netdev_alloc_frag->page_frag_alloc may cause memory corruption in
the following process:
1. A netdev_alloc_frag function call need alloc 200 Bytes to build a skb.
2. Insufficient memory to alloc PAGE_FRAG_CACHE_MAX_ORDER(32K) in
__page_frag_cache_refill to fill frag cache, then one page(eg:4K)
is allocated, now current frag cache is 4K, alloc is success,
nc->pagecnt_bias--.
3. Then this 200 bytes skb in step 1 is freed, page->_refcount--.
4. Another netdev_alloc_frag function call need alloc 5k, page->_refcount
is equal to nc->pagecnt_bias, reset page count bias and offset to
start of new frag. page_frag_alloc will return the 4K memory for a
5K memory request.
5. The caller write on the extra 1k memory which is not actual allocated
will cause memory corruption.
page_frag_alloc is for fragmented allocation. We should warn the caller
to avoid memory corruption.
Signed-off-by: Chen Lin <[email protected]>
---
mm/page_alloc.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e008a3d..6c0db52 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5574,6 +5574,11 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
struct page *page;
int offset;
+ /* frag_alloc is not suitable for memory alloc which fragsz
+ * is bigger than PAGE_SIZE, use kmalloc or alloc_pages instead.
+ */
+ WARN_ON(fragsz > PAGE_SIZE);
+
if (unlikely(!nc->va)) {
refill:
page = __page_frag_cache_refill(nc, gfp_mask);
--
1.7.9.5
On Sat, 28 May 2022 23:39:33 +0800 Chen Lin <[email protected]> wrote:
> netdev_alloc_frag->page_frag_alloc may cause memory corruption in
> the following process:
>
> 1. A netdev_alloc_frag function call need alloc 200 Bytes to build a skb.
>
> 2. Insufficient memory to alloc PAGE_FRAG_CACHE_MAX_ORDER(32K) in
> __page_frag_cache_refill to fill frag cache, then one page(eg:4K)
> is allocated, now current frag cache is 4K, alloc is success,
> nc->pagecnt_bias--.
>
> 3. Then this 200 bytes skb in step 1 is freed, page->_refcount--.
>
> 4. Another netdev_alloc_frag function call need alloc 5k, page->_refcount
> is equal to nc->pagecnt_bias, reset page count bias and offset to
> start of new frag. page_frag_alloc will return the 4K memory for a
> 5K memory request.
>
> 5. The caller write on the extra 1k memory which is not actual allocated
> will cause memory corruption.
>
> page_frag_alloc is for fragmented allocation. We should warn the caller
> to avoid memory corruption.
>
Let's cc Alexander and the networking developers.
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5574,6 +5574,11 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
> struct page *page;
> int offset;
>
> + /* frag_alloc is not suitable for memory alloc which fragsz
> + * is bigger than PAGE_SIZE, use kmalloc or alloc_pages instead.
> + */
> + WARN_ON(fragsz > PAGE_SIZE);
> +
> if (unlikely(!nc->va)) {
> refill:
> page = __page_frag_cache_refill(nc, gfp_mask);
Odd. All this does is generate a warning. If the kernel is corrupting
memory, that's a bug which needs fixing?
netdev_alloc_frag->page_frag_alloc may cause memory corruption in
the following process:
1. A netdev_alloc_frag function call need alloc 200 Bytes to build a skb.
2. Insufficient memory to alloc PAGE_FRAG_CACHE_MAX_ORDER(32K) in
__page_frag_cache_refill to fill frag cache, then one page(eg:4K)
is allocated, now current frag cache is 4K, alloc is success,
nc->pagecnt_bias--.
3. Then this 200 bytes skb in step 1 is freed, page->_refcount--.
4. Another netdev_alloc_frag function call need alloc 5k, page->_refcount
is equal to nc->pagecnt_bias, reset page count bias and offset to
start of new frag. page_frag_alloc will return the 4K memory for a
5K memory request.
5. The caller write on the extra 1k memory which is not actual allocated
will cause memory corruption.
page_frag_alloc is for fragmented allocation. We should warn the caller
to avoid memory corruption.
When fragsz is larger than one page, we report the failure and return.
I don't think it is a good idea to make efforts to support the
allocation of more than one page in this function because the total
frag cache size(PAGE_FRAG_CACHE_MAX_SIZE 32768) is relatively small.
When the request is larger than one page, the caller should switch to
use other kernel interfaces, such as kmalloc and alloc_Pages.
This bug is mainly caused by the reuse of the previously allocated
frag cache memory by the following LARGER allocations. This bug existed
before page_frag_alloc was ported from __netdev_alloc_frag in
net/core/skbuff.c, so most Linux versions have this problem.
Signed-off-by: Chen Lin <[email protected]>
---
mm/page_alloc.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e008a3d..1e9e2c4 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5574,6 +5574,16 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
struct page *page;
int offset;
+ /* frag_alloc is not suitable for memory alloc which fragsz
+ * is bigger than PAGE_SIZE, use kmalloc or alloc_pages instead.
+ */
+ if (unlikely(fragsz > PAGE_SIZE)) {
+ WARN(1, "alloc fragsz(%d) > PAGE_SIZE(%ld) not supported,
+ alloc fail\n", fragsz, PAGE_SIZE);
+
+ return NULL;
+ }
+
if (unlikely(!nc->va)) {
refill:
page = __page_frag_cache_refill(nc, gfp_mask);
--
1.7.9.5
On Tue, May 31, 2022 at 8:47 AM Jakub Kicinski <[email protected]> wrote:
>
> On Tue, 31 May 2022 23:36:22 +0800 Chen Lin wrote:
> > At 2022-05-31 22:14:12, "Jakub Kicinski" <[email protected]> wrote:
> > >On Tue, 31 May 2022 22:41:12 +0800 Chen Lin wrote:
> > >> The sample code above cannot completely solve the current problem.
> > >> For example, when fragsz is greater than PAGE_FRAG_CACHE_MAX_SIZE(32768),
> > >> __page_frag_cache_refill will return a memory of only 32768 bytes, so
> > >> should we continue to expand the PAGE_FRAG_CACHE_MAX_SIZE? Maybe more
> > >> work needs to be done
> > >
> > >Right, but I can think of two drivers off the top of my head which will
> > >allocate <=32k frags but none which will allocate more.
> >
> > In fact, it is rare to apply for more than one page, so is it necessary to
> > change it to support?
>
> I don't really care if it's supported TBH, but I dislike adding
> a branch to the fast path just to catch one or two esoteric bad
> callers.
>
> Maybe you can wrap the check with some debug CONFIG_ so it won't
> run on production builds?
Also the example used here to define what is triggering the behavior
is seriously flawed. The code itself is meant to allow for order0 page
reuse, and the 32K page was just an optimization. So the assumption
that you could request more than 4k is a bad assumption in the driver
that is making this call.
So I am in agreement with Kuba. We shouldn't be needing to add code in
the fast path to tell users not to shoot themselves in the foot.
We already have code in place in __netdev_alloc_skb that is calling
the slab allocator if "len > SKB_WITH_OVERHEAD(PAGE_SIZE)". We could
probably just add a DEBUG wrapped BUG_ON to capture those cases where
a driver is making that mistake with __netdev_alloc_frag_align.
On Mon, 30 May 2022 21:39:02 +0800 Chen Lin <[email protected]> wrote:
> netdev_alloc_frag->page_frag_alloc may cause memory corruption in
> the following process:
>
> 1. A netdev_alloc_frag function call need alloc 200 Bytes to build a skb.
>
> 2. Insufficient memory to alloc PAGE_FRAG_CACHE_MAX_ORDER(32K) in
> __page_frag_cache_refill to fill frag cache, then one page(eg:4K)
> is allocated, now current frag cache is 4K, alloc is success,
> nc->pagecnt_bias--.
>
> 3. Then this 200 bytes skb in step 1 is freed, page->_refcount--.
>
> 4. Another netdev_alloc_frag function call need alloc 5k, page->_refcount
> is equal to nc->pagecnt_bias, reset page count bias and offset to
> start of new frag. page_frag_alloc will return the 4K memory for a
> 5K memory request.
>
> 5. The caller write on the extra 1k memory which is not actual allocated
> will cause memory corruption.
>
> page_frag_alloc is for fragmented allocation. We should warn the caller
> to avoid memory corruption.
>
> When fragsz is larger than one page, we report the failure and return.
> I don't think it is a good idea to make efforts to support the
> allocation of more than one page in this function because the total
> frag cache size(PAGE_FRAG_CACHE_MAX_SIZE 32768) is relatively small.
> When the request is larger than one page, the caller should switch to
> use other kernel interfaces, such as kmalloc and alloc_Pages.
>
> This bug is mainly caused by the reuse of the previously allocated
> frag cache memory by the following LARGER allocations. This bug existed
> before page_frag_alloc was ported from __netdev_alloc_frag in
> net/core/skbuff.c, so most Linux versions have this problem.
>
I won't attempt to address the large issues here (like, should
networking be changed to support this). But I can nitpick :)
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5574,6 +5574,16 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
> struct page *page;
> int offset;
>
> + /* frag_alloc is not suitable for memory alloc which fragsz
Like this please:
/*
* frag_alloc...
> + * is bigger than PAGE_SIZE, use kmalloc or alloc_pages instead.
> + */
> + if (unlikely(fragsz > PAGE_SIZE)) {
> + WARN(1, "alloc fragsz(%d) > PAGE_SIZE(%ld) not supported,
> + alloc fail\n", fragsz, PAGE_SIZE);
It's neater to do
if (WARN(fragsz > PAGE_SIZE, "alloc fragsz(%d...", ...))
return NULL;
Also, you have a newline and a bunch of tabs in that string.
Also, please consider WARN_ONCE. We don't want to provide misbehaved
or malicious userspace with the ability to flood the logs with
warnings.
On Wed, Jun 1, 2022 at 5:33 AM 愚树 <[email protected]> wrote:
>
> At 2022-06-01 01:28:59, "Alexander Duyck" <[email protected]> wrote:
> >On Tue, May 31, 2022 at 8:47 AM Jakub Kicinski <[email protected]> wrote:
> >>
> >> On Tue, 31 May 2022 23:36:22 +0800 Chen Lin wrote:
> >> > At 2022-05-31 22:14:12, "Jakub Kicinski" <[email protected]> wrote:
> >> > >On Tue, 31 May 2022 22:41:12 +0800 Chen Lin wrote:
> >> > >> The sample code above cannot completely solve the current problem.
> >> > >> For example, when fragsz is greater than PAGE_FRAG_CACHE_MAX_SIZE(32768),
> >> > >> __page_frag_cache_refill will return a memory of only 32768 bytes, so
> >> > >> should we continue to expand the PAGE_FRAG_CACHE_MAX_SIZE? Maybe more
> >> > >> work needs to be done
> >> > >
> >> > >Right, but I can think of two drivers off the top of my head which will
> >> > >allocate <=32k frags but none which will allocate more.
> >> >
> >> > In fact, it is rare to apply for more than one page, so is it necessary to
> >> > change it to support?
> >>
> >> I don't really care if it's supported TBH, but I dislike adding
> >> a branch to the fast path just to catch one or two esoteric bad
> >> callers.
> >>
> >> Maybe you can wrap the check with some debug CONFIG_ so it won't
> >> run on production builds?
> >
> >Also the example used here to define what is triggering the behavior
> >is seriously flawed. The code itself is meant to allow for order0 page
> >reuse, and the 32K page was just an optimization. So the assumption
> >that you could request more than 4k is a bad assumption in the driver
> >that is making this call.
> >
> >So I am in agreement with Kuba. We shouldn't be needing to add code in
> >the fast path to tell users not to shoot themselves in the foot.
> >
> >We already have code in place in __netdev_alloc_skb that is calling
> >the slab allocator if "len > SKB_WITH_OVERHEAD(PAGE_SIZE)". We could
> >probably just add a DEBUG wrapped BUG_ON to capture those cases where
> >a driver is making that mistake with __netdev_alloc_frag_align.
>
> Thanks for the clear explanation.
> The reality is that it is not easy to capture the drivers that make such mistake.
> Because memory corruption usually leads to errors on other unrelated modules.
> Not long ago, we have spent a lot of time and effort to locate a issue that
> occasionally occurs in different kernel modules, and finally find the root cause is
> the improper use of this netdev_alloc_frag interface in DPAA net driver from NXP.
> It's a miserable process.
>
> I also found that some net drivers in the latest Linux version have this issue.
> Like:
> 1. netdev_alloc_frag "len" may larger than PAGE_SIZE
> #elif (PAGE_SIZE >= E1000_RXBUFFER_4096)
> adapter->rx_buffer_len = PAGE_SIZE;
> #endif
>
> static unsigned int e1000_frag_len(const struct e1000_adapter *a)
> {
> return SKB_DATA_ALIGN(a->rx_buffer_len + E1000_HEADROOM) +
> SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> }
>
> static void *e1000_alloc_frag(const struct e1000_adapter *a)
> {
> unsigned int len = e1000_frag_len(a);
> u8 *data = netdev_alloc_frag(len);
> }
> "./drivers/net/ethernet/intel/e1000/e1000_main.c" 5316 --38%--
So there isn't actually a bug in this code. Specifically the code is
split up between two paths. The first code block comes from the jumbo
frames path which creates a fraglist skb and will memcpy the header
out if I recall correctly. The code from the other two functions is
from the non-jumbo frames path which has restricted the length to
MAXIMUM_ETHERNET_VLAN_SIZE.
> 2. netdev_alloc_frag "ring->frag_size" may larger than (4096 * 3)
>
> #define MTK_MAX_LRO_RX_LENGTH (4096 * 3)
> if (rx_flag == MTK_RX_FLAGS_HWLRO) {
> rx_data_len = MTK_MAX_LRO_RX_LENGTH;
> rx_dma_size = MTK_HW_LRO_DMA_SIZE;
> } else {
> rx_data_len = ETH_DATA_LEN;
> rx_dma_size = MTK_DMA_SIZE;
> }
>
> ring->frag_size = mtk_max_frag_size(rx_data_len);
>
> for (i = 0; i < rx_dma_size; i++) {
> ring->data[i] = netdev_alloc_frag(ring->frag_size);
> if (!ring->data[i])
> return -ENOMEM;
> }
> "drivers/net/ethernet/mediatek/mtk_eth_soc.c" 3344 --50%--
>
> I will try to fix these drivers later.
This one I don't know as much about, and it does appear to contain a
bug. What it should be doing is a check before doing the
netdev_alloc_frag call to verify if it is less than 4K then it uses
netdev_alloc_frag, if it is greater then it needs to use alloc_pages.
> Even experienced driver engineers may use this netdev_alloc_frag
> interface incorrectly.
> So I thought it is best to provide some prompt information of usage
> error inside the netdev_alloc_frag, or it's OK to report such mistake
> during system running which may caused by fragsz varies(exceeded page size).
>
> Now, as you and Kuba mentioned earlier, "do not add code in fast path".
>
> Can we just add code to the relatively slow path to capture the mistake
> before it lead to memory corruption?
> Like:
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index e6f211d..ac60a97 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5580,6 +5580,7 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
> /* reset page count bias and offset to start of new frag */
> nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
> offset = size - fragsz;
> + BUG_ON(offset < 0);
> }
>
> nc->pagecnt_bias--;
>
I think I could be onboard with a patch like this. The test shouldn't
add more than 1 instruction since it is essentially just a jump if
signed test which will be performed after the size - fragsz check.
> Additional, we may modify document to clearly indicate the limits of the
> input parameter fragsz.
> Like:
> diff --git a/Documentation/vm/page_frags.rst b/Documentation/vm/page_frags.rst
> index 7d6f938..61b2805 100644
> --- a/Documentation/vm/page_frags.rst
> +++ b/Documentation/vm/page_frags.rst
> @@ -4,7 +4,7 @@
> Page fragments
> ==============
>
> -A page fragment is an arbitrary-length arbitrary-offset area of memory
> +A page fragment is an arbitrary-length(must <= PAGE_SIZE) arbitrary-offset area of memory
> which resides within a 0 or higher order compound page.
The main thing I would call out about the page fragment is that it
should be less than an order 0 page in size, ideally at least half a
page to allow for reuse even in the case of order 0 pages. Otherwise
it is really an abuse of the interface as it isn't really meant to be
allocating 1 fragment per page since the efficiency will drop pretty
significantly as memory becomes fragmented and it becomes harder to
allocate higher order pages. It would essentially just become
alloc_page with more overhead.
At 2022-06-01 01:28:59, "Alexander Duyck" <[email protected]> wrote:
>On Tue, May 31, 2022 at 8:47 AM Jakub Kicinski <[email protected]> wrote:
>>
>> On Tue, 31 May 2022 23:36:22 +0800 Chen Lin wrote:
>> > At 2022-05-31 22:14:12, "Jakub Kicinski" <[email protected]> wrote:
>> > >On Tue, 31 May 2022 22:41:12 +0800 Chen Lin wrote:
>> > >> The sample code above cannot completely solve the current problem.
>> > >> For example, when fragsz is greater than PAGE_FRAG_CACHE_MAX_SIZE(32768),
>> > >> __page_frag_cache_refill will return a memory of only 32768 bytes, so
>> > >> should we continue to expand the PAGE_FRAG_CACHE_MAX_SIZE? Maybe more
>> > >> work needs to be done
>> > >
>> > >Right, but I can think of two drivers off the top of my head which will
>> > >allocate <=32k frags but none which will allocate more.
>> >
>> > In fact, it is rare to apply for more than one page, so is it necessary to
>> > change it to support?
>>
>> I don't really care if it's supported TBH, but I dislike adding
>> a branch to the fast path just to catch one or two esoteric bad
>> callers.
>>
>> Maybe you can wrap the check with some debug CONFIG_ so it won't
>> run on production builds?
>
>Also the example used here to define what is triggering the behavior
>is seriously flawed. The code itself is meant to allow for order0 page
>reuse, and the 32K page was just an optimization. So the assumption
>that you could request more than 4k is a bad assumption in the driver
>that is making this call.
>
>So I am in agreement with Kuba. We shouldn't be needing to add code in
>the fast path to tell users not to shoot themselves in the foot.
>
>We already have code in place in __netdev_alloc_skb that is calling
>the slab allocator if "len > SKB_WITH_OVERHEAD(PAGE_SIZE)". We could
>probably just add a DEBUG wrapped BUG_ON to capture those cases where
>a driver is making that mistake with __netdev_alloc_frag_align.
Thanks for the clear explanation.
The reality is that it is not easy to capture the drivers that make such mistake.
Because memory corruption usually leads to errors on other unrelated modules.
Not long ago, we have spent a lot of time and effort to locate a issue that
occasionally occurs in different kernel modules, and finally find the root cause is
the improper use of this netdev_alloc_frag interface in DPAA net driver from NXP.
It's a miserable process.
I also found that some net drivers in the latest Linux version have this issue.
Like:
1. netdev_alloc_frag "len" may larger than PAGE_SIZE
#elif (PAGE_SIZE >= E1000_RXBUFFER_4096)
adapter->rx_buffer_len = PAGE_SIZE;
#endif
static unsigned int e1000_frag_len(const struct e1000_adapter *a)
{
return SKB_DATA_ALIGN(a->rx_buffer_len + E1000_HEADROOM) +
SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
}
static void *e1000_alloc_frag(const struct e1000_adapter *a)
{
unsigned int len = e1000_frag_len(a);
u8 *data = netdev_alloc_frag(len);
}
"./drivers/net/ethernet/intel/e1000/e1000_main.c" 5316 --38%--
2. netdev_alloc_frag "ring->frag_size" may larger than (4096 * 3)
#define MTK_MAX_LRO_RX_LENGTH (4096 * 3)
if (rx_flag == MTK_RX_FLAGS_HWLRO) {
rx_data_len = MTK_MAX_LRO_RX_LENGTH;
rx_dma_size = MTK_HW_LRO_DMA_SIZE;
} else {
rx_data_len = ETH_DATA_LEN;
rx_dma_size = MTK_DMA_SIZE;
}
ring->frag_size = mtk_max_frag_size(rx_data_len);
for (i = 0; i < rx_dma_size; i++) {
ring->data[i] = netdev_alloc_frag(ring->frag_size);
if (!ring->data[i])
return -ENOMEM;
}
"drivers/net/ethernet/mediatek/mtk_eth_soc.c" 3344 --50%--
I will try to fix these drivers later.
Even experienced driver engineers may use this netdev_alloc_frag
interface incorrectly.
So I thought it is best to provide some prompt information of usage
error inside the netdev_alloc_frag, or it's OK to report such mistake
during system running which may caused by fragsz varies(exceeded page size).
Now, as you and Kuba mentioned earlier, "do not add code in fast path".
Can we just add code to the relatively slow path to capture the mistake
before it lead to memory corruption?
Like:
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e6f211d..ac60a97 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5580,6 +5580,7 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
/* reset page count bias and offset to start of new frag */
nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
offset = size - fragsz;
+ BUG_ON(offset < 0);
}
nc->pagecnt_bias--;
Additional, we may modify document to clearly indicate the limits of the
input parameter fragsz.
Like:
diff --git a/Documentation/vm/page_frags.rst b/Documentation/vm/page_frags.rst
index 7d6f938..61b2805 100644
--- a/Documentation/vm/page_frags.rst
+++ b/Documentation/vm/page_frags.rst
@@ -4,7 +4,7 @@
Page fragments
==============
-A page fragment is an arbitrary-length arbitrary-offset area of memory
+A page fragment is an arbitrary-length(must <= PAGE_SIZE) arbitrary-offset area of memory
which resides within a 0 or higher order compound page.
Thanks
On Mon, 30 May 2022 12:27:05 -0700 Jakub Kicinski wrote:
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index e008a3df0485..360a545ee5e8 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5537,6 +5537,7 @@ EXPORT_SYMBOL(free_pages);
> * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
> */
> static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
> + unsigned int fragsz,
> gfp_t gfp_mask)
> {
> struct page *page = NULL;
> @@ -5549,7 +5550,7 @@ static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
> PAGE_FRAG_CACHE_MAX_ORDER);
> nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
> #endif
> - if (unlikely(!page))
> + if (unlikely(!page && fragsz <= PAGE_SIZE))
> page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
>
> nc->va = page ? page_address(page) : NULL;
> @@ -5576,7 +5577,7 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
>
> if (unlikely(!nc->va)) {
> refill:
> - page = __page_frag_cache_refill(nc, gfp_mask);
> + page = __page_frag_cache_refill(nc, fragsz, gfp_mask);
> if (!page)
> return NULL;
Oh, well, the reuse also needs an update. We can slap a similar
condition next to the pfmemalloc check.
netdev_alloc_frag->page_frag_alloc may cause memory corruption in
the following process:
1. A netdev_alloc_frag function call need alloc 200 Bytes to build a skb.
2. Insufficient memory to alloc PAGE_FRAG_CACHE_MAX_ORDER(32K) in
__page_frag_cache_refill to fill frag cache, then one page(eg:4K)
is allocated, now current frag cache is 4K, alloc is success,
nc->pagecnt_bias--.
3. Then this 200 bytes skb in step 1 is freed, page->_refcount--.
4. Another netdev_alloc_frag function call need alloc 5k, page->_refcount
is equal to nc->pagecnt_bias, reset page count bias and offset to
start of new frag. page_frag_alloc will return the 4K memory for a
5K memory request.
5. The caller write on the extra 1k memory which is not actual allocated
will cause memory corruption.
page_frag_alloc is for fragmented allocation. We should warn the caller
to avoid memory corruption.
When fragsz is larger than one page, we report the failure and return.
I don't think it is a good idea to make efforts to support the
allocation of more than one page in this function because the total
frag cache size(PAGE_FRAG_CACHE_MAX_SIZE 32768) is relatively small.
When the request is larger than one page, the caller should switch to
use other kernel interfaces, such as kmalloc and alloc_Pages.
This bug is mainly caused by the reuse of the previously allocated
frag cache memory by the following LARGER allocations. This bug existed
before page_frag_alloc was ported from __netdev_alloc_frag in
net/core/skbuff.c, so most Linux versions have this problem.
Signed-off-by: Chen Lin <[email protected]>
---
mm/page_alloc.c | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e008a3d..ffc42b5 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5574,6 +5574,15 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
struct page *page;
int offset;
+ /*
+ * frag_alloc is not suitable for memory alloc which fragsz
+ * is bigger than PAGE_SIZE, use kmalloc or alloc_pages instead.
+ */
+ if (WARN_ONCE(fragz > PAGE_SIZE,
+ "alloc fragsz(%d) > PAGE_SIZE(%ld) not supported, alloc fail\n",
+ fragsz, PAGE_SIZE))
+ return NULL;
+
if (unlikely(!nc->va)) {
refill:
page = __page_frag_cache_refill(nc, gfp_mask);
--
1.7.9.5
On Mon, 30 May 2022 21:39:02 +0800 Chen Lin wrote:
> netdev_alloc_frag->page_frag_alloc may cause memory corruption in
> the following process:
>
> 1. A netdev_alloc_frag function call need alloc 200 Bytes to build a skb.
>
> 2. Insufficient memory to alloc PAGE_FRAG_CACHE_MAX_ORDER(32K) in
> __page_frag_cache_refill to fill frag cache, then one page(eg:4K)
> is allocated, now current frag cache is 4K, alloc is success,
> nc->pagecnt_bias--.
>
> 3. Then this 200 bytes skb in step 1 is freed, page->_refcount--.
>
> 4. Another netdev_alloc_frag function call need alloc 5k, page->_refcount
> is equal to nc->pagecnt_bias, reset page count bias and offset to
> start of new frag. page_frag_alloc will return the 4K memory for a
> 5K memory request.
>
> 5. The caller write on the extra 1k memory which is not actual allocated
> will cause memory corruption.
>
> page_frag_alloc is for fragmented allocation. We should warn the caller
> to avoid memory corruption.
>
> When fragsz is larger than one page, we report the failure and return.
> I don't think it is a good idea to make efforts to support the
> allocation of more than one page in this function because the total
> frag cache size(PAGE_FRAG_CACHE_MAX_SIZE 32768) is relatively small.
> When the request is larger than one page, the caller should switch to
> use other kernel interfaces, such as kmalloc and alloc_Pages.
>
> This bug is mainly caused by the reuse of the previously allocated
> frag cache memory by the following LARGER allocations. This bug existed
> before page_frag_alloc was ported from __netdev_alloc_frag in
> net/core/skbuff.c, so most Linux versions have this problem.
>
> Signed-off-by: Chen Lin <[email protected]>
> ---
> mm/page_alloc.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index e008a3d..1e9e2c4 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5574,6 +5574,16 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
> struct page *page;
> int offset;
>
> + /* frag_alloc is not suitable for memory alloc which fragsz
> + * is bigger than PAGE_SIZE, use kmalloc or alloc_pages instead.
> + */
> + if (unlikely(fragsz > PAGE_SIZE)) {
> + WARN(1, "alloc fragsz(%d) > PAGE_SIZE(%ld) not supported,
> + alloc fail\n", fragsz, PAGE_SIZE);
> +
> + return NULL;
> + }
> +
> if (unlikely(!nc->va)) {
> refill:
> page = __page_frag_cache_refill(nc, gfp_mask);
Let's see what Alex says (fixing his email now). It seems a little too
drastic to me. I'd go with something like:
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e008a3df0485..360a545ee5e8 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5537,6 +5537,7 @@ EXPORT_SYMBOL(free_pages);
* sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
*/
static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
+ unsigned int fragsz,
gfp_t gfp_mask)
{
struct page *page = NULL;
@@ -5549,7 +5550,7 @@ static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
PAGE_FRAG_CACHE_MAX_ORDER);
nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
#endif
- if (unlikely(!page))
+ if (unlikely(!page && fragsz <= PAGE_SIZE))
page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
nc->va = page ? page_address(page) : NULL;
@@ -5576,7 +5577,7 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
if (unlikely(!nc->va)) {
refill:
- page = __page_frag_cache_refill(nc, gfp_mask);
+ page = __page_frag_cache_refill(nc, fragsz, gfp_mask);
if (!page)
return NULL;
st 1. 6. 2022 v 17:05 odesílatel Alexander Duyck
<[email protected]> napsal:
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index e6f211d..ac60a97 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -5580,6 +5580,7 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
> > /* reset page count bias and offset to start of new frag */
> > nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
> > offset = size - fragsz;
> > + BUG_ON(offset < 0);
> > }
> >
> > nc->pagecnt_bias--;
> >
>
>
> I think I could be onboard with a patch like this. The test shouldn't
> add more than 1 instruction since it is essentially just a jump if
> signed test which will be performed after the size - fragsz check.
FYI, I hit this problem a few days ago with the nfp network driver, it uses
page_frag_alloc() with a frag size larger than PAGE_SIZE when MTU is
set to 9000,
this may result in memory corruptions when the system runs out of memory.
The solution I was working on was something like the following, this
makes the allocation
fail if fragsz is greater than the cache size.
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 4dc0d333279f..c6b40b85c55d 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5544,12 +5544,17 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
/* if size can vary use size else just use PAGE_SIZE */
size = nc->size;
#endif
- /* OK, page count is 0, we can safely set it */
- set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
-
/* reset page count bias and offset to start of new frag */
nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
offset = size - fragsz;
+ if (unlikely(offset < 0)) {
+ free_the_page(page, compound_order(page));
+ nc->va = NULL;
+ return NULL;
+ }
+
+ /* OK, page count is 0, we can safely set it */
+ set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
}
nc->pagecnt_bias--;
Maurizio
st 1. 6. 2022 v 14:49 odesílatel 愚树 <[email protected]> napsal:
> Can we just add code to the relatively slow path to capture the mistake
> before it lead to memory corruption?
> Like:
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index e6f211d..ac60a97 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5580,6 +5580,7 @@ void *page_frag_alloc_align(struct page_frag_cache *nc,
> /* reset page count bias and offset to start of new frag */
> nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
> offset = size - fragsz;
> + BUG_ON(offset < 0);
> }
>
Personally, I'm not really convinced this is the best solution.
The next time a driver abuses the page_frag_alloc() interface, the
bug may go unnoticed for a long time...
until a server in production runs into OOM and crashes because it hits
the BUG_ON().
And why should the kernel panic? It's perfectly able to handle this
condition by failing
the allocation and returning NULL, and printing a warning maybe.
Maurizio