Skip to contentSkip to conversation

libbpf: handle producer position overflow

PATCH Superseded Andrew Werner 4 participants 8 messages Jul 24, 2023 – Aug 24, 2023

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 libbpf: handle producer position overflow

Before this patch, the producer position could overflow unsigned long, in which case libbpf would forever stop processing new writes to the ringbuf. This patch addresses that bug by avoiding ordered comparison between the consumer and producer position. If the consumer position is greater than the producer position, the assumption is that the producer has overflowed.

A more defensive check could be to ensure that the delta is within the allowed range, but such defensive checks are neither present in the kernel side code nor in libbpf. The overflow that this patch handles can occur while the producer and consumer follow a correct protocol.

A selftest was written to demonstrate the bug, and indeed this patch allows the test to continue to make progress past the overflow. However, the author was unable to create a testing environment on a 32-bit machine, and the test requires substantial memory and over 4 hours to hit the overflow point on a 64-bit machine. Thus, the test is not included in this patch because of the impracticality of running it.

Additionally, this patch adds commentary around a separate point to note that the modular arithmetic is valid in the face of overflows, as that fact may not be obvious to future readers.

---
 tools/lib/bpf/ringbuf.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/lib/bpf/ringbuf.c +10 -1
index 02199364db13..6271757bc3d2 100644
@@ -237,7 +237,11 @@ static int64_t ringbuf_process_ring(struct ring *r)
 	do {
 		got_new_data = false;
 		prod_pos = smp_load_acquire(r->producer_pos);
-		while (cons_pos < prod_pos) {
+
+		/* Avoid signed comparisons between the positions; the producer
+		 * position can overflow before the consumer position.
+		 */
+		while (cons_pos != prod_pos) {
 			len_ptr = r->data + (cons_pos & r->mask);
 			len = smp_load_acquire(len_ptr);
 
@@ -498,6 +502,11 @@ void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
 	prod_pos = smp_load_acquire(rb->producer_pos);
 
 	max_size = rb->mask + 1;
+
+	/* Note that this formulation in the face of overflow of prod_pos
+	 * so long as the delta between prod_pos and cons_pos remains no
+	 * greater than max_size.
+	 */
 	avail_size = max_size - (prod_pos - cons_pos);
 	/* Round up total size to a multiple of 8. */
 	total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
Signature · 1 line
2.39.2
#1 lore raw
Quoted 8 lines · Andrew Werner

Before this patch, the producer position could overflow unsigned long, in which case libbpf would forever stop processing new writes to the ringbuf. This patch addresses that bug by avoiding ordered comparison between the consumer and producer position. If the consumer position is greater than the producer position, the assumption is that the producer has overflowed.

Quoted 5 lines

A more defensive check could be to ensure that the delta is within the allowed range, but such defensive checks are neither present in the kernel side code nor in libbpf. The overflow that this patch handles can occur while the producer and consumer follow a correct protocol.

Yep, great find!

Quoted 5 lines

A selftest was written to demonstrate the bug, and indeed this patch allows the test to continue to make progress past the overflow. However, the author was unable to create a testing environment on a 32-bit machine, and the test requires substantial memory and over 4

2GB of memory for ringbuf, right? Perhaps it would be good to just outline the repro, even if we won't have it implemented in selftests. Something along the lines of: a) set up ringbuf of 2GB size and reserve+commit maximum-sized record (UMAX/4) constantly as fast as possible. With 1 million records per second repro time should be about 4.7 hours. Can you please update the commit with something like that instead of a vague "there is repro, but I won't show it ;)" ? Thanks!

Quoted 18 lines

hours to hit the overflow point on a 64-bit machine. Thus, the test is not included in this patch because of the impracticality of running it.

Additionally, this patch adds commentary around a separate point to note that the modular arithmetic is valid in the face of overflows, as that fact may not be obvious to future readers.

---
 tools/lib/bpf/ringbuf.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/lib/bpf/ringbuf.c +0 -0
index 02199364db13..6271757bc3d2 100644

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 241

Quoted 6 lines
tools/lib/bpf/ringbuf.c +2 -1
@@ -237,7 +237,11 @@ static int64_t ringbuf_process_ring(struct ring *r)
        do {
                got_new_data = false;
                prod_pos = smp_load_acquire(r->producer_pos);
-               while (cons_pos < prod_pos) {
+
+               /* Avoid signed comparisons between the positions; the producer

"signed comparisons" is confusing and invalid, as cons_pos and prod_pos are unsigned and comparison here is unsigned. What you meant is inequality comparison, which is invalid when done naively (like it is done in libbpf right now, sigh...), if counters can wrap around.

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 244

Quoted 3 lines
tools/lib/bpf/ringbuf.c +3 -0
+                * position can overflow before the consumer position.
+                */
+               while (cons_pos != prod_pos) {

I'm wondering if we should preserve the "consumer pos is before producer pos" check just for clarity's sake (with a comment about wrapping around of counters, of course) like:

if ((long)(cons_pos - prod_pos) < 0) ?

BTW, I think kernel code needs fixing as well in __bpf_user_ringbuf_peek (we should compare consumer/producer positions directly, only through subtraction and casting to signed long as above), would you be able to fix it at the same time with libbpf? Would be good to also double-check the rest of kernel/bpf/ringbuf.c to make sure we don't directly compare positions anywhere else.

Quoted 19 lines
tools/lib/bpf/ringbuf.c +5 -1
                        len_ptr = r->data + (cons_pos & r->mask);
                        len = smp_load_acquire(len_ptr);
 
@@ -498,6 +502,11 @@ void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
        prod_pos = smp_load_acquire(rb->producer_pos);
 
        max_size = rb->mask + 1;
+
+       /* Note that this formulation in the face of overflow of prod_pos
+        * so long as the delta between prod_pos and cons_pos remains no
+        * greater than max_size.
+        */
        avail_size = max_size - (prod_pos - cons_pos);
        /* Round up total size to a multiple of 8. */
        total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
--

2.39.2

#2 lore raw

Hi,

Quoted 19 lines · Andrii Nakryiko
Quoted 7 lines · Andrew Werner

Before this patch, the producer position could overflow unsigned long, in which case libbpf would forever stop processing new writes to the ringbuf. This patch addresses that bug by avoiding ordered comparison between the consumer and producer position. If the consumer position is greater than the producer position, the assumption is that the producer has overflowed.

Quoted 5 lines

A more defensive check could be to ensure that the delta is within the allowed range, but such defensive checks are neither present in the kernel side code nor in libbpf. The overflow that this patch handles can occur while the producer and consumer follow a correct protocol.

Yep, great find!

Quoted 4 lines

A selftest was written to demonstrate the bug, and indeed this patch allows the test to continue to make progress past the overflow. However, the author was unable to create a testing environment on a 32-bit machine, and the test requires substantial memory and over 4

Quoted 7 lines

2GB of memory for ringbuf, right? Perhaps it would be good to just outline the repro, even if we won't have it implemented in selftests. Something along the lines of: a) set up ringbuf of 2GB size and reserve+commit maximum-sized record (UMAX/4) constantly as fast as possible. With 1 million records per second repro time should be about 4.7 hours. Can you please update the commit with something like that instead of a vague "there is repro, but I won't show it ;)" ? Thanks!

I think it would be great that the commit message can elaborate about the repo. Andrew had already posted an external link to the reproducer in v2 [0]: https://github.com/ajwerner/bpf/commit/85e1240e7713

[0]:
https://lore.kernel.org/bpf/[email protected]/

Quoted 36 lines
Quoted 18 lines

hours to hit the overflow point on a 64-bit machine. Thus, the test is not included in this patch because of the impracticality of running it.

Additionally, this patch adds commentary around a separate point to note that the modular arithmetic is valid in the face of overflows, as that fact may not be obvious to future readers.

---
 tools/lib/bpf/ringbuf.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/lib/bpf/ringbuf.c +0 -0
index 02199364db13..6271757bc3d2 100644

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 241

Quoted 6 lines
tools/lib/bpf/ringbuf.c +2 -1
@@ -237,7 +237,11 @@ static int64_t ringbuf_process_ring(struct ring *r)
        do {
                got_new_data = false;
                prod_pos = smp_load_acquire(r->producer_pos);
-               while (cons_pos < prod_pos) {
+
+               /* Avoid signed comparisons between the positions; the producer

"signed comparisons" is confusing and invalid, as cons_pos and prod_pos are unsigned and comparison here is unsigned. What you meant is inequality comparison, which is invalid when done naively (like it is done in libbpf right now, sigh...), if counters can wrap around.

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 244

Quoted 3 lines
tools/lib/bpf/ringbuf.c +3 -0
+                * position can overflow before the consumer position.
+                */
+               while (cons_pos != prod_pos) {

I'm wondering if we should preserve the "consumer pos is before producer pos" check just for clarity's sake (with a comment about wrapping around of counters, of course) like:

Quoted 8 lines

if ((long)(cons_pos - prod_pos) < 0) ?

BTW, I think kernel code needs fixing as well in __bpf_user_ringbuf_peek (we should compare consumer/producer positions directly, only through subtraction and casting to signed long as above), would you be able to fix it at the same time with libbpf? Would be good to also double-check the rest of kernel/bpf/ringbuf.c to make sure we don't directly compare positions anywhere else.

I missed __bpf_user_ringbuf_peek() when reviewed the patch. I also can help to double-check kernel/bpf/ringbuf.c.

Quoted 21 lines
Quoted 17 lines
tools/lib/bpf/ringbuf.c +5 -1
                        len_ptr = r->data + (cons_pos & r->mask);
                        len = smp_load_acquire(len_ptr);
 
@@ -498,6 +502,11 @@ void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
        prod_pos = smp_load_acquire(rb->producer_pos);
 
        max_size = rb->mask + 1;
+
+       /* Note that this formulation in the face of overflow of prod_pos
+        * so long as the delta between prod_pos and cons_pos remains no
+        * greater than max_size.
+        */
        avail_size = max_size - (prod_pos - cons_pos);
        /* Round up total size to a multiple of 8. */
        total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
--
Quoted 3 lines

2.39.2

.

#3 lore raw
Quoted 31 lines · Hou Tao

Hi,

Quoted 19 lines · Andrii Nakryiko
Quoted 7 lines · Andrew Werner

Before this patch, the producer position could overflow unsigned long, in which case libbpf would forever stop processing new writes to the ringbuf. This patch addresses that bug by avoiding ordered comparison between the consumer and producer position. If the consumer position is greater than the producer position, the assumption is that the producer has overflowed.

Quoted 5 lines

A more defensive check could be to ensure that the delta is within the allowed range, but such defensive checks are neither present in the kernel side code nor in libbpf. The overflow that this patch handles can occur while the producer and consumer follow a correct protocol.

Yep, great find!

Quoted 4 lines

A selftest was written to demonstrate the bug, and indeed this patch allows the test to continue to make progress past the overflow. However, the author was unable to create a testing environment on a 32-bit machine, and the test requires substantial memory and over 4

Quoted 7 lines

2GB of memory for ringbuf, right? Perhaps it would be good to just outline the repro, even if we won't have it implemented in selftests. Something along the lines of: a) set up ringbuf of 2GB size and reserve+commit maximum-sized record (UMAX/4) constantly as fast as possible. With 1 million records per second repro time should be about 4.7 hours. Can you please update the commit with something like that instead of a vague "there is repro, but I won't show it ;)" ? Thanks!

Quoted 3 lines

I think it would be great that the commit message can elaborate about the repo. Andrew had already posted an external link to the reproducer in v2 [0]: https://github.com/ajwerner/bpf/commit/85e1240e7713

Yeah, of course, I saw it. I didn't mean that he actually refuses to show repro, it's just that the commit message is vague about it and we should improve that. I think we are on the same page. Including the link to full repro in addition to more detailed description is also good.

Quoted 48 lines

[0]:
https://lore.kernel.org/bpf/[email protected]/

Quoted 36 lines
Quoted 18 lines

hours to hit the overflow point on a 64-bit machine. Thus, the test is not included in this patch because of the impracticality of running it.

Additionally, this patch adds commentary around a separate point to note that the modular arithmetic is valid in the face of overflows, as that fact may not be obvious to future readers.

---
 tools/lib/bpf/ringbuf.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/lib/bpf/ringbuf.c +0 -0
index 02199364db13..6271757bc3d2 100644

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 241

Quoted 6 lines
tools/lib/bpf/ringbuf.c +2 -1
@@ -237,7 +237,11 @@ static int64_t ringbuf_process_ring(struct ring *r)
        do {
                got_new_data = false;
                prod_pos = smp_load_acquire(r->producer_pos);
-               while (cons_pos < prod_pos) {
+
+               /* Avoid signed comparisons between the positions; the producer

"signed comparisons" is confusing and invalid, as cons_pos and prod_pos are unsigned and comparison here is unsigned. What you meant is inequality comparison, which is invalid when done naively (like it is done in libbpf right now, sigh...), if counters can wrap around.

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 244

Quoted 3 lines
tools/lib/bpf/ringbuf.c +3 -0
+                * position can overflow before the consumer position.
+                */
+               while (cons_pos != prod_pos) {

I'm wondering if we should preserve the "consumer pos is before producer pos" check just for clarity's sake (with a comment about wrapping around of counters, of course) like:

Quoted 8 lines

if ((long)(cons_pos - prod_pos) < 0) ?

BTW, I think kernel code needs fixing as well in __bpf_user_ringbuf_peek (we should compare consumer/producer positions directly, only through subtraction and casting to signed long as above), would you be able to fix it at the same time with libbpf? Would be good to also double-check the rest of kernel/bpf/ringbuf.c to make sure we don't directly compare positions anywhere else.

Quoted 2 lines

I missed __bpf_user_ringbuf_peek() when reviewed the patch. I also can help to double-check kernel/bpf/ringbuf.c.

great, thanks!

Quoted 23 lines
Quoted 21 lines
Quoted 17 lines
tools/lib/bpf/ringbuf.c +5 -1
                        len_ptr = r->data + (cons_pos & r->mask);
                        len = smp_load_acquire(len_ptr);
 
@@ -498,6 +502,11 @@ void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
        prod_pos = smp_load_acquire(rb->producer_pos);
 
        max_size = rb->mask + 1;
+
+       /* Note that this formulation in the face of overflow of prod_pos
+        * so long as the delta between prod_pos and cons_pos remains no
+        * greater than max_size.
+        */
        avail_size = max_size - (prod_pos - cons_pos);
        /* Round up total size to a multiple of 8. */
        total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
--
Quoted 3 lines

2.39.2

.

#4 lore raw
Quoted 37 lines · Andrii Nakryiko
Quoted 31 lines · Hou Tao

Hi,

Quoted 19 lines · Andrii Nakryiko
Quoted 7 lines · Andrew Werner

Before this patch, the producer position could overflow unsigned long, in which case libbpf would forever stop processing new writes to the ringbuf. This patch addresses that bug by avoiding ordered comparison between the consumer and producer position. If the consumer position is greater than the producer position, the assumption is that the producer has overflowed.

Quoted 5 lines

A more defensive check could be to ensure that the delta is within the allowed range, but such defensive checks are neither present in the kernel side code nor in libbpf. The overflow that this patch handles can occur while the producer and consumer follow a correct protocol.

Yep, great find!

Quoted 4 lines

A selftest was written to demonstrate the bug, and indeed this patch allows the test to continue to make progress past the overflow. However, the author was unable to create a testing environment on a 32-bit machine, and the test requires substantial memory and over 4

Quoted 7 lines

2GB of memory for ringbuf, right? Perhaps it would be good to just outline the repro, even if we won't have it implemented in selftests. Something along the lines of: a) set up ringbuf of 2GB size and reserve+commit maximum-sized record (UMAX/4) constantly as fast as possible. With 1 million records per second repro time should be about 4.7 hours. Can you please update the commit with something like that instead of a vague "there is repro, but I won't show it ;)" ? Thanks!

Quoted 3 lines

I think it would be great that the commit message can elaborate about the repo. Andrew had already posted an external link to the reproducer in v2 [0]: https://github.com/ajwerner/bpf/commit/85e1240e7713

Quoted 5 lines

Yeah, of course, I saw it. I didn't mean that he actually refuses to show repro, it's just that the commit message is vague about it and we should improve that. I think we are on the same page. Including the link to full repro in addition to more detailed description is also good.

I'll update the message as you suggest in v3.

Quoted 33 lines
Quoted 33 lines

[0]:
https://lore.kernel.org/bpf/[email protected]/

Quoted 29 lines
Quoted 18 lines

hours to hit the overflow point on a 64-bit machine. Thus, the test is not included in this patch because of the impracticality of running it.

Additionally, this patch adds commentary around a separate point to note that the modular arithmetic is valid in the face of overflows, as that fact may not be obvious to future readers.

---
 tools/lib/bpf/ringbuf.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/lib/bpf/ringbuf.c +0 -0
index 02199364db13..6271757bc3d2 100644

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 241

Quoted 6 lines
tools/lib/bpf/ringbuf.c +2 -1
@@ -237,7 +237,11 @@ static int64_t ringbuf_process_ring(struct ring *r)
        do {
                got_new_data = false;
                prod_pos = smp_load_acquire(r->producer_pos);
-               while (cons_pos < prod_pos) {
+
+               /* Avoid signed comparisons between the positions; the producer

"signed comparisons" is confusing and invalid, as cons_pos and prod_pos are unsigned and comparison here is unsigned. What you meant is inequality comparison, which is invalid when done naively (like it is done in libbpf right now, sigh...), if counters can wrap around.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 244

Quoted 3 lines
tools/lib/bpf/ringbuf.c +3 -0
+                * position can overflow before the consumer position.
+                */
+               while (cons_pos != prod_pos) {

I'm wondering if we should preserve the "consumer pos is before producer pos" check just for clarity's sake (with a comment about wrapping around of counters, of course) like:

if ((long)(cons_pos - prod_pos) < 0) ?

Will do.

Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

BTW, I think kernel code needs fixing as well in __bpf_user_ringbuf_peek (we should compare consumer/producer positions directly, only through subtraction and casting to signed long as above), would you be able to fix it at the same time with libbpf? Would be good to also double-check the rest of kernel/bpf/ringbuf.c to make sure we don't directly compare positions anywhere else.

Quoted 4 lines
Quoted 2 lines

I missed __bpf_user_ringbuf_peek() when reviewed the patch. I also can help to double-check kernel/bpf/ringbuf.c.

great, thanks!

I'll update the code in __bpf_user_ringbuf_peek. One observation is that the code there differs from the regular (non-user) ringbuf and the libbpf code in that it uses u64 for the positions whereas everywhere else (including in the struct definition) the types of these positions are unsigned long. I get that on 64-bit architectures, practically speaking, these are the same. What I'm less clear on is whether there's anything that prevents this code from running (perhaps with bugs) on 32-bit machines. I suppose that on little-endian machines things will just work out until the positions overflow 32 bits.

Would it make sense for me to make the change to always use unsigned long for these values in the same change?

Quoted 24 lines
Quoted 24 lines
Quoted 22 lines
Quoted 18 lines
tools/lib/bpf/ringbuf.c +5 -1
                        len_ptr = r->data + (cons_pos & r->mask);
                        len = smp_load_acquire(len_ptr);
 
@@ -498,6 +502,11 @@ void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
        prod_pos = smp_load_acquire(rb->producer_pos);
 
        max_size = rb->mask + 1;
+
+       /* Note that this formulation in the face of overflow of prod_pos
+        * so long as the delta between prod_pos and cons_pos remains no
+        * greater than max_size.
+        */
        avail_size = max_size - (prod_pos - cons_pos);
        /* Round up total size to a multiple of 8. */
        total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
--
Quoted 3 lines

2.39.2

.

#5 lore raw
Quoted 116 lines · Andrew Werner
Quoted 37 lines · Andrii Nakryiko
Quoted 31 lines · Hou Tao

Hi,

Quoted 19 lines · Andrii Nakryiko
Quoted 7 lines · Andrew Werner

Before this patch, the producer position could overflow unsigned long, in which case libbpf would forever stop processing new writes to the ringbuf. This patch addresses that bug by avoiding ordered comparison between the consumer and producer position. If the consumer position is greater than the producer position, the assumption is that the producer has overflowed.

Quoted 5 lines

A more defensive check could be to ensure that the delta is within the allowed range, but such defensive checks are neither present in the kernel side code nor in libbpf. The overflow that this patch handles can occur while the producer and consumer follow a correct protocol.

Yep, great find!

Quoted 4 lines

A selftest was written to demonstrate the bug, and indeed this patch allows the test to continue to make progress past the overflow. However, the author was unable to create a testing environment on a 32-bit machine, and the test requires substantial memory and over 4

Quoted 7 lines

2GB of memory for ringbuf, right? Perhaps it would be good to just outline the repro, even if we won't have it implemented in selftests. Something along the lines of: a) set up ringbuf of 2GB size and reserve+commit maximum-sized record (UMAX/4) constantly as fast as possible. With 1 million records per second repro time should be about 4.7 hours. Can you please update the commit with something like that instead of a vague "there is repro, but I won't show it ;)" ? Thanks!

Quoted 3 lines

I think it would be great that the commit message can elaborate about the repo. Andrew had already posted an external link to the reproducer in v2 [0]: https://github.com/ajwerner/bpf/commit/85e1240e7713

Quoted 5 lines

Yeah, of course, I saw it. I didn't mean that he actually refuses to show repro, it's just that the commit message is vague about it and we should improve that. I think we are on the same page. Including the link to full repro in addition to more detailed description is also good.

I'll update the message as you suggest in v3.

Quoted 33 lines
Quoted 33 lines

[0]:
https://lore.kernel.org/bpf/[email protected]/

Quoted 29 lines
Quoted 18 lines

hours to hit the overflow point on a 64-bit machine. Thus, the test is not included in this patch because of the impracticality of running it.

Additionally, this patch adds commentary around a separate point to note that the modular arithmetic is valid in the face of overflows, as that fact may not be obvious to future readers.

---
 tools/lib/bpf/ringbuf.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/lib/bpf/ringbuf.c +0 -0
index 02199364db13..6271757bc3d2 100644

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 241

Quoted 6 lines
tools/lib/bpf/ringbuf.c +2 -1
@@ -237,7 +237,11 @@ static int64_t ringbuf_process_ring(struct ring *r)
        do {
                got_new_data = false;
                prod_pos = smp_load_acquire(r->producer_pos);
-               while (cons_pos < prod_pos) {
+
+               /* Avoid signed comparisons between the positions; the producer

"signed comparisons" is confusing and invalid, as cons_pos and prod_pos are unsigned and comparison here is unsigned. What you meant is inequality comparison, which is invalid when done naively (like it is done in libbpf right now, sigh...), if counters can wrap around.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

tools/lib/bpf/ringbuf.c static int64_t ringbuf_process_ring(struct ring *r) · line 244

Quoted 3 lines
tools/lib/bpf/ringbuf.c +3 -0
+                * position can overflow before the consumer position.
+                */
+               while (cons_pos != prod_pos) {

I'm wondering if we should preserve the "consumer pos is before producer pos" check just for clarity's sake (with a comment about wrapping around of counters, of course) like:

if ((long)(cons_pos - prod_pos) < 0) ?

Will do.

Quoted 8 lines
Quoted 8 lines
Quoted 7 lines

BTW, I think kernel code needs fixing as well in __bpf_user_ringbuf_peek (we should compare consumer/producer positions directly, only through subtraction and casting to signed long as above), would you be able to fix it at the same time with libbpf? Would be good to also double-check the rest of kernel/bpf/ringbuf.c to make sure we don't directly compare positions anywhere else.

Quoted 4 lines
Quoted 2 lines

I missed __bpf_user_ringbuf_peek() when reviewed the patch. I also can help to double-check kernel/bpf/ringbuf.c.

great, thanks!

I'll update the code in __bpf_user_ringbuf_peek. One observation is that the code there differs from the regular (non-user) ringbuf and the libbpf code in that it uses u64 for the positions whereas everywhere else (including in the struct definition) the types of these positions are unsigned long. I get that on 64-bit architectures, practically speaking, these are the same. What I'm less clear on is whether there's anything that prevents this code from running (perhaps with bugs) on 32-bit machines. I suppose that on little-endian machines things will just work out until the positions overflow 32 bits.

Quoted 2 lines

Would it make sense for me to make the change to always use unsigned long for these values in the same change?

I'm not sure, but I suspect that u64 usage is not intentional, so we probably do want to use unsigned long consistently.

cc'ing David as well

Quoted 25 lines
Quoted 25 lines
Quoted 25 lines
Quoted 23 lines
Quoted 19 lines
tools/lib/bpf/ringbuf.c +5 -1
                        len_ptr = r->data + (cons_pos & r->mask);
                        len = smp_load_acquire(len_ptr);
 
@@ -498,6 +502,11 @@ void *user_ring_buffer__reserve(struct user_ring_buffer *rb, __u32 size)
        prod_pos = smp_load_acquire(rb->producer_pos);
 
        max_size = rb->mask + 1;
+
+       /* Note that this formulation in the face of overflow of prod_pos
+        * so long as the delta between prod_pos and cons_pos remains no
+        * greater than max_size.
+        */
        avail_size = max_size - (prod_pos - cons_pos);
        /* Round up total size to a multiple of 8. */
        total_size = (size + BPF_RINGBUF_HDR_SZ + 7) / 8 * 8;
--
Quoted 3 lines

2.39.2

.

#6 lore raw

On Wed, Aug 23, 2023 at 05:07:42PM -0700, Andrii Nakryiko wrote:

[...]

Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 9 lines
Quoted 6 lines

BTW, I think kernel code needs fixing as well in __bpf_user_ringbuf_peek (we should compare consumer/producer positions directly, only through subtraction and casting to signed long as above), would you be able to fix it at the same time with libbpf? Would be good to also double-check the rest of kernel/bpf/ringbuf.c to make sure we don't directly compare positions anywhere else.

I missed __bpf_user_ringbuf_peek() when reviewed the patch. I also can help to double-check kernel/bpf/ringbuf.c.

great, thanks!

Good catch, and thanks!

Quoted 12 lines
Quoted 12 lines

I'll update the code in __bpf_user_ringbuf_peek. One observation is that the code there differs from the regular (non-user) ringbuf and the libbpf code in that it uses u64 for the positions whereas everywhere else (including in the struct definition) the types of these positions are unsigned long. I get that on 64-bit architectures, practically speaking, these are the same. What I'm less clear on is whether there's anything that prevents this code from running (perhaps with bugs) on 32-bit machines. I suppose that on little-endian machines things will just work out until the positions overflow 32 bits.

Quoted 5 lines
Quoted 2 lines

Would it make sense for me to make the change to always use unsigned long for these values in the same change?

I'm not sure, but I suspect that u64 usage is not intentional, so we probably do want to use unsigned long consistently.

Yeah, that was an oversight and should be fixed. Thanks for pointing it out Andrew. FYI, we'd also need to fix user_ring_buffer__reserve() in libbpf.

Andrew -- I'm happy to send a patch that fixes this. If you'd like to send it out, please feel free to do so. Just let me know.

Thanks,
David

#7 lore raw
Quoted 37 lines · David Vernet

On Wed, Aug 23, 2023 at 05:07:42PM -0700, Andrii Nakryiko wrote:

[...]

Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 9 lines
Quoted 6 lines

BTW, I think kernel code needs fixing as well in __bpf_user_ringbuf_peek (we should compare consumer/producer positions directly, only through subtraction and casting to signed long as above), would you be able to fix it at the same time with libbpf? Would be good to also double-check the rest of kernel/bpf/ringbuf.c to make sure we don't directly compare positions anywhere else.

I missed __bpf_user_ringbuf_peek() when reviewed the patch. I also can help to double-check kernel/bpf/ringbuf.c.

great, thanks!

Good catch, and thanks!

Quoted 12 lines
Quoted 12 lines

I'll update the code in __bpf_user_ringbuf_peek. One observation is that the code there differs from the regular (non-user) ringbuf and the libbpf code in that it uses u64 for the positions whereas everywhere else (including in the struct definition) the types of these positions are unsigned long. I get that on 64-bit architectures, practically speaking, these are the same. What I'm less clear on is whether there's anything that prevents this code from running (perhaps with bugs) on 32-bit machines. I suppose that on little-endian machines things will just work out until the positions overflow 32 bits.

Quoted 5 lines
Quoted 2 lines

Would it make sense for me to make the change to always use unsigned long for these values in the same change?

I'm not sure, but I suspect that u64 usage is not intentional, so we probably do want to use unsigned long consistently.

Quoted 6 lines

Yeah, that was an oversight and should be fixed. Thanks for pointing it out Andrew. FYI, we'd also need to fix user_ring_buffer__reserve() in libbpf.

Andrew -- I'm happy to send a patch that fixes this. If you'd like to send it out, please feel free to do so. Just let me know.

I handled it in v3 which I just posted here:
https://lore.kernel.org/bpf/[email protected]/T/#u

Quoted 3 lines

Thanks,
David