Hi Alex, Xiaogang, thanks for looking at this quickly.
Replying here because this patch instance carries my Tested-by tag. One clarification: my runtime testing covers the reproduced split-tail fault on the RX 7600 XT. It does not validate the head-side boundary condition in this exact patch. The remaining concern is the head-side condition: this version still uses head->last there. In svm_range_split_head(), the generated head range is [old_start, new_start - 1], so the split boundary is new_start, i.e. head->last + 1. I replied with the concrete boundary examples here: https://lore.kernel.org/all/[email protected]/ For that reason I think the head-side condition should use the exclusive split boundary, e.g. new_start as in v2: https://lore.kernel.org/all/[email protected]/ To avoid my Tested-by tag being read as covering the head-side boundary logic, please either use an exclusive-boundary version or drop my Tested-by from this exact version. My test result covers the reproduced split-tail regression, not the head-side boundary logic. Thanks, Gerhard
