From owner-svn-src-all@freebsd.org Mon Dec 2 22:44:36 2019 Return-Path: Delivered-To: svn-src-all@mailman.nyi.freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2610:1c1:1:606c::19:1]) by mailman.nyi.freebsd.org (Postfix) with ESMTP id 363971BB148; Mon, 2 Dec 2019 22:44:36 +0000 (UTC) (envelope-from jeff@FreeBSD.org) Received: from mxrelay.nyi.freebsd.org (mxrelay.nyi.freebsd.org [IPv6:2610:1c1:1:606c::19:3]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) server-signature RSA-PSS (4096 bits) client-signature RSA-PSS (4096 bits) client-digest SHA256) (Client CN "mxrelay.nyi.freebsd.org", Issuer "Let's Encrypt Authority X3" (verified OK)) by mx1.freebsd.org (Postfix) with ESMTPS id 47RgCm1QB3z4VyR; Mon, 2 Dec 2019 22:44:36 +0000 (UTC) (envelope-from jeff@FreeBSD.org) Received: from repo.freebsd.org (repo.freebsd.org [IPv6:2610:1c1:1:6068::e6a:0]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (Client did not present a certificate) by mxrelay.nyi.freebsd.org (Postfix) with ESMTPS id 13ACA6387; Mon, 2 Dec 2019 22:44:36 +0000 (UTC) (envelope-from jeff@FreeBSD.org) Received: from repo.freebsd.org ([127.0.1.37]) by repo.freebsd.org (8.15.2/8.15.2) with ESMTP id xB2MiZTN045508; Mon, 2 Dec 2019 22:44:35 GMT (envelope-from jeff@FreeBSD.org) Received: (from jeff@localhost) by repo.freebsd.org (8.15.2/8.15.2/Submit) id xB2MiYDH045503; Mon, 2 Dec 2019 22:44:34 GMT (envelope-from jeff@FreeBSD.org) Message-Id: <201912022244.xB2MiYDH045503@repo.freebsd.org> X-Authentication-Warning: repo.freebsd.org: jeff set sender to jeff@FreeBSD.org using -f From: Jeff Roberson Date: Mon, 2 Dec 2019 22:44:34 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: svn commit: r355315 - in head/sys: kern sys vm X-SVN-Group: head X-SVN-Commit-Author: jeff X-SVN-Commit-Paths: in head/sys: kern sys vm X-SVN-Commit-Revision: 355315 X-SVN-Commit-Repository: base MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit X-BeenThere: svn-src-all@freebsd.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: "SVN commit messages for the entire src tree \(except for " user" and " projects" \)" List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Mon, 02 Dec 2019 22:44:36 -0000 Author: jeff Date: Mon Dec 2 22:44:34 2019 New Revision: 355315 URL: https://svnweb.freebsd.org/changeset/base/355315 Log: Use a precise bit count for the slab free items in UMA. This significantly shrinks embedded slab structures. Reviewed by: markj, rlibby (prior version) Differential Revision: https://reviews.freebsd.org/D22584 Modified: head/sys/kern/subr_vmem.c head/sys/sys/bitset.h head/sys/vm/uma_core.c head/sys/vm/uma_int.h head/sys/vm/vm_page.c Modified: head/sys/kern/subr_vmem.c ============================================================================== --- head/sys/kern/subr_vmem.c Mon Dec 2 22:42:05 2019 (r355314) +++ head/sys/kern/subr_vmem.c Mon Dec 2 22:44:34 2019 (r355315) @@ -670,8 +670,8 @@ int vmem_startup_count(void) { - return (howmany(BT_MAXALLOC, - UMA_SLAB_SPACE / sizeof(struct vmem_btag))); + return (howmany(BT_MAXALLOC, slab_ipers(sizeof(struct vmem_btag), + UMA_ALIGN_PTR))); } #endif Modified: head/sys/sys/bitset.h ============================================================================== --- head/sys/sys/bitset.h Mon Dec 2 22:42:05 2019 (r355314) +++ head/sys/sys/bitset.h Mon Dec 2 22:44:34 2019 (r355315) @@ -246,10 +246,11 @@ #define BITSET_FSET(n) \ [ 0 ... ((n) - 1) ] = (-1L) +#define BITSET_SIZE(_s) (__bitset_words((_s)) * sizeof(long)) + /* * Dynamically allocate a bitset. */ -#define BITSET_ALLOC(_s, mt, mf) \ - malloc(__bitset_words(_s) * sizeof(long), mt, (mf)) +#define BITSET_ALLOC(_s, mt, mf) malloc(BITSET_SIZE((_s)), mt, (mf)) #endif /* !_SYS_BITSET_H_ */ Modified: head/sys/vm/uma_core.c ============================================================================== --- head/sys/vm/uma_core.c Mon Dec 2 22:42:05 2019 (r355314) +++ head/sys/vm/uma_core.c Mon Dec 2 22:44:34 2019 (r355315) @@ -1200,9 +1200,9 @@ keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int dom slab->us_freecount = keg->uk_ipers; slab->us_flags = sflags; slab->us_domain = domain; - BIT_FILL(SLAB_SETSIZE, &slab->us_free); + BIT_FILL(keg->uk_ipers, &slab->us_free); #ifdef INVARIANTS - BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); + BIT_ZERO(SLAB_MAX_SETSIZE, &slab->us_debugfree); #endif if (keg->uk_init != NULL) { @@ -1486,6 +1486,46 @@ zero_init(void *mem, int size, int flags) } /* + * Actual size of embedded struct slab (!OFFPAGE). + */ +size_t +slab_sizeof(int nitems) +{ + size_t s; + + s = sizeof(struct uma_slab) + BITSET_SIZE(nitems); + return (roundup(s, UMA_ALIGN_PTR + 1)); +} + +/* + * Size of memory for embedded slabs (!OFFPAGE). + */ +size_t +slab_space(int nitems) +{ + return (UMA_SLAB_SIZE - slab_sizeof(nitems)); +} + +/* + * Compute the number of items that will fit in an embedded (!OFFPAGE) slab + * with a given size and alignment. + */ +int +slab_ipers(size_t size, int align) +{ + int rsize; + int nitems; + + /* + * Compute the ideal number of items that will fit in a page and + * then compute the actual number based on a bitset nitems wide. + */ + rsize = roundup(size, align + 1); + nitems = UMA_SLAB_SIZE / rsize; + return (slab_space(nitems) / rsize); +} + +/* * Finish creating a small uma keg. This calculates ipers, and the keg size. * * Arguments @@ -1519,20 +1559,25 @@ keg_small_init(uma_keg_t keg) * allocation bits for we round it up. */ rsize = keg->uk_size; - if (rsize < slabsize / SLAB_SETSIZE) - rsize = slabsize / SLAB_SETSIZE; + if (rsize < slabsize / SLAB_MAX_SETSIZE) + rsize = slabsize / SLAB_MAX_SETSIZE; if (rsize & keg->uk_align) - rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); + rsize = roundup(rsize, keg->uk_align + 1); keg->uk_rsize = rsize; KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || keg->uk_rsize < UMA_PCPU_ALLOC_SIZE, ("%s: size %u too large", __func__, keg->uk_rsize)); + /* + * Use a pessimistic bit count for shsize. It may be possible to + * squeeze one more item in for very particular sizes if we were + * to loop and reduce the bitsize if there is waste. + */ if (keg->uk_flags & UMA_ZONE_OFFPAGE) shsize = 0; else - shsize = SIZEOF_UMA_SLAB; + shsize = slab_sizeof(slabsize / rsize); if (rsize <= slabsize - shsize) keg->uk_ipers = (slabsize - shsize) / rsize; @@ -1543,7 +1588,7 @@ keg_small_init(uma_keg_t keg) ("%s: size %u greater than slab", __func__, keg->uk_size)); keg->uk_ipers = 1; } - KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, + KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); memused = keg->uk_ipers * rsize + shsize; @@ -1571,7 +1616,7 @@ keg_small_init(uma_keg_t keg) if ((wastedspace >= slabsize / UMA_MAX_WASTE) && (keg->uk_ipers < (slabsize / keg->uk_rsize))) { keg->uk_ipers = slabsize / keg->uk_rsize; - KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, + KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_MAX_SETSIZE, ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); CTR6(KTR_UMA, "UMA decided we need offpage slab headers for " "keg: %s(%p), calculated wastedspace = %d, " @@ -1620,7 +1665,8 @@ keg_large_init(uma_keg_t keg) /* Check whether we have enough space to not do OFFPAGE. */ if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0 && - PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < SIZEOF_UMA_SLAB) { + PAGE_SIZE * keg->uk_ppera - keg->uk_rsize < + slab_sizeof(SLAB_MIN_SETSIZE)) { /* * We can't do OFFPAGE if we're internal, in which case * we need an extra page per allocation to contain the @@ -1667,7 +1713,7 @@ keg_cachespread_init(uma_keg_t keg) keg->uk_ppera = pages; keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; - KASSERT(keg->uk_ipers <= SLAB_SETSIZE, + KASSERT(keg->uk_ipers <= SLAB_MAX_SETSIZE, ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, keg->uk_ipers)); } @@ -1730,7 +1776,7 @@ keg_ctor(void *mem, int size, void *udata, int flags) if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { keg_cachespread_init(keg); } else { - if (keg->uk_size > UMA_SLAB_SPACE) + if (keg->uk_size > slab_space(SLAB_MIN_SETSIZE)) keg_large_init(keg); else keg_small_init(keg); @@ -1770,11 +1816,14 @@ keg_ctor(void *mem, int size, void *udata, int flags) /* * If we're putting the slab header in the actual page we need to - * figure out where in each page it goes. See SIZEOF_UMA_SLAB - * macro definition. + * figure out where in each page it goes. See slab_sizeof + * definition. */ if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { - keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - SIZEOF_UMA_SLAB; + size_t shsize; + + shsize = slab_sizeof(keg->uk_ipers); + keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - shsize; /* * The only way the following is possible is if with our * UMA_ALIGN_PTR adjustments we are now bigger than @@ -1782,8 +1831,7 @@ keg_ctor(void *mem, int size, void *udata, int flags) * mathematically possible for all cases, so we make * sure here anyway. */ - KASSERT(keg->uk_pgoff + sizeof(struct uma_slab) <= - PAGE_SIZE * keg->uk_ppera, + KASSERT(keg->uk_pgoff + shsize <= PAGE_SIZE * keg->uk_ppera, ("zone %s ipers %d rsize %d size %d slab won't fit", zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size)); } @@ -2245,6 +2293,7 @@ int uma_startup_count(int vm_zones) { int zones, pages; + size_t space, size; ksize = sizeof(struct uma_keg) + (sizeof(struct uma_domain) * vm_ndomains); @@ -2265,39 +2314,29 @@ uma_startup_count(int vm_zones) zones = UMA_BOOT_ZONES + vm_zones; vm_zones = 0; #endif + size = slab_sizeof(SLAB_MAX_SETSIZE); + space = slab_space(SLAB_MAX_SETSIZE); /* Memory for the rest of startup zones, UMA and VM, ... */ - if (zsize > UMA_SLAB_SPACE) { + if (zsize > space) { /* See keg_large_init(). */ u_int ppera; ppera = howmany(roundup2(zsize, UMA_BOOT_ALIGN), PAGE_SIZE); - if (PAGE_SIZE * ppera - roundup2(zsize, UMA_BOOT_ALIGN) < - SIZEOF_UMA_SLAB) + if (PAGE_SIZE * ppera - roundup2(zsize, UMA_BOOT_ALIGN) < size) ppera++; pages += (zones + vm_zones) * ppera; - } else if (roundup2(zsize, UMA_BOOT_ALIGN) > UMA_SLAB_SPACE) + } else if (roundup2(zsize, UMA_BOOT_ALIGN) > space) /* See keg_small_init() special case for uk_ppera = 1. */ pages += zones; else pages += howmany(zones, - UMA_SLAB_SPACE / roundup2(zsize, UMA_BOOT_ALIGN)); + space / roundup2(zsize, UMA_BOOT_ALIGN)); /* ... and their kegs. Note that zone of zones allocates a keg! */ pages += howmany(zones + 1, - UMA_SLAB_SPACE / roundup2(ksize, UMA_BOOT_ALIGN)); + space / roundup2(ksize, UMA_BOOT_ALIGN)); - /* - * Most of startup zones are not going to be offpages, that's - * why we use UMA_SLAB_SPACE instead of UMA_SLAB_SIZE in all - * calculations. Some large bucket zones will be offpage, and - * thus will allocate hashes. We take conservative approach - * and assume that all zones may allocate hash. This may give - * us some positive inaccuracy, usually an extra single page. - */ - pages += howmany(zones, UMA_SLAB_SPACE / - (sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT)); - return (pages); } @@ -2355,7 +2394,7 @@ uma_startup(void *mem, int npages) /* Now make a zone for slab headers */ slabzone = uma_zcreate("UMA Slabs", - sizeof(struct uma_slab), + slab_sizeof(SLAB_MAX_SETSIZE), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); @@ -3056,8 +3095,8 @@ slab_alloc_item(uma_keg_t keg, uma_slab_t slab) KEG_LOCK_ASSERT(keg); - freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; - BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); + freei = BIT_FFS(keg->uk_ipers, &slab->us_free) - 1; + BIT_CLR(keg->uk_ipers, freei, &slab->us_free); item = slab->us_data + (keg->uk_rsize * freei); slab->us_freecount--; keg->uk_free--; @@ -3569,7 +3608,7 @@ slab_free_item(uma_zone_t zone, uma_slab_t slab, void /* Slab management. */ freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; - BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); + BIT_SET(keg->uk_ipers, freei, &slab->us_free); slab->us_freecount++; /* Keg statistics. */ @@ -4450,10 +4489,10 @@ uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void * keg = zone->uz_keg; freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; - if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) + if (BIT_ISSET(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree)) panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); - BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); + BIT_SET_ATOMIC(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree); return; } @@ -4486,11 +4525,11 @@ uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *i panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); - if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) + if (!BIT_ISSET(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree)) panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); - BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); + BIT_CLR_ATOMIC(SLAB_MAX_SETSIZE, freei, &slab->us_debugfree); } #endif /* INVARIANTS */ Modified: head/sys/vm/uma_int.h ============================================================================== --- head/sys/vm/uma_int.h Mon Dec 2 22:42:05 2019 (r355314) +++ head/sys/vm/uma_int.h Mon Dec 2 22:44:34 2019 (r355315) @@ -140,19 +140,6 @@ #define UMA_MAX_WASTE 10 /* - * Actual size of uma_slab when it is placed at an end of a page - * with pointer sized alignment requirement. - */ -#define SIZEOF_UMA_SLAB ((sizeof(struct uma_slab) & UMA_ALIGN_PTR) ? \ - (sizeof(struct uma_slab) & ~UMA_ALIGN_PTR) + \ - (UMA_ALIGN_PTR + 1) : sizeof(struct uma_slab)) - -/* - * Size of memory in a not offpage single page slab available for actual items. - */ -#define UMA_SLAB_SPACE (PAGE_SIZE - SIZEOF_UMA_SLAB) - -/* * I doubt there will be many cases where this is exceeded. This is the initial * size of the hash table for uma_slabs that are managed off page. This hash * does expand by powers of two. Currently it doesn't get smaller. @@ -273,8 +260,10 @@ typedef struct uma_keg * uma_keg_t; /* * Free bits per-slab. */ -#define SLAB_SETSIZE (PAGE_SIZE / UMA_SMALLEST_UNIT) -BITSET_DEFINE(slabbits, SLAB_SETSIZE); +#define SLAB_MAX_SETSIZE (PAGE_SIZE / UMA_SMALLEST_UNIT) +#define SLAB_MIN_SETSIZE _BITSET_BITS +BITSET_DEFINE(slabbits, SLAB_MAX_SETSIZE); +BITSET_DEFINE(noslabbits, 0); /* * The slab structure manages a single contiguous allocation from backing @@ -284,13 +273,13 @@ struct uma_slab { LIST_ENTRY(uma_slab) us_link; /* slabs in zone */ SLIST_ENTRY(uma_slab) us_hlink; /* Link for hash table */ uint8_t *us_data; /* First item */ - struct slabbits us_free; /* Free bitmask. */ -#ifdef INVARIANTS - struct slabbits us_debugfree; /* Debug bitmask. */ -#endif uint16_t us_freecount; /* How many are free? */ uint8_t us_flags; /* Page flags see uma.h */ uint8_t us_domain; /* Backing NUMA domain. */ +#ifdef INVARIANTS + struct slabbits us_debugfree; /* Debug bitmask. */ +#endif + struct noslabbits us_free; /* Free bitmask. */ }; #if MAXMEMDOM >= 255 @@ -298,6 +287,11 @@ struct uma_slab { #endif typedef struct uma_slab * uma_slab_t; + +/* These three functions are for embedded (!OFFPAGE) use only. */ +size_t slab_sizeof(int nitems); +size_t slab_space(int nitems); +int slab_ipers(size_t size, int align); TAILQ_HEAD(uma_bucketlist, uma_bucket); Modified: head/sys/vm/vm_page.c ============================================================================== --- head/sys/vm/vm_page.c Mon Dec 2 22:42:05 2019 (r355314) +++ head/sys/vm/vm_page.c Mon Dec 2 22:44:34 2019 (r355315) @@ -600,7 +600,7 @@ vm_page_startup(vm_offset_t vaddr) boot_pages += vmem_startup_count(); /* vm_map_startup() calls uma_prealloc(). */ boot_pages += howmany(MAX_KMAP, - UMA_SLAB_SPACE / sizeof(struct vm_map)); + slab_ipers(sizeof(struct vm_map), UMA_ALIGN_PTR)); /* * Before going fully functional kmem_init() does allocation