1 / 15

EM64T ‘fast’ system-calls

EM64T ‘fast’ system-calls. A look at the requirements for using Intel’s ‘syscall’ and ‘sysret’ instructions in 64-bit mode. Privilege-levels.

Jimmy
Download Presentation

EM64T ‘fast’ system-calls

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. EM64T ‘fast’ system-calls A look at the requirements for using Intel’s ‘syscall’ and ‘sysret’ instructions in 64-bit mode

  2. Privilege-levels • Although the x86 processor supports four distinct privilege-levels in protected-mode, only two are actually used in the popular Windows and Linux operating-systems Ring3 (for applications) Ring2 (not used) Ring1 (not used) Ring0 (for the OS kernel)

  3. Opportunity for optimization • Just as the suppression of ‘segmentation’ in 64-bit memory-addressing has offered extra execution-speed and programming simplicity, there is opportunity for faster privilege-level transitions by eliminating references to system-tables in memory ring3 ring0 system-call application program kernel services return

  4. Sacrifice ‘flexibility’ for ‘speed’ • The ‘syscall’ and ‘sysret’ instructions allow much faster ring-transitions during normal system-calls – by keeping all the required information in special CPU registers – but accepting some limitations: • Transitions are only between ring3 and ring0 • Only one ‘entry-point’ to all kernel-services • Some formerly ‘general-purpose’ registers would have to acquire a dedicated function

  5. Layout of GDT descriptors • Use of ‘syscall’ requires a pair of global descriptors to be adjacently placed: • Use of ‘sysret’ requires a triple of global descriptors to be adjacently placed: 64-bit code DPL=0 ring0 data DPL=0 GDTR 16/32-bit code DPL=3 ring3 data DPL=3 64-bit code DPL=3 GDTR

  6. Model-Specific Registers • Some MSRs must be suitably initialized: 0xC0000080: IA32_MSR_EFER 0xC0000081: IA32_MSR_STAR 0xC0000082: IA32_MSR_LSTAR 0xC0000083: IA32_MSR_CSTAR 0xC0000084: IA32_MSR_FMASK • The Intel processor must be executing in 64-bit mode to use ‘syscall’ and ‘sysret’

  7. Extended Feature Enable Register • This Model-Specific Register (MSR) was introduced in the AMD64 architecture and perpetuated by EM64T (for compatibility) 63 11 10 8 0 N X E L M A L M E S C E Legend: SCE = SysCall/sysret is Enabled (1=yes, 0=no) LME = Long-Mode is Enabled (1=yes, 0=no) LMA = Long-Mode is Active (1=yes, 0=no) NXE = Non-eXecutable pages Enabled (1=yes, 0=no) NOTE: The MSR address-index for EFER = 0xC0000080, and this register is accessed using RDMSR or WRMSR instructions

  8. The MSR_STAR register for ‘sysret’ for ‘syscall’ 63 48 47 32 31 0 selector for ‘compatibility’ ring3 code-descriptor selector for 64-bit ring0 code-descriptor unused This selector is for the first in a pair of adjacently-placed GDT descriptors for the CS and SS registers, respectively, upon the transition from ring3 to ring0 This selector is for the first in a triple of adjacently-placed GDT descriptors for the CS and SS (or SS and CS) registers, respectively, upon the transition from ring0 to ring3 (depending on whether ‘sysret’ or ‘sysretq’ is used)

  9. The MSR_LSTAR register 63 0 Linear-address of the system-call entry-point This is the 64-bit address which will go into the RIP register when the ‘syscall’ instruction is ececuted by ring3 code The former value from the RIP register (i.e., the ‘return-address’) will be saved in the RCX general-purpose register, to be used later by the ‘sysret’ instruction (so therefore it must be preserved)

  10. The MSR_CSTAR register It’s a mystery … 63 0 The function of this register is unknown This register is observed to exist – Linux x86_64 writes a value into this register in fact – although current Intel documentation omits mention or explanation of this Model-Specific Register (We did find an obsolete Intel document online which referred to this register, but did not make clear its past purpose or function)

  11. The MSR_FMASK register 63 31 0 Reserved (unused) Flags Mask This register can be programmed by an Operating System with a bitmask that will be used by the processor to automatically ‘clear’ a specified selection of bits in the RFLAGS register when ‘syscall’ is executed (the former value of RFLAGS is saved in the general-purpose R11 register)

  12. ‘fastcall.s’ • We created a demo-program that shows the use of ‘syscall’ and ‘sysret’, indicating what setup-steps are needed: • Page-mapping tables (user-accessible frames) • Global Descriptor Table layout • Task-State Segment (needs ESP0 value) • EFER (needs LME=1 and SCE=1) • CR4 needs PAE=1, • CR3 needs physical address of page-map level4 • CR0 needs PE=1 and PG=1

  13. Transitions in ‘fastcall.s’ real mode Ia-32e compatibility mode 64-bit mode (Ring3) mov lret syscall The main purpose of this demo-program is to illustrate the use of ‘syscall’ and ‘sysret’ (but we have to get to 64-bit mode to do it) 64-bit mode (Ring0) sysret real mode Ia-32e compatibility mode 64-bit mode (Ring0) 64-bit mode (Ring3) ljmp mov lcall

  14. In-class exercise • In our ‘fastcall.s’ demo-program there are two transitions from ring3 to ring0 (in one case via ‘syscall’ and in the other via ‘lcall’ through a call-gate • Can you measure which of these is faster? (for example, by using the processor’s TimeStamp Counter, accessible with the ‘rdtsc’ instruction)

  15. TimeStamp Counter 63 0 64-bit register automatically increments with every cpu clock-cycle The ‘rdtsc’ instruction returns this register’s current value: rdtsc # EAX = least-significant 32-bits from TSC # EDX = most-significant 32-bits from TSC (This 64-bit register is initialized to zero at system-startup)

More Related