From 196f5066228d908cbac060396a1aba557efe1ef3 Mon Sep 17 00:00:00 2001 From: Pedro Martelletto Date: Sat, 26 May 2007 20:26:52 +0000 Subject: Dynamic buffer cache. Initial diff from mickey@, okay art@ beck@ toby@ deraadt@ dlg@. --- sys/arch/amd64/amd64/machdep.c | 164 ++++------------------------------------- 1 file changed, 14 insertions(+), 150 deletions(-) (limited to 'sys/arch/amd64') diff --git a/sys/arch/amd64/amd64/machdep.c b/sys/arch/amd64/amd64/machdep.c index a49db688c75..823f5d96da0 100644 --- a/sys/arch/amd64/amd64/machdep.c +++ b/sys/arch/amd64/amd64/machdep.c @@ -1,4 +1,4 @@ -/* $OpenBSD: machdep.c,v 1.56 2007/05/23 20:33:46 pvalchev Exp $ */ +/* $OpenBSD: machdep.c,v 1.57 2007/05/26 20:26:50 pedro Exp $ */ /* $NetBSD: machdep.c,v 1.3 2003/05/07 22:58:18 fvdl Exp $ */ /*- @@ -178,12 +178,6 @@ int kbd_reset; struct vm_map *exec_map = NULL; struct vm_map *phys_map = NULL; -#ifdef NBUF -int nbuf = NBUF; -#else -int nbuf = 0; -#endif - #ifndef BUFCACHEPERCENT #define BUFCACHEPERCENT 10 #endif @@ -248,7 +242,7 @@ phys_ram_seg_t mem_clusters[VM_PHYSSEG_MAX]; int mem_cluster_cnt; vaddr_t allocsys(vaddr_t); -void setup_buffers(vaddr_t *); +void setup_buffers(void); int cpu_dump(void); int cpu_dumpsize(void); u_long cpu_dump_mempagecnt(void); @@ -319,11 +313,7 @@ cpu_startup(void) if (allocsys(v) - v != sz) panic("startup: table size inconsistency"); - /* - * Now allocate buffers proper. They are different than the above - * in that they usually occupy more virtual memory than physical. - */ - setup_buffers(&maxaddr); + setup_buffers(); /* * Allocate a submap for exec arguments. This map effectively @@ -342,8 +332,6 @@ cpu_startup(void) printf("avail mem = %lu (%luMB)\n", ptoa(uvmexp.free), ptoa(uvmexp.free)/1024/1024); - printf("using %u buffers containing %u bytes (%uK) of memory\n", - nbuf, bufpages * PAGE_SIZE, bufpages * PAGE_SIZE / 1024); bufinit(); @@ -359,18 +347,6 @@ cpu_startup(void) x86_bus_space_mallocok(); } - -/* - * The following defines are for the code in setup_buffers that tries to - * ensure that enough ISA DMAable memory is still left after the buffercache - * has been allocated. - */ -#define CHUNKSZ (3 * 1024 * 1024) -#define ISADMA_LIMIT (16 * 1024 * 1024) /* XXX wrong place */ -#define ALLOC_PGS(sz, limit, pgs) \ - uvm_pglistalloc((sz), 0, (limit), PAGE_SIZE, 0, &(pgs), 1, 0) -#define FREE_PGS(pgs) uvm_pglistfree(&(pgs)) - /* * Allocate space for system data structures. We are given * a starting virtual address and we return a final virtual @@ -394,136 +370,24 @@ allocsys(vaddr_t v) valloc(msqids, struct msqid_ds, msginfo.msgmni); #endif - /* - * Determine how many buffers to allocate. We use 10% of the - * first 2MB of memory, and 10% of the rest, with a minimum of 16 - * buffers. We allocate 1/2 as many swap buffer headers as file - * i/o buffers. - */ - if (bufpages == 0) { - bufpages = (btoc(2 * 1024 * 1024) + physmem) * - bufcachepercent / 100; - } - if (nbuf == 0) { - nbuf = bufpages; - if (nbuf < 16) - nbuf = 16; - } - - /* Restrict to at most 35% filled kvm */ - /* XXX - This needs UBC... */ - if (nbuf > - (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / MAXBSIZE * 35 / 100) - nbuf = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / - MAXBSIZE * 35 / 100; - - /* More buffer pages than fits into the buffers is senseless. */ - if (bufpages > nbuf * MAXBSIZE / PAGE_SIZE) - bufpages = nbuf * MAXBSIZE / PAGE_SIZE; - - valloc(buf, struct buf, nbuf); return v; } void -setup_buffers(vaddr_t *maxaddr) +setup_buffers() { - vsize_t size; - vaddr_t addr; - int base, residual, left, chunk, i; - struct pglist pgs, saved_pgs; - struct vm_page *pg; - int rv; - - size = MAXBSIZE * nbuf; - addr = vm_map_min(kernel_map); - if ((rv = uvm_map(kernel_map, &addr, round_page(size), - NULL, UVM_UNKNOWN_OFFSET, 0, - UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, - UVM_ADV_NORMAL, 0)))) - panic("cpu_startup: cannot allocate VM for buffers %d", rv); - buffers = (char *)addr; - - base = bufpages / nbuf; - residual = bufpages % nbuf; - if (base >= MAXBSIZE / PAGE_SIZE) { - /* don't want to alloc more physical mem than needed */ - base = MAXBSIZE / PAGE_SIZE; - residual = 0; - } - /* - * In case we might need DMA bouncing we have to make sure there - * is some memory below 16MB available. On machines with many - * pages reserved for the buffer cache we risk filling all of that - * area with buffer pages. We still want much of the buffers - * reside there as that lowers the probability of them needing to - * bounce, but we have to set aside some space for DMA buffers too. - * - * The current strategy is to grab hold of one 3MB chunk below 16MB - * first, which we are saving for DMA buffers, then try to get - * one chunk at a time for fs buffers, until that is not possible - * anymore, at which point we get the rest wherever we may find it. - * After that we give our saved area back. That will guarantee at - * least 3MB below 16MB left for drivers' attach routines, among - * them isadma. However we still have a potential problem of PCI - * devices attached earlier snatching that memory. This can be - * solved by making the PCI DMA memory allocation routines go for - * memory above 16MB first. + * Determine how many buffers to allocate. + * We allocate bufcachepercent% of memory for buffer space. */ - - left = bufpages; - - /* - * First, save ISA DMA bounce buffer area so we won't lose that - * capability. - */ - TAILQ_INIT(&saved_pgs); - TAILQ_INIT(&pgs); - if (!ALLOC_PGS(CHUNKSZ, ISADMA_LIMIT, saved_pgs)) { - /* - * Then, grab as much ISA DMAable memory as possible - * for the buffer cache as it is nice to not need to - * bounce all buffer I/O. - */ - for (left = bufpages; left > 0; left -= chunk) { - chunk = min(left, CHUNKSZ / PAGE_SIZE); - if (ALLOC_PGS(chunk * PAGE_SIZE, ISADMA_LIMIT, pgs)) - break; - } - } - - /* - * If we need more pages for the buffer cache, get them from anywhere. - */ - if (left > 0 && ALLOC_PGS(left * PAGE_SIZE, avail_end, pgs)) - panic("cannot get physical memory for buffer cache"); - - /* - * Finally, give back the ISA DMA bounce buffer area, so it can be - * allocated by the isadma driver later. - */ - if (!TAILQ_EMPTY(&saved_pgs)) - FREE_PGS(saved_pgs); - - pg = TAILQ_FIRST(&pgs); - for (i = 0; i < nbuf; i++) { - /* - * First buffers get (base+1) physical pages - * allocated for them. The rest get (base) physical pages. - * - * The rest of each buffer occupies virtual space, - * but has no physical memory allocated for it. - */ - addr = (vaddr_t)buffers + i * MAXBSIZE; - for (size = PAGE_SIZE * (i < residual ? base + 1 : base); - size > 0; size -= PAGE_SIZE, addr += PAGE_SIZE) { - pmap_kenter_pa(addr, VM_PAGE_TO_PHYS(pg), - VM_PROT_READ|VM_PROT_WRITE); - pg = TAILQ_NEXT(pg, pageq); - } - } - pmap_update(pmap_kernel()); + if (bufpages == 0) + bufpages = physmem * bufcachepercent / 100; + + /* Restrict to at most 25% filled kvm */ + if (bufpages > + (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) + bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / + PAGE_SIZE / 4; } /* -- cgit v1.2.3