Re: [PATCH bpf-next v2 05/20] bpf: implement bpf_seq_read() for bpf iterator

[Yonghong Song <yhs@fb.com>]

On 5/5/20 12:56 PM, Andrii Nakryiko wrote:
> On Sun, May 3, 2020 at 11:26 PM Yonghong Song <yhs@fb.com> wrote:
>>
>> bpf iterator uses seq_file to provide a lossless
>> way to transfer data to user space. But we want to call
>> bpf program after all objects have been traversed, and
>> bpf program may write additional data to the
>> seq_file buffer. The current seq_read() does not work
>> for this use case.
>>
>> Besides allowing stop() function to write to the buffer,
>> the bpf_seq_read() also fixed the buffer size to one page.
>> If any single call of show() or stop() will emit data
>> more than one page to cause overflow, -E2BIG error code
>> will be returned to user space.
>>
>> Signed-off-by: Yonghong Song <yhs@fb.com>
>> ---
>>   kernel/bpf/bpf_iter.c | 128 ++++++++++++++++++++++++++++++++++++++++++
>>   1 file changed, 128 insertions(+)
>>
>> diff --git a/kernel/bpf/bpf_iter.c b/kernel/bpf/bpf_iter.c
>> index 05ae04ac1eca..2674c9cbc3dc 100644
>> --- a/kernel/bpf/bpf_iter.c
>> +++ b/kernel/bpf/bpf_iter.c
>> @@ -26,6 +26,134 @@ static DEFINE_MUTEX(targets_mutex);
>>   /* protect bpf_iter_link changes */
>>   static DEFINE_MUTEX(link_mutex);
>>
>> +/* bpf_seq_read, a customized and simpler version for bpf iterator.
>> + * no_llseek is assumed for this file.
>> + * The following are differences from seq_read():
>> + *  . fixed buffer size (PAGE_SIZE)
>> + *  . assuming no_llseek
>> + *  . stop() may call bpf program, handling potential overflow there
>> + */
>> +static ssize_t bpf_seq_read(struct file *file, char __user *buf, size_t size,
>> +                           loff_t *ppos)
>> +{
>> +       struct seq_file *seq = file->private_data;
>> +       size_t n, offs, copied = 0;
>> +       int err = 0;
>> +       void *p;
>> +
>> +       mutex_lock(&seq->lock);
>> +
>> +       if (!seq->buf) {
>> +               seq->size = PAGE_SIZE;
>> +               seq->buf = kmalloc(seq->size, GFP_KERNEL);
>> +               if (!seq->buf)
>> +                       goto Enomem;
> 
> Why not just mutex_unlock and exit with -ENOMEM? Less goto'ing, more
> straightforward.
> 
>> +       }
>> +
>> +       if (seq->count) {
>> +               n = min(seq->count, size);
>> +               err = copy_to_user(buf, seq->buf + seq->from, n);
>> +               if (err)
>> +                       goto Efault;
>> +               seq->count -= n;
>> +               seq->from += n;
>> +               copied = n;
>> +               goto Done;
>> +       }
>> +
>> +       seq->from = 0;
>> +       p = seq->op->start(seq, &seq->index);
>> +       if (!p || IS_ERR(p))
> 
> IS_ERR_OR_NULL?

Ack.

> 
>> +               goto Stop;
>> +
>> +       err = seq->op->show(seq, p);
>> +       if (seq_has_overflowed(seq)) {
>> +               err = -E2BIG;
>> +               goto Error_show;
>> +       } else if (err) {
>> +               /* < 0: go out, > 0: skip */
>> +               if (likely(err < 0))
>> +                       goto Error_show;
>> +               seq->count = 0;
>> +       }
> 
> This seems a bit more straightforward:
> 
> if (seq_has_overflowed(seq))
>      err = -E2BIG;
> if (err < 0)
>      goto Error_show;
> else if (err > 0)
>      seq->count = 0;
> 
> Also, I wonder if err > 0 (so skip was requested), should we ignore
> overflow? So something like:

Think about overflow vs. err > 0 case, I double checked seq_file()
implementation again, yes, it is skipped. So your suggestion below
looks reasonable.

> 
> if (err > 0) {
>      seq->count = 0;
> } else {
>      if (seq_has_overflowed(seq))
>          err = -E2BIG;
>      if (err)
>          goto Error_show;
> }
> 
>> +
>> +       while (1) {
>> +               loff_t pos = seq->index;
>> +
>> +               offs = seq->count;
>> +               p = seq->op->next(seq, p, &seq->index);
>> +               if (pos == seq->index) {
>> +                       pr_info_ratelimited("buggy seq_file .next function %ps "
>> +                               "did not updated position index\n",
>> +                               seq->op->next);
>> +                       seq->index++;
>> +               }
>> +
>> +               if (!p || IS_ERR(p)) {
> 
> Same, IS_ERR_OR_NULL.

Ack.

> 
>> +                       err = PTR_ERR(p);
>> +                       break;
>> +               }
>> +               if (seq->count >= size)
>> +                       break;
>> +
>> +               err = seq->op->show(seq, p);
>> +               if (seq_has_overflowed(seq)) {
>> +                       if (offs == 0) {
>> +                               err = -E2BIG;
>> +                               goto Error_show;
>> +                       }
>> +                       seq->count = offs;
>> +                       break;
>> +               } else if (err) {
>> +                       /* < 0: go out, > 0: skip */
>> +                       seq->count = offs;
>> +                       if (likely(err < 0)) {
>> +                               if (offs == 0)
>> +                                       goto Error_show;
>> +                               break;
>> +                       }
>> +               }
> 
> Same question here about ignoring overflow if skip was requested.

Yes, we should prioritize err > 0 over overflow.

> 
>> +       }
>> +Stop:
>> +       offs = seq->count;
>> +       /* may call bpf program */
>> +       seq->op->stop(seq, p);
>> +       if (seq_has_overflowed(seq)) {
>> +               if (offs == 0)
>> +                       goto Error_stop;
>> +               seq->count = offs;
> 
> just want to double-check, because it's not clear from the code. If
> all the start()/show()/next() succeeded, but stop() overflown. Would
> stop() be called again on subsequent read? Would start/show/next
> handle this correctly as well?

I am supposed to handle this unless there is a bug...
The idea is:
    - if start()/show()/next() is fine and stop() overflow,
      we will skip stop() output and move on.
      (if we found out, we skip to the beginning of the
       buffer, we will return -E2BIG. Otherwise, we will return
       0 here, the user read() may just exit.)
    - next time, when read() called again, the start() will return
      NULL (since previous next() returns NULL) and the control
      will jump to stop(), which will try to do another dump().

> 
>> +       }
>> +
>> +       n = min(seq->count, size);
>> +       err = copy_to_user(buf, seq->buf, n);
>> +       if (err)
>> +               goto Efault;
>> +       copied = n;
>> +       seq->count -= n;
>> +       seq->from = n;
>> +Done:
>> +       if (!copied)
>> +               copied = err;
>> +       else
>> +               *ppos += copied;
>> +       mutex_unlock(&seq->lock);
>> +       return copied;
>> +
>> +Error_show:
>> +       seq->op->stop(seq, p);
>> +Error_stop:
>> +       seq->count = 0;
>> +       goto Done;
>> +
>> +Enomem:
>> +       err = -ENOMEM;
>> +       goto Done;
>> +
>> +Efault:
>> +       err = -EFAULT;
>> +       goto Done;
> 
> Enomem and Efault seem completely redundant and just add goto
> complexity to this algorithm. Let's just inline `err =
> -E(NOMEM|FAULT); goto Done;` instead?

We can do this. This is kind of original seq_read() coding
style. Agree that we do not need to follow them.

> 
>> +}
>> +
>>   int bpf_iter_reg_target(struct bpf_iter_reg *reg_info)
>>   {
>>          struct bpf_iter_target_info *tinfo;
>> --
>> 2.24.1
>>