IRC channel logs
2024-01-26.log
back to list of logs
<oriansj>(sorry mean to express less likely for the disk read/write block to not be in memory than a 2 page buffer) but yes that block miss would only occur a couple times max per stage0 step. <muurkha>oriansj: agreed, though at some point you have to set ebx to the buffer you pick, and with three or mour buffers I think you might as well write a loop (mov ecx, nbufs; search: cmp eax, [bufbase + ecx]; je found; loop search; mov ecx, [nextbuf]; inc ecx; and ecx, nbufs-1; mov [nextbuf], ecx; found:) 8 instructions instead of 10 and accommodates any power-of-2 number of buffers, though with enough <muurkha>of them it gets slower to check the buffer cache than just loading the page <lrvick>it isn't obvious to me how this could cause a determinisim bug, but it is also the only related change in a couple years as far as I can tell <lrvick>I'll apply the patch and see wht happens <oriansj>muurkha: fair point but one only needs 2 blocks per file buffer <oriansj>only if one is providing one shared buffer for multiple processes talking to multiple files would one need something much more carefully designed.