/* Copyright (C) 2014 Robert Kausch Copyright (C) 2008 Monty This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program. If not, see . */ /* we can ask most drives what their various caches' sizes are, but no drive will tell if it caches redbook data. None should, many do, and there's no way in (eg) MMC/ATAPI to tell a cdrom drive not to cache when accessing audio. SCSI drives have a FUA facility, but it's not clear how many ignore it. MMC does specify some cache side effect as part of SET READ AHEAD, but it's not clear we can rely on them. For that reason, we need to empirically determine cache size and strategy used for reads. */ #include #include #include #include #include #include #include #include "cachetest.h" /* not strictly just seeks, but also recapture and read resume when reading/readahead is suspended and idling */ #define MIN_SEEK_MS 6 #define reportC(...) {if(progress){fprintf(progress, __VA_ARGS__);} \ if(log){fprintf(log, __VA_ARGS__);}} #define printC(...) {if(progress){fprintf(progress, __VA_ARGS__);}} #define logC(...) {if(log){fprintf(log, __VA_ARGS__);}} static int time_drive(cdrom_drive_t *d, FILE *progress, FILE *log, int lba, int len, int initial_seek){ int i,x; int latency=0; double sum=0; double sumsq=0; int sofar; logC("\n"); for(i=0,sofar=0;sofar9999)x=9999; if(x<0)x=0; logC("%d:%d:%d ",lba+sofar,ret,x); sofar+=ret; if(i || !initial_seek){ sum+=x; sumsq+= x*x /(float)ret; }else latency=x; } /* we count even the upper outliers because the drive is almost certainly reading ahead and that will work itself out as we keep reading to catch up. Besides-- the tests would rather see too slow a timing than too fast; the timing data is used as an optimization when sleeping. */ { double mean = sum/(float)(len-1); double stddev = sqrt( (sumsq/(float)(len-1) - mean*mean)); if(initial_seek){ printC("%4dms seek, %.2fms/sec read [%.1fx]",latency,mean,1000./75./mean); logC("\n\tInitial seek latency (%d sectors): %dms",len,latency); } logC("\n\tAverage read latency: %.2fms/sector (raw speed: %.1fx)",mean,1000./75./mean); logC("\n\tRead latency standard deviation: %.2fms/sector",stddev); return sum; } } static float retime_drive(cdrom_drive_t *d, FILE *progress, FILE *log, int lba, int readahead, float oldmean){ int sectors = 2000; int total; float newmean; if(sectors*oldmean > 5000) sectors=5000/oldmean; readahead*=10; readahead/=9; if(readahead>sectors)sectors=readahead; printC("\bo"); logC("\n\tRetiming drive... "); total = time_drive(d,NULL,log,lba,sectors,1); newmean = total/(float)sectors; logC("\n\tOld mean=%.2fms/sec, New mean=%.2fms/sec\n",oldmean,newmean); printC("\b"); if(newmean>oldmean)return newmean; return oldmean; } int analyze_cache(cdrom_drive_t *d, FILE *progress, FILE *log, int speed){ /* Some assumptions about timing: We can't perform cache determination timing based on looking at average transfer times; on slow setups, the speed of a drive reading sectors via PIO will not be reliably distinguishable from the same drive returning data from the cache via pio. We need something even more noticable and reliable: the seek time. It is unlikely we'd ever see a seek latency of under ~10ms given the synchronization requirements of a CD and the maximum possible rotational velocity. A cache hit would always be faster, even with PIO. Further complicating things, we have to watch the data collection carefully as we're not always going to be on an unloaded system, and we even have to guard against other apps accessing the drive (something that should never happen on purpose, but could happen by accident). As we know in our testing when seeks should never occur, a sudden seek-sized latency popping up in the middle of a collection is an indication that collection is possibly invalid. A second cause of 'spurious latency' would be media damage; if we're consistently hitting latency on the same sector during initial collection, may need to move past it. */ int i,j,ret=0,x; int firstsector=-1; int lastsector=-1; int firsttest=-1; int lasttest=-1; int offset; int warn=0; int current=1000; int hi=15000; int cachesize=0; int readahead=0; int rollbehind=0; int cachegran=0; float mspersector=0; if(speed<=0)speed=-1; reportC("\n=================== Checking drive cache/timing behavior ===================\n"); d->error_retry=0; /* verify the lib and cache analysis match */ if(strcmp(LIBCDIO_PARANOIA_VERSION,paranoia_version())){ reportC("\nWARNING: cdparanoia application (and thus the cache tests) does not match the" "\ninstalled (or in use) libcdda_paranoia.so library. The final verdict of this" "\ntesting may or may not be accurate for the actual version of the paranoia" "library. Continuing anyway...\n\n"); } /* find the longest stretch of available audio data */ for(i=0;itracks;i++){ if(cdda_track_audiop(d,i+1)==1){ if(firsttest == -1) firsttest=cdda_track_firstsector(d,i+1); lasttest=cdda_track_lastsector(d,i+1); if(lasttest-firsttest > lastsector-firstsector){ firstsector=firsttest; lastsector=lasttest; } }else{ firsttest=-1; lasttest=-1; } } if(firstsector==-1){ reportC("\n\tNo audio on disc; Cannot determine timing behavior..."); return -1; } /* Dump some initial timing data to give a little context for human eyes. Take readings ten minutes apart (45000 sectors) and at end of disk. */ { int best=0; int bestcount=0; int iterating=0; offset = lastsector-firstsector-current-1; reportC("\nSeek/read timing:\n"); while(offset>=firstsector){ int m = offset/4500; int s = (offset-m*4500)/75; int f = offset-m*4500-s*75; int sofar; if(iterating){ reportC("\n"); }else{ printC("\r"); logC("\n"); } reportC("\t[%02d:%02d.%02d]: ",m,s,f); /* initial seek to put at at a small offset past end of upcoming reads */ if((ret=cdda_read(d,NULL,offset+current+1,1))<0){ /* media error! grr! retry elsewhere */ if(ret==-404)return -1; reportC("\n\tWARNING: media error during read; continuing at next offset..."); offset = (offset-firstsector+44999)/45000*45000+firstsector; offset-=45000; continue; } sofar=time_drive(d,progress, log, offset, current, 1); if(offset==firstsector)mspersector = sofar/(float)current; if(sofar==-404) return -1; else if(sofar<0){ reportC("\n\tWARNING: media error during read; continuing at next offset..."); offset = (offset-firstsector+44999)/45000*45000+firstsector; offset-=45000; continue; }else{ if(!iterating){ if(best==0 || sofar*1.01sofar && bestcount>4000) iterating=1; } } } if(iterating){ offset = (offset-firstsector+44999)/45000*45000+firstsector; offset-=45000; printC(" "); }else{ offset--; printC(" spinning up... "); } } } reportC("\n\nAnalyzing cache behavior...\n"); /* search on cache size; cache hits are fast, seeks are not, so a linear search through cache hits up to a miss are faster than a bisection */ { int under=1; int onex=0; current=0; offset = firstsector+10; while(current <= hi && under){ int i,j; under=0; current++; if(onex){ if(speed==-1){ logC("\tAttempting to reset read speed to full... "); }else{ logC("\tAttempting to reset read speed to %dx... ",speed); } if(cdda_speed_set(d,speed)){ logC("failed.\n"); }else{ logC("drive said OK\n"); } onex=0; } printC("\r"); reportC("\tFast search for approximate cache size... %d sectors ",current-1); logC("\n"); for(i=0;i<25 && !under;i++){ for(j=0;;j++){ int ret1=0,ret2=0; if(i>=15){ int sofar=0; if(i==15){ printC("\r"); reportC("\tSlow verify for approximate cache size... %d sectors",current-1); logC("\n"); logC("\tAttempting to reduce read speed to 1x... "); if(cdda_speed_set(d,1)){ logC("failed.\n"); }else{ logC("drive said OK\n"); } onex=1; } printC("."); logC("\t\t>>> "); while(sofar>> fast_read=%d:%d:%d ",offset+current-1,ret1,x); /* Some drives are 'easily distracted' when readahead is running ahead of the read cursor, causing accesses of the earliest sectors in the cache to show bursty latency. If there's no seek here after a borderline long read of the earliest sector in the cache, then the cache must not have been dumped yet. */ if (ret==1 && i && xlastsector){ reportC("\n\tToo many read errors while performing drive cache checks;" "\n\t aborting test.\n\n"); return(-1); } reportC("\n\tRead error while performing drive cache checks;" "\n\t choosing new offset and trying again.\n"); }else{ if(x==-1){ reportC("\n\tTiming error while performing drive cache checks; aborting test.\n"); return(-1); }else{ if(x paranoia_cachemodel_size(p,-1)){ reportC("\nWARNING: This drive appears to be caching more sectors of\n" " readahead than Paranoia can currently handle!\n"); warn=1; } paranoia_free(p); } if(speed==-1){ logC("\tAttempting to reset read speed to full... "); }else{ logC("\tAttempting to reset read speed to %d... ",speed); } if(cdda_speed_set(d,speed)){ logC("failed.\n"); }else{ logC("drive said OK\n"); } /* This is similar to the Fast search above, but just in case the cache is being tracked as multiple areas that are treated differently if non-contiguous.... */ { int seekoff = cachesize*3; int under=0; reportC("\tVerifying that cache is contiguous..."); for(i=0;i<20 && !under;i++){ printC("."); for(j=0;;j++){ int ret1,ret2; if(offset+seekoff>lastsector){ reportC("\n\tOut of readable space on CDROM while performing drive checks;" "\n\t aborting test.\n\n"); return(-1); } ret1 = cdda_read_timed(d,NULL,offset+seekoff,1,&x); logC("\t\t>>> %d:%d:%d ",offset+seekoff,ret1,x); ret2 = cdda_read_timed(d,NULL,offset,1,&x); logC("seek_read:%d:%d:%d\n",offset,ret2,x); if(ret1<=0 || ret2<=0){ offset+=cachesize+100; if(j==10){ reportC("\n\tToo many read errors while performing drive cache checks;" "\n\t aborting test.\n\n"); return(-1); } reportC("\n\tRead error while performing drive cache checks;" "\n\t choosing new offset and trying again.\n"); }else{ if(x==-1){ reportC("\n\tTiming error while performing drive cache checks; aborting test.\n"); return(-1); }else{ if(x1 || under){ tests++; if(tests>8 && gran<64){ gran<<=3; tests=0; it=3; } if(gran && !under){ gran>>=3; tests=0; if(gran==1)it=10; } under=0; readahead=lower+gran; printC("\r"); logC("\n"); reportC("\tTesting background readahead past read cursor... %d",readahead); printC(" \b\b\b\b\b\b\b\b\b\b\b"); for (i=0;i>> ",i); while (sofarmax)usec=max; logC("sleep=%dus ",usec); usleep(usec); } /* seek to offset+cachesize+readahead */ ret = cdda_read_timed(d,NULL,offset+cachesize+readahead-1,1,&x); if(ret<=0)break; logC("seek=%d:%d:%d",offset+cachesize+readahead-1,ret,x); if(x>> "); printC("."); while(sofar0){ sofar--; ret = cdda_read_timed(d,NULL,offset+sofar,1,&x); if(ret<=0)break; logC("%d:%d:%d ",sofar,ret,x); if(x>=MIN_SEEK_MS){ if(rollbehind != sofar+1){ rollbehind=sofar+1; i=0; }else{ i++; } break; } if(sofar==0)rollbehind=0; } error: if(ret<=0){ offset+=cachesize; retry++; if(retry>10 || offset+cachesize>lastsector){ reportC("\n\tToo many read errors while performing drive cache checks;" "\n\t aborting test.\n\n"); return(-1); } reportC("\n\tRead error while performing drive cache checks;" "\n\t choosing new offset and trying again.\n"); continue; } } /* verify that the drive timing didn't suddenly change */ { float newms=retime_drive(d, progress, log, offset, readahead, mspersector); if(newms > mspersector*1.2){ mspersector=newms; printC("\r"); reportC("\tDrive timing changed during test; retrying..."); continue; } } break; } logC("\n"); printC("\r"); if(rollbehind==0){ reportC("\tCache tail cursor tied to read cursor \n"); }else{ reportC("\tCache tail rollbehind: %d sector(s) \n",rollbehind); } reportC("\tTesting granularity of cache tail"); while(1){ cachegran=cachesize+1; for(i=0;i<10 && cachegran;){ int sofar=0,ret=0,retry=0; logC("\n\t\t>>> "); printC("."); while(sofar=MIN_SEEK_MS){ if(cachegran == sofar+1){ i++; }else{ cachegran=sofar+1; i=0; } break; } if(sofar==0)cachegran=0; } error2: if(ret<=0){ offset+=cachesize; retry++; if(retry>10 || offset+cachesize>lastsector){ reportC("\n\tToo many read errors while performing drive cache checks;" "\n\t aborting test.\n\n"); return(-1); } reportC("\n\tRead error while performing drive cache checks;" "\n\t choosing new offset and trying again.\n"); continue; } } /* verify that the drive timing didn't suddenly change */ { float newms=retime_drive(d, progress, log, offset, readahead, mspersector); if(newms > mspersector*1.2){ mspersector=newms; printC("\r"); reportC("\tDrive timing changed during test; retrying..."); continue; } } break; } cachegran -= rollbehind; logC("\n"); printC("\r"); reportC("\tCache tail granularity: %d sector(s) \n",cachegran); /* Verify that a read that begins before the cached readahead dumps the entire readahead cache */ /* This is tricky because we can't simply read a one sector back seek, then rely on timing/seeking of subsequent sectors; the drive may well not seek ahead if reading linearly would be faster (and it often will be), and simply reading ahead after the seek and watching timing will be inaccurate because the drive may roll some readahead into the initial seek/read before returning the first block. */ /* we will need to use the timing of reading from media in one form or another and thus need to guard against slow bus transfer times [eg, no DMA] swamping the actual read time from media. */ /* sample cache access for five realtime seconds. */ { float cachems; float readms; int readsize = cachesize-rollbehind-8; int retry=0; if(readsize>cachesize-1)readsize=cachesize-1; if(readsize<7){ reportC("\tCache size (considering rollbehind) too small to test cache speed.\n"); }else{ reportC("\tTesting cache transfer speed..."); /* cache timing isn't dependent on rotational speed, so get a good read and then just hammer the cache; we will only need to do it once */ /* we need to time the cache using the most conservative possible read pattern; many drives will flush cache on *any* nonlinear access, but not if the read starts from the same point. The original cache size verification algo knows this, and we need to do it the same way here (this the '0' for 'initial_seek' on time_drve */ while(1){ int ret=time_drive(d, NULL, log, offset, readsize, 0); if(ret==-404) return -1; if(ret>0)break; retry++; if(retry==10){ reportC("\n\tToo many read errors while performing drive cache checks;" "\n\t aborting test.\n\n"); return(-1); } reportC("\n\tRead error while performing drive cache checks;" "\n\t choosing new offset and trying again.\n"); } { int elapsed=0; int sectors=0; int spinner=0; while(elapsed<5000){ sectors += readsize; elapsed += time_drive(d, NULL, log, offset, readsize, 0); spinner = elapsed*5/1000%4; printC("\b%c",(spinner==0?'o':(spinner==1?'O':(spinner==2?'o':'.')))); } printC("\r"); logC("\n"); cachems = elapsed/(float)sectors; reportC("\t Cache read speed: %.2fms/sector [%dx]\n", cachems,(int)(1000./75./cachems)); } if(cachems*3>mspersector){ reportC("\tCache access insufficiently faster than media access to\n" "\t\tperform cache backseek tests\n\n"); }else{ /* now do the read/backseek combo */ reportC("\tTesting that backseek flushes cache..."); { int total=0; int elapsed=0; int sectors=0; int spinner=0; int retry=0; while(elapsed<5000 && total<25){ /* don't kill the drive */ int ret; while(1){ /* need to force seek/flush, but don't kill the drive */ int seekpos = offset+cachesize+20000; if(seekpos>lastsector-150)seekpos=lastsector-150; ret=cdda_read(d, NULL, seekpos, 1); if(ret>0) ret=time_drive(d, NULL, log, offset+1, readsize, 1); if(ret>=0) ret=time_drive(d, NULL, log, offset, readsize, 1); if(ret<=0){ retry++; if(retry==10){ reportC("\n\tToo many read errors while performing drive cache checks;" "\n\t aborting test.\n\n"); return(-1); } reportC("\n\tRead error while performing drive cache checks; retrying..."); }else break; } sectors += (readsize-1); elapsed += ret; total++; spinner = elapsed*5/1000%4; printC("\b%c",(spinner==0?'o':(spinner==1?'O':(spinner==2?'o':'.')))); } printC("\r"); logC("\n"); readms = elapsed/(float)sectors; reportC("\t Access speed after backseek: %.2fms/sector [%dx]\n", readms,(int)(1000./75./readms)); if(readms*2. < mspersector || cachems*2. > readms){ reportC("\tWARNING: Read timing after backseek faster than expected!\n" "\t It's possible/likely that this drive is not\n" "\t flushing the readahead cache on backward seeks!\n\n"); warn=1; }else{ reportC("\tBackseek flushes the cache as expected\n"); } } } } } return warn; }