Skip to contentSkip to conversation

drm/amd/display: Fix missing HF-EEODB blocks in EDID copies

PATCH In review Timo Prömer 4 participants 9 messages Jul 13 – Sep 18

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

v2 cover drm/amd/display: Fix missing HF-EEODB blocks in EDID copies

Fix an issue in amd/display where devices with HF-EEODB blocks would be missing these additional blocks during EDID reads.

The driver previously used edid->extensions + 1 to calculate the number of blocks to copy, but the base extension flag does not include HF-EEODB blocks.

Use drm_edid_block_count() directly to get the true number of blocks, ensuring that HF-EEODB blocks are properly copied.

Changes in v2:
- Fixed a bug in patch 2 where the bounds check still used the old extension count.
- Patch 1 remains unchanged.

Timo Prömer (2):
drm/edid: Export drm_edid_block_count()
drm/amd/display: Use drm_edid_block_count() instead of raw extensions

 .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c | 11 ++++++++---
 drivers/gpu/drm/drm_edid.c                            |  3 ++-
 include/drm/drm_edid.h                                |  1 +
 3 files changed, 11 insertions(+), 4 deletions(-)
Signature · 1 line
2.55.0
#1 lore raw

PATCH v2 1/2 drm/edid: Export drm_edid_block_count()

From: Timoyoungster <[email protected]>

Drivers currently calculating EDID size by reading the extensions field of the raw EDID structure (e.g., edid->extensions + 1) will calculate the wrong size if the EDID contains an HF-EEODB (HDMI Forum EDID Extension Override Data Block). The base extension flag does not account for these override blocks, leading to truncated EDIDs.

Remove the static declaration and export drm_edid_block_count() so drivers can safely query the true block count. This allows drivers to leverage the core DRM's proper handling of HF-EEODB and other edge cases without having to parse the raw EDID fields themselves.

---
 drivers/gpu/drm/drm_edid.c | 3 ++-
 include/drm/drm_edid.h     | 1 +
 2 files changed, 3 insertions(+), 1 deletion(-)
drivers/gpu/drm/drm_edid.c +2 -1
index df3c25bac..34560b33a 100644
@@ -1698,12 +1698,13 @@ static int __drm_edid_block_count(const struct drm_edid *drm_edid)
 }
 
 /* EDID block count, limited by allocated size */
-static int drm_edid_block_count(const struct drm_edid *drm_edid)
+int drm_edid_block_count(const struct drm_edid *drm_edid)
 {
 	/* Limit by allocated size */
 	return min(__drm_edid_block_count(drm_edid),
 		   (int)drm_edid->size / EDID_LENGTH);
 }
+EXPORT_SYMBOL(drm_edid_block_count);
 
 /* EDID extension block count, limited by allocated size */
 static int drm_edid_extension_block_count(const struct drm_edid *drm_edid)
include/drm/drm_edid.h +1 -0
index 04f7a7f1f..4a990bf87 100644
@@ -481,6 +481,7 @@ const struct drm_edid *drm_edid_read_switcheroo(struct drm_connector *connector,
 int drm_edid_connector_update(struct drm_connector *connector,
 			      const struct drm_edid *edid);
 int drm_edid_connector_add_modes(struct drm_connector *connector);
+int drm_edid_block_count(const struct drm_edid *drm_edid);
 bool drm_edid_is_digital(const struct drm_edid *drm_edid);
 void drm_edid_get_product_id(const struct drm_edid *drm_edid,
 			     struct drm_edid_product_id *id);
Signature · 1 line
2.55.0
#3 lore raw

PATCH v2 2/2 drm/amd/display: Use drm_edid_block_count() instead of raw extensions

In reply to Timo Prömer

From: Timoyoungster <[email protected]>

Instead of manually calculating the EDID block count by reading the extensions field from the raw edid structure (edid->extensions + 1), utilize the core DRM helper drm_edid_block_count().

This now includes possible HF-EEODB extension blocks, which are not included in edid->extensions and were previously truncated with the memmove.

---
 .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c +8 -3
index c6f94eb71..d6d6ea719 100644
@@ -1160,6 +1160,7 @@ enum dc_edid_status dm_helpers_read_local_edid(
 	struct drm_connector *connector = &aconnector->base;
 	struct i2c_adapter *ddc;
 	int retry = 25;
+	int block_count;
 	enum dc_edid_status edid_status = EDID_NO_RESPONSE;
 	const struct drm_edid *drm_edid;
 	const struct edid *edid;
@@ -1201,11 +1202,15 @@ enum dc_edid_status dm_helpers_read_local_edid(
 			continue;
 
 		edid = drm_edid_raw(drm_edid); // FIXME: Get rid of drm_edid_raw()
-		if (!edid ||
-		    edid->extensions >= sizeof(sink->dc_edid.raw_edid) / EDID_LENGTH)
+		if (!edid)
 			return EDID_BAD_INPUT;
 
-		sink->dc_edid.length = EDID_LENGTH * (edid->extensions + 1);
+		block_count = drm_edid_block_count(drm_edid);
+
+		if (block_count > sizeof(sink->dc_edid.raw_edid) / EDID_LENGTH)
+			return EDID_BAD_INPUT;
+
+		sink->dc_edid.length = EDID_LENGTH * block_count;
 		memmove(sink->dc_edid.raw_edid, (uint8_t *)edid, sink->dc_edid.length);
 
 		/* We don't need the original edid anymore */
Signature · 1 line
2.55.0
#4 lore raw

Re: [PATCH v2 2/2] drm/amd/display: Use drm_edid_block_count() instead of raw extensions

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [High] Kernel heap out-of-bounds read due to mismatch between HF-EEODB block count and raw edid->extensions byte.

Pre-existing issues:
- [Critical] Heap buffer overflow in amdgpu_dm_connector_funcs_force() when copying EDID. --

--- Patch [2]: [PATCH v2 2/2] drm/amd/display: Use drm_edid_block_count() instead of raw extensions --- commit bb5e357c1182326e11f3ce4e756df91c3df854a6
Author: Timoyoungster <[email protected]>

drm/amd/display: Use drm_edid_block_count() instead of raw extensions

Instead of manually calculating the EDID block count by reading the extensions field from the raw edid structure (edid->extensions + 1), utilize the core DRM helper drm_edid_block_count().

This now includes possible HF-EEODB extension blocks, which are not included in edid->extensions and were previously truncated with the memmove.

Quoted 11 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c +1 -2
index 71e2627f9a9d4..82f3cce870a3c 100644
@@ -1220,11 +1221,15 @@ enum dc_edid_status dm_helpers_read_local_edid(
 			continue;
 
 		edid = drm_edid_raw(drm_edid); // FIXME: Get rid of drm_edid_raw()
-		if (!edid ||
-		    edid->extensions >= sizeof(sink->dc_edid.raw_edid) / EDID_LENGTH)
+		if (!edid)

drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c enum dc_edid_status dm_helpers_read_local_edid( · line 1233

Quoted 10 lines
drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_helpers.c +6 -1
 			return EDID_BAD_INPUT;
 
-		sink->dc_edid.length = EDID_LENGTH * (edid->extensions + 1);
+		block_count = drm_edid_block_count(drm_edid);
+
+		if (block_count > sizeof(sink->dc_edid.raw_edid) / EDID_LENGTH)
+			return EDID_BAD_INPUT;
+
+		sink->dc_edid.length = EDID_LENGTH * block_count;
 		memmove(sink->dc_edid.raw_edid, (uint8_t *)edid, sink->dc_edid.length);

[Severity: High]
Does this create an out-of-bounds read when the copied EDID is parsed downstream?

If a malicious display provides an EDID with an extension count of 255 but an HF-EEODB block specifying a count of 2, drm_edid_block_count() will return 2. The bounds check passes, and memmove() copies 256 bytes into raw_edid.

However, the raw extensions byte in the copied data remains 255. When this is later parsed in dm_helpers_parse_edid_caps() which delegates to drm_edid_is_valid(), doesn't that legacy helper use the unmodified extensions byte to iterate up to 256 times, reading past the end of the 2048-byte raw_edid array on the kernel heap?

Should the raw extensions byte in the copied buffer be manually updated to match block_count - 1, preventing legacy functions from performing out-of-bounds reads?

Also, looking at related EDID copying code in the same subsystem:

[Severity: Critical]
This isn't a bug introduced by this patch, but does the following code in amdgpu_dm_connector.c:amdgpu_dm_connector_funcs_force() have a heap buffer overflow?

amdgpu_dm_connector_funcs_force() {
    ...
    memmove(dc_em_sink->dc_edid.raw_edid, edid, (edid->extensions + 1) * EDID_LENGTH);
    ...
}

If a malicious display or I2C emulator provides an EDID with extensions set to 255 and valid block checksums, and userspace with DRM privileges forces a connector status update, it appears this could copy up to 32KB into the 2KB dc_em_sink->dc_edid.raw_edid buffer, corrupting the heap.

Is a bounds check needed here similar to the one performed in dm_helpers_read_local_edid()?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
#5 lore raw

Re: [PATCH v2 2/2] drm/amd/display: Use drm_edid_block_count() instead of raw extensions

Hi Timo,

Thanks for the patch. Heads-up that this overlaps with a fix already in amd-staging-drm-next:

21590e1adc9e ("drm/amd/display: Fix 8K Mode Not Parsed by EDID")

That commit addresses the same underlying problem your patch targets -- edid->extensions (byte 0x7e) not accounting for HF-EEODB blocks, so only 2 of 4 blocks of an HDMI 2.1 EDID get copied into sink->dc_edid and the 8K modes in the DisplayID extension blocks are lost. It takes a slightly different approach: instead of drm_edid_block_count(), it populates edid_blob_ptr with a full, HF-EEODB-aware blob and copies edid_blob_ptr->length. So on current amd-staging the dm_helpers_read_local_edid() hunk here no longer applies cleanly.

Could you rebase on the latest amd-staging-drm-next and see what, if anything, is still needed on top? There may still be value in the helper conversion for readability, but we should avoid re-fixing the same HF-EEODB issue two different ways.

Thanks,
Jerry

#6 lore raw

Re: [PATCH v2 2/2] drm/amd/display: Use drm_edid_block_count() instead of raw extensions

Hi Jerry,

Thank you for the response.

I see that the patch you referenced was already released with rc4. I tested my use case with that release and the EDID is read correctly now, so I believe my patch can be dropped.

This is my very first interaction with the kernel process, so I am sure the released patch handles things much better than my attempt anyway!

Thank you for your time,
Timo

#2 lore raw

Re: [PATCH v2 1/2] drm/edid: Export drm_edid_block_count()

In reply to Timo Prömer
Quoted 25 lines · Timo Prömer

From: Timoyoungster <[email protected]>

Drivers currently calculating EDID size by reading the extensions field of the raw EDID structure (e.g., edid->extensions + 1) will calculate the wrong size if the EDID contains an HF-EEODB (HDMI Forum EDID Extension Override Data Block). The base extension flag does not account for these override blocks, leading to truncated EDIDs.

Remove the static declaration and export drm_edid_block_count() so drivers can safely query the true block count. This allows drivers to leverage the core DRM's proper handling of HF-EEODB and other edge cases without having to parse the raw EDID fields themselves.

---
 drivers/gpu/drm/drm_edid.c | 3 ++-
 include/drm/drm_edid.h     | 1 +
 2 files changed, 3 insertions(+), 1 deletion(-)
drivers/gpu/drm/drm_edid.c +0 -0
index df3c25bac..34560b33a 100644
@@ -1698,12 +1698,13 @@ static int __drm_edid_block_count(const struct drm_edid *drm_edid)
 }

drivers/gpu/drm/drm_edid.c static int __drm_edid_block_count(const struct drm_edid *drm_edid) · line 1707

Quoted 10 lines
drivers/gpu/drm/drm_edid.c +2 -1
 
 /* EDID block count, limited by allocated size */
-static int drm_edid_block_count(const struct drm_edid *drm_edid)
+int drm_edid_block_count(const struct drm_edid *drm_edid)
 {
 	/* Limit by allocated size */
 	return min(__drm_edid_block_count(drm_edid),
 		   (int)drm_edid->size / EDID_LENGTH);
 }
+EXPORT_SYMBOL(drm_edid_block_count);

Not everything that's inside drm_edid.c is supposed to be looked into. It's abstracted and hidden for a reason. Please don't hack into this.

BR,
Jani.

Quoted 15 lines
drivers/gpu/drm/drm_edid.c +0 -0
 /* EDID extension block count, limited by allocated size */
 static int drm_edid_extension_block_count(const struct drm_edid *drm_edid)
include/drm/drm_edid.h +1 -0
index 04f7a7f1f..4a990bf87 100644
@@ -481,6 +481,7 @@ const struct drm_edid *drm_edid_read_switcheroo(struct drm_connector *connector,
 int drm_edid_connector_update(struct drm_connector *connector,
 			      const struct drm_edid *edid);
 int drm_edid_connector_add_modes(struct drm_connector *connector);
+int drm_edid_block_count(const struct drm_edid *drm_edid);
 bool drm_edid_is_digital(const struct drm_edid *drm_edid);
 void drm_edid_get_product_id(const struct drm_edid *drm_edid,
 			     struct drm_edid_product_id *id);
Signature · 1 line
Jani Nikula, Intel
#7 lore raw

Re: [PATCH v2 2/2] drm/amd/display: Use drm_edid_block_count() instead of raw extensions

In reply to Fangzhi Zuo
Quoted 6 lines · Fangzhi Zuo

Hi Timo,

Thanks for the patch. Heads-up that this overlaps with a fix already in amd-staging-drm-next:

21590e1adc9e ("drm/amd/display: Fix 8K Mode Not Parsed by EDID")

More precisely commit 11a90eaf5c80 ("drm/amd/display: Fix 8K Mode Not Parsed by EDID") upstream.

Quoted 8 lines

That commit addresses the same underlying problem your patch targets -- edid->extensions (byte 0x7e) not accounting for HF-EEODB blocks, so only 2 of 4 blocks of an HDMI 2.1 EDID get copied into sink->dc_edid and the 8K modes in the DisplayID extension blocks are lost. It takes a slightly different approach: instead of drm_edid_block_count(), it populates edid_blob_ptr with a full, HF-EEODB-aware blob and copies edid_blob_ptr->length. So on current amd-staging the dm_helpers_read_local_edid() hunk here no longer applies cleanly.

That's an awful, ugly hack. Please work to remove it. Drivers are not supposed to look at or use connector->edid_blob_ptr at all. See the documentation.

I really did put in a *lot* of effort into abstracting HF-EEODB properly and neatly for everyone, across the subsystem. The opaque struct drm_edid is the abstraction. Please embrace it, and you'll get HF-EEODB for free, with no hacks.

If you face issues, please talk to me instead of hacking stuff and intentionally working around the the abstractions.

Quoted 4 lines

Could you rebase on the latest amd-staging-drm-next and see what, if anything, is still needed on top? There may still be value in the helper conversion for readability, but we should avoid re-fixing the same HF-EEODB issue two different ways.

The one and only fix for HF-EEODB is to use struct drm_edid everywhere.

BR,
Jani.

Signature · 1 line
Jani Nikula, Intel
#8 lore raw

RE: [PATCH v2 2/2] drm/amd/display: Use drm_edid_block_count() instead of raw extensions

AMD General

Quoted 46 lines

-----Original Message-----
From: Jani Nikula <[email protected]>
Sent: Friday, September 18, 2026 10:30
To: Zuo, Jerry <[email protected]>; [email protected]; Wentland, Harry <[email protected]>; Li, Sun peng (Leo) <[email protected]>; Deucher, Alexander <[email protected]>
Cc: Zuo, Jerry <[email protected]>; [email protected]; amd- [email protected]
Subject: Re: [PATCH v2 2/2] drm/amd/display: Use drm_edid_block_count() instead of raw extensions

Quoted 7 lines · Fangzhi Zuo

Hi Timo,

Thanks for the patch. Heads-up that this overlaps with a fix already in
amd-staging-drm-next:

21590e1adc9e ("drm/amd/display: Fix 8K Mode Not Parsed by EDID")

More precisely commit 11a90eaf5c80 ("drm/amd/display: Fix 8K Mode Not Parsed by EDID") upstream.

Quoted 10 lines

That commit addresses the same underlying problem your patch targets --
edid->extensions (byte 0x7e) not accounting for HF-EEODB blocks, so edid->only
2 of 4 blocks of an HDMI 2.1 EDID get copied into sink->dc_edid and the 8K modes in the DisplayID extension blocks are lost. It takes a slightly different approach: instead of drm_edid_block_count(), it populates edid_blob_ptr with a full, HF-EEODB-aware blob and copies edid_blob_ptr->length. So on current amd-staging the dm_helpers_read_local_edid() hunk here no longer applies cleanly.

That's an awful, ugly hack. Please work to remove it. Drivers are not supposed to look at or use connector->edid_blob_ptr at all. See the documentation.

I really did put in a *lot* of effort into abstracting HF-EEODB properly and neatly for everyone, across the subsystem. The opaque struct drm_edid is the abstraction. Please embrace it, and you'll get HF-EEODB for free, with no hacks.

If you face issues, please talk to me instead of hacking stuff and intentionally working around the the abstractions.

Quoted 6 lines
Quoted 4 lines

Could you rebase on the latest amd-staging-drm-next and see what, if anything, is still needed on top? There may still be value in the helper conversion for readability, but we should avoid re-fixing the same HF-EEODB issue two different ways.

The one and only fix for HF-EEODB is to use struct drm_edid everywhere.

Thanks, I totally agree. The fact is the dc_edid to drm_edid transition/refactor is not fully completed yet. That leads to w/a is coming up as a temporary fix. Will remove once the transition is complete.

Quoted 8 lines

BR,
Jani.

--
Jani Nikula, Intel