perfmon.c

这个linux源代码是很全面的~基本完整了~使用c编译的~由于时间问题我没有亲自测试~但就算用来做参考资料也是非常好的

 * Generates a unique (per CPU) timestamp
 */
static inline unsigned long
pfm_get_stamp(void)
{
	/*
	 * XXX: must find something more efficient
	 */
	return ia64_get_itc();
}

/* Given PGD from the address space's page table, return the kernel
 * virtual mapping of the physical memory mapped at ADR.
 */
static inline unsigned long
uvirt_to_kva(pgd_t *pgd, unsigned long adr)
{
	unsigned long ret = 0UL;
	pmd_t *pmd;
	pte_t *ptep, pte;

	if (!pgd_none(*pgd)) {
		pmd = pmd_offset(pgd, adr);
		if (!pmd_none(*pmd)) {
			ptep = pte_offset(pmd, adr);
			pte = *ptep;
			if (pte_present(pte)) {
				ret = (unsigned long) page_address(pte_page(pte));
				ret |= (adr & (PAGE_SIZE - 1));
			}
		}
	}
	DBprintk(("[%d] uv2kva(%lx-->%lx)\n", current->pid, adr, ret));
	return ret;
}

/* Here we want the physical address of the memory.
 * This is used when initializing the contents of the
 * area and marking the pages as reserved.
 */
static inline unsigned long
pfm_kvirt_to_pa(unsigned long adr)
{
	__u64 pa = ia64_tpa(adr);
	//DBprintk(("kv2pa(%lx-->%lx)\n", adr, pa));
	return pa;
}

static void *
pfm_rvmalloc(unsigned long size)
{
	void *mem;
	unsigned long adr, page;

	mem=vmalloc(size);
	if (mem) {
		//printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
		memset(mem, 0, size); /* Clear the ram out, no junk to the user */
		adr=(unsigned long) mem;
		while (size > 0) {
			page = pfm_kvirt_to_pa(adr);
			mem_map_reserve(virt_to_page(__va(page)));
			adr  += PAGE_SIZE;
			size -= PAGE_SIZE;
		}
	}
	return mem;
}

static void
pfm_rvfree(void *mem, unsigned long size)
{
	unsigned long adr, page = 0;

	if (mem) {
		adr=(unsigned long) mem;
		while (size > 0) {
			page = pfm_kvirt_to_pa(adr);
			mem_map_unreserve(virt_to_page(__va(page)));
			adr+=PAGE_SIZE;
			size-=PAGE_SIZE;
		}
		vfree(mem);
	}
	return;
}

/*
 * This function gets called from mm/mmap.c:exit_mmap() only when there is a sampling buffer
 * attached to the context AND the current task has a mapping for it, i.e., it is the original
 * creator of the context.
 *
 * This function is used to remember the fact that the vma describing the sampling buffer
 * has now been removed. It can only be called when no other tasks share the same mm context.
 *
 */
static void 
pfm_vm_close(struct vm_area_struct *vma)
{
	pfm_smpl_buffer_desc_t *psb = (pfm_smpl_buffer_desc_t *)vma->vm_private_data;

	if (psb == NULL) {
		printk("perfmon: psb is null in [%d]\n", current->pid);
		return;
	}
	/*
	 * Add PSB to list of buffers to free on release_thread() when no more users
	 *
	 * This call is safe because, once the count is zero is cannot be modified anymore.
	 * This is not because there is no more user of the mm context, that the sampling
	 * buffer is not being used anymore outside of this task. In fact, it can still
	 * be accessed from within the kernel by another task (such as the monitored task).
	 *
	 * Therefore, we only move the psb into the list of buffers to free when we know
	 * nobody else is using it.
	 * The linked list if independent of the perfmon context, because in the case of
	 * multi-threaded processes, the last thread may not have been involved with
	 * monitoring however it will be the one removing the vma and it should therefore
	 * also remove the sampling buffer. This buffer cannot be removed until the vma
	 * is removed.
	 *
	 * This function cannot remove the buffer from here, because exit_mmap() must first
	 * complete. Given that there is no other vma related callback in the generic code,
	 * we have created our own with the linked list of sampling buffers to free. The list
	 * is part of the thread structure. In release_thread() we check if the list is
	 * empty. If not we call into perfmon to free the buffer and psb. That is the only
	 * way to ensure a safe deallocation of the sampling buffer which works when
	 * the buffer is shared between distinct processes or with multi-threaded programs.
	 *
	 * We need to lock the psb because the refcnt test and flag manipulation must
	 * looked like an atomic operation vis a vis pfm_context_exit()
	 */
	LOCK_PSB(psb);

	if (psb->psb_refcnt == 0) {

		psb->psb_next = current->thread.pfm_smpl_buf_list;
		current->thread.pfm_smpl_buf_list = psb;

		DBprintk(("[%d] add smpl @%p size %lu to smpl_buf_list psb_flags=0x%x\n", 
			current->pid, psb->psb_hdr, psb->psb_size, psb->psb_flags));
	}
	DBprintk(("[%d] clearing psb_flags=0x%x smpl @%p size %lu\n", 
			current->pid, psb->psb_flags, psb->psb_hdr, psb->psb_size));
	/*
	 * decrement the number vma for the buffer
	 */
	psb->psb_flags &= ~PSB_HAS_VMA;

	UNLOCK_PSB(psb);
}

/*
 * This function is called from pfm_destroy_context() and also from pfm_inherit()
 * to explicitely remove the sampling buffer mapping from the user level address space.
 */
static int
pfm_remove_smpl_mapping(struct task_struct *task)
{
	pfm_context_t *ctx = task->thread.pfm_context;
	pfm_smpl_buffer_desc_t *psb;
	int r;

	/*
	 * some sanity checks first
	 */
	if (ctx == NULL || task->mm == NULL || ctx->ctx_smpl_vaddr == 0 || ctx->ctx_psb == NULL) {
		printk("perfmon: invalid context mm=%p\n", task->mm);
		return -1;
	}
	psb = ctx->ctx_psb;

	down_write(&task->mm->mmap_sem);

	r = do_munmap(task->mm, ctx->ctx_smpl_vaddr, psb->psb_size);

	up_write(&task->mm->mmap_sem);
	if (r !=0) {
		printk("perfmon: pid %d unable to unmap sampling buffer @0x%lx size=%ld\n", 
				task->pid, ctx->ctx_smpl_vaddr, psb->psb_size);
	}

	DBprintk(("[%d] do_unmap(0x%lx, %ld)=%d refcnt=%lu psb_flags=0x%x\n", 
		task->pid, ctx->ctx_smpl_vaddr, psb->psb_size, r, psb->psb_refcnt, psb->psb_flags));

	return 0;
}

static pfm_context_t *
pfm_context_alloc(void)
{
	pfm_context_t *ctx;

	/* allocate context descriptor */
	ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
	if (ctx) memset(ctx, 0, sizeof(pfm_context_t));
	
	return ctx;
}

static void
pfm_context_free(pfm_context_t *ctx)
{
	if (ctx) kfree(ctx);
}

static int
pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
{
	unsigned long page;

	DBprintk(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));

	while (size > 0) {
		page = pfm_kvirt_to_pa(buf);

		if (remap_page_range(vma, addr, page, PAGE_SIZE, PAGE_READONLY)) return -ENOMEM;
		
		addr  += PAGE_SIZE;
		buf   += PAGE_SIZE;
		size  -= PAGE_SIZE;
	}
	return 0;
}

/*
 * counts the number of PMDS to save per entry.
 * This code is generic enough to accomodate more than 64 PMDS when they become available
 */
static unsigned long
pfm_smpl_entry_size(unsigned long *which, unsigned long size)
{
	unsigned long res = 0;
	int i;

	for (i=0; i < size; i++, which++) res += hweight64(*which);

	DBprintk(("weight=%ld\n", res));

	return res;
}

/*
 * Allocates the sampling buffer and remaps it into caller's address space
 */
static int
pfm_smpl_buffer_alloc(pfm_context_t *ctx, unsigned long *which_pmds, unsigned long entries, 
		      void **user_vaddr)
{
	struct mm_struct *mm = current->mm;
	struct vm_area_struct *vma = NULL;
	unsigned long size, regcount;
	void *smpl_buf;
	pfm_smpl_buffer_desc_t *psb;


	/* note that regcount might be 0, in this case only the header for each
	 * entry will be recorded.
	 */
	regcount = pfm_smpl_entry_size(which_pmds, 1);

	if ((sizeof(perfmon_smpl_hdr_t)+ entries*sizeof(perfmon_smpl_entry_t)) <= entries) {
		DBprintk(("requested entries %lu is too big\n", entries));
		return -EINVAL;
	}

	/*
	 * 1 buffer hdr and for each entry a header + regcount PMDs to save
	 */
	size = PAGE_ALIGN(  sizeof(perfmon_smpl_hdr_t)
			  + entries * (sizeof(perfmon_smpl_entry_t) + regcount*sizeof(u64)));

	DBprintk(("sampling buffer size=%lu bytes\n", size));

	/*
	 * check requested size to avoid Denial-of-service attacks
	 * XXX: may have to refine this test	
	 * Check against address space limit.
	 *
	 * if ((mm->total_vm << PAGE_SHIFT) + len> current->rlim[RLIMIT_AS].rlim_cur) 
	 * 	return -ENOMEM;
	 */
	if (size > current->rlim[RLIMIT_MEMLOCK].rlim_cur) return -EAGAIN;

	/*
	 * We do the easy to undo allocations first.
 	 *
	 * pfm_rvmalloc(), clears the buffer, so there is no leak
	 */
	smpl_buf = pfm_rvmalloc(size);
	if (smpl_buf == NULL) {
		DBprintk(("Can't allocate sampling buffer\n"));
		return -ENOMEM;
	}

	DBprintk(("smpl_buf @%p\n", smpl_buf));

	/* allocate sampling buffer descriptor now */
	psb = kmalloc(sizeof(*psb), GFP_KERNEL);
	if (psb == NULL) {
		DBprintk(("Can't allocate sampling buffer descriptor\n"));
		goto error_kmalloc;
	}

	/* allocate vma */
	vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
	if (!vma) {
		DBprintk(("Cannot allocate vma\n"));
		goto error_kmem;
	}
	/*
	 * partially initialize the vma for the sampling buffer
	 *
	 * The VM_DONTCOPY flag is very important as it ensures that the mapping
	 * will never be inherited for any child process (via fork()) which is always 
	 * what we want.
	 */
	vma->vm_mm	     = mm;
	vma->vm_flags	     = VM_READ| VM_MAYREAD |VM_RESERVED|VM_DONTCOPY;
	vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
	vma->vm_ops	     = &pfm_vm_ops; /* necesarry to get the close() callback */
	vma->vm_pgoff	     = 0;
	vma->vm_file	     = NULL;
	vma->vm_raend	     = 0;
	vma->vm_private_data = psb;	/* information needed by the pfm_vm_close() function */

	/*
	 * Now we have everything we need and we can initialize
	 * and connect all the data structures
	 */

	psb->psb_hdr	 = smpl_buf;
	psb->psb_addr    = ((char *)smpl_buf)+sizeof(perfmon_smpl_hdr_t); /* first entry */
	psb->psb_size    = size; /* aligned size */
	psb->psb_index   = 0;
	psb->psb_entries = entries;
	psb->psb_refcnt  = 1;
	psb->psb_flags   = PSB_HAS_VMA;

	spin_lock_init(&psb->psb_lock);

	/*
	 * XXX: will need to do cacheline alignment to avoid false sharing in SMP mode and
	 * multitask monitoring.
	 */
	psb->psb_entry_size = sizeof(perfmon_smpl_entry_t) + regcount*sizeof(u64);

	DBprintk(("psb @%p entry_size=%ld hdr=%p addr=%p refcnt=%lu psb_flags=0x%x\n", 
		  (void *)psb,psb->psb_entry_size, (void *)psb->psb_hdr, 
		  (void *)psb->psb_addr, psb->psb_refcnt, psb->psb_flags));

	/* initialize some of the fields of user visible buffer header */
	psb->psb_hdr->hdr_version    = PFM_SMPL_VERSION;
	psb->psb_hdr->hdr_entry_size = psb->psb_entry_size;
	psb->psb_hdr->hdr_pmds[0]    = which_pmds[0];

	/*
	 * Let's do the difficult operations next.
	 *
	 * now we atomically find some area in the address space and
	 * remap the buffer in it.
	 */
	down_write(&current->mm->mmap_sem);


	/* find some free area in address space, must have mmap sem held */
	vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
	if (vma->vm_start == 0UL) {
		DBprintk(("Cannot find unmapped area for size %ld\n", size));
		up_write(&current->mm->mmap_sem);
		goto error;
	}
	vma->vm_end = vma->vm_start + size;

	DBprintk(("entries=%ld aligned size=%ld, unmapped @0x%lx\n", entries, size, vma->vm_start));
		
	/* can only be applied to current, need to have the mm semaphore held when called */
	if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
		DBprintk(("Can't remap buffer\n"));
		up_write(&current->mm->mmap_sem);
		goto error;
	}

	/*
	 * now insert the vma in the vm list for the process, must be
	 * done with mmap lock held
	 */
	insert_vm_struct(mm, vma);

	mm->total_vm  += size >> PAGE_SHIFT;

	up_write(&current->mm->mmap_sem);

	/* store which PMDS to record */
	ctx->ctx_smpl_regs[0] = which_pmds[0];


	/* link to perfmon context */
	ctx->ctx_psb        = psb;

	/*
	 * keep track of user level virtual address 
	 */
	ctx->ctx_smpl_vaddr = *(unsigned long *)user_vaddr = vma->vm_start;

	return 0;

error:
	kmem_cache_free(vm_area_cachep, vma);
error_kmem:
	kfree(psb);
error_kmalloc:
	pfm_rvfree(smpl_buf, size);
	return -ENOMEM;
}

static int
pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned long cpu_mask)
{
	unsigned long m, undo_mask;
	unsigned int n, i;

	/*
	 * validy checks on cpu_mask have been done upstream
	 */
	LOCK_PFS();

	if (is_syswide) {
		/* 
		 * cannot mix system wide and per-task sessions
		 */
		if (pfm_sessions.pfs_task_sessions > 0UL) {
			DBprintk(("system wide not possible, %u conflicting task_sessions\n", 
			  	pfm_sessions.pfs_task_sessions));
			goto abort;
		}

		m = cpu_mask; undo_mask = 0UL; n = 0;
		for(i=0; m; i++, m>>=1) {

			if ((m & 0x1) == 0UL) continue;

			if (pfm_sessions.pfs_sys_session[i]) goto undo;

			DBprintk(("reserving CPU%d currently on CPU%d\n", i, smp_processor_id()));

			pfm_sessions.pfs_sys_session[i] = task;
			undo_mask |= 1UL << i;
			n++;
		}
		pfm_sessions.pfs_sys_sessions += n;
	} else {
		if (pfm_sessions.pfs_sys_sessions) goto abort;
		pfm_sessions.pfs_task_sessions++;
	}
	UNLOCK_PFS();
	return 0;
undo:
	DBprintk(("system wide not possible, conflicting session [%d] on CPU%d\n",
  		pfm_sessions.pfs_sys_session[i]->pid, i));

	for(i=0; undo_mask; i++, undo_mask >>=1) {
		pfm_sessions.pfs_sys_session[i] = NULL;
	}
abort: