SMP implementation for OpenBSD/sgi
Download as KEY, PDF0 likes1,242 views
The document describes the process of implementing SMP support for OpenBSD on a SGI Octane 2 machine. Key steps included restructuring per-processor data, implementing locking primitives, handling hardware aspects like spinning up secondary processors, and debugging challenges like detecting deadlocks. Debugging was made difficult by timing issues but was aided by tools like JTAG, DDB, printfs, and modifying locks to record stuck locations. Interrupts could block inter-processor communication so the clock handler was modified to re-enable interrupts during locking.
![Complex case - pmap.c:
original and modified code
uint pmap_alloc_tlbpid(struct proc *p)
...
tlbpid_cnt = id + 1;
pmap->pm_tlbpid = id;
uint pmap_alloc_tlbpid(struct proc *p)
...
tlbpid_cnt[cpuid] = id + 1;
pmap->pm_tlbpid[cpuid] = id;](https://image.slidesharecdn.com/asiabsdcon2010201003012-100314124558-phpapp02/85/SMP-implementation-for-OpenBSD-sgi-23-320.jpg)
![void pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
...
arg.pmap = pmap;
arg.va = va;
smp_rendezvous(0, pmap_invalidate_page_action, 0,
(void *)&arg);
void pmap_invalidate_page_action(void *arg)
...
pmap_t pmap = ((struct pmap_invalidate_page_arg *)arg)->pmap;
vm_offset_t va = ((struct pmap_invalidate_page_arg *)arg)->va;
if (is_kernel_pmap(pmap)) {
pmap_TLB_invalidate_kernel(va);
return;
}
if (pmap->pm_asid[PCPU_GET(cpuid)].gen
!= PCPU_GET(asid_generation))
return;
else if (!(pmap->pm_active & PCPU_GET(cpumask))) {
pmap->pm_asid[PCPU_GET(cpuid)].gen = 0;
return;
}
va = pmap_va_asid(pmap, (va & ~PGOFSET));
mips_TBIS(va);](https://image.slidesharecdn.com/asiabsdcon2010201003012-100314124558-phpapp02/85/SMP-implementation-for-OpenBSD-sgi-60-320.jpg)
![CPU_INFO_FOREACH(cii, ci)
if (cpuset_isset(&cpus_running, ci)) {
unsigned int i = ci->ci_cpuid;
unsigned int m = 1 << i;
if (pmap->pm_asid[i].pma_asidgen !=
pmap_asid_info[i].pma_asidgen)
continue;
else if (ci->ci_curpmap != pmap) {
pmap->pm_asid[i].pma_asidgen = 0;
continue;
}
cpumask |= m;
}
if (cpumask == 1 << cpuid) {
u_long asid;
asid = pmap->pm_asid[cpuid].pma_asid << VMTLB_PID_SHIFT;
tlb_flush_addr(va | asid);
} else if (cpumask) {
struct pmap_invalidate_page_arg arg;
arg.pmap = pmap;
arg.va = va;
smp_rendezvous_cpus(cpumask, pmap_invalidate_user_page_action,
&arg);
}](https://image.slidesharecdn.com/asiabsdcon2010201003012-100314124558-phpapp02/85/SMP-implementation-for-OpenBSD-sgi-61-320.jpg)
SMP implementation for OpenBSD/sgi
- 1. SMP Implementation for OpenBSD/sgi Takuya ASADA<[email protected]>
- 2. Introduction • I was working to add SMP & 64bit support to a BSD-based embedded OS at The target device was MIPS64 • There’s no complete *BSD/MIPS SMP implementation at that time, so I implemented it • The implementation was proprietary, but I wanted to contribute something to BSD community • I decided to implement SMP from scratch, tried to find a suitable MIPS SMP machine
- 3. Finding MIPS/SMP Machine(1) • Broadcom SiByte BCM1250 looks nice - 2core 1GHz MIPS64, DDR DRAM, GbE, PCI, HT • Cisco 3845 integrated this processor • How much is it?
- 4. Finding MIPS/SMP Machine(1) • Broadcom SiByte BCM1250 looks nice - 2core 1GHz MIPS64, DDR DRAM, GbE, PCI, HT • Cisco 3845 integrated this processor • How much is it? from $14,200! Totally unacceptable!!!
- 5. Finding MIPS/SMP Machine(2) • Some antique SGI machines are available on eBay • These are basically very cheap • I realized SGI Octane 2 has 2 cores, already supported by OpenBSD
- 6. Finding MIPS/SMP Machine(2) • Some antique SGI machines are available on eBay • These are basically very cheap • I realized SGI Octane 2 has 2 cores, already supported by OpenBSD Just $33! Do you remember how much was it?
- 7. This is my Octane2 Processors MIPS R12000 400MHz x2 Memory 1GB SDRAM Graphics 3D Graphics Card Sound Integrated Digital Audio Storage 35GB SCSI HDD Ethernet 100BASE-T
- 12. Become an OpenBSD developer • I started working on Octane2 since Apr 2009, wrote about on my blog • Miod Vallat discovered it, suggested my code be merged into OpenBSD main tree • I become an OpenBSD developer in Sep 2009, started merging the code, joined hackathon, worked with Miod and Joel Sing
- 13. NOW IT WORKS!!!! • Merged into OpenBSD main tree • You can try it now! • It seems stable - I tried “make build” again and again over a day, it keeps working
- 14. Demo
- 15. Performance - kernel build benchmark '(&!!" 4% overhead '(%!!" '($!!" '(#!!" 46% speed up '(!!!" ,-.-/01)02+!" &!!" ,-.-/01)02+!342" %!!" $!!" #!!" !" (sec) ")*'" ")*#" ")*+"
- 16. Performance - OpenBSD/amd64 for comparison &!!" %#!" 2% overhead %!!" 44% speed up $#!" )*+*,-." )*+*,-./01" $!!" #!" !" "'($" "'(%" "'(&" (sec)
- 17. What did we need to do for the SMP work There were lots of works... • Support multiple cpu_info • Secondary processor entry and processor related point macros • IPI: Inter-Processor Interrupt • Move per-processor data into cpu_info • Per-processor ASID management • Implement lock primitives • TLB shootdown • Acquiring giant lock • Lazy FPU handling • Implement atomic operations • Per-processor clock • Spin up secondary processors
- 18. Describe it more simply We can classify tasks into three kind of problems: • Restructuring per-processor informations • Implement and use lock/atomic primitives • Hardware related problems
- 19. Restructuring per- processor informations • In the original sgi port, some informations which related the processor are allocated only one • Simple case: Information is stored into global variable Move it into “cpu_info”, per-processor structure • Complex case: In pmap, we need to maintain some information per-processor * per-process
- 20. Simple case - clock.c: original code // defined as global variables u_int32_t cpu_counter_last; u_int32_t cpu_counter_interval; u_int32_t pendingticks; uint32_t clock_int5(uint32_t mask, struct trap_frame *tf) ... clkdiff = cp0_get_count() - cpu_counter_last; while (clkdiff >= cpu_counter_interval) { cpu_counter_last += cpu_counter_interval; clkdiff = cp0_get_count() - cpu_counter_last; pendingticks++;
- 21. Simple case - clock.c: modified code uint32_t clock_int5(uint32_t mask, struct trap_frame *tf) ... struct cpu_info *ci = curcpu(); clkdiff = cp0_get_count() - ci->ci_cpu_counter_last; while (clkdiff >= ci->ci_cpu_counter_interval) { ci->ci_cpu_counter_last += ci->ci_cpu_counter_interval; clkdiff = cp0_get_count() - ci->ci_cpu_counter_last; ci->ci_pendingticks++;
- 22. Complex case: pmap • MIPS TLB entries are tagged with 8bit process id called ASID, used for improve performance MMU skips different process TLB entries on lookup We won’t need to flush TLB every context switches • Need to maintain process:ASID assign information because it’s smaller than PID, we need to rotate it • The information should keep beyond context switch • We maintain ASID individually per-processor • What we need is: ASID information per-processor * per-process
- 23. Complex case - pmap.c: original and modified code uint pmap_alloc_tlbpid(struct proc *p) ... tlbpid_cnt = id + 1; pmap->pm_tlbpid = id; uint pmap_alloc_tlbpid(struct proc *p) ... tlbpid_cnt[cpuid] = id + 1; pmap->pm_tlbpid[cpuid] = id;
- 24. Implement and use lock/atomic primitives • Needed to implement • lock primitives: mutex, mp_lock • atomic primitives: CAS, 64bit add, etc.. • Acquiring giant lock prior to entering the kernel context • hardware interrupts • software interrupts • trap()
- 25. Acquiring giant lock on clock interrupt handler uint32_t clock_int5(uint32_t mask, struct trap_frame *tf) ... if (tf->ipl < IPL_CLOCK) { #ifdef MULTIPROCESSOR __mp_lock(&kernel_lock); #endif while (ci->ci_pendingticks) { clk_count.ec_count++; hardclock(tf); ci->ci_pendingticks--; } #ifdef MULTIPROCESSOR __mp_unlock(&kernel_lock); #endif
- 26. Acquiring giant lock on clock interrupt handler uint32_t clock_int5(uint32_t mask, struct trap_frame *tf) ... if (tf->ipl < IPL_CLOCK) { #ifdef MULTIPROCESSOR __mp_lock(&kernel_lock); #endif while (ci->ci_pendingticks) { clk_count.ec_count++; hardclock(tf); ci->ci_pendingticks--; } #ifdef MULTIPROCESSOR __mp_unlock(&kernel_lock); #endif Actually, it causes a bug... described later
- 27. Hardware related problems • Spin up secondary processor • Keeping TLB consistency by software • Cache coherency
- 28. Spin up secondary processor • We need to launch secondary processor • Access hardware register on controller device to power on the processor • Secondary processor needs own bootstrap code • Has to be similar to primary processor’s one • But we won’t need kernel, memory, and device initialization • Because primary processor already did it
- 29. Keeping TLB consistency by software • MIPS TLB doesn’t have mechanism to keep consistency, we need to do it by software Just like the other typical architectures • Invalidate/update other processor’s TLB using TLB shootdown • It implemented by IPI (Inter-Processor Interrupt)
- 30. Cache coherency • MIPS R10000/R12000 processors have full cache coherency • We don’t have to care about it on Octane • But, some other processors haven’t full cache coherency
- 31. Ideas implementing SMP • We have faced a number of issues while implementing SMP • Fighting against deadlock • Dynamic memory allocation without using virtual address • Reduce frequency of TLB shootdown
- 32. Fighting against deadlock • It’s hard to find the cause of deadlocks because both processors runs concurrently • It causes timing bugs - conditions are depend on timing • Need to be able to determine what happened on both processors at that time • There are tools for debugging it
- 33. JTAG ICE • Very useful for debugging • We can get any data for debugging on desired timing, even after kernel hangs • I used it when I was implementing SMP for the embedded OS • Not for Octane, there’s no way to connect
- 34. ddb • OpenBSD kernel has in-kernel debugger, named ddb • We can get similar data for JTAG ICE, but kernel need to alive - because it’s part of the kernel • Missing features: We hadn’t implemented “machine ddbcpu<#>” - which is processor switching function on ddb Without this, we can only debug one processor which invoked ddb on a breakpoint • Not always useful
- 35. printf() • Most popular kernel debugging tool • Just write printf(message) on your code, Easy to use ;) • Unfortunately, it has some problems • printf() wastes lot of cycles, changes timing between processors We may miss timing bug because of it • Some point of the code are printf() unsafe causes kernel hang • We use it anyway
- 36. Divide printf output for two serial port • There’s two serial port and two processors • If we have lots of debug print, it’s hard to understand which lines belongs to what processor • I implemented dirty hack code which output strings directly to secondary serial port, named combprintf() • Rewrite debug print to primary processor outputs primary serial port, secondary processor outputs secondary serial port
- 37. What do we need to print for debugging? • To know where the kernel running roughly • put debug print everywhere the point kernel may running through • dump all system call using SYSCALL_DEBUG • How can we determine a deadlock point?
- 38. Determine a deadlock point • Deadlocks are occurring on spinlocks • It loops permanently until a condition become available, but that condition never comes up • At least to know which lock primitives we stuck on, we need to stop permanent loop by implementing timeout counter and print debug message
- 39. Adding timeout counter into mutex void mtx_enter(struct mutex *mtx) ... for (;;) { if (mtx->mtx_wantipl != IPL_NONE) s = splraise(mtx->mtx_wantipl); if (try_lock(mtx)) { if (mtx->mtx_wantipl != IPL_NONE) mtx->mtx_oldipl = s; mtx->mtx_owner = curcpu(); return; } if (mtx->mtx_wantipl != IPL_NONE) splx(s); if (++i > MTX_TIMEOUT) panic("mtx deadlockedn”); }
- 40. Adding timeout counter into mutex void mtx_enter(struct mutex *mtx) ... for (;;) { if (mtx->mtx_wantipl != IPL_NONE) s = splraise(mtx->mtx_wantipl); if (try_lock(mtx)) { if (mtx->mtx_wantipl != IPL_NONE) mtx->mtx_oldipl = s; mtx->mtx_owner = curcpu(); return; } if (mtx->mtx_wantipl != IPL_NONE) splx(s); if (++i > MTX_TIMEOUT) panic("mtx deadlockedn”); } But, this is not enough
- 41. Why it’s not enough CPU A CPU B Acquires lock Acquires lock Lock Lock A B Spins until released Spins until released
- 42. Why it’s not enough CPU A CPU B Acquires lock Acquires lock Lock Lock A B Spins until released Spins until released We can break here
- 43. Why it’s not enough CPU A CPU B Acquires lock Acquires lock We wanna know which Lock Lock function acquired it A B Spins until released Spins until released We can break here
- 44. Remember who acquired it void mtx_enter(struct mutex *mtx) ... for (;;) { if (mtx->mtx_wantipl != IPL_NONE) s = splraise(mtx->mtx_wantipl); if (try_lock(mtx)) { if (mtx->mtx_wantipl != IPL_NONE) mtx->mtx_oldipl = s; mtx->mtx_owner = curcpu(); mtx->mtx_ra = __builtin_return_address(0); return; } if (mtx->mtx_wantipl != IPL_NONE) splx(s); if (++i > MTX_TIMEOUT) panic("mtx deadlocked ra:%pn", mtx->mtx_ra); }
- 45. Interrupt blocks IPI CPU A CPU B Interrupt Fault Disable interrupt Acquire lock Lock Wait until released TLB shootdown Blocked IPI Wait ACK
- 46. Interrupt blocks IPI CPU A CPU B Interrupt Fault Disable interrupt Acquire lock Lock Wait until released TLB shootdown IPI Wait ACK
- 47. Interrupt blocks IPI CPU A CPU B Interrupt Fault Disable interrupt Acquire lock Re-enable interrupt Lock Wait until released TLB shootdown Accept Interrupt IPI Wait ACK
- 48. Interrupt blocks IPI CPU A CPU B Interrupt Fault Disable interrupt Acquire lock Re-enable interrupt Lock Wait until released TLB shootdown Accept Interrupt IPI Wait ACK ACK for rendezvous
- 49. Interrupt blocks IPI CPU A CPU B Interrupt Fault Solution: Enable IPI interrupt on Disable interrupt Acquire lock interrupt handlers Re-enable interrupt Lock Wait until released TLB shootdown Accept Interrupt IPI Wait ACK ACK for rendezvous
- 50. Fixing clock interrupt handler uint32_t clock_int5(uint32_t mask, struct trap_frame *tf) ... if (tf->ipl < IPL_CLOCK) { #ifdef MULTIPROCESSOR u_int32_t sr; sr = getsr(); ENABLEIPI(); __mp_lock(&kernel_lock); #endif while (ci->ci_pendingticks) { clk_count.ec_count++; hardclock(tf); ci->ci_pendingticks--; } #ifdef MULTIPROCESSOR __mp_unlock(&kernel_lock); setsr(sr); #endif
- 51. splhigh() blocks IPI CPU A CPU B splhigh() Fault Acquire lock Lock Wait until released TLB shootdown Blocked IPI Wait ACK
- 52. splhigh() blocks IPI CPU A CPU B splhigh() Fault Acquire lock Lock Wait until released TLB shootdown IPI Wait ACK
- 53. splhigh() blocks IPI CPU A CPU B splhigh() Fault Acquire lock Lock Wait until released TLB shootdown Accept Interrupt IPI Wait ACK
- 54. splhigh() blocks IPI CPU A CPU B splhigh() Fault Acquire lock Lock Wait until released TLB shootdown Accept Interrupt IPI Wait ACK ACK for rendezvous
- 55. splhigh() blocks IPI CPU A CPU B splhigh() Solution: defined new interrupt priority Fault level named IPL_IPI to be higher than Acquire lock IPL_HIGH Lock Wait until released TLB shootdown Accept Interrupt IPI Wait ACK ACK for rendezvous
- 56. Adding new interrupt priority level #define IPL_TTY 4 /* terminal */ #define IPL_VM 5 /* memory allocation */ #define IPL_CLOCK 6 /* clock */ #define IPL_HIGH 7 /* everything */ #define NIPLS 8 /* Number of levels */ #define IPL_TTY 4 /* terminal */ #define IPL_VM 5 /* memory allocation */ #define IPL_CLOCK 6 /* clock */ #define IPL_HIGH 7 /* everything */ #define IPL_IPI 8 /* ipi */ #define NIPLS 9 /* Number of levels */
- 57. Dynamic memory allocation without using virtual address • To support N(>2) processors, the cpu_info structure and the bootstrap kernel stack for secondary processors should be allocated dynamically • But we can’t use virtual address for them • stack may used before TLB initialization, thus causing the processor fault • MIPS has Software TLB, need to maintain TLB by software TLB miss handler is the code to handle it This handler refers cpu_info, it cause TLB miss loop • To avoid these problems, we implemented wrapper function to allocate memory dynamically, then get the physical address and return it
- 58. The wrapper function vaddr_t smp_malloc(size_t size) ... if (size < PAGE_SIZE) { va = (vaddr_t)malloc(size, M_DEVBUF, M_NOWAIT); if (va == NULL) return NULL; error = pmap_extract(pmap_kernel(), va, &pa); if (error == FALSE) return NULL; } else { TAILQ_INIT(&mlist); error = uvm_pglistalloc(size, 0, -1L, 0, 0, &mlist, 1, UVM_PLA_NOWAIT); if (error) return NULL; m = TAILQ_FIRST(&mlist); pa = VM_PAGE_TO_PHYS(m); } return PHYS_TO_XKPHYS(pa, CCA_CACHED);
- 59. Reduce frequency of TLB shootdown • There’s a condition we can skip TLB shootdown in invalidate/ update: using the pagetable not using kernel mode need need user mode need don’t need • In user mode, if shootee processor doesn’t using the pagetable, we won’t need shootdown; just changing ASID assign is enough • In reference pmap implementation, a TLB shootdown performed non-conditionally, even in case it isn’t really needed • We added the condition to reduce frequency of it
- 60. void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) ... arg.pmap = pmap; arg.va = va; smp_rendezvous(0, pmap_invalidate_page_action, 0, (void *)&arg); void pmap_invalidate_page_action(void *arg) ... pmap_t pmap = ((struct pmap_invalidate_page_arg *)arg)->pmap; vm_offset_t va = ((struct pmap_invalidate_page_arg *)arg)->va; if (is_kernel_pmap(pmap)) { pmap_TLB_invalidate_kernel(va); return; } if (pmap->pm_asid[PCPU_GET(cpuid)].gen != PCPU_GET(asid_generation)) return; else if (!(pmap->pm_active & PCPU_GET(cpumask))) { pmap->pm_asid[PCPU_GET(cpuid)].gen = 0; return; } va = pmap_va_asid(pmap, (va & ~PGOFSET)); mips_TBIS(va);
- 61. CPU_INFO_FOREACH(cii, ci) if (cpuset_isset(&cpus_running, ci)) { unsigned int i = ci->ci_cpuid; unsigned int m = 1 << i; if (pmap->pm_asid[i].pma_asidgen != pmap_asid_info[i].pma_asidgen) continue; else if (ci->ci_curpmap != pmap) { pmap->pm_asid[i].pma_asidgen = 0; continue; } cpumask |= m; } if (cpumask == 1 << cpuid) { u_long asid; asid = pmap->pm_asid[cpuid].pma_asid << VMTLB_PID_SHIFT; tlb_flush_addr(va | asid); } else if (cpumask) { struct pmap_invalidate_page_arg arg; arg.pmap = pmap; arg.va = va; smp_rendezvous_cpus(cpumask, pmap_invalidate_user_page_action, &arg); }
- 62. Future works • Commit machine ddbcpu<#> • New port for Cavium OCTEON, if it’s acceptable for OpenBSD project • Maybe SMP support for SGI Origin 350 • Also interested in MI part of SMP