@@ -61,6 +61,15 @@ static int all_symbols;
static int absolute_percpu;
static int base_relative;
+/*
+ * Each entry in the symbol table consists of the symbol type and the symbol
+ * itself. To optimize the performance of finding or traversing symbols in
+ * kernel, do not compress the symbol type. In this way, when looking for a
+ * symbol of unknown type, we can first compress the searched symbol and then
+ * make a quick comparison based on the compressed length and content.
+ */
+static int sym_start_idx = 1;
+
static int token_profit[0x10000];
/* the table that holds the result of the compression */
@@ -511,7 +520,7 @@ static void learn_symbol(const unsigned char *symbol, int len)
{
int i;
- for (i = 0; i < len - 1; i++)
+ for (i = sym_start_idx; i < len - 1; i++)
token_profit[ symbol[i] + (symbol[i + 1] << 8) ]++;
}
@@ -520,7 +529,7 @@ static void forget_symbol(const unsigned char *symbol, int len)
{
int i;
- for (i = 0; i < len - 1; i++)
+ for (i = sym_start_idx; i < len - 1; i++)
token_profit[ symbol[i] + (symbol[i + 1] << 8) ]--;
}
@@ -538,7 +547,7 @@ static unsigned char *find_token(unsigned char *str, int len,
{
int i;
- for (i = 0; i < len - 1; i++) {
+ for (i = sym_start_idx; i < len - 1; i++) {
if (str[i] == token[0] && str[i+1] == token[1])
return &str[i];
}
Currently, to search for a symbol, we need to expand the symbols in 'kallsyms_names' one by one, and then use the expanded string for comparison. Because we do not know the symbol type, and the symbol type may be combined with the following characters to form a token. So if we don't compress the symbol type, we can first compress the searched symbol and then make a quick comparison based on the compressed length and content. In this way, for entries with mismatched lengths, there is no need to expand and compare strings. And for those matching lengths, there's no need to expand the symbol. This saves a lot of time. According to my test results, the average performance of kallsyms_lookup_name() can be improved by 20 to 30 times. Of course, because the symbol type is forcibly not compressed, the compression rate also decreases. Here are the test results with defconfig: arm64: <<<<<< --------------------------------------------------------------- | ALL | nr_symbols | compressed size | original size | ratio(%) | -----|---------------------------------------------------------| Before | Y | 174094 | 1884938 | 3750653 | 50.25 | After | Y | 174099 | 1960154 | 3750756 | 52.26 | Before | N | 61744 | 725507 | 1222737 | 59.33 | After | N | 61747 | 745733 | 1222801 | 60.98 | --------------------------------------------------------------- The memory overhead is increased by: 73.5KiB and 4.0% if CONFIG_KALLSYMS_ALL=y. 19.8KiB and 2.8% if CONFIG_KALLSYMS_ALL=n. x86: <<<<<<<< --------------------------------------------------------------- | ALL | nr_symbols | compressed size | original size | ratio(%) | -----|---------------------------------------------------------| Before | Y | 131415 | 1697542 | 3161216 | 53.69 | After | Y | 131540 | 1747769 | 3163933 | 55.24 | Before | N | 60695 | 737627 | 1283046 | 57.49 | After | N | 60699 | 754797 | 1283149 | 58.82 | --------------------------------------------------------------- The memory overhead is increased by: 49.0KiB and 3.0% if CONFIG_KALLSYMS_ALL=y. 16.8KiB and 2.3% if CONFIG_KALLSYMS_ALL=n. This additional memory overhead is worth it compared to the performance improvement, I think. Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> --- scripts/kallsyms.c | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-)