PROBLEM: relay - stale data copied to user space
Hi,
Tracing block I/Os using blktrace and blkiomon results in "bad trace" messages, meaning that user space has come across a trace which is broken:
bad trace magic 100
bad trace magic 10e7
bad trace magic 1
bad trace magic 1
bad trace magic 374
bad trace magic 7000010
bad trace magic 11000010
bad trace magic 1000
bad trace magic 10e3
bad trace magic 0
bad trace magic 1000
bad trace magic 10de
bad trace magic 8b3
bad trace magic 0
This condition can't be recovered from because blktrace has lost track of where the next valid trace is.
I gathered hexdumps of blktrace data in user space. Looking at the data as read by blktrace from relay files, I found fragments of old traces partially overlaying other traces:
(In the hexdump-ed trace below, I have added "|"s as delimiters between traces. The corrupted part of the trace - a fragment containing the sequence number 0x4d1d3 - is in parenthesis.)
12d8ca0 |6561 7407 0005 6440 0000 00fb ef9d 6496 -> sequence 0x56440
12d8cb0 0000 0000 000f 4790 0002 0000 4001 0011 12d8cc0 0000 1445 0080 0020 0000 0001 0000 0018 12d8cd0 0000 0001 0001 007d 0000 0000 0001 3e40 12d8ce0 0000 0000 001b badc|6561 7407 0005 6441 -> sequence 0x56441 12d8cf0 0000 00fb ef9d 79e3 0000 0000 000f 4790 12d8d00 0002 0000 0181 0008 0000 1445 0080 0020 12d8d10 0000 0001 0000 0000|6561 7407 0005 6442 -> sequence 0x56442
12d8d20 (0000 1444 0080 0020 0000 0001 0000 0000
12d8d30 |6561 7407 0004 d1d3 0000 00e3 4a8b 1544 -> sequence 0x4d1d3
12d8d40 0000 0000 0015 5728 0003 0000 4001 0011 old trace!!
12d8d50 0000 1447 0080 0020 0000 0001 0000 0018)
12d8d60 |6561 7407 0005 6443 0000 00fb f04d c32a -> sequence 0x56443
12d8d70 0000 0000 000f 4928 0000 8000 4001 0011 12d8d80 0000 1445 0080 0020 0000 0001 0000 0018 12d8d90 0000 0001 0001 007f 0000 0000 0000 55d8 12d8da0 0000 0000 0008 1070|6561 7407 0005 6444 -> sequence 0x56444
The same fragment containing sequence 0x4d1d3 originally appeared in this context:
10d8d70 0000 0000 0013 970e|6561 7407 0004 d1d2 -> sequence 0x4d1d2 10d8d80 0000 00e3 4a82 8773 0000 0000 0039 4d20 10d8d90 0001 0000 0181 0008(0000 1444 0080 0020 10d8da0 0000 0001 0000 0000|6561 7407 0004 d1d3 -> sequence 0x4d1d3 10d8db0 0000 00e3 4a8b 1544 0000 0000 0015 5728 10d8dc0 0003 0000 4001 0011 0000 1447 0080 0020 10d8dd0 0000 0001 0000 0018)0000 0001 0001 0080 10d8de0 0000 0000 0001 156c 0000 0000 001a 10c8
10d8df0 |6561 7407 0004 d1d4 0000 00e3 4a8b 239c -> sequence 0x4d1d4 10d8e00 0000 0000 0015 5728 0003 0000 0181 0008
10d8e10 0000 1447 0080 0020 0000 0001 0000 0000
Blktrace data is handled by relay, which uses ringbuffers...
I decided to use the following debug patch:
debug patch: poison various kinds of unused relay buffers
--- kernel/relay.c | 5 +++++ 1 file changed, 5 insertions(+)
kernel/relay.c +5 -0
@@ -735,6 +735,8 @@ size_t relay_switch_subbuf(struct rchan_ old_subbuf = buf->subbufs_produced % buf->chan->n_subbufs; buf->padding[old_subbuf] = buf->prev_padding; buf->subbufs_produced++; +memset ((char *)buf->data + buf->offset,0x66 , + buf->prev_padding);if (buf->dentry) buf->dentry->d_inode->i_size += buf->chan->subbuf_size -
@@ -767,6 +769,8 @@ size_t relay_switch_subbuf(struct rchan_if (unlikely (length + buf->offset > buf->chan->subbuf_size)) goto toobig; +memset (buf->data,0x55 , buf->chan->subbuf_size); +return length; toobig:
@@ -1112,6 +1116,7 @@ static int subbuf_read_actor(size_t read desc->error = -EFAULT ; ret =0 ; } +memset (from,0x77 , ret); desc->arg.data += ret; desc->written += ret; desc->count -= ret;
I found out that the trace data is already corrupted when it is being copied to user space by copy_to_user() in kernel/relay.c. My 0x55 patterns marking invalidated buffers showed up in user space:
magic 0x65617407
sequence 0x00003baf
time 0x5555555555555555 sector 0x5555555555555555 bytes 0x55555555 action 0x55555555 pid 0x55555555 device 0x55555555 cpu 0x55555555 error 0x5555 pdu_len 0x5555
65617407 00003baf 55555555 55555555 55555555 55555555 55555555 55555555 55555555 55555555 55555555 55555555
Next I added another debug patch to catch poisoned buffers, that is, stale data prior to copy_to_user().
debug patch: detect relay buffer poisoning when copying data to user space
--- kernel/relay.c | 45 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+)
kernel/relay.c +40 -0
@@ -1098,6 +1098,45 @@ static size_t relay_file_read_end_pos(streturn end_pos; } +static inline void hexdump (char *buf,int len,char *descr) +{ +int i; + +printk ("<----- %s, buffer at %p, length %d, first <=4096 bytes:" , + descr, buf, len); +for (i =0 ; i <min (len,4096 ); i++) { +if (i %4 ==0 ) +printk (" " ); +if (i %32 ==0 ) +printk ("\n" ); +printk ("%02x" , buf[i]); + } +printk ("----->\n\n" ); +} + +static inline void *memsearch (char *buf,char c,size_t buf_len,size_t c_len) +{ +char *p = buf; +int i; + + do { + p = memchr(p, c, buf_len - (p - buf)); + if (!p) + return NULL; + for (i = 0; i < c_len - 1; i++, p++) { + if (p >= buf + buf_len) + return NULL; + if (*p != c) + break; + } + if (i == c_len - 1) { + hexdump(p - (c_len + 32), c_len + 64, "memsearch"); +/* * subbuf_read_actor - read up to one subbuf's worth of data */
@@ -1112,6 +1151,12 @@ static int subbuf_read_actor(size_t read from = buf->start + read_start; ret = avail; + +if (memsearch (from,0x55 , avail,16 )/* || +memsearch(from, 0x66, avail, 8) || +memsearch(from, 0x77, avail, 8)*/ ) +hexdump (from, avail,"relay: found buffer poisoning" ); +if (copy_to_user (desc->arg.buf, from, avail)) { desc->error = -EFAULT ; ret =0 ;return (p - c_len);
After some tests with and without detection of buffer poison in kernel space, I think that the time required by my kernel debug patch to scan for buffer poison reduces the chances of hitting bad traces in user space.
Another observation: when scanning buffers that are about to be copied to user space, hexdumps of poisoned buffers do not show buffer poison where it has been found just a split second before. A typical example:
77777777 77777777 77777777 77777777 77777777 77777777 77656174 070001c3 37000001 26a13312 80000000 00003bf9 d0000080 00014100 0700000f 46008000 10000000 00000000 00555555 55555555
The second line should contain a 0x55 pattern. It's gone.
What else do we see.... Scanning backward there are bytes already (recently?) copied to user space (0x77 pattern). Scanning forward there is room for more traces (0x55 pattern).
This is my theory:
Timing matters. It's a race caused by improper protection of critical sections in a producer-consumer scenario. A bug in the bookkeeping allows a reader to read at a position that is just being written to.
I have got no real evidence. I don't understand yet, how reader-writer serialisation is implemented in relay.
I use FCP disks attached to a Linux guest on a System z and an internal tool for I/O workload generation. I have stripped my environment down to 1 disk with 4 paths. Data is written to and read from a device mapper device. Blktrace is run for the 4 underlying SCSI devices:
blktrace -a issue -a complete -o - /dev/sda /dev/sdu /dev/sdao \/dev/sdbi | blkiomon -I 100 -b blkiomon.out
I run a recent git kernel and a recent git blktrace. The problem has also been seen with older code, though.
Any comment is welcome.
Please keep me on to: or cc: since I am not subscribed to linux-kernel.
Thanks,
Martin
Hi,
Quoted 6 lines · Martin Peschke
This is my theory:
Timing matters. It's a race caused by improper protection of critical sections in a producer-consumer scenario. A bug in the bookkeeping allows a reader to read at a position that is just being written to.
It does look consistent with a reader reading an event that's been reserved but not yet written, or partially written e.g. if an event being written on one cpu was read by another before the first one finished. Can you see if the below patch to blktrace userspace helps?
Or failing that, explicitly using gettid() in place of getpid() in sched_setaffinity(). Or, failing that, you had mentioned previously that you would try to reproduce the problem on your laptop - were you able to do that? If so, it would help in debugging it further...
Tom
blktrace.c +1 -1
@@ -610,7 +610,7 @@ static int lock_on_cpu(int cpu)CPU_ZERO (&cpu_mask);CPU_SET (cpu, &cpu_mask); -if (sched_setaffinity (getpid (),sizeof (cpu_mask), &cpu_mask) <0 ) +if (sched_setaffinity (0 ,sizeof (cpu_mask), &cpu_mask) <0 )return errno;return 0 ;
Quoted 11 lines · Tom Zanussi
On Wed, 2009-03-18 at 16:07 +0100, Martin Peschke wrote
Quoted 5 lines
This is my theory:
Timing matters. It's a race caused by improper protection of critical sections in a producer-consumer scenario. A bug in the bookkeeping allows a reader to read at a position that is just being written to.It does look consistent with a reader reading an event that's been reserved but not yet written, or partially written e.g. if an event being written on one cpu was read by another before the first one finished.
So this is part of relay's design, and it's up to user space to make sure that reader and writer are on the same CPU?
Quoted 1 line
Can you see if the below patch to blktrace userspace helps?
It appears to fix it. I will give it more testing in a larger environment.
Quoted 4 lines
Or failing that, explicitly using
gettid()in place ofgetpid()insched_setaffinity(). Or, failing that, you had mentioned previously that you would try to reproduce the problem on your laptop - were you able to do that? If so, it would help in debugging it further...
This didn't work out. But then, it's a single-CPU machine.
Thanks,
Martin
Hi,
Quoted 8 lines · Martin Peschke
Quoted 7 lines · Tom Zanussi
On Wed, 2009-03-18 at 16:07 +0100, Martin Peschke wrote
Quoted 5 lines
This is my theory:
Timing matters. It's a race caused by improper protection of critical sections in a producer-consumer scenario. A bug in the bookkeeping allows a reader to read at a position that is just being written to.
Quoted 8 lines
Quoted 4 lines
It does look consistent with a reader reading an event that's been reserved but not yet written, or partially written e.g. if an event being written on one cpu was read by another before the first one finished.
So this is part of relay's design, and it's up to user space to make sure that reader and writer are on the same CPU?
No, that wasn't originally part of relay's design, but for all practical purposes that's what you need to do if you're using the read() interface.
The original design was meant to be used with mmap() and poll() on complete sub-buffers i.e. you grab complete sub-buffers from the mmaped buffer only when poll() says they're ready. If used this way, it doesn't matter which cpu is doing the reading or writing. relay still works this way if you use mmap() as in e.g. the klog example.
The read() was added later, on top of something that wasn't designed with it in mind, the main difference being that a read() can be done at any time and it expects whatever's available to be returned, regardless of whether it's reading from a complete sub-buffer or not (if you could only read from complete sub-buffers, you might as well use the mmap() method).
So if a reader can read from a sub-buffer at any time, and there can simultaneously be a writer in that sub-buffer, if they're on different cpus, you can get what you were seeing. One way to prevent that is to keep the reader on the same cpu, and disable preemption for the writer, so when the reader on that cpu finally reads, it can be sure there are no in-progress events up to the endpoint of what it's reading.
Hope that helps,
Tom
Quoted 16 lines
Quoted 1 line
Can you see if the below patch to blktrace userspace helps?
It appears to fix it. I will give it more testing in a larger environment.
Quoted 4 lines
Or failing that, explicitly using
gettid()in place ofgetpid()insched_setaffinity(). Or, failing that, you had mentioned previously that you would try to reproduce the problem on your laptop - were you able to do that? If so, it would help in debugging it further...This didn't work out. But then, it's a single-CPU machine.
Thanks,
Martin