.h b/include/exec/cpu-defs.h
index 98b9cff..a67f295 100644
--- a/include/exec/cpu-defs.h
+++ b/include/exec/cpu-defs.h
@@ -27,6 +27,7 @@
 #include <inttypes.h>
 #include "qemu/osdep.h"
 #include "qemu/queue.h"
+#include "qemu/range.h"
 #include "tcg-target.h"
 #ifndef CONFIG_USER_ONLY
 #include "exec/hwaddr.h"
@@ -150,5 +151,16 @@ typedef struct CPUIOTLBEntry {
 #define CPU_COMMON                                                      \
     /* soft mmu support */                                              \
     CPU_COMMON_TLB                                                      \
+                                                                        \
+    /* Used by the atomic insn translation backend. */                  \
+    int ll_sc_context;                                                  \
+    /* vCPU current exclusive addresses range.
+     * The address is set to EXCLUSIVE_RESET_ADDR if the vCPU is not.
+     * in the middle of a LL/SC. */                                     \
+    struct Range excl_protected_range;                                  \
+    /* Used to carry the SC result but also to flag a normal (legacy)
+     * store access made by a stcond (see softmmu_template.h). */       \
+    int excl_succeeded;                                                 \
+
 
 #endif
diff --git a/softmmu_template.h b/softmmu_template.h
index d42d89d..e4431e8 100644
--- a/softmmu_template.h
+++ b/softmmu_template.h
@@ -409,19 +409,53 @@ void helper_le_st_name(CPUArchState *env, target_ulong addr, DATA_TYPE val,
         tlb_addr = env->tlb_table[mmu_idx][index].addr_write;
     }
 
-    /* Handle an IO access.  */
+    /* Handle an IO access or exclusive access.  */
     if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
-        CPUIOTLBEntry *iotlbentry;
-        if ((addr & (DATA_SIZE - 1)) != 0) {
-            goto do_unaligned_access;
+        CPUIOTLBEntry *iotlbentry = &env->iotlb[mmu_idx][index];
+
+        if ((tlb_addr & ~TARGET_PAGE_MASK) == TLB_EXCL) {
+            /* The slow-path has been forced since we are writing to
+             * exclusive-protected memory. */
+            hwaddr hw_addr = (iotlbentry->addr & TARGET_PAGE_MASK) + addr;
+
+            /* The function lookup_and_reset_cpus_ll_addr could have reset the
+             * exclusive address. Fail the SC in this case.
+             * N.B.: Here excl_succeeded == 0 means that helper_le_st_name has
+             * not been called by a softmmu_llsc_template.h. */
+            if(env->excl_succeeded) {
+                if (env->excl_protected_range.begin != hw_addr) {
+                    /* The vCPU is SC-ing to an unprotected address. */
+                    env->excl_protected_range.begin = EXCLUSIVE_RESET_ADDR;
+                    env->excl_succeeded = 0;
+
+                    return;
+                }
+
+                cpu_physical_memory_set_excl_dirty(hw_addr, ENV_GET_CPU(env)->cpu_index);
+            }
+
+            haddr = addr + env->tlb_table[mmu_idx][index].addend;
+        #if DATA_SIZE == 1
+            glue(glue(st, SUFFIX), _p)((uint8_t *)haddr, val);
+        #else
+            glue(glue(st, SUFFIX), _le_p)((uint8_t *)haddr, val);
+        #endif
+
+            lookup_and_reset_cpus_ll_addr(hw_addr, DATA_SIZE);
+
+            return;
+        } else {
+            if ((addr & (DATA_SIZE - 1)) != 0) {
+                goto do_unaligned_access;
+            }
+            iotlbentry = &env->iotlb[mmu_idx][index];
+
+            /* ??? Note that the io helpers always read data in the target
+               byte ordering.  We should push the LE/BE request down into io.  */
+            val = TGT_LE(val);
+            glue(io_write, SUFFIX)(env, iotlbentry, val, addr, retaddr);
+            return;
         }
-        iotlbentry = &env->iotlb[mmu_idx][index];
-
-        /* ??? Note that the io helpers always read data in the target
-           byte ordering.  We should push the LE/BE request down into io.  */
-        val = TGT_LE(val);
-        glue(io_write, SUFFIX)(env, iotlbentry, val, addr, retaddr);
-        return;
     }
 
     /* Handle slow unaligned access (it spans two pages or IO).  */
@@ -489,19 +523,53 @@ void helper_be_st_name(CPUArchState *env, target_ulong addr, DATA_TYPE val,
         tlb_addr = env->tlb_table[mmu_idx][index].addr_write;
     }
 
-    /* Handle an IO access.  */
+    /* Handle an IO access or exclusive access.  */
     if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
-        CPUIOTLBEntry *iotlbentry;
-        if ((addr & (DATA_SIZE - 1)) != 0) {
-            goto do_unaligned_access;
+        CPUIOTLBEntry *iotlbentry = &env->iotlb[mmu_idx][index];
+
+        if ((tlb_addr & ~TARGET_PAGE_MASK) == TLB_EXCL) {
+            /* The slow-path has been forced since we are writing to
+             * exclusive-protected memory. */
+            hwaddr hw_addr = (iotlbentry->addr & TARGET_PAGE_MASK) + addr;
+
+            /* The function lookup_and_reset_cpus_ll_addr could have reset the
+             * exclusive address. Fail the SC in this case.
+             * N.B.: Here excl_succeeded == 0 means that helper_le_st_name has
+             * not been called by a softmmu_llsc_template.h. */
+            if(env->excl_succeeded) {
+                if (env->excl_protected_range.begin != hw_addr) {
+                    /* The vCPU is SC-ing to an unprotected address. */
+                    env->excl_protected_range.begin = EXCLUSIVE_RESET_ADDR;
+                    env->excl_succeeded = 0;
+
+                    return;
+                }
+
+                cpu_physical_memory_set_excl_dirty(hw_addr, ENV_GET_CPU(env)->cpu_index);
+            }
+
+            haddr = addr + env->tlb_table[mmu_idx][index].addend;
+        #if DATA_SIZE == 1
+            glue(glue(st, SUFFIX), _p)((uint8_t *)haddr, val);
+        #else
+            glue(glue(st, SUFFIX), _le_p)((uint8_t *)haddr, val);
+        #endif
+
+            lookup_and_reset_cpus_ll_addr(hw_addr, DATA_SIZE);
+
+            return;
+        } else {
+            if ((addr & (DATA_SIZE - 1)) != 0) {
+                goto do_unaligned_access;
+            }
+            iotlbentry = &env->iotlb[mmu_idx][index];
+
+            /* ??? Note that the io helpers always read data in the target
+               byte ordering.  We should push the LE/BE request down into io.  */
+            val = TGT_BE(val);
+            glue(io_write, SUFFIX)(env, iotlbentry, val, addr, retaddr);
+            return;
         }
-        iotlbentry = &env->iotlb[mmu_idx][index];
-
-        /* ??? Note that the io helpers always read data in the target
-           byte ordering.  We should push the LE/BE request down into io.  */
-        val = TGT_BE(val);
-        glue(io_write, SUFFIX)(env, iotlbentry, val, addr, retaddr);
-        return;
     }
 
     /* Handle slow unaligned access (it spans two pages or IO).  */
-- 
2.5.0