Front end hardening: bound the descriptor-chain walk, thin query and configure log lines

未關閉
#61 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
3/5
預估耗時
1-2 天
新手友好度
64/100
Issue 類型
缺陷
描述清晰度
描述清楚
活躍度
活躍
技術堆疊
cpp

研究方向

Start in clean/fsr/src/frontend.cpp: the pNext walk in CreateContext (~line 460) and CreateUpscale (~533), and the logging in Query (~1445) and Configure (~1195), which should use ShouldLogDispatch (~446). Read config.h:113 and log_win.h:137 for the log_dispatches and size rules. Done when a self-referencing chain returns FFX_API_RETURN_ERROR_PARAMETER in a CPU test, and 1200 jitter queries with log_dispatches=1 leave about 2 query lines.

由索引模型根據 Issue 內容生成。

描述

area: front-end enhancement

Two small, independent changes in clean/fsr/src/frontend.cpp. Found by reading the code; neither has been reproduced.

1. The descriptor-chain walk is not bounded

What happens

The walk over the pNext chain in CreateContext (:460) has no bound, and neither has the walk in CreateUpscale (:533). A chain that points back to itself would hang the calling thread. DescribeChain (clean/fsr/src/ffx_describe.cpp:337) stops at 16 nodes and CreateAmdWithoutPrivateVersion (frontend.cpp:192) stops at 64.

When it can happen

Only for a caller that breaks the API contract. A cyclic chain is malformed input, and AMD's runtime walks the same chains when a call is forwarded to it. No working game or host sends one.

To do

At the top of CreateContext, walk the chain once with a bound of 64 (the existing limit) and return FFX_API_RETURN_ERROR_PARAMETER if it is exceeded. CreateUpscale then only sees chains that passed.

Test

CPU. Build a two-node chain that points back to itself, call fe->CreateContext, and expect ERROR_PARAMETER. Without the change the test hangs, so the runner's timeout catches it.

2. Query and configure log lines are not thinned with log_dispatches = 1

What happens

log_dispatches is documented to thin dispatch lines to one in 600 above the log size limit (clean/fsr/src/config.h:113, clean/fsr/README.md:124). ShouldLogDispatch (frontend.cpp:446) does that. Query (:1445) and Configure (:1195) use seen < 3 || config_.logDispatches instead and never thin. A game that queries the jitter every frame keeps writing about two lines per frame after the log passes log_max_mb. The text log is only moved aside when it is opened, above 8 MiB (clean/fsr/src/log_win.h:137).

When it can happen

Only with log_dispatches = 1, a diagnostic setting that is off by default, over a long session. With the default only the first 3 calls of each type are logged.

To do

In both places, use seen < 3 || (config_.logDispatches && ShouldLogDispatch(seen)). Failed calls are still always logged. A live size cap for the text log is optional; if added, it has to coordinate with the writer's write_mutex_.

Test

CPU. Give the front end a stand-in platform_.logBytes that returns more than log_max_mb, set log_dispatches = 1, run 1200 jitter queries, and check that the log holds about 2 query lines, not 1200.

主要語言
C++
星號
12
分支
2
平均合併
3 小時 10 分鐘
30 天內合併 PR
171

環境準備

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

phantoslayer21/wronggpu 的其他 Issue

查看 phantoslayer21/wronggpu 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。