Regression, "Installing bootloader" does take about 30seconds for me now, while it used to take only about 2seconds

"Installing bootloader" does take about 30seconds for me now, while it used to take only about 2seconds. bootloader.png

I report it against 0.4.15-dev-2689-ge70df4c.

It is a regression:
reactos-bootcd-0.4.15-dev-2689-ge70df4c-x86-gcc-lin-dbg.iso slow
reactos-bootcd-0.4.15-dev-2534-geba00d1-x86-gcc-lin-dbg.iso slow
reactos-bootcd-0.4.15-dev-2460-ga825523-x86-gcc-lin-dbg.iso slow
reactos-bootcd-0.4.15-dev-2400-g0995b35-x86-gcc-lin-dbg.iso slow
reactos-bootcd-0.4.15-dev-2022-g44eecfa-x86-gcc-lin-dbg.iso still fast
ReactOS-0.4.14-RC-71-gcfc2431-iso still fast

The issue has been confirmed already by binarymaster and hbelusca and ThePhysicist.

Sidenote
I am pretty sure, that it is caused by some MM/FS-write-caching-changes, because I also noticed that the copy-performance during 1st stage flie-copy is a bit slower than before and especially less linear in progressing, I do even get some workqueue deadlocks detected:

(ntoskrnl/mm/balance.c:144) MM BALANCER: Removing access bit!

(ntoskrnl/mm/balance.c:144) MM BALANCER: Removing access bit!

(ntoskrnl/mm/balance.c:144) MM BALANCER: Removing access bit!

(ntoskrnl/mm/balance.c:144) MM BALANCER: Removing access bit!

(ntoskrnl/ex/work.c:348) EX: Work Queue Deadlock detected: 1

(ntoskrnl/ex/work.c:350) Dynamic threads queued 2

(ntoskrnl/mm/balance.c:144) MM BALANCER: Removing access bit!

(ntoskrnl/ex/work.c:348) EX: Work Queue Deadlock detected: 1

(ntoskrnl/ex/work.c:350) Dynamic threads queued 3

(ntoskrnl/mm/balance.c:144) MM BALANCER: Removing access bit!

(ntoskrnl/mm/balance.c:144) MM BALANCER: Removing access bit!

I guess, that when the bootsector low-level-access happens, then all the queued outstanding stuff from the 1st-stage-file-copying has to be persisted at once and that may cause the "hang".

Therefore the question: Can we explicitly disable that kind of caching during 1st stage copying to regain the speed when writing bootsector and also to make the copy-progress more linear again?