Changeset 104858 in vbox for trunk/src/VBox/VMM/include

Timestamp:

Jun 5, 2024 6:10:20 PM (8 months ago)

Author:

vboxsync

Message:

VMM/IEM: Optimize executable memory allocation on macOS by removing the need for calling RTMemProtect() to switch between RW and RX memory, bugref:10555.

On macOS it is impossible to have memory allocated RWX which is the reason for the current RTMemProtect() trickery which induces additional overhead. However the mach
VM API allows remapping the physical memory backed by a virtual address into another region which can have different protection flags. This allows having a virtual
memory region with readable/writeable permissions and a second region with readable/executable permissions both backed by the same physical memory.
A profiling build before this optimization took 76 ticks on average (taken before any memory pruning started to take place because the maximum amount of executable
memory was reached) when allocating executable memory, which translates to 3166.7ns given the 24MHz frequency of CNTVCT_EL0 used as the time source.
With the optimization in place the average is now 15 ticks, or 625ns.

File:

• trunk/src/VBox/VMM/include/IEMInternal.h (modified) (1 diff)

trunk/src/VBox/VMM/include/IEMInternal.h

@@ -6253 +6253 @@
 DECLHIDDEN(void)    iemNativeDisassembleTb(PCIEMTB pTb, PCDBGFINFOHLP pHlp) RT_NOEXCEPT;
 int                 iemExecMemAllocatorInit(PVMCPU pVCpu, uint64_t cbMax, uint64_t cbInitial, uint32_t cbChunk) RT_NOEXCEPT;
-DECLHIDDEN(void *)  iemExecMemAllocatorAlloc(PVMCPU pVCpu, uint32_t cbReq, PIEMTB pTb) RT_NOEXCEPT;
+DECLHIDDEN(void *)  iemExecMemAllocatorAlloc(PVMCPU pVCpu, uint32_t cbReq, PIEMTB pTb, void **ppvExec) RT_NOEXCEPT;
 DECLHIDDEN(void)    iemExecMemAllocatorReadyForUse(PVMCPUCC pVCpu, void *pv, size_t cb) RT_NOEXCEPT;
 void                iemExecMemAllocatorFree(PVMCPU pVCpu, void *pv, size_t cb) RT_NOEXCEPT;