Re: [PATCH 2/6] kallsyms: Emit symbol at the holes in the text

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

 



On 2024/7/20 22:14, Masahiro Yamada wrote:
On Thu, Jul 18, 2024 at 12:45 PM Zheng Yejian <zhengyejian1@xxxxxxxxxx> wrote:

On 2024/7/16 16:33, Masahiro Yamada wrote:
On Thu, Jun 13, 2024 at 10:36 PM Zheng Yejian <zhengyejian1@xxxxxxxxxx> wrote:

When a weak type function is overridden, its symbol will be removed
from the symbol table, but its code will not be removed. Besides,
due to lacking of size for kallsyms, kernel compute function size by
substracting its symbol address from its next symbol address (see
kallsyms_lookup_size_offset()). These will cause that size of some
function is computed to be larger than it actually is, just because
symbol of its following weak function is removed.

This issue also causes multiple __fentry__ locations to be counted in
the some function scope, and eventually causes ftrace_location() to find
wrong __fentry__ location. It was reported in
Link: https://lore.kernel.org/all/20240607115211.734845-1-zhengyejian1@xxxxxxxxxx/

Peter suggested to change scipts/kallsyms.c to emit readily
identifiable symbol names for all the weak junk, eg:

    __weak_junk_NNNNN

The name of this kind symbol needs some discussion, but it's temporarily
called "__hole_symbol_XXXXX" in this patch:
1. Pass size info to scripts/kallsyms  (see mksysmap());
2. Traverse sorted function symbols, if one function address plus its
     size less than next function address, it means there's a hole, then
     emit a symbol "__hole_symbol_XXXXX" there which type is 't'.

After this patch, the effect is as follows:

    $ cat /proc/kallsyms | grep -A 3 do_one_initcall
    ffffffff810021e0 T do_one_initcall
    ffffffff8100245e t __hole_symbol_XXXXX
    ffffffff810024a0 t __pfx_rootfs_init_fs_context

Suggested-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Signed-off-by: Zheng Yejian <zhengyejian1@xxxxxxxxxx>



With my quick test, "t__hole_symbol_XXXXX" was encoded
into the following 10-byte stream.

.byte 0x09, 0x94, 0xbf, 0x18, 0xf3, 0x3d, 0xce, 0xd1, 0xd1, 0x58



Now "t__hole_symbol_XXXXX" is the most common symbol name.
However, 10 byte is consumed for every instance of
"t__hole_symbol_XXXXX".

This is much less efficient thanI had expected,
although I did not analyze the logic of this inefficiency.

Hi, Masahiro!

In my local test, "t__hole_symbol_XXXXX" was finally encoded
into just one byte. See "kallsyms_token_table" in the .tmp_vmlinux.kallsyms2.S:

    kallsyms_token_table:
          [...]
          .asciz  "t__hole_symbol_XXXXX"
          .asciz  "hole_symbol_XXXXX"
          .asciz  "e_symbol_XXXXX"
          .asciz  "XXXXX"
          .asciz  "XXX"
          .asciz  "e_symbol_"
          .asciz  "ymbol_"
          .asciz  "ymb"
          .asciz  "hol"
          .asciz  "ol_"
          .asciz  "pfx"
          .asciz  "pf"
          .asciz  "e_s"
          .asciz  "ym"
          .asciz  "t__"
          .asciz  "_s"
          .asciz  "ol"
          .asciz  "__"
          .asciz  "XX"

But it would still takes up several tokens due to substrings of
"t__hole_symbol_XXXXX" would also become the most common ones.
After this patch, the number of "t__hole_symbol_XXXXX" will be ~30% of the total.







---
   scripts/kallsyms.c      | 101 +++++++++++++++++++++++++++++++++++++++-
   scripts/link-vmlinux.sh |   4 +-
   scripts/mksysmap        |   2 +-
   3 files changed, 102 insertions(+), 5 deletions(-)

diff --git a/scripts/kallsyms.c b/scripts/kallsyms.c
index 6559a9802f6e..5c4cde864a04 100644
--- a/scripts/kallsyms.c
+++ b/scripts/kallsyms.c
@@ -35,6 +35,7 @@
   struct sym_entry {
          struct sym_entry *next;
          unsigned long long addr;
+       unsigned long long size;
          unsigned int len;
          unsigned int seq;
          unsigned int start_pos;
@@ -74,6 +75,7 @@ static int token_profit[0x10000];
   static unsigned char best_table[256][2];
   static unsigned char best_table_len[256];

+static const char hole_symbol[] = "__hole_symbol_XXXXX";

   static void usage(void)
   {
@@ -130,8 +132,16 @@ static struct sym_entry *read_symbol(FILE *in, char **buf, size_t *buf_len)
          size_t len;
          ssize_t readlen;
          struct sym_entry *sym;
+       unsigned long long size = 0;

          errno = 0;
+       /*
+        * Example of expected symbol format:
+        * 1. symbol with size info:
+        *    ffffffff81000070 00000000000001d7 T __startup_64
+        * 2. symbol without size info:
+        *    0000000002a00000 A text_size
+        */
          readlen = getline(buf, buf_len, in);
          if (readlen < 0) {
                  if (errno) {
@@ -145,9 +155,24 @@ static struct sym_entry *read_symbol(FILE *in, char **buf, size_t *buf_len)
                  (*buf)[readlen - 1] = 0;

          addr = strtoull(*buf, &p, 16);
+       if (*buf == p || *p++ != ' ') {
+               fprintf(stderr, "line format error: unable to parse address\n");
+               exit(EXIT_FAILURE);
+       }
+
+       if (*p == '0') {
+               char *str = p;

-       if (*buf == p || *p++ != ' ' || !isascii((type = *p++)) || *p++ != ' ') {
-               fprintf(stderr, "line format error\n");
+               size = strtoull(str, &p, 16);
+               if (str == p || *p++ != ' ') {
+                       fprintf(stderr, "line format error: unable to parse size\n");
+                       exit(EXIT_FAILURE);
+               }
+       }
+
+       type = *p++;
+       if (!isascii(type) || *p++ != ' ') {
+               fprintf(stderr, "line format error: unable to parse type\n");
                  exit(EXIT_FAILURE);
          }

@@ -182,6 +207,7 @@ static struct sym_entry *read_symbol(FILE *in, char **buf, size_t *buf_len)
                  exit(EXIT_FAILURE);
          }
          sym->addr = addr;
+       sym->size = size;
          sym->len = len;
          sym->sym[0] = type;
          strcpy(sym_name(sym), name);
@@ -795,6 +821,76 @@ static void sort_symbols(void)
          qsort(table, table_cnt, sizeof(table[0]), compare_symbols);
   }

+static int may_exist_hole_after_symbol(const struct sym_entry *se)


The return type should be bool.


Yes!



+{
+       char type = se->sym[0];
+
+       /* Only check text symbol or weak symbol */
+       if (type != 't' && type != 'T' &&
+           type != 'w' && type != 'W')
+               return 0;
+       /* Symbol without size has no hole */
+       return se->size != 0;
+}
+
+static struct sym_entry *gen_hole_symbol(unsigned long long addr)
+{
+       struct sym_entry *sym;
+       static size_t len = sizeof(hole_symbol);
+
+       /* include type field */
+       sym = malloc(sizeof(*sym) + len + 1);
+       if (!sym) {
+               fprintf(stderr, "unable to allocate memory for hole symbol\n");
+               exit(EXIT_FAILURE);
+       }
+       sym->addr = addr;
+       sym->size = 0;
+       sym->len = len;
+       sym->sym[0] = 't';
+       strcpy(sym_name(sym), hole_symbol);
+       sym->percpu_absolute = 0;
+       return sym;
+}
+
+static void emit_hole_symbols(void)
+{
+       unsigned int i, pos, nr_emit;
+       struct sym_entry **new_table;
+       unsigned int new_cnt;
+
+       nr_emit = 0;
+       for (i = 0; i < table_cnt - 1; i++) {
+               if (may_exist_hole_after_symbol(table[i]) &&
+                   table[i]->addr + table[i]->size < table[i+1]->addr)
+                       nr_emit++;
+       }
+       if (!nr_emit)
+               return;
+
+       new_cnt = table_cnt + nr_emit;
+       new_table = malloc(sizeof(*new_table) * new_cnt);


Do you need to allocate another huge table?

You can use realloc() to append the room for nr_emit
if you iterate the table in the reverse order.


Yes, it would be much better. If it turns out to be the
"emit hole symbol" solution, I'll change it to that in the next version,
actually, I forgot to mark this series as "RFC".


"__hole_symbol_XXXXX" is too much.

You can use the empty symbol type/name as a special case
to represent the hole.


Ok, I'll try it in v2.



--
Thanks,
Zheng Yejian





[Index of Archives]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Big List of Linux Books]

  Powered by Linux