在我们内核里添加调试信息时，很多时候需要记录一段时间内有哪些进程曾经活动过。
自然就需要有一个集合来完成此功能。
用数组来实现集合比较麻烦，而且查找效率较低。因此抄袭内核本身的PID HASH结构，
用hash链表外加一个list结构，来实现一个集合。hash可以提高查找速度，list结构可以
得到集合的概念。
修改方案：
1. kernel/pid.c
#if 1
static __cacheline_aligned_in_smp DEFINE_SPINLOCK(test_pidmap_lock);
static struct hlist_head *test_pid_hash;
struct test_pid_data
{
int magic;
long load;
char comm[TASK_COMM_LEN];
};
struct test_pid
{
int nr;
struct list_head list;
struct test_pid_data data;
struct hlist_node pid_chain;
struct rcu_head rcu;
};
#define test_pid_hashfn(nr) hash_long((unsigned long)nr, pidhash_shift)
struct list_head test_pid_list;
static void init_test_pid_hash_list(void)
{
int pidhash_size,i;
pidhash_size = 1 << pidhash_shift;
test_pid_hash = alloc_bootmem(pidhash_size*sizeof(*(test_pid_hash)));
if (!test_pid_hash)
panic("Could not alloc test_pidhash!\n");
for (i = 0; i < pidhash_size; i++)
INIT_HLIST_HEAD(&test_pid_hash[i]);
INIT_LIST_HEAD(&test_pid_list);
}
static struct test_pid * fastcall test_find_pid(int nr)
{
struct hlist_node *elem;
struct test_pid *pid;
hlist_for_each_entry_rcu(pid, elem,
&test_pid_hash[test_pid_hashfn(nr)], pid_chain) {
if (pid->nr == nr)
return pid;
}
return NULL;
}
static struct test_pid * fastcall test_alloc_pid(int nr)
{
struct test_pid *pid;
pid = (struct test_pid *)kzalloc(sizeof(struct test_pid),GFP_KERNEL);
if(NULL==pid){
return NULL;
}
pid->nr = nr;
spin_lock(&test_pidmap_lock);
hlist_add_head_rcu(&pid->pid_chain, &test_pid_hash[test_pid_hashfn(pid->nr)]);
list_add_rcu(&pid->list,&test_pid_list);
spin_unlock(&test_pidmap_lock);
return pid;
}
typedef void (*test_fn)(struct test_pid * pid);
void test_traverse_pid_list(test_fn callback)
{
struct list_head *p;
struct test_pid *pid;
rcu_read_lock();
list_for_each_rcu(p, &test_pid_list) {
pid = list_entry(p, struct test_pid, list);
callback(pid);
}
rcu_read_unlock();
}
/*
* If not exit,create it and add to set then assign with data
*/
int test_set_pid_data(int nr,struct test_pid_data* test_data)
{
struct test_pid *pid;
pid = test_find_pid(nr);
if(NULL==pid){
pid = test_alloc_pid(nr);
if(NULL==pid)
return -ENOMEM;
}
memcpy(&pid->data,test_data,sizeof(struct test_pid_data));
return 0;
}
struct test_pid_data* test_get_pid_data(int nr)
{
struct test_pid *pid;
pid = test_find_pid(nr);
if(NULL==pid)
return NULL;
return &pid->data;
}
static void test_safe_free_pid(struct rcu_head *rp)
{
struct test_pid *pid = container_of(rp, struct test_pid, rcu);
kfree(pid);
}
void test_release_pid(struct test_pid *pid)
{
spin_lock(&test_pidmap_lock);
hlist_del_rcu(&pid->pid_chain);
list_del_rcu(&pid->list);
spin_unlock(&test_pidmap_lock);
call_rcu(&pid->rcu,test_safe_free_pid);
}
void test_remove_pid_all(void)
{
test_traverse_pid_list(test_release_pid);
}
#endif
void __init pidhash_init(void)
{
#if 1
init_test_pid_hash_list();
#endif
}
2. 下面是该集合的一个使用场景：记录某段时间内，smp下，各进程消耗虚拟内存的大小。
用户通过系统调用通知开始和结束记录
mm/swap.c
#if 1
struct test_pid_data
{
int magic;
long load;
char comm[TASK_COMM_LEN];
};
struct test_pid
{
int nr;
struct list_head list;
struct test_pid_data data;
struct hlist_node pid_chain;
struct rcu_head rcu;
};
typedef void (*test_fn)(struct test_pid* pid);
extern int test_set_pid_data(int nr,struct test_pid_data* test_data);
extern struct test_pid_data* test_get_pid_data(int nr);
extern void test_traverse_pid_list(test_fn callback);
extern void test_remove_pid_all(void);
static atomic_t test_flag = ATOMIC_INIT(0);
void test_print_pid_data(struct test_pid *pid)
{
printk("(pid:%d,name:%s) mmap len:%ld",
pid->nr,pid->data.comm,(unsigned long)pid->data.load);
if(pid->data.magic)
printk(",NO ALLOC!");
printk("\n");
}
void dump_pid_result(void)
{
test_traverse_pid_list(test_print_pid_data);
}
static void update_process_mmap_len(long delta)
{
int pid = 0;
struct test_pid_data data;
struct test_pid_data* datap;
if(0 == atomic_read(&test_flag))
return;
pid = (int)current->pid;
rcu_read_lock();
datap = test_get_pid_data(pid);
if(datap == NULL){
strcpy(data.comm,current->comm);
data.load = delta;
data.magic = 0;
test_set_pid_data(pid,&data);
}else
datap->load +=delta;
rcu_read_unlock();
}
void update_process_mmap_flag(unsigned long flag)
{
int pid = 0;
struct test_pid_data data;
struct test_pid_data* datap;
if(0 == atomic_read(&test_flag))
return;
if(!flag)
return;
pid = (int)current->pid;
rcu_read_lock();
datap = test_get_pid_data(pid);
if(datap == NULL){
strcpy(data.comm,current->comm);
data.load = 0;
data.magic = 0;
test_set_pid_data(pid,&data);
}
datap->magic = 1;
rcu_read_unlock();
}
asmlinkage void sys_notify_record(int data)
{
int flag = (int)data;
if(1 == flag){
printk("start record\n");
atomic_set(&test_flag,1);
smp_wmb();
}
else if(0 == flag){
atomic_set(&test_flag,0);
smp_wmb();
printk("end record\n");
dump_pid_result();
test_remove_pid_all();
}
}
void test_start_mmap(void)
{
printk("start record\n");
sys_notify_record(1);
}
void test_end_mmap(void)
{
printk("end record\n");
sys_notify_record(0);
//dump_pid_result();
}
#endif
#define ACCT_THRESHOLD max(16, NR_CPUS * 2)
static DEFINE_PER_CPU(long, committed_space) = 0;
void vm_acct_memory(long pages)
{
long *local;
preempt_disable();
local = &__get_cpu_var(committed_space);
*local += pages;
if (*local > ACCT_THRESHOLD || *local < -ACCT_THRESHOLD) {
#if 1
update_process_mmap_len(*local);
#endif
atomic_add(*local, &vm_committed_space);
*local = 0;
}
preempt_enable();
}
mm/mmap.c
#if 1
extern void dump_pid_result(void);
extern void test_remove_pid_all(void);
#endif
int __vm_enough_memory(long pages, int cap_sys_admin)
{
if (atomic_read(&vm_committed_space) < (long)allowed)
return 0;
#if 1
dump_pid_result();
test_remove_pid_all();
#endif
error:
vm_unacct_memory(pages);
include/asm-x86_64/unistd.h
#define __NR_unify_syscall 280
__SYSCALL(__NR_unify_syscall, sys_unify_syscall)
#define __NR_notify_record_syscall 281
__SYSCALL(__NR_notify_record_syscall, sys_notify_record)
//#define __NR_syscall_max __NR_unify_syscall
#define __NR_syscall_max __NR_notify_record_syscall
测试：
用户态malloc一块10000000字节的内存，然后fork
通过syscall(281,1)启动记录，syscall(281,0)关闭记录并打印结果，最后清除记录
pid自带的信息我们用struct test_pid_data来表示，用户可根据情况自行定义结构类型。
上面例子中magic是用来指示该进程是否曾经用MAP_NOALLOC分配过内存。
(pid:338,name:fork_malloc) mmap len:5004 (pid:337,name:fork_malloc) mmap len:5004 (pid:336,name:fork_malloc) mmap len:5004 (pid:335,name:fork_malloc) mmap len:5004 (pid:334,name:fork_malloc) mmap len:5021