qemu-devel
[Top][All Lists]
Advanced

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

Re: [PATCH v3 17/36] plugins: force slow path when plugins instrument me


From: Alex Bennée
Subject: Re: [PATCH v3 17/36] plugins: force slow path when plugins instrument memory ops
Date: Thu, 29 Jun 2023 12:59:50 +0100
User-agent: mu4e 1.11.7; emacs 29.0.92

Richard Henderson <richard.henderson@linaro.org> writes:

> On 6/28/23 11:06, Alex Bennée wrote:
>>> I thought we dropped this patch until we could do something with TLB
>>> accesses.
>> I did suggest something like:
>> --8<---------------cut here---------------start------------->8---
>> modified   include/hw/core/cpu.h
>> @@ -80,10 +80,24 @@ DECLARE_CLASS_CHECKERS(CPUClass, CPU,
>>       typedef struct ArchCPU CpuInstanceType; \
>>       OBJECT_DECLARE_TYPE(ArchCPU, CpuClassType, CPU_MODULE_OBJ_NAME);
>>   +/**
>> + * typedef MMUAccessType - describe the type of access for cputlb
>> + *
>> + * When handling the access to memory we need to know the type of
>> + * access we are doing. Loads and store rely on read and write page
>> + * permissions where as the instruction fetch relies on execute
>> + * permissions. Additional bits are used for TLB access so we can
>> + * suppress instrumentation of memory when the CPU is probing.
>> + */
>>   typedef enum MMUAccessType {
>>       MMU_DATA_LOAD  = 0,
>>       MMU_DATA_STORE = 1,
>> -    MMU_INST_FETCH = 2
>> +    MMU_INST_FETCH = 2,
>> +    /* MMU Mask */
>> +    MMU_VALID_MASK = (MMU_DATA_LOAD | MMU_DATA_STORE | MMU_INST_FETCH),
>> +    /* Represents the CPU walking the page table */
>> +    MMU_TLB_ACCESS = 0x4,
>> +    MMU_TLB_LOAD = MMU_DATA_LOAD | MMU_TLB_ACCESS
>>   } MMUAccessType;
>>     typedef struct CPUWatchpoint CPUWatchpoint;
>> modified   accel/tcg/cputlb.c
>> @@ -1503,11 +1503,12 @@ static void notdirty_write(CPUState *cpu, vaddr 
>> mem_vaddr, unsigned size,
>>   }
>>     static int probe_access_internal(CPUArchState *env, target_ulong
>> addr,
>> -                                 int fault_size, MMUAccessType access_type,
>> +                                 int fault_size, MMUAccessType 
>> full_access_type,
>>                                    int mmu_idx, bool nonfault,
>>                                    void **phost, CPUTLBEntryFull **pfull,
>>                                    uintptr_t retaddr)
>>   {
>> +    MMUAccessType access_type = full_access_type & MMU_VALID_MASK;
>>       uintptr_t index = tlb_index(env, mmu_idx, addr);
>>       CPUTLBEntry *entry = tlb_entry(env, mmu_idx, addr);
>>       target_ulong tlb_addr = tlb_read_idx(entry, access_type);
>> @@ -1546,7 +1547,9 @@ static int probe_access_internal(CPUArchState *env, 
>> target_ulong addr,
>>       /* Fold all "mmio-like" bits into TLB_MMIO.  This is not RAM.  */
>>       if (unlikely(flags & ~(TLB_WATCHPOINT | TLB_NOTDIRTY))
>>           ||
>> -        (access_type != MMU_INST_FETCH && 
>> cpu_plugin_mem_cbs_enabled(env_cpu(env)))) {
>> +        (access_type != MMU_INST_FETCH &&
>> +         !(full_access_type & MMU_TLB_ACCESS) &&
>> +         cpu_plugin_mem_cbs_enabled(env_cpu(env)))) {
>>           *phost = NULL;
>>           return TLB_MMIO;
>>       }
>> --8<---------------cut here---------------end--------------->8---
>> and then we can apply MMU_TLB_LOAD as the type in the page walking
>> code.
>> I wanted to know if that was the sort of thing you where thinking off or
>> if that is too ugly.
>
> It's not implausible, but probably not ideal.
>
>> The other option is a specific probe_access_* function for TLB type
>> operations.
>
> Or that, yes.
>
> I'm confused that you'd simply re-include this patch as-is when it has
> known errors.

I was hoping for some testing from the people that reported the fault.
Anyway this is what I have now:

--8<---------------cut here---------------start------------->8---
plugins: force slow path when plugins instrument memory ops

The lack of SVE memory instrumentation has been an omission in plugin
handling since it was introduced. Fortunately we can utilise the
probe_* functions to force all all memory access to follow the slow
path. We do this by checking the access type and presence of plugin
memory callbacks and if set return the TLB_MMIO flag.

We have to jump through a few hoops in user mode to re-use the flag
but it was the desired effect:

 ./qemu-system-aarch64 -display none -serial mon:stdio \
   -M virt -cpu max -semihosting-config enable=on \
   -kernel ./tests/tcg/aarch64-softmmu/memory-sve \
   -plugin ./contrib/plugins/libexeclog.so,ifilter=st1w,afilter=0x40001808 -d 
plugin

gives (disas doesn't currently understand st1w):

  0, 0x40001808, 0xe54342a0, ".byte 0xa0, 0x42, 0x43, 0xe5", store, 0x40213010, 
RAM, store, 0x40213014, RAM, store, 0x40213018, RAM

And for user-mode:

  ./qemu-aarch64 \
    -plugin contrib/plugins/libexeclog.so,afilter=0x4007c0 \
    -d plugin \
    ./tests/tcg/aarch64-linux-user/sha512-sve

gives:

  1..10
  ok 1 - do_test(&tests[i])
  0, 0x4007c0, 0xa4004b80, ".byte 0x80, 0x4b, 0x00, 0xa4", load, 0x5500800370, 
load, 0x5500800371, load, 0x5500800372, load, 0x5500800373, load, 0x5500800374, 
load, 0x5500800375, load, 0x5500800376, load, 0x5500800377, load, 0x5500800378, 
load, 0x5500800379, load, 0x550080037a, load, 0x550080037b, load, 0x550080037c, 
load, 0x550080037d, load, 0x550080037e, load, 0x550080037f, load, 0x5500800380, 
load, 0x5500800381, load, 0x5500800382, load, 0x5500800383, load, 0x5500800384, 
load, 0x5500800385, load, 0x5500800386, lo
  ad, 0x5500800387, load, 0x5500800388, load, 0x5500800389, load, 0x550080038a, 
load, 0x550080038b, load, 0x550080038c, load, 0x550080038d, load, 0x550080038e, 
load, 0x550080038f, load, 0x5500800390, load, 0x5500800391, load, 0x5500800392, 
load, 0x5500800393, load, 0x5500800394, load, 0x5500800395, load, 0x5500800396, 
load, 0x5500800397, load, 0x5500800398, load, 0x5500800399, load, 0x550080039a, 
load, 0x550080039b, load, 0x550080039c, load, 0x550080039d, load, 0x550080039e, 
load, 0x550080039f, load, 0x55008003a0, load, 0x55008003a1, load, 0x55008003a2, 
load, 0x55008003a3, load, 0x55008003a4, load, 0x55008003a5, load, 0x55008003a6, 
load, 0x55008003a7, load, 0x55008003a8, load, 0x55008003a9, load, 0x55008003aa, 
load, 0x55008003ab, load, 0x55008003ac, load, 0x55008003ad, load, 0x55008003ae, 
load, 0x55008003af

(4007c0 is the ld1b in the sha512-sve)

Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Cc: Robert Henry <robhenry@microsoft.com>
Cc: Aaron Lindsay <aaron@os.amperecomputing.com>

---
v2
 - allow TLB_MMIO to appear in user-mode probe_access
v3
 - checkpatch cleanups
v4
 - add new probe helper for mmu's, add check_mem_cbs to probe_internal

8 files changed, 94 insertions(+), 21 deletions(-)
include/exec/cpu-all.h            |  2 +-
include/exec/exec-all.h           | 23 ++++++++++++++++++++++
include/hw/core/cpu.h             | 17 +++++++++++++++++
accel/tcg/cputlb.c                | 40 ++++++++++++++++++++++++++++++++-------
accel/tcg/user-exec.c             |  8 ++++++--
target/arm/ptw.c                  | 13 ++++++-------
target/arm/tcg/sve_helper.c       |  4 ----
tests/tcg/aarch64/Makefile.target |  8 ++++++++

modified   include/exec/cpu-all.h
@@ -301,7 +301,7 @@ CPUArchState *cpu_copy(CPUArchState *env);
  * be signaled by probe_access_flags().
  */
 #define TLB_INVALID_MASK    (1 << (TARGET_PAGE_BITS_MIN - 1))
-#define TLB_MMIO            0
+#define TLB_MMIO            (1 << (TARGET_PAGE_BITS_MIN - 2))
 #define TLB_WATCHPOINT      0
 
 #else
modified   include/exec/exec-all.h
@@ -464,6 +464,29 @@ int probe_access_full(CPUArchState *env, vaddr addr, int 
size,
                       MMUAccessType access_type, int mmu_idx,
                       bool nonfault, void **phost,
                       CPUTLBEntryFull **pfull, uintptr_t retaddr);
+
+/**
+ * probe_access_mmu() - Like probe_access_full except cannot fault and
+ * doesn't trigger instrumentation.
+ *
+ * @env: CPUArchState
+ * @vaddr: virtual address to probe
+ * @size: size of the probe
+ * @access_type: read, write or execute permission
+ * @mmu_idx: softmmu index
+ * @phost: ptr to return value host address or NULL
+ * @pfull: ptr to return value CPUTLBEntryFull structure or NULL
+ *
+ * The CPUTLBEntryFull structure returned via @pfull is transient
+ * and must be consumed or copied immediately, before any further
+ * access or changes to TLB @mmu_idx.
+ *
+ * Returns: TLB flags as per probe_access_flags()
+ */
+int probe_access_full_mmu(CPUArchState *env, vaddr addr, int size,
+                          MMUAccessType access_type, int mmu_idx,
+                          void **phost, CPUTLBEntryFull **pfull);
+
 #endif
 
 /* Hide the qatomic_read to make code a little easier on the eyes */
modified   include/hw/core/cpu.h
@@ -976,6 +976,23 @@ void cpu_watchpoint_remove_by_ref(CPUState *cpu, 
CPUWatchpoint *watchpoint);
 void cpu_watchpoint_remove_all(CPUState *cpu, int mask);
 #endif
 
+/**
+ * cpu_plugin_mem_cbs_enabled() - are plugin memory callbacks enabled?
+ * @cs: CPUState pointer
+ *
+ * The memory callbacks are installed if a plugin has instrumented an
+ * instruction for memory. This can be useful to know if you want to
+ * force a slow path for a series of memory accesses.
+ */
+static inline bool cpu_plugin_mem_cbs_enabled(const CPUState *cpu)
+{
+#ifdef CONFIG_PLUGIN
+    return !!cpu->plugin_mem_cbs;
+#else
+    return false;
+#endif
+}
+
 /**
  * cpu_get_address_space:
  * @cpu: CPU to get address space from
modified   accel/tcg/cputlb.c
@@ -1513,13 +1513,14 @@ static int probe_access_internal(CPUArchState *env, 
vaddr addr,
                                  int fault_size, MMUAccessType access_type,
                                  int mmu_idx, bool nonfault,
                                  void **phost, CPUTLBEntryFull **pfull,
-                                 uintptr_t retaddr)
+                                 uintptr_t retaddr, bool check_mem_cbs)
 {
     uintptr_t index = tlb_index(env, mmu_idx, addr);
     CPUTLBEntry *entry = tlb_entry(env, mmu_idx, addr);
     uint64_t tlb_addr = tlb_read_idx(entry, access_type);
     vaddr page_addr = addr & TARGET_PAGE_MASK;
     int flags = TLB_FLAGS_MASK & ~TLB_FORCE_SLOW;
+    bool force_mmio = check_mem_cbs && 
cpu_plugin_mem_cbs_enabled(env_cpu(env));
     CPUTLBEntryFull *full;
 
     if (!tlb_hit_page(tlb_addr, page_addr)) {
@@ -1553,7 +1554,9 @@ static int probe_access_internal(CPUArchState *env, vaddr 
addr,
     flags |= full->slow_flags[access_type];
 
     /* Fold all "mmio-like" bits into TLB_MMIO.  This is not RAM.  */
-    if (unlikely(flags & ~(TLB_WATCHPOINT | TLB_NOTDIRTY))) {
+    if (unlikely(flags & ~(TLB_WATCHPOINT | TLB_NOTDIRTY))
+        ||
+        (access_type != MMU_INST_FETCH && force_mmio)) {
         *phost = NULL;
         return TLB_MMIO;
     }
@@ -1569,7 +1572,7 @@ int probe_access_full(CPUArchState *env, vaddr addr, int 
size,
                       uintptr_t retaddr)
 {
     int flags = probe_access_internal(env, addr, size, access_type, mmu_idx,
-                                      nonfault, phost, pfull, retaddr);
+                                      nonfault, phost, pfull, retaddr, true);
 
     /* Handle clean RAM pages.  */
     if (unlikely(flags & TLB_NOTDIRTY)) {
@@ -1580,6 +1583,29 @@ int probe_access_full(CPUArchState *env, vaddr addr, int 
size,
     return flags;
 }
 
+int probe_access_full_mmu(CPUArchState *env, vaddr addr, int size,
+                          MMUAccessType access_type, int mmu_idx,
+                          void **phost, CPUTLBEntryFull **pfull)
+{
+    void *discard_phost;
+    CPUTLBEntryFull *discard_tlb;
+
+    /* privately handle users that don't need full results */
+    phost = phost ? phost : &discard_phost;
+    pfull = pfull ? pfull : &discard_tlb;
+
+    int flags = probe_access_internal(env, addr, size, access_type, mmu_idx,
+                                      true, phost, pfull, 0, false);
+
+    /* Handle clean RAM pages.  */
+    if (unlikely(flags & TLB_NOTDIRTY)) {
+        notdirty_write(env_cpu(env), addr, 1, *pfull, 0);
+        flags &= ~TLB_NOTDIRTY;
+    }
+
+    return flags;
+}
+
 int probe_access_flags(CPUArchState *env, vaddr addr, int size,
                        MMUAccessType access_type, int mmu_idx,
                        bool nonfault, void **phost, uintptr_t retaddr)
@@ -1590,7 +1616,7 @@ int probe_access_flags(CPUArchState *env, vaddr addr, int 
size,
     g_assert(-(addr | TARGET_PAGE_MASK) >= size);
 
     flags = probe_access_internal(env, addr, size, access_type, mmu_idx,
-                                  nonfault, phost, &full, retaddr);
+                                  nonfault, phost, &full, retaddr, true);
 
     /* Handle clean RAM pages. */
     if (unlikely(flags & TLB_NOTDIRTY)) {
@@ -1611,7 +1637,7 @@ void *probe_access(CPUArchState *env, vaddr addr, int 
size,
     g_assert(-(addr | TARGET_PAGE_MASK) >= size);
 
     flags = probe_access_internal(env, addr, size, access_type, mmu_idx,
-                                  false, &host, &full, retaddr);
+                                  false, &host, &full, retaddr, true);
 
     /* Per the interface, size == 0 merely faults the access. */
     if (size == 0) {
@@ -1644,7 +1670,7 @@ void *tlb_vaddr_to_host(CPUArchState *env, abi_ptr addr,
     int flags;
 
     flags = probe_access_internal(env, addr, 0, access_type,
-                                  mmu_idx, true, &host, &full, 0);
+                                  mmu_idx, true, &host, &full, 0, false);
 
     /* No combination of flags are expected by the caller. */
     return flags ? NULL : host;
@@ -1667,7 +1693,7 @@ tb_page_addr_t get_page_addr_code_hostp(CPUArchState 
*env, vaddr addr,
     void *p;
 
     (void)probe_access_internal(env, addr, 1, MMU_INST_FETCH,
-                                cpu_mmu_index(env, true), false, &p, &full, 0);
+                                cpu_mmu_index(env, true), false, &p, &full, 0, 
false);
     if (p == NULL) {
         return -1;
     }
modified   accel/tcg/user-exec.c
@@ -745,6 +745,10 @@ static int probe_access_internal(CPUArchState *env, vaddr 
addr,
     if (guest_addr_valid_untagged(addr)) {
         int page_flags = page_get_flags(addr);
         if (page_flags & acc_flag) {
+            if ((acc_flag == PAGE_READ || acc_flag == PAGE_WRITE)
+                && cpu_plugin_mem_cbs_enabled(env_cpu(env))) {
+                return TLB_MMIO;
+            }
             return 0; /* success */
         }
         maperr = !(page_flags & PAGE_VALID);
@@ -767,7 +771,7 @@ int probe_access_flags(CPUArchState *env, vaddr addr, int 
size,
 
     g_assert(-(addr | TARGET_PAGE_MASK) >= size);
     flags = probe_access_internal(env, addr, size, access_type, nonfault, ra);
-    *phost = flags ? NULL : g2h(env_cpu(env), addr);
+    *phost = (flags & TLB_INVALID_MASK) ? NULL : g2h(env_cpu(env), addr);
     return flags;
 }
 
@@ -778,7 +782,7 @@ void *probe_access(CPUArchState *env, vaddr addr, int size,
 
     g_assert(-(addr | TARGET_PAGE_MASK) >= size);
     flags = probe_access_internal(env, addr, size, access_type, false, ra);
-    g_assert(flags == 0);
+    g_assert((flags & ~TLB_MMIO) == 0);
 
     return size ? g2h(env_cpu(env), addr) : NULL;
 }
modified   target/arm/ptw.c
@@ -489,9 +489,9 @@ static bool S1_ptw_translate(CPUARMState *env, S1Translate 
*ptw,
         int flags;
 
         env->tlb_fi = fi;
-        flags = probe_access_full(env, addr, 0, MMU_DATA_LOAD,
-                                  arm_to_core_mmu_idx(s2_mmu_idx),
-                                  true, &ptw->out_host, &full, 0);
+        flags = probe_access_full_mmu(env, addr, 0, MMU_DATA_LOAD,
+                                      arm_to_core_mmu_idx(s2_mmu_idx),
+                                      &ptw->out_host, &full);
         env->tlb_fi = NULL;
 
         if (unlikely(flags & TLB_INVALID_MASK)) {
@@ -644,12 +644,11 @@ static uint64_t arm_casq_ptw(CPUARMState *env, uint64_t 
old_val,
      */
     if (unlikely(!ptw->out_rw)) {
         int flags;
-        void *discard;
 
         env->tlb_fi = fi;
-        flags = probe_access_flags(env, ptw->out_virt, 0, MMU_DATA_STORE,
-                                   arm_to_core_mmu_idx(ptw->in_ptw_idx),
-                                   true, &discard, 0);
+        flags = probe_access_full_mmu(env, ptw->out_virt, 0,
+                                      MMU_DATA_STORE, 
arm_to_core_mmu_idx(ptw->in_ptw_idx),
+                                      NULL, NULL);
         env->tlb_fi = NULL;
 
         if (unlikely(flags & TLB_INVALID_MASK)) {
modified   target/arm/tcg/sve_helper.c
@@ -5688,9 +5688,6 @@ void sve_ldN_r(CPUARMState *env, uint64_t *vg, const 
target_ulong addr,
 
     flags = info.page[0].flags | info.page[1].flags;
     if (unlikely(flags != 0)) {
-#ifdef CONFIG_USER_ONLY
-        g_assert_not_reached();
-#else
         /*
          * At least one page includes MMIO.
          * Any bus operation can fail with cpu_transaction_failed,
@@ -5727,7 +5724,6 @@ void sve_ldN_r(CPUARMState *env, uint64_t *vg, const 
target_ulong addr,
             memcpy(&env->vfp.zregs[(rd + i) & 31], &scratch[i], reg_max);
         }
         return;
-#endif
     }
 
     /* The entire operation is in RAM, on valid pages. */
--8<---------------cut here---------------end--------------->8---

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro



reply via email to

[Prev in Thread] Current Thread [Next in Thread]