某SPI设备驱动引起的开关机压力测试死机问题一例
作者：互联网
环境
硬件平台：某ARM SoC
软件平台：Linux
问题现象：产品做开关机压力测试，发生死机。
分析
用crash工具解析两次死机dump信息，得到死机前的log如下。两次死机的backtrace略有不同，但死机原因类似：最后都是在调用 complete 的过程中访问空指针导致 kernel panic。
log 1: [ 1.092790] c2 Unable to handle kernel NULL pointer dereference at virtual address 00000004 [ 1.092796] c2 pgd = c0004000 [ 1.092802] c0 [00000004] *pgd=00000000 [ 1.092812] c2 Internal error: Oops: 5 [#1] PREEMPT SMP ARM ... [ 1.094412] c0 Backtrace: [ 1.094430] c2 [<c07e971c>] (_raw_spin_lock_irqsave) from [<c006e2a8>] (complete+0x1c/0x4c) [ 1.094446] c2 [<c006e28c>] (complete) from [<c04318d4>] (spi_complete+0x10/0x14) [ 1.094468] c2 [<c04318c4>] (spi_complete) from [<c0432328>] (spi_finalize_current_message+0x228/0x26c) [ 1.094479] c2 [<c0432100>] (spi_finalize_current_message) from [<c04327c8>] (spi_transfer_one_message+0x45c/0x484) [ 1.094503] c2 [<c043236c>] (spi_transfer_one_message) from [<c0432e98>] (__spi_pump_messages+0x6a8/0x6e8) [ 1.094531] c2 [<c04327f0>] (__spi_pump_messages) from [<c04330fc>] (__spi_sync+0x208/0x270) [ 1.094559] c2 [<c0432ef4>] (__spi_sync) from [<c0433178>] (spi_sync+0x14/0x18) [ 1.094586] c2 [<c0433164>] (spi_sync) from [<c0440744>] (dm9051_r_reg+0x4c/0x6c) [ 1.094595] c2 [<c04406f8>] (dm9051_r_reg) from [<c0441e30>] (dm9051_probe+0x6e4/0x82c) [ 1.094605] c2 [<c044174c>] (dm9051_probe) from [<c0431764>] (spi_drv_probe+0x8c/0xa8) log 2: [ 1.271300] c3 Unable to handle kernel NULL pointer dereference at virtual address 00000004 [ 1.271305] c3 pgd = c0004000 [ 1.271312] c0 [00000004] *pgd=00000000 [ 1.271323] c3 Internal error: Oops: 5 [#1] PREEMPT SMP ARM ... [ 1.414585] c0 Backtrace: [ 1.414603] c3 [<c07e97f4>] (_raw_spin_lock_irqsave) from [<c006e2a8>] (complete+0x1c/0x4c) [ 1.414618] c3 [<c006e28c>] (complete) from [<c0431fb8>] (spi_complete+0x28/0x34) [ 1.414643] c3 [<c0431f90>] (spi_complete) from [<c0432350>] (spi_finalize_current_message+0x230/0x278) [ 1.414666] c3 [<c0432120>] (spi_finalize_current_message) from [<c04327f4>] (spi_transfer_one_message+0x45c/0x484) [ 1.414693] c3 [<c0432398>] (spi_transfer_one_message) from [<c0432ec4>] (__spi_pump_messages+0x6a8/0x6e8) [ 1.414725] c3 [<c043281c>] (__spi_pump_messages) from [<c0432f1c>] (spi_pump_messages+0x18/0x1c) [ 1.414759] c3 [<c0432f04>] (spi_pump_messages) from [<c0042abc>] (kthread_worker_fn+0xc0/0x14c) [ 1.414771] c3 [<c00429fc>] (kthread_worker_fn) from [<c00429e8>] (kthread+0x110/0x124) [ 1.414798] c3 [<c00428d8>] (kthread) from [<c000f208>] (ret_from_fork+0x14/0x2c)
梳理kernel spi驱动代码：drivers/spi/spi.c
//调用 complete
static void spi_complete(void *arg)
{
complete(arg);
}
//给 spi_message 的 complete 与 context 成员赋值
static int __spi_sync(struct spi_device *spi, struct spi_message *message,
int bus_locked)
{
DECLARE_COMPLETION_ONSTACK(done);
...
message->complete = spi_complete;
message->context = &done;
message->spi = spi;
if (status == 0) {
...
__spi_pump_messages(master, false);
...
wait_for_completion(&done);
status = message->status;
}
message->context = NULL;
return status;
}
//回调 spi_message 的 complete 函数，即 spi_complete，传入的参数是 spi_message 的 context，即 __spi_sync 函数里面定义的 completion 变量 “done”。
void spi_finalize_current_message(struct spi_master *master)
{
...
spin_lock_irqsave(&master->queue_lock, flags);
mesg = master->cur_msg;
spin_unlock_irqrestore(&master->queue_lock, flags);
...
if (mesg->complete)
mesg->complete(mesg->context);
}
因此基本可以确定是执行 mesg->complete(mesg->context); 时，传入的参数 mesg->context 为 NULL 导致问题。继续看死机前的部分log：
//cpu2 上发起一次 spi 传输并完成
[ 1.271151] c2 70a00000.spi: __spi_sync [ 1.271172] c2 spi_master spi0: spi_finalize_current_message [ 1.271180] c2 [spi_complete]
//cpu2 上再次发起一次 spi 传输并完成
[ 1.271192] c2 70a00000.spi: __spi_sync [ 1.271212] c2 spi_master spi0: spi_finalize_current_message [ 1.271219] c2 [spi_complete]
//cpu2 上再次发起一次 spi 传输，但在完成前，cpu1 发起了新的 spi 传输
[ 1.271227] c2 70a00000.spi: __spi_sync [ 1.271247] c2 spi_master spi0: spi_finalize_current_message [ 1.271249] c1 70a00000.spi: __spi_sync //cpu1 发起新的传输 [ 1.271261] c2 [spi_complete] [ 1.271281] c1 dm9051_r_reg: spi_sync() failed
//cpu3 上执行传输完成的 completion 时，传入的 mesg->context 为 NULL 导致死机
[ 1.271288] c3 spi_master spi0: spi_finalize_current_message [ 1.271293] c3 [spi_complete] [ 1.271300] c3 Unable to handle kernel NULL pointer dereference at virtual address 00000004
__spi_sync 函数中，有如下语句：
__spi_pump_messages(master, false); // 处理spi message，发起传输 wait_for_completion(&done); // 等待传输完成 message->context = NULL; //将该 spi message 的 context 置为 NULL
也就是说，__spi_sync 等到 completion 后会将本次已传输完成的 spi message 的 context 置为NULL。
如果在cpu1上发起新的传输时，传入的 spi message 变量地址与cpu2上的是同一个，那么cpu1就有可能访问到这个 NULL context。
spi message 变量是dm9051驱动传递下来的，查看dm9051驱动，其调用spi的代码如下，可以看到它使用了全局变量 dm->spi_msg1 来构造 spi message。因此造成了问题。
static u8 dm9051_r_reg(struct dm9051_net *dm, u16 reg_addr)
{
struct spi_transfer *xfer = &dm->spi_xfer1;
struct spi_message *msg = &dm->spi_msg1;
u8 r_cmd[2] = {reg_addr, 0x00};
u8 r_data[2] = {0x00, 0x00};
int ret;
xfer->tx_buf = r_cmd;
xfer->rx_buf = r_data;
xfer->len = 2;
ret = spi_sync(dm->spidev, msg);
if (ret < 0)
DM_MSG0("dm9051_r_reg: spi_sync() failed\n");
return r_data[1];
}
static int dm9051_probe(struct spi_device *spi)
{
...
spi_message_init(&dm->spi_msg1);
spi_message_add_tail(&dm->spi_xfer1, &dm->spi_msg1);
...
}
解决
仿照 include/linux/spi/spi.h 中 spi_read 和 spi_write 的实现，每次读写前都构造新的 spi message。
static inline int
spi_read(struct spi_device *spi, void *buf, size_t len)
{
struct spi_transfer t = {
.rx_buf = buf,
.len = len,
};
struct spi_message m;
spi_message_init(&m);
spi_message_add_tail(&t, &m);
return spi_sync(spi, &m);
}
-------------------------------------------------
作者：bigfish99
博客：https://www.cnblogs.com/bigfish0506/
公众号：大鱼嵌入式
标签：开关机,__,complete,SPI,死机,sync,spi,c2,message 来源： https://www.cnblogs.com/bigfish0506/p/14831710.html