Sharing and protection in Multics Landon Cox February 5, 2016

Layer 0 Layer 1 Layer 2 Layer 3 Layer 4 Scheduler Pager Console I/O devices User programs Superviso r code

User/kernel translation data Virtual memory 0GB 4GB 3GB (0xc ) Kernel data (same for all page tables) User data (different for every process)

Where to store translation data? 1. Could be kept in physical memory How the MMU accesses it On x86, PTBR (aka CR3 register) refers to physical memory 2. Could be in kernel virtual address space A little weird, but what is done in practice How to keep translation data in a place that must be translated? Translation for user address space is in kernel virtual memory Kernel’s translation data always in physical memory (pinned) Does anything else need to be pinned (and why)? Kernel’s page fault handler code also has to be pinned.

Where to store translation data? How does kernel access data in user’s address space? If for currently loaded process, access using current page table If for another process, have to change page table

Kernel vs. user mode Who sets up the data used by the MMU? Can’t be the user process Otherwise could access anything Only kernel is allowed to modify any memory Processor must know to allow kernel To update the translator To execute privileged instructions (halt, do I/O)

Kernel vs. user mode How does machine know kernel is running? This requires hardware support Process supports two modes, kernel and user Mode is indicated by a hardware register Mode bit

Protection 1. All memory accesses go through a translator Who can modify translator’s data? 2. Only kernel can modify translator’s data How do we know if kernel is running? 3. Mode bit indicates if kernel is running Who can modify the mode bit? Making progress: the amount of protected data is down to a bit GBs of protected data One bit of protected data

Protecting the mode bit Can kernel change the mode bit? Yes. Kernel is completely trusted. Can user process change the mode bit? Not directly User programs need to invoke the kernel Must be able to initiate a change

When to transition from user to kernel? 1. Exceptions (interrupts, traps) Access something out of your valid address space (e.g., segmentation fault) Disk I/O finishes, causes interrupt Timer pre-emption, causes interrupt Page faults 2. System calls Similar in purpose to a function call Kernel as software library

Example system calls Process management Fork/exec (start a new process), exit, getpid Signals Kill (send a signal), sigaction (set up handler) Memory management Brk (extend the valid address space), mmap File I/O Open, close, read, write Network I/O Accept, send, receive System management Reboot

System call implementation Syscalls are like functions, but different Implemented by special instruction syscall Execution of syscall traps to kernel Processor safely transfers control to kernel Hardware invokes kernel’s syscall trap handler

System call implementation Libc wraps systems calls C++/Java make calls into libc

Tracing through “ cin >> a; ” C++: cin >> a C: read () libc: syscall(SYS_read,filenum,offset,numbytes) Java: in.read() Processor: kernel’s syscall handler kernel: sys_read kernel: // perform I/O Crucial step

Kernel trap details Hardware must atomically 1. Set processor mode bit to “kernel” 2. Save current registers (SP, PC, general purpose) 3. Change execution stack to kernel (SP) 4. Jump to exception handler in kernel (PC) What does this look a lot like? Switching between threads (see previous lecture)

Kernel trap details User process can initiate mode switch But only transfers control in limited way (i.e., to predefined kernel code) How does processor know where to jump? Stored in hardware “interrupt vector table” Who can update the interrupt vector table? Kernel does this at startup Code that runs first gets control of machine

Syscall arguments, return values For args passed in memory, i n which address space do they reside? User (programs can’t access kernel’s portion of address space) Linux passes parameters via registers Before syscall Push ebp, edx, ecx vals onto user stack Copy esp (stack pointer) into ebp Each process has kernel stack CPU knows kernel stack via “Task Segment State” CPU sets esp using TSS Kernel finds user stack in ebp register

Tracing through read() C: read (int fd, void *buf, size_t size) libc: push fd, buf, size onto user’s stack Note: pointer! kernel: sys_read () { verify fd is an open file verify [buf – buf+size) valid, writable read file from data (e.g. disk) find physical location of buf copy data to buf’s physical location } Verify argument s!

Layer 0 Layer 1 Layer 2 Layer 3 Layer 4 Scheduler Pager Console I/O devices User programs

THE Were THE processes protected from each other? Yes, processes had private segment spaces Those spaces were isolated via the compiler and Layer 1 Isolation/protection was the motivation for layering What is the flip side of protection? Sharing! There are often many benefits to sharing (i.e., “features”) Sharing also creates complexity Ongoing issue: balancing features and complexity

THE Could THE processes share in-core data? Only within the structure of the hierarchy Lower layers were like shared libraries User processes could not share with each other What about persistent data? Not really, “no common data base.”

THE Important idea not supported by THE The notion of a user “not intended as a multi-access system” What is a “user” in a computer system? Identifier assigned to processes and storage objects Unit of access control in most systems Often associated with a person, but doesn’t have to Do modern machines support users? Desktops/servers do, mobile devices kind of do

Multics Multi-user operating system Primary goal was to allow efficient, safe sharing btw users What was the central data abstraction in Multics? A segment All data was contained within a segment No distinction between files and memory Means everything is persistent What is a segment? Named range of contiguous data + associated meta-data Accessed through loads/stores in memory Think of a segment as an mmapped region of memory

Multics segments Implications of segments Difference between a program and process? We think of a process as an execution of a program In Multics, there was very little distinction What does this force programmers to do? Must explicitly manage segments Process cannot just exit and return to a “known” state Places a lot of burden on programmers Beware, complexity creep …

Segment access Address [s,i] Core|L d Core|L s |Acc|F Descriptor Segment s Segment Descriptor Word Segment word of data i Descriptor Base Register if(DBR.L < S) fault sdw = DW[s] if(sdw.F || sdw.L<i) fault if(!verify(sdw.Acc)) fault return sdw.Core + i Lookup algorithm? LdLd LsLs

What about paging? How do we make sure that segments are in- core? Segments are broken into fixed-size pages (1k) Another structure to map parts of a segment to pages Page tables describe where pieces are in core Descriptor segment is a segment too Must locate the pages on which it is located Can then walk data structure to locate data

Paged segment access [s,i]  [sp,sw,ip,iw] Core| L Core|L|Acc|F Page sp of DS sw Segment Descriptor Word Page ip of Segment S word of data iwiw Descriptor Base Register Page Table of Descriptor Segment Core|L sp Page Table of Segment S Core|L ip This is a lot of memory accesses What’s the problem with this? Caching in hardware: TLB How do we make it faster? (1) Find page table for descriptor segment (2) Page holds part of descriptor segment (3) Use descriptor segment to find page table for data segment (4) Use page table to locate data we want to access

Multics supervisor code Location of THE supervisor state In separate processes at low-level layers In Multics, where did supervisor state reside? In segments mapped into every process Supervisor segments were at top of address space How was supervisor state protected? Hardware protection provided by processor Hardware supported 8 protection rings Idea was to enforce layering via hardware Exactly like mode bit (kernel/user) you are used to

Multics supervisor code How did programs invoke the supervisor? Just a procedure call Calls always reside in segments at top of addr. sp. Just jump into code in those segments What else has to happen? Have to change hardware protection mode Who/what changed protection modes? This was hardware enforced All segments were assigned a ring level (including code) Mode for an instruction was set in descriptor on which it resided

Segment sharing To facilitate sharing need common namespace for segments Multics uses file hierarchy to name all segments Populating address space like a bunch of mmap calls Why is this a nice abstraction? Maps human-readable names (easy to program) to data Makes it easy to specify what data you want to operate on Who manages the namespace? Must be privileged code Common data structure used by all processes Only code trusted by all should be allowed to modify directly

Segment sharing What operations can a process perform on the namespace? Create a segment Delete a segment Change a segment’s name Change the access policy of a segment Read the content of a directory Who prevents collisions in the namespace? The supervisor code On a request to create a segment, checks to see if it already exists

Segment sharing How does a process populate its address space? Must know or compute parts of the namespace Invokes the OS to map named segment into memory OS updates the Known Segment Table for process KST maps segments to pathnames OS returns beginning of segment, length to process Once mapped, how is data loaded into memory? Data is demand loaded Process accesses address (e.g., [s,i]) This triggers a page fault and a trap to the OS The OS uses the KST to locate data on disk

Discussion Are Multics segments a good idea or bad? Programmers must do a lot of garbage collection Temporary state must be re-initialized by hand Adds complexity Danger of accessing persistent data via loads/stores? Buggy programs can lead to stray writes Stray writes become permanent Why are permanent stray writes bad? Important data structures could become corrupted “Restarting” program just puts you back in a bad state

Discussion How do most modern systems present persistent storage? Must access persistent storage through file system File system interface: open, read, write Why is an explicit file system interface safer? Write system call is similar to a “commit” An explicit acknowledgement that data is ready to be made persistent Buggy programs much less unlikely to generate a spurious write call Spurious loads and stores are very common in buggy programs We actually see this observation in Multics itself. Where? In how it handles the namespace Important, persistent data structure Can only be modified in a controlled way, through narrow interface Nice to offer same protections for user data

Topic we will revisit Interfaces to persistent storage Periodic source of new research In the 90s: battery-backed RAM Memory persists across reboots Rio-Vista from Michigan (SOSP ‘97) In the 00s: phase-change memory Fast, persistent substrate BPFS from Microsoft and UCLA (SOSP ‘09)