We discovered a security-related issue in the DMA pool allocator.

V1 of our RFC was submitted to the Linux kernel security team. They
recommended submitting it to the relevant subsystem maintainers and the
hardening mailing list instead, as they did not consider this an explicit
security issue. Their rationale was that Linux implicitly assumes hardware
can be trusted.

**Threat Model**: While Linux drivers typically trust their hardware, there
may be specific drivers that do not operate under this assumption. Hence,
this threat model assumes a malicious peripheral device capable of
corrupting DMA data to exploit the kernel. In this scenario, the device
manipulates kernel-initialized data (similar to the attack described in the
Thunderclap paper [0]) to achieve arbitrary kernel memory corruption. 

**DMA pool background**. A DMA pool aims to reduce the overhead of DMA
allocations by creating a large DMA buffer --- the "pool" --- from which
smaller buffers are allocated as needed. Fundamentally, a DMA pool
functions like a heap: it is a structure composed of linked memory
"blocks", which, in this context, are DMA buffers. When a driver employs a
DMA pool, it grants the device access not only to these blocks but also to
the pointers linking them.

**Vulnerability**. Similar to traditional heap corruption vulnerabilities
--- where a malicious program corrupts heap metadata to e.g., hijack
control flow --- a malicious device may corrupt DMA pool metadata. This
corruption can trivially lead to arbitrary kernel memory corruption from
any driver that uses it. Indeed, because the DMA pool API is extensively
used, this vulnerability is not confined to a single instance. In fact,
every usage of the DMA pool API is potentially vulnerable. An exploit
proceeds with the following steps:

1. The DMA `pool` initializes its list of blocks, then points to the first
block.
2. The malicious device overwrites the first 8 bytes of the first block ---
which contain its `next_block` pointer --- to an arbitrary kernel address,
`kernel_addr`.
3. The driver makes its first call to `dma_pool_alloc()`, after which, the
pool should point to the second block. However, it instead points to
`kernel_addr`.
4. The driver again calls `dma_pool_alloc()`, which incorrectly returns
`kernel_addr`. Therefore, anytime the driver writes to this "block", it may
corrupt sensitive kernel data.

I have a PDF document that illustrates how these steps work. Please let me
know if you would like me to share it with you.

**Proposed mitigation**. To mitigate the corruption of DMA pool metadata
(i.e., the pointers linking the blocks), the metadata should be moved into
non-DMA memory, ensuring it cannot be altered by a device. I have included
a patch series that implements this change. Since I am not deeply familiar
with the DMA pool internals, I would appreciate any feedback on the
patches. I have tested the patches with the `DMAPOOL_TEST` test and my own
basic unit tests that ensure the DMA pool allocator is not vulnerable.

**Performance**. I evaluated the patch set's performance by running the
`DMAPOOL_TEST` test with `DMAPOOL_DEBUG` enabled and with/without the
patches applied. Here is its output *without* the patches applied:
```
dmapool test: size:16   align:16   blocks:8192 time:3194110
dmapool test: size:64   align:64   blocks:8192 time:4730440
dmapool test: size:256  align:256  blocks:8192 time:5489630
dmapool test: size:1024 align:1024 blocks:2048 time:517150
dmapool test: size:4096 align:4096 blocks:1024 time:399616
dmapool test: size:68   align:32   blocks:8192 time:6156527
```

And here is its output *with* the patches applied:
```
dmapool test: size:16   align:16   blocks:8192 time:3541031
dmapool test: size:64   align:64   blocks:8192 time:4227262
dmapool test: size:256  align:256  blocks:8192 time:4890273
dmapool test: size:1024 align:1024 blocks:2048 time:515775
dmapool test: size:4096 align:4096 blocks:1024 time:523096
dmapool test: size:68   align:32   blocks:8192 time:3450830
```

Based on my interpretation of the output, the patch set does not appear to
negatively impact performance. In fact, it shows slight performance
improvements in some tests (i.e., for sizes 64, 256, 1024, and 68).

I speculate that these performance gains may be due to improved spatial
locality of the `next_block` pointers. With the patches applied, the
`next_block` pointers are consistently spaced 24 bytes apart, matching the
new size of `struct dma_block`. Previously, the spacing between
`next_block` pointers depended on the block size, so for 1024-byte blocks,
the pointers were spaced 1024 bytes apart. However, I am still unsure why
the performance improvement for 68-byte blocks is so significant.

[0] Link: https://www.csl.sri.com/~neumann/ndss-iommu.pdf

Brian Johannesmeyer (2):
  dmapool: Move pool metadata into non-DMA memory
  dmapool: Use pool_find_block() in pool_block_err()

 mm/dmapool.c | 96 ++++++++++++++++++++++++++++++++++------------------
 1 file changed, 63 insertions(+), 33 deletions(-)

-- 
2.34.1

On Tue, Nov 19, 2024 at 09:55:27PM +0100, Brian Johannesmeyer wrote:
> We discovered a security-related issue in the DMA pool allocator.
> 
> V1 of our RFC was submitted to the Linux kernel security team. They
> recommended submitting it to the relevant subsystem maintainers and the
> hardening mailing list instead, as they did not consider this an explicit
> security issue. Their rationale was that Linux implicitly assumes hardware
> can be trusted.

You should probably Cc Keith as the person who most recently did major
work on the dmpool code and might still remember how it works.

> 
> **Threat Model**: While Linux drivers typically trust their hardware, there
> may be specific drivers that do not operate under this assumption. Hence,
> this threat model assumes a malicious peripheral device capable of
> corrupting DMA data to exploit the kernel. In this scenario, the device
> manipulates kernel-initialized data (similar to the attack described in the
> Thunderclap paper [0]) to achieve arbitrary kernel memory corruption. 
> 
> **DMA pool background**. A DMA pool aims to reduce the overhead of DMA
> allocations by creating a large DMA buffer --- the "pool" --- from which
> smaller buffers are allocated as needed. Fundamentally, a DMA pool
> functions like a heap: it is a structure composed of linked memory
> "blocks", which, in this context, are DMA buffers. When a driver employs a
> DMA pool, it grants the device access not only to these blocks but also to
> the pointers linking them.
> 
> **Vulnerability**. Similar to traditional heap corruption vulnerabilities
> --- where a malicious program corrupts heap metadata to e.g., hijack
> control flow --- a malicious device may corrupt DMA pool metadata. This
> corruption can trivially lead to arbitrary kernel memory corruption from
> any driver that uses it. Indeed, because the DMA pool API is extensively
> used, this vulnerability is not confined to a single instance. In fact,
> every usage of the DMA pool API is potentially vulnerable. An exploit
> proceeds with the following steps:
> 
> 1. The DMA `pool` initializes its list of blocks, then points to the first
> block.
> 2. The malicious device overwrites the first 8 bytes of the first block ---
> which contain its `next_block` pointer --- to an arbitrary kernel address,
> `kernel_addr`.
> 3. The driver makes its first call to `dma_pool_alloc()`, after which, the
> pool should point to the second block. However, it instead points to
> `kernel_addr`.
> 4. The driver again calls `dma_pool_alloc()`, which incorrectly returns
> `kernel_addr`. Therefore, anytime the driver writes to this "block", it may
> corrupt sensitive kernel data.
> 
> I have a PDF document that illustrates how these steps work. Please let me
> know if you would like me to share it with you.
> 
> **Proposed mitigation**. To mitigate the corruption of DMA pool metadata
> (i.e., the pointers linking the blocks), the metadata should be moved into
> non-DMA memory, ensuring it cannot be altered by a device. I have included
> a patch series that implements this change. Since I am not deeply familiar
> with the DMA pool internals, I would appreciate any feedback on the
> patches. I have tested the patches with the `DMAPOOL_TEST` test and my own
> basic unit tests that ensure the DMA pool allocator is not vulnerable.
> 
> **Performance**. I evaluated the patch set's performance by running the
> `DMAPOOL_TEST` test with `DMAPOOL_DEBUG` enabled and with/without the
> patches applied. Here is its output *without* the patches applied:
> ```
> dmapool test: size:16   align:16   blocks:8192 time:3194110
> dmapool test: size:64   align:64   blocks:8192 time:4730440
> dmapool test: size:256  align:256  blocks:8192 time:5489630
> dmapool test: size:1024 align:1024 blocks:2048 time:517150
> dmapool test: size:4096 align:4096 blocks:1024 time:399616
> dmapool test: size:68   align:32   blocks:8192 time:6156527
> ```
> 
> And here is its output *with* the patches applied:
> ```
> dmapool test: size:16   align:16   blocks:8192 time:3541031
> dmapool test: size:64   align:64   blocks:8192 time:4227262
> dmapool test: size:256  align:256  blocks:8192 time:4890273
> dmapool test: size:1024 align:1024 blocks:2048 time:515775
> dmapool test: size:4096 align:4096 blocks:1024 time:523096
> dmapool test: size:68   align:32   blocks:8192 time:3450830
> ```
> 
> Based on my interpretation of the output, the patch set does not appear to
> negatively impact performance. In fact, it shows slight performance
> improvements in some tests (i.e., for sizes 64, 256, 1024, and 68).
> 
> I speculate that these performance gains may be due to improved spatial
> locality of the `next_block` pointers. With the patches applied, the
> `next_block` pointers are consistently spaced 24 bytes apart, matching the
> new size of `struct dma_block`. Previously, the spacing between
> `next_block` pointers depended on the block size, so for 1024-byte blocks,
> the pointers were spaced 1024 bytes apart. However, I am still unsure why
> the performance improvement for 68-byte blocks is so significant.
> 
> [0] Link: https://www.csl.sri.com/~neumann/ndss-iommu.pdf
> 
> Brian Johannesmeyer (2):
>   dmapool: Move pool metadata into non-DMA memory
>   dmapool: Use pool_find_block() in pool_block_err()
> 
>  mm/dmapool.c | 96 ++++++++++++++++++++++++++++++++++------------------
>  1 file changed, 63 insertions(+), 33 deletions(-)
> 
> -- 
> 2.34.1
> 
> 
---end quoted text---

On Wed, Nov 20, 2024 at 01:29:19AM -0800, Christoph Hellwig wrote:
> On Tue, Nov 19, 2024 at 09:55:27PM +0100, Brian Johannesmeyer wrote:
> > **Performance**. I evaluated the patch set's performance by running the
> > `DMAPOOL_TEST` test with `DMAPOOL_DEBUG` enabled and with/without the
> > patches applied. Here is its output *without* the patches applied:

Could you rerun your tests without DMAPOOL_DEBUG enabled? That's the
more interesting kernel setup for performance comparisions.

> You should probably Cc Keith as the person who most recently did major
> work on the dmpool code and might still remember how it works.

Thank you for adding him, and apologies for not including him initially.

> The intrusive list was overlayed in the freed blocks for spatial
> optimizations. If you're moving these field outside of it (I'll have to
> review the patch on lore), you can probably relax the minimum dma block
> size too since we don't need to hold the data structure information in
> it.

I see. AFAICT, relaxing the minimum DMA block size would just mean
removing these lines from `dma_pool_create()`:
```
    if (size < sizeof(struct dma_block))
        size = sizeof(struct dma_block);
```

> Could you rerun your tests without DMAPOOL_DEBUG enabled? That's the
> more interesting kernel setup for performance comparisions.

Sure, that makes sense. Here are the results with DMAPOOL_DEBUG disabled:

**Without the patches applied:**
```
dmapool test: size:16   align:16   blocks:8192 time:11860
dmapool test: size:64   align:64   blocks:8192 time:11951
dmapool test: size:256  align:256  blocks:8192 time:12287
dmapool test: size:1024 align:1024 blocks:2048 time:3134
dmapool test: size:4096 align:4096 blocks:1024 time:1686
dmapool test: size:68   align:32   blocks:8192 time:12050
```

**With the patches applied:**
```
dmapool test: size:16   align:16   blocks:8192 time:34432
dmapool test: size:64   align:64   blocks:8192 time:62262
dmapool test: size:256  align:256  blocks:8192 time:238137
dmapool test: size:1024 align:1024 blocks:2048 time:61386
dmapool test: size:4096 align:4096 blocks:1024 time:75342
dmapool test: size:68   align:32   blocks:8192 time:88243
```

These results are consistent across multiple runs. It seems that with
DMAPOOL_DEBUG disabled, the patches introduce a significant
performance hit. Let me know if you have any suggestions or further
tests you'd like me to run.

Thanks,

Brian Johannesmeyer

On Wed, Nov 20, 2024 at 02:58:54PM -0700, Brian Johannesmeyer wrote:
> These results are consistent across multiple runs. It seems that with
> DMAPOOL_DEBUG disabled, the patches introduce a significant
> performance hit. Let me know if you have any suggestions or further
> tests you'd like me to run.

That's what I was afraid of. I was working on the dma pool because it
showed significant lock contention on the pool for storage heavy
workloads, so cutting down the critical section was priority. With the
current kernel, the dma pool doesn't even register on the profiles
anymore, so it'd be great to keep it that way.

The idea for embedding the links in freed blocks was assuming a driver
wouldn't ask the kernel to free a dma block if the mapped device was
still using it. Untrustworthy hardware is why we can't have nice
things...

Here's my quick thoughts at this late hour, though I might have
something else tomorrow. If the links are external to the dma addr
being freed, then I think you need to change the entire alloc/free API
to replace the dma_addr_t handle with a new type, like a tuple of

  { dma_addr_t, priv_list_link }

That should make it possible to preserve the low constant time to alloc
and free in the critical section, which I think is a good thing to have.
I found 170 uses of dma_pool_alloc, and 360 dma_pool_free in the kernel,
so changing the API is no small task. :(

> Here's my quick thoughts at this late hour, though I might have
> something else tomorrow. If the links are external to the dma addr
> being freed, then I think you need to change the entire alloc/free API
> to replace the dma_addr_t handle with a new type, like a tuple of
>
>   { dma_addr_t, priv_list_link }
>
> That should make it possible to preserve the low constant time to alloc
> and free in the critical section, which I think is a good thing to have.
> I found 170 uses of dma_pool_alloc, and 360 dma_pool_free in the kernel,
> so changing the API is no small task. :(

Indeed, an API change like this might be the only way to move metadata
out of the DMA blocks while maintaining its current performance.

Regarding the performance hit of the submitted patches:
- *Allocations* remain constant time (`O(1)`), as they simply check
the `pool->next_block` pointer.
- *Frees* are no longer constant time. Previously, `dma_pool_free()`
converted a `vaddr` to its corresponding `block` by type casting, but
now it calls `pool_find_block()`, which iterates over all pages
(`O(n)`).

Therefore, optimizing `dma_pool_free()` is key. While restructuring
the API is the "best" solution, I implemented a compromise:
introducing a `struct maple_tree block_map` field in `struct dma_pool`
to save mappings of a `vaddr` to its corresponding `block`. A maple
tree isn’t constant time, but it offers `O(log n)` performance, which
is a significant improvement over the current `O(n)` iteration.

Here are the performance results. I've already reported the first two
sets of numbers, but I'll repeat them here:

**Without no patches applied:**
```
dmapool test: size:16   align:16   blocks:8192 time:11860
dmapool test: size:64   align:64   blocks:8192 time:11951
dmapool test: size:256  align:256  blocks:8192 time:12287
dmapool test: size:1024 align:1024 blocks:2048 time:3134
dmapool test: size:4096 align:4096 blocks:1024 time:1686
dmapool test: size:68   align:32   blocks:8192 time:12050
```

**With the submitted patches applied:**
```
dmapool test: size:16   align:16   blocks:8192 time:34432
dmapool test: size:64   align:64   blocks:8192 time:62262
dmapool test: size:256  align:256  blocks:8192 time:238137
dmapool test: size:1024 align:1024 blocks:2048 time:61386
dmapool test: size:4096 align:4096 blocks:1024 time:75342
dmapool test: size:68   align:32   blocks:8192 time:88243
```

**With the submitted patches applied AND using a maple tree to improve
the performance of vaddr-to-block translations:**
```
dmapool test: size:16   align:16   blocks:8192 time:43668
dmapool test: size:64   align:64   blocks:8192 time:44746
dmapool test: size:256  align:256  blocks:8192 time:45434
dmapool test: size:1024 align:1024 blocks:2048 time:11013
dmapool test: size:4096 align:4096 blocks:1024 time:5250
dmapool test: size:68   align:32   blocks:8192 time:45900
```

The maple tree optimization reduces the performance hit for most block
sizes, especially for larger blocks. While the performance is not
fully back to baseline, it gives a reasonable trade-off between
protection, runtime performance, and ease of deployment (i.e., not
requiring an API change).

If this approach looks acceptable, I can submit it as a V3 patch
series for further review and discussion.

Thanks,

Brian Johannesmeyer

On Thu, Nov 21, 2024 at 10:31:11AM -0700, Brian Johannesmeyer wrote:
> **With the submitted patches applied AND using a maple tree to improve
> the performance of vaddr-to-block translations:**

If you have the time, could you compare with using xarray instead?

On Thu, Nov 21, 2024 at 11:06 AM Keith Busch <kbusch@kernel.org> wrote:
> If you have the time, could you compare with using xarray instead?

Sure. Good idea.

**With the submitted patches applied AND using an xarray for
vaddr-to-block translations:**
```
dmapool test: size:16   align:16   blocks:8192 time:37954
dmapool test: size:64   align:64   blocks:8192 time:40036
dmapool test: size:256  align:256  blocks:8192 time:41942
dmapool test: size:1024 align:1024 blocks:2048 time:10964
dmapool test: size:4096 align:4096 blocks:1024 time:6101
dmapool test: size:68   align:32   blocks:8192 time:41307
```

The xarray approach shows a slight improvement in performance compared
to the maple tree approach.

FWIW, I implemented the two with slightly different semantics:
- In the maple tree implementation, I saved the `block`'s entire
`vaddr` range, allowing any `vaddr` within the `block` to be passed to
`dma_pool_free()`.
- In the xarray implementation, I saved only the `block's` base
`vaddr`, requiring `dma_pool_free()` to be called with the exact
`vaddr` returned by `dma_pool_alloc()`. This aligns with the DMA pool
API documentation, which specifies that the `vaddr` returned by
`dma_pool_alloc()` should be passed to `dma_pool_free()`.

Let me know if you'd like further adjustments.

Thanks,

Brian Johannesmeyer

I’ll go ahead and prepare a V3 patch series with the following updates:

- Using an xarray for vaddr-to-block translations, which improves the
performance of free operations.
- Removing the minimum DMA block size constraint, as it is no longer necessary.

Let me know if there are any additional suggestions or concerns to
address before submission.

Thanks,

Brian

On Thu, Nov 21, 2024 at 12:07 PM Brian Johannesmeyer
<bjohannesmeyer@gmail.com> wrote:
>
> On Thu, Nov 21, 2024 at 11:06 AM Keith Busch <kbusch@kernel.org> wrote:
> > If you have the time, could you compare with using xarray instead?
>
> Sure. Good idea.
>
> **With the submitted patches applied AND using an xarray for
> vaddr-to-block translations:**
> ```
> dmapool test: size:16   align:16   blocks:8192 time:37954
> dmapool test: size:64   align:64   blocks:8192 time:40036
> dmapool test: size:256  align:256  blocks:8192 time:41942
> dmapool test: size:1024 align:1024 blocks:2048 time:10964
> dmapool test: size:4096 align:4096 blocks:1024 time:6101
> dmapool test: size:68   align:32   blocks:8192 time:41307
> ```
>
> The xarray approach shows a slight improvement in performance compared
> to the maple tree approach.
>
> FWIW, I implemented the two with slightly different semantics:
> - In the maple tree implementation, I saved the `block`'s entire
> `vaddr` range, allowing any `vaddr` within the `block` to be passed to
> `dma_pool_free()`.
> - In the xarray implementation, I saved only the `block's` base
> `vaddr`, requiring `dma_pool_free()` to be called with the exact
> `vaddr` returned by `dma_pool_alloc()`. This aligns with the DMA pool
> API documentation, which specifies that the `vaddr` returned by
> `dma_pool_alloc()` should be passed to `dma_pool_free()`.
>
> Let me know if you'd like further adjustments.
>
> Thanks,
>
> Brian Johannesmeyer

On Wed, Nov 20, 2024 at 01:29:19AM -0800, Christoph Hellwig wrote:
> On Tue, Nov 19, 2024 at 09:55:27PM +0100, Brian Johannesmeyer wrote:
> > We discovered a security-related issue in the DMA pool allocator.
> > 
> > V1 of our RFC was submitted to the Linux kernel security team. They
> > recommended submitting it to the relevant subsystem maintainers and the
> > hardening mailing list instead, as they did not consider this an explicit
> > security issue. Their rationale was that Linux implicitly assumes hardware
> > can be trusted.
> 
> You should probably Cc Keith as the person who most recently did major
> work on the dmpool code and might still remember how it works.

Thanks.

The intrusive list was overlayed in the freed blocks for spatial
optimizations. If you're moving these field outside of it (I'll have to
review the patch on lore), you can probably relax the minimum dma block
size too since we don't need to hold the data structure information in
it.
 
> > **Threat Model**: While Linux drivers typically trust their hardware, there
> > may be specific drivers that do not operate under this assumption. Hence,
> > this threat model assumes a malicious peripheral device capable of
> > corrupting DMA data to exploit the kernel. In this scenario, the device
> > manipulates kernel-initialized data (similar to the attack described in the
> > Thunderclap paper [0]) to achieve arbitrary kernel memory corruption. 
> > 
> > **DMA pool background**. A DMA pool aims to reduce the overhead of DMA
> > allocations by creating a large DMA buffer --- the "pool" --- from which
> > smaller buffers are allocated as needed. Fundamentally, a DMA pool
> > functions like a heap: it is a structure composed of linked memory
> > "blocks", which, in this context, are DMA buffers. When a driver employs a
> > DMA pool, it grants the device access not only to these blocks but also to
> > the pointers linking them.
> > 
> > **Vulnerability**. Similar to traditional heap corruption vulnerabilities
> > --- where a malicious program corrupts heap metadata to e.g., hijack
> > control flow --- a malicious device may corrupt DMA pool metadata. This
> > corruption can trivially lead to arbitrary kernel memory corruption from
> > any driver that uses it. Indeed, because the DMA pool API is extensively
> > used, this vulnerability is not confined to a single instance. In fact,
> > every usage of the DMA pool API is potentially vulnerable. An exploit
> > proceeds with the following steps:
> > 
> > 1. The DMA `pool` initializes its list of blocks, then points to the first
> > block.
> > 2. The malicious device overwrites the first 8 bytes of the first block ---
> > which contain its `next_block` pointer --- to an arbitrary kernel address,
> > `kernel_addr`.
> > 3. The driver makes its first call to `dma_pool_alloc()`, after which, the
> > pool should point to the second block. However, it instead points to
> > `kernel_addr`.
> > 4. The driver again calls `dma_pool_alloc()`, which incorrectly returns
> > `kernel_addr`. Therefore, anytime the driver writes to this "block", it may
> > corrupt sensitive kernel data.
> > 
> > I have a PDF document that illustrates how these steps work. Please let me
> > know if you would like me to share it with you.
> > 
> > **Proposed mitigation**. To mitigate the corruption of DMA pool metadata
> > (i.e., the pointers linking the blocks), the metadata should be moved into
> > non-DMA memory, ensuring it cannot be altered by a device. I have included
> > a patch series that implements this change. Since I am not deeply familiar
> > with the DMA pool internals, I would appreciate any feedback on the
> > patches. I have tested the patches with the `DMAPOOL_TEST` test and my own
> > basic unit tests that ensure the DMA pool allocator is not vulnerable.
> > 
> > **Performance**. I evaluated the patch set's performance by running the
> > `DMAPOOL_TEST` test with `DMAPOOL_DEBUG` enabled and with/without the
> > patches applied. Here is its output *without* the patches applied:
> > ```
> > dmapool test: size:16   align:16   blocks:8192 time:3194110
> > dmapool test: size:64   align:64   blocks:8192 time:4730440
> > dmapool test: size:256  align:256  blocks:8192 time:5489630
> > dmapool test: size:1024 align:1024 blocks:2048 time:517150
> > dmapool test: size:4096 align:4096 blocks:1024 time:399616
> > dmapool test: size:68   align:32   blocks:8192 time:6156527
> > ```
> > 
> > And here is its output *with* the patches applied:
> > ```
> > dmapool test: size:16   align:16   blocks:8192 time:3541031
> > dmapool test: size:64   align:64   blocks:8192 time:4227262
> > dmapool test: size:256  align:256  blocks:8192 time:4890273
> > dmapool test: size:1024 align:1024 blocks:2048 time:515775
> > dmapool test: size:4096 align:4096 blocks:1024 time:523096
> > dmapool test: size:68   align:32   blocks:8192 time:3450830
> > ```
> > 
> > Based on my interpretation of the output, the patch set does not appear to
> > negatively impact performance. In fact, it shows slight performance
> > improvements in some tests (i.e., for sizes 64, 256, 1024, and 68).
> > 
> > I speculate that these performance gains may be due to improved spatial
> > locality of the `next_block` pointers. With the patches applied, the
> > `next_block` pointers are consistently spaced 24 bytes apart, matching the
> > new size of `struct dma_block`. Previously, the spacing between
> > `next_block` pointers depended on the block size, so for 1024-byte blocks,
> > the pointers were spaced 1024 bytes apart. However, I am still unsure why
> > the performance improvement for 68-byte blocks is so significant.
> > 
> > [0] Link: https://www.csl.sri.com/~neumann/ndss-iommu.pdf
> > 
> > Brian Johannesmeyer (2):
> >   dmapool: Move pool metadata into non-DMA memory
> >   dmapool: Use pool_find_block() in pool_block_err()
> > 
> >  mm/dmapool.c | 96 ++++++++++++++++++++++++++++++++++------------------

On Tue, Nov 19, 2024 at 09:55:27PM +0100, Brian Johannesmeyer wrote:
> We discovered a security-related issue in the DMA pool allocator.
> 
> V1 of our RFC was submitted to the Linux kernel security team. They
> recommended submitting it to the relevant subsystem maintainers and the
> hardening mailing list instead, as they did not consider this an explicit
> security issue. Their rationale was that Linux implicitly assumes hardware
> can be trusted.
> 
> **Threat Model**: While Linux drivers typically trust their hardware, there
> may be specific drivers that do not operate under this assumption. Hence,
> this threat model assumes a malicious peripheral device capable of
> corrupting DMA data to exploit the kernel. In this scenario, the device
> manipulates kernel-initialized data (similar to the attack described in the
> Thunderclap paper [0]) to achieve arbitrary kernel memory corruption. 
> 
> **DMA pool background**. A DMA pool aims to reduce the overhead of DMA
> allocations by creating a large DMA buffer --- the "pool" --- from which
> smaller buffers are allocated as needed. Fundamentally, a DMA pool
> functions like a heap: it is a structure composed of linked memory
> "blocks", which, in this context, are DMA buffers. When a driver employs a
> DMA pool, it grants the device access not only to these blocks but also to
> the pointers linking them.
> 
> **Vulnerability**. Similar to traditional heap corruption vulnerabilities
> --- where a malicious program corrupts heap metadata to e.g., hijack
> control flow --- a malicious device may corrupt DMA pool metadata. This
> corruption can trivially lead to arbitrary kernel memory corruption from
> any driver that uses it. Indeed, because the DMA pool API is extensively
> used, this vulnerability is not confined to a single instance. In fact,
> every usage of the DMA pool API is potentially vulnerable. An exploit
> proceeds with the following steps:
> 
> 1. The DMA `pool` initializes its list of blocks, then points to the first
> block.
> 2. The malicious device overwrites the first 8 bytes of the first block ---
> which contain its `next_block` pointer --- to an arbitrary kernel address,
> `kernel_addr`.
> 3. The driver makes its first call to `dma_pool_alloc()`, after which, the
> pool should point to the second block. However, it instead points to
> `kernel_addr`.
> 4. The driver again calls `dma_pool_alloc()`, which incorrectly returns
> `kernel_addr`. Therefore, anytime the driver writes to this "block", it may
> corrupt sensitive kernel data.
> 
> I have a PDF document that illustrates how these steps work. Please let me
> know if you would like me to share it with you.

I know I said it privately, but I'll say it here in public, very cool
finding, this is nice work!

> **Proposed mitigation**. To mitigate the corruption of DMA pool metadata
> (i.e., the pointers linking the blocks), the metadata should be moved into
> non-DMA memory, ensuring it cannot be altered by a device. I have included
> a patch series that implements this change. Since I am not deeply familiar
> with the DMA pool internals, I would appreciate any feedback on the
> patches. I have tested the patches with the `DMAPOOL_TEST` test and my own
> basic unit tests that ensure the DMA pool allocator is not vulnerable.
> 
> **Performance**. I evaluated the patch set's performance by running the
> `DMAPOOL_TEST` test with `DMAPOOL_DEBUG` enabled and with/without the
> patches applied. Here is its output *without* the patches applied:
> ```
> dmapool test: size:16   align:16   blocks:8192 time:3194110
> dmapool test: size:64   align:64   blocks:8192 time:4730440
> dmapool test: size:256  align:256  blocks:8192 time:5489630
> dmapool test: size:1024 align:1024 blocks:2048 time:517150
> dmapool test: size:4096 align:4096 blocks:1024 time:399616
> dmapool test: size:68   align:32   blocks:8192 time:6156527
> ```
> 
> And here is its output *with* the patches applied:
> ```
> dmapool test: size:16   align:16   blocks:8192 time:3541031
> dmapool test: size:64   align:64   blocks:8192 time:4227262
> dmapool test: size:256  align:256  blocks:8192 time:4890273
> dmapool test: size:1024 align:1024 blocks:2048 time:515775
> dmapool test: size:4096 align:4096 blocks:1024 time:523096
> dmapool test: size:68   align:32   blocks:8192 time:3450830
> ```

You had mentioned that the size:68 numbers were going to be re-run, has
that happened and this really is that much of a boost to that size?  Or
is this the original numbers?

thanks,

greg k-h

On Tue, Nov 19, 2024 at 3:15 PM Greg KH <gregkh@linuxfoundation.org> wrote:
> I know I said it privately, but I'll say it here in public, very cool
> finding, this is nice work!

Thanks! I appreciate your earlier feedback as well.

> You had mentioned that the size:68 numbers were going to be re-run, has
> that happened and this really is that much of a boost to that size?  Or
> is this the original numbers?

I re-ran the test, and the numbers are consistent across multiple
runs. I’m also surprised by how significant the improvement is for the
68-byte block size.

Thanks,

Brian Johannesmeyer