On Jan 11, 2013, at 5:48 PM, Christian Peron <c...@sqrt.ca> wrote:
> Guy,
>
> Can you please describe the race and how to reproduce it? When we introduced
> zerocopy-bpf, we also introduced bpf_bufheld() and bpf_bufreclaimed() which
> are effectively nops for the regular buffer mode. Maybe we could maintain
> state there as well? In any case, I would like to understand the race/and
> reproduce it
>
> Thanks!
Related to rwatson's comment from r175903, dropping the bpf descriptor lock
while performing the uiomove() in bpfread() enables a couple of races: 1)
simultaneous reads on a bpf device (not great), and 2) races between bpfread()
and catchpacket() (panics due to NULL pointer dereference).
The symptom I encountered was a panic in catchpacket() where the save buffer
was NULL. I added diagnostic code and found that one of the two buffers was
being lost. As I reviewed the code, it seemed that the only way this could
happen was if the hold buffer was being invalidated (rotated) out from under
these lines in bpfread():
BPFD_UNLOCK(d);
/*
* Move data from hold buffer into user space.
* We know the entire buffer is transferred since
* we checked above that the read buffer is bpf_bufsize bytes.
*
* XXXRW: More synchronization needed here: what if a second thread
* issues a read on the same fd at the same time? Don't want this
* getting invalidated.
*/
error = bpf_uiomove(d, d->bd_hbuf, d->bd_hlen, uio);
BPFD_LOCK(d);
d->bd_fbuf = d->bd_hbuf; /* if other code in the driver simultaneously
sets hbuf to NULL, this loses a buffer! */
d->bd_hbuf = NULL;
d->bd_hlen = 0;
bpf_buf_reclaimed(d);
BPFD_UNLOCK(d);
My instrumented code validated this hypothesis, and I added the bd_hbuf_in_use
flag to allow dropping the descriptor lock over the bpf_uiomove() call while
protecting the hold buffer from rotation anywhere else in the driver. In my own
tree, I have left some additional diagnostic code to make sure hbuf is still
valid after uiomove completes, and that code has not been triggered since I
added the bd_hbuf_in_use flag and associated code.
The way I was able to trigger the panic in catchpacket() was to call
pcap_setfilter() more than one time on a live, promiscuous pcap descriptor
while reading from a very busy network. I was able to temporarily work around
this issue by using the BIOCSETFNR ioctl directly on the bpf descriptor rather
than using pcap_setfilter(), which avoided calling reset_d() in the ioctl
handler and thus avoided clearing bd_hbuf in a race with bpfread().
Just as I'm writing this, I realized the code that triggered the problem may
have called pcap_setfilter() in a separate thread from the packet reading
thread. I wrote a unit test for this issue, but as it was single-threaded, I
was never able to cause the crash with it.
Regarding the zero-copy code, I'm not sure how bpf_bufheld() could be used as a
general approach to solving this race. It appears to me that a condition
variable is necessary to work around the necessity of releasing the bpf
descriptor lock over the uiomove() call.
Hope this helps!
Guy
>
> On 2012-11-13, at 3:40 PM, Guy Helmer wrote:
>
>> To try to completely resolve the race in bpfread(), I have put together
>> these changes to add a flag to indicate when the hold buffer cannot be
>> modified because it is in use. Since it's my first time using mtx_sleep()
>> and wakeup(), I wanted to run these past the list to see if I can get any
>> feedback on the approach.
>>
>>
>> Index: bpf.c
>> ===================================================================
>> --- bpf.c (revision 242997)
>> +++ bpf.c (working copy)
>> @@ -819,6 +819,7 @@ bpfopen(struct cdev *dev, int flags, int fmt, stru
>> * particular buffer method.
>> */
>> bpf_buffer_init(d);
>> + d->bd_hbuf_in_use = 0;
>> d->bd_bufmode = BPF_BUFMODE_BUFFER;
>> d->bd_sig = SIGIO;
>> d->bd_direction = BPF_D_INOUT;
>> @@ -872,6 +873,9 @@ bpfread(struct cdev *dev, struct uio *uio, int iof
>> callout_stop(&d->bd_callout);
>> timed_out = (d->bd_state == BPF_TIMED_OUT);
>> d->bd_state = BPF_IDLE;
>> + while (d->bd_hbuf_in_use)
>> + mtx_sleep(&d->bd_hbuf_in_use, &d->bd_lock,
>> + PRINET|PCATCH, "bd_hbuf", 0);
>> /*
>> * If the hold buffer is empty, then do a timed sleep, which
>> * ends when the timeout expires or when enough packets
>> @@ -940,27 +944,27 @@ bpfread(struct cdev *dev, struct uio *uio, int iof
>> /*
>> * At this point, we know we have something in the hold slot.
>> */
>> + d->bd_hbuf_in_use = 1;
>> BPFD_UNLOCK(d);
>>
>> /*
>> * Move data from hold buffer into user space.
>> * We know the entire buffer is transferred since
>> * we checked above that the read buffer is bpf_bufsize bytes.
>> - *
>> - * XXXRW: More synchronization needed here: what if a second thread
>> - * issues a read on the same fd at the same time? Don't want this
>> - * getting invalidated.
>> + *
>> + * We do not have to worry about simultaneous reads because
>> + * we waited for sole access to the hold buffer above.
>> */
>> error = bpf_uiomove(d, d->bd_hbuf, d->bd_hlen, uio);
>>
>> BPFD_LOCK(d);
>> - if (d->bd_hbuf != NULL) {
>> - /* Free the hold buffer only if it is still valid. */
>> - d->bd_fbuf = d->bd_hbuf;
>> - d->bd_hbuf = NULL;
>> - d->bd_hlen = 0;
>> - bpf_buf_reclaimed(d);
>> - }
>> + KASSERT(d->bd_hbuf != NULL, ("bpfread: lost bd_hbuf"));
>> + d->bd_fbuf = d->bd_hbuf;
>> + d->bd_hbuf = NULL;
>> + d->bd_hlen = 0;
>> + bpf_buf_reclaimed(d);
>> + d->bd_hbuf_in_use = 0;
>> + wakeup(&d->bd_hbuf_in_use);
>> BPFD_UNLOCK(d);
>>
>> return (error);
>> @@ -1114,6 +1118,9 @@ reset_d(struct bpf_d *d)
>>
>> BPFD_LOCK_ASSERT(d);
>>
>> + while (d->bd_hbuf_in_use)
>> + mtx_sleep(&d->bd_hbuf_in_use, &d->bd_lock, PRINET,
>> + "bd_hbuf", 0);
>> if ((d->bd_hbuf != NULL) &&
>> (d->bd_bufmode != BPF_BUFMODE_ZBUF || bpf_canfreebuf(d))) {
>> /* Free the hold buffer. */
>> @@ -1254,6 +1261,9 @@ bpfioctl(struct cdev *dev, u_long cmd, caddr_t add
>>
>> BPFD_LOCK(d);
>> n = d->bd_slen;
>> + while (d->bd_hbuf_in_use)
>> + mtx_sleep(&d->bd_hbuf_in_use, &d->bd_lock,
>> + PRINET, "bd_hbuf", 0);
>> if (d->bd_hbuf)
>> n += d->bd_hlen;
>> BPFD_UNLOCK(d);
>> @@ -1967,6 +1977,9 @@ filt_bpfread(struct knote *kn, long hint)
>> ready = bpf_ready(d);
>> if (ready) {
>> kn->kn_data = d->bd_slen;
>> + while (d->bd_hbuf_in_use)
>> + mtx_sleep(&d->bd_hbuf_in_use, &d->bd_lock,
>> + PRINET, "bd_hbuf", 0);
>> if (d->bd_hbuf)
>> kn->kn_data += d->bd_hlen;
>> } else if (d->bd_rtout > 0 && d->bd_state == BPF_IDLE) {
>> @@ -2299,6 +2312,9 @@ catchpacket(struct bpf_d *d, u_char *pkt, u_int pk
>> * spot to do it.
>> */
>> if (d->bd_fbuf == NULL && bpf_canfreebuf(d)) {
>> + while (d->bd_hbuf_in_use)
>> + mtx_sleep(&d->bd_hbuf_in_use, &d->bd_lock,
>> + PRINET, "bd_hbuf", 0);
>> d->bd_fbuf = d->bd_hbuf;
>> d->bd_hbuf = NULL;
>> d->bd_hlen = 0;
>> @@ -2341,6 +2357,9 @@ catchpacket(struct bpf_d *d, u_char *pkt, u_int pk
>> ++d->bd_dcount;
>> return;
>> }
>> + while (d->bd_hbuf_in_use)
>> + mtx_sleep(&d->bd_hbuf_in_use, &d->bd_lock,
>> + PRINET, "bd_hbuf", 0);
>> ROTATE_BUFFERS(d);
>> do_wakeup = 1;
>> curlen = 0;
>> Index: bpf.h
>> ===================================================================
>> --- bpf.h (revision 242997)
>> +++ bpf.h (working copy)
>> @@ -1235,7 +1235,8 @@ SYSCTL_DECL(_net_bpf);
>> /*
>> * Rotate the packet buffers in descriptor d. Move the store buffer into the
>> * hold slot, and the free buffer ino the store slot. Zero the length of the
>> - * new store buffer. Descriptor lock should be held.
>> + * new store buffer. Descriptor lock should be held. Hold buffer must
>> + * not be marked "in use".
>> */
>> #define ROTATE_BUFFERS(d) do {
>> \
>> (d)->bd_hbuf = (d)->bd_sbuf; \
>> Index: bpf_buffer.c
>> ===================================================================
>> --- bpf_buffer.c (revision 242997)
>> +++ bpf_buffer.c (working copy)
>> @@ -189,6 +189,9 @@ bpf_buffer_ioctl_sblen(struct bpf_d *d, u_int *i)
>> return (EINVAL);
>> }
>>
>> + while (d->bd_hbuf_in_use)
>> + mtx_sleep(&d->bd_hbuf_in_use, &d->bd_lock,
>> + PRINET, "bd_hbuf", 0);
>> /* Free old buffers if set */
>> if (d->bd_fbuf != NULL)
>> free(d->bd_fbuf, M_BPF);
>> Index: bpfdesc.h
>> ===================================================================
>> --- bpfdesc.h (revision 242997)
>> +++ bpfdesc.h (working copy)
>> @@ -63,6 +63,7 @@ struct bpf_d {
>> caddr_t bd_sbuf; /* store slot */
>> caddr_t bd_hbuf; /* hold slot */
>> caddr_t bd_fbuf; /* free slot */
>> + int bd_hbuf_in_use; /* don't rotate buffers */
>> int bd_slen; /* current length of store buffer */
>> int bd_hlen; /* current length of hold buffer */
>>
>> _______________________________________________
>> freebsd-net@freebsd.org mailing list
>> http://lists.freebsd.org/mailman/listinfo/freebsd-net
>> To unsubscribe, send any mail to "freebsd-net-unsubscr...@freebsd.org"
>
_______________________________________________
freebsd-net@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-net
To unsubscribe, send any mail to "freebsd-net-unsubscr...@freebsd.org"
