Re: [PATCH] mtd_blkdevs: add mtd_table_mutex lock back to blktrans_{open, release} to avoid race condition

From: Liwei Song
Date: Tue Aug 09 2022 - 05:05:00 EST




On 8/9/22 16:08, ChristophHellwig wrote:
> On Tue, Aug 09, 2022 at 03:57:53PM +0800, Liwei Song wrote:
>> without lock mtd_table_mutex in blktrans_{open, release}, there will
>> be a race condition when access devices /dev/mtd1 and /dev/mtdblock1
>> at the same time with a high frequency open and close test:
>>
>> kernel BUG at drivers/mtd/mtdcore.c:1221!
>> lr : blktrans_release+0xb0/0x120
>
> This does not seem on a current mainline kernel and seems to be
> a somewhat incomplete backtrace. Can you send the full dmesg of
> a latest mainline run and maybe share the reproducer?

Yes, the kernel I used is 5.15, unfortunately this is the latest version
that worked on my board, the whole log is:

[ 31.301343] ------------[ cut here ]------------
[ 31.301343] ------------[ cut here ]------------
[ 31.301365] kernel BUG at drivers/mtd/mtdcore.c:1221!
[ 31.314981] kernel BUG at drivers/mtd/mtdcore.c:1221!
[ 31.329328] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP
[ 31.374117] Modules linked in: 8021q sch_fq_codel openvswitch nsh nf_conncount nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 fuse
[ 31.395820] CPU: 2 PID: 390 Comm: a.out Not tainted 5.15.58-yocto-standard #1
[ 31.412672] Hardware name: SoCFPGA Agilex SoCDK (DT)
[ 31.427372] pstate: 60000005 (nZCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[ 31.444058] pc : __put_mtd_device+0x4c/0x84
[ 31.457977] lr : put_mtd_device+0x3c/0x5c
[ 31.464122] sp : ffff80000c26bc50
[ 31.466126] x29: ffff80000c26bc50 x28: ffff000285785100 x27: 0000000000000000
[ 31.471945] x26: 0000000045585401 x25: 0000000000000000 x24: ffff000285785768
[ 31.477762] x23: ffff0002803ee520 x22: ffff0002804f8900 x21: ffff000281956800
[ 31.483580] x20: ffff000281956800 x19: ffff000281955080 x18: 0000000000000000
[ 31.489402] x17: 0000000000000000 x16: 0000000000000000 x15: 0000000000000000
[ 31.495219] x14: 0000000000000000 x13: 0000000000000000 x12: 0000000000000000
[ 31.501037] x11: 0000000000000000 x10: 0000000000000000 x9 : ffff8000088a79b0
[ 31.506855] x8 : ffff80000c26bcb0 x7 : 0000000000000000 x6 : 0000000000000001
[ 31.512673] x5 : ffff000280959488 x4 : 0000000000000000 x3 : 0000000000000000
[ 31.518491] x2 : ffff000281956800 x1 : 00000000ffffffff x0 : ffff000281955080
[ 31.524310] Call trace:
[ 31.525450] __put_mtd_device+0x4c/0x84
[ 31.527979] put_mtd_device+0x3c/0x5c
[ 31.530333] mtdchar_close+0x3c/0x84
[ 31.532604] __fput+0x78/0x220
[ 31.534357] ____fput+0x1c/0x30
[ 31.536193] task_work_run+0x88/0xc0
[ 31.538467] do_notify_resume+0x384/0x12a0
[ 31.541261] el0_svc+0x6c/0x80
[ 31.543015] el0t_64_sync_handler+0xa4/0x130
[ 31.545977] el0t_64_sync+0x1a0/0x1a4
[ 31.548338] Code: b9448841 51000421 b9048841 36ffff41 (d4210000)
[ 31.553115] ---[ end trace 9652b26ea1d7daa1 ]---
[ 31.556420] Internal error: Oops - BUG: 0 [#2] PREEMPT SMP
[ 31.556420] note: a.out[390] exited with preempt_count 1
[ 31.560588] Modules linked in: 8021q sch_fq_codel openvswitch nsh nf_conncount nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 fuse
[ 31.575265] CPU: 3 PID: 391 Comm: a.out Tainted: G D 5.15.58-yocto-standard #1
[ 31.582466] Hardware name: SoCFPGA Agilex SoCDK (DT)
[ 31.586115] pstate: 40000005 (nZcv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[ 31.591757] pc : __put_mtd_device+0x4c/0x84
[ 31.594642] lr : blktrans_release+0xb0/0x120
[ 31.597603] sp : ffff80000c22bc20
[ 31.599608] x29: ffff80000c22bc20 x28: ffff000285785e80 x27: 0000000000000000
[ 31.605428] x26: 0000000045585401 x25: 0000000000000000 x24: ffff0002857864e8
[ 31.611247] x23: ffff0002803ee520 x22: ffff0002803e3470 x21: ffff0002803e3400
[ 31.617066] x20: ffff0002803e3020 x19: ffff000281955080 x18: 0000000000000000
[ 31.622884] x17: 0000000000000000 x16: 0000000000000000 x15: 0000000000000000
[ 31.628702] x14: 0000000000000000 x13: 0000000000000000 x12: 0000000000000000
[ 31.634519] x11: 0000000000000000 x10: 0000000000000000 x9 : ffff8000088b0230
[ 31.640337] x8 : ffff80000c22bb90 x7 : 0000000000000000 x6 : 0000000000000001
[ 31.646155] x5 : 0000000000000000 x4 : 0000000000000000 x3 : 0000000000000000
[ 31.651973] x2 : ffff000281956800 x1 : 00000000ffffffff x0 : ffff000281955080
[ 31.657792] Call trace:
[ 31.658933] __put_mtd_device+0x4c/0x84
[ 31.661462] blktrans_release+0xb0/0x120
[ 31.664077] blkdev_put+0xd4/0x210
[ 31.666175] blkdev_close+0x34/0x50
[ 31.668355] __fput+0x78/0x220
[ 31.670108] ____fput+0x1c/0x30
[ 31.671943] task_work_run+0x88/0xc0
[ 31.674217] do_notify_resume+0x384/0x12a0
[ 31.677009] el0_svc+0x6c/0x80
[ 31.678762] el0t_64_sync_handler+0xa4/0x130
[ 31.681723] el0t_64_sync+0x1a0/0x1a4
[ 31.684082] Code: b9448841 51000421 b9048841 36ffff41 (d4210000)
[ 31.688857] ---[ end trace 9652b26ea1d7daa2 ]---
[ 31.692161] note: a.out[391] exited with preempt_count 1

the testcase a.out is compiled from below code:
when run the case /dev/mtd1 and /dev/mtdblock1 will be used for open and close test.

#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <string.h>
#include <signal.h>
#include <unistd.h>

int main(int argc, char *argv[])
{
pid_t pid, pid1, pid2;
int fd,ret = 0;
int status = 0;
char device_char[12]="/dev/mtd";
char device_block[17]="/dev/mtdblock";

strcat(device_char, argv[1]);
strcat(device_block, argv[1]);

pid1 = fork();
if(pid1 == 0){
while(1){
fd = open(device_char, O_SYNC|O_RDWR);
ret = close(fd);
}
}
pid2 = fork();
if(pid2 == 0){
while(1){
fd = open(device_block, O_SYNC|O_RDWR);
ret = close(fd);
}
}
sleep(10);
kill(pid1, SIGKILL);
kill(pid2, SIGKILL);
pid = wait(&status);
pid = wait(&status);
return 0;
}

>
>> diff --git a/drivers/mtd/mtd_blkdevs.c b/drivers/mtd/mtd_blkdevs.c
>> index b8ae1ec14e17..147e4a11dfe4 100644
>> --- a/drivers/mtd/mtd_blkdevs.c
>> +++ b/drivers/mtd/mtd_blkdevs.c
>> @@ -188,6 +188,8 @@ static int blktrans_open(struct block_device *bdev, fmode_t mode)
>>
>> kref_get(&dev->ref);
>>
>> + if (!mutex_trylock(&mtd_table_mutex))
>> + return ret;
>
> No, that's not really the solution.
>
> Turning the kref_get above into a kref_get_unless_zero might be better
> path to look into.

Thanks, I will have a look at this.

Liwei.


>