* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define DEBUG_SUBSYSTEM S_LLITE
-//#include <lustre_mdc.h>
#include <lustre_lite.h>
#include "llite_internal.h"
#include <linux/lustre_compat25.h>
-#define VMA_DEBUG(vma, fmt, arg...) \
- CDEBUG(D_MMAP, "vma(%p) start(%ld) end(%ld) pgoff(%ld) inode(%p) " \
- "ino(%lu) iname(%s): " fmt, vma, vma->vm_start, vma->vm_end, \
- vma->vm_pgoff, vma->vm_file->f_dentry->d_inode, \
- vma->vm_file->f_dentry->d_inode->i_ino, \
- vma->vm_file->f_dentry->d_iname, ## arg); \
-
struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
int *type);
+static struct vm_operations_struct ll_file_vm_ops;
+
void policy_from_vma(ldlm_policy_data_t *policy,
struct vm_area_struct *vma, unsigned long addr,
size_t count)
spin_lock(&mm->page_table_lock);
for(vma = find_vma(mm, addr);
vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
- if (vma->vm_ops && vma->vm_ops->nopage == ll_nopage &&
+ if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
vma->vm_flags & VM_SHARED) {
ret = vma;
break;
}
/**
+ * API independent part for page fault initialization.
+ * \param vma - virtual memory area addressed to page fault
+ * \param env - corespondent lu_env to processing
+ * \param nest - nested level
+ * \param index - page index corespondent to fault.
+ * \parm ra_flags - vma readahead flags.
+ *
+ * \return allocated and initialized env for fault operation.
+ * \retval EINVAL if env can't allocated
+ * \return other error codes from cl_io_init.
+ */
+struct cl_io *ll_fault_io_init(struct vm_area_struct *vma,
+ struct lu_env **env_ret,
+ struct cl_env_nest *nest,
+ pgoff_t index, unsigned long *ra_flags)
+{
+ struct file *file = vma->vm_file;
+ struct inode *inode = file->f_dentry->d_inode;
+ struct cl_io *io;
+ struct cl_fault_io *fio;
+ struct lu_env *env;
+ ENTRY;
+
+ *env_ret = NULL;
+ if (ll_file_nolock(file))
+ RETURN(ERR_PTR(-EOPNOTSUPP));
+
+ /*
+ * page fault can be called when lustre IO is
+ * already active for the current thread, e.g., when doing read/write
+ * against user level buffer mapped from Lustre buffer. To avoid
+ * stomping on existing context, optionally force an allocation of a new
+ * one.
+ */
+ env = cl_env_nested_get(nest);
+ if (IS_ERR(env))
+ RETURN(ERR_PTR(-EINVAL));
+
+ *env_ret = env;
+
+ io = ccc_env_thread_io(env);
+ io->ci_obj = ll_i2info(inode)->lli_clob;
+ LASSERT(io->ci_obj != NULL);
+
+ fio = &io->u.ci_fault;
+ fio->ft_index = index;
+ fio->ft_executable = vma->vm_flags&VM_EXEC;
+
+ /*
+ * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
+ * the kernel will not read other pages not covered by ldlm in
+ * filemap_nopage. we do our readahead in ll_readpage.
+ */
+ if (ra_flags != NULL)
+ *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
+ vma->vm_flags &= ~VM_SEQ_READ;
+ vma->vm_flags |= VM_RAND_READ;
+
+ CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
+ fio->ft_index, fio->ft_executable);
+
+ if (cl_io_init(env, io, CIT_FAULT, io->ci_obj) == 0) {
+ struct ccc_io *cio = ccc_env_io(env);
+ struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
+
+ LASSERT(cio->cui_cl.cis_io == io);
+
+ /* mmap lock must be MANDATORY
+ * it has to cache pages. */
+ io->ci_lockreq = CILR_MANDATORY;
+
+ cio->cui_fd = fd;
+ }
+
+ return io;
+}
+
+/* Sharing code of page_mkwrite method for rhel5 and rhel6 */
+static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
+ bool *retry)
+{
+ struct lu_env *env;
+ struct cl_io *io;
+ struct vvp_io *vio;
+ struct cl_env_nest nest;
+ int result;
+ ENTRY;
+
+ LASSERT(vmpage != NULL);
+
+ io = ll_fault_io_init(vma, &env, &nest, vmpage->index, NULL);
+ if (IS_ERR(io))
+ GOTO(out, result = PTR_ERR(io));
+
+ result = io->ci_result;
+ if (result < 0)
+ GOTO(out, result);
+
+ /* Don't enqueue new locks for page_mkwrite().
+ * If the lock has been cancelled then page must have been
+ * truncated, in that case, kernel will handle it.
+ */
+ io->ci_lockreq = CILR_PEEK;
+ io->u.ci_fault.ft_mkwrite = 1;
+ io->u.ci_fault.ft_writable = 1;
+
+ vio = vvp_env_io(env);
+ vio->u.fault.ft_vma = vma;
+ vio->u.fault.ft_vmpage = vmpage;
+
+ result = cl_io_loop(env, io);
+
+ if (result == -ENODATA) /* peek failed, no lock caching. */
+ CDEBUG(D_MMAP, "race on page_mkwrite: %lx (%lu %p)\n",
+ vma->vm_flags, io->u.ci_fault.ft_index, vmpage);
+
+ if (result == 0 || result == -ENODATA) {
+ lock_page(vmpage);
+ if (vmpage->mapping == NULL) {
+ unlock_page(vmpage);
+
+ /* page was truncated and lock was cancelled, return
+ * ENODATA so that VM_FAULT_NOPAGE will be returned
+ * to handle_mm_fault(). */
+ if (result == 0)
+ result = -ENODATA;
+ } else if (result == -ENODATA) {
+ /* Invalidate it if the cl_lock is being revoked.
+ * This piece of code is definitely needed for RHEL5,
+ * otherwise, SIGBUS will be wrongly returned to
+ * applications. */
+ write_one_page(vmpage, 1);
+ lock_page(vmpage);
+ if (vmpage->mapping != NULL) {
+ ll_invalidate_page(vmpage);
+ LASSERT(vmpage->mapping == NULL);
+ }
+ unlock_page(vmpage);
+ } else if (!PageDirty(vmpage)) {
+ /* race, the page has been cleaned by ptlrpcd after
+ * it was unlocked, it has to be added into dirty
+ * cache again otherwise this soon-to-dirty page won't
+ * consume any grants, even worse if this page is being
+ * transferred because it will break RPC checksum.
+ */
+ unlock_page(vmpage);
+
+ CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
+ "been written out, retry.\n",
+ vmpage, vmpage->index);
+
+ *retry = true;
+ result = -EAGAIN;
+ }
+ }
+ EXIT;
+
+out:
+ cl_io_fini(env, io);
+ cl_env_nested_put(&nest, env);
+
+ CDEBUG(D_MMAP, "%s mkwrite with %d\n", cfs_current()->comm, result);
+
+ LASSERT(ergo(result == 0, PageLocked(vmpage)));
+ return(result);
+}
+
+
+#ifndef HAVE_VM_OP_FAULT
+/**
* Lustre implementation of a vm_operations_struct::nopage() method, called by
* VM to server page fault (both in kernel and user space).
*
* \param address - address when hit fault
* \param type - of fault
*
- * XXX newer 2.6 kernels provide vm_operations_struct::fault() method with
- * slightly different semantics instead.
- *
- * \return allocated and filled page for address
+ * \return allocated and filled _unlocked_ page for address
* \retval NOPAGE_SIGBUS if page not exist on this address
* \retval NOPAGE_OOM not have memory for allocate new page
*/
struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
int *type)
{
- struct file *file = vma->vm_file;
- struct inode *inode = file->f_dentry->d_inode;
- struct lu_env *env;
- struct cl_io *io;
- struct page *page = NULL;
- struct cl_env_nest nest;
- int result;
+ struct lu_env *env;
+ struct cl_env_nest nest;
+ struct cl_io *io;
+ struct page *page = NOPAGE_SIGBUS;
+ struct vvp_io *vio = NULL;
+ unsigned long ra_flags;
+ pgoff_t pg_offset;
+ int result;
+ const unsigned long writable = VM_SHARED|VM_WRITE;
+ ENTRY;
+
+ pg_offset = ((address - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
+ io = ll_fault_io_init(vma, &env, &nest, pg_offset, &ra_flags);
+ if (IS_ERR(io))
+ return NOPAGE_SIGBUS;
+
+ result = io->ci_result;
+ if (result < 0)
+ goto out_err;
+
+ io->u.ci_fault.ft_writable = (vma->vm_flags&writable) == writable;
+ vio = vvp_env_io(env);
+ vio->u.fault.ft_vma = vma;
+ vio->u.fault.nopage.ft_address = address;
+ vio->u.fault.nopage.ft_type = type;
+ vio->u.fault.ft_vmpage = NULL;
+
+ result = cl_io_loop(env, io);
+ page = vio->u.fault.ft_vmpage;
+ if (result != 0 && page != NULL)
+ page_cache_release(page);
+
+out_err:
+ if (result == -ENOMEM)
+ page = NOPAGE_OOM;
+
+ vma->vm_flags &= ~VM_RAND_READ;
+ vma->vm_flags |= ra_flags;
+
+ cl_io_fini(env, io);
+ cl_env_nested_put(&nest, env);
+
+ RETURN(page);
+}
+
+#else
+/**
+ * Lustre implementation of a vm_operations_struct::fault() method, called by
+ * VM to server page fault (both in kernel and user space).
+ *
+ * \param vma - is virtiual area struct related to page fault
+ * \param vmf - structure which describe type and address where hit fault
+ *
+ * \return allocated and filled _locked_ page for address
+ * \retval VM_FAULT_ERROR on general error
+ * \retval NOPAGE_OOM not have memory for allocate new page
+ */
+static int ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
+{
+ struct lu_env *env;
+ struct cl_io *io;
+ struct vvp_io *vio = NULL;
+ struct page *vmpage;
+ unsigned long ra_flags;
+ struct cl_env_nest nest;
+ int result;
+ int fault_ret = 0;
ENTRY;
- if (ll_file_nolock(file))
- RETURN(ERR_PTR(-EOPNOTSUPP));
+ io = ll_fault_io_init(vma, &env, &nest, vmf->pgoff, &ra_flags);
+ if (IS_ERR(io))
+ RETURN(VM_FAULT_ERROR);
- /*
- * vm_operations_struct::nopage() can be called when lustre IO is
- * already active for the current thread, e.g., when doing read/write
- * against user level buffer mapped from Lustre buffer. To avoid
- * stomping on existing context, optionally force an allocation of a new
- * one.
- */
- env = cl_env_nested_get(&nest);
- if (!IS_ERR(env)) {
- pgoff_t pg_offset;
- const unsigned long writable = VM_SHARED|VM_WRITE;
- unsigned long ra_flags;
- struct cl_fault_io *fio;
-
- io = &ccc_env_info(env)->cti_io;
- memset(io, 0, sizeof(*io));
- io->ci_obj = ll_i2info(inode)->lli_clob;
- LASSERT(io->ci_obj != NULL);
-
- fio = &io->u.ci_fault;
- pg_offset = (address - vma->vm_start) >> PAGE_SHIFT;
- fio->ft_index = pg_offset + vma->vm_pgoff;
- fio->ft_writable = (vma->vm_flags&writable) == writable;
- fio->ft_executable = vma->vm_flags&VM_EXEC;
-
- /*
- * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
- * the kernel will not read other pages not covered by ldlm in
- * filemap_nopage. we do our readahead in ll_readpage.
- */
- ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
- vma->vm_flags &= ~VM_SEQ_READ;
- vma->vm_flags |= VM_RAND_READ;
-
- CDEBUG(D_INFO, "vm_flags: %lx (%lu %i %i)\n", vma->vm_flags,
- fio->ft_index, fio->ft_writable, fio->ft_executable);
-
- if (cl_io_init(env, io, CIT_FAULT, io->ci_obj) == 0) {
- struct vvp_io *vio = vvp_env_io(env);
- struct ccc_io *cio = ccc_env_io(env);
- struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
-
- LASSERT(cio->cui_cl.cis_io == io);
-
- /* mmap lock must be MANDATORY. */
- io->ci_lockreq = CILR_MANDATORY;
- vio->u.fault.ft_vma = vma;
- vio->u.fault.ft_address = address;
- vio->u.fault.ft_type = type;
- cio->cui_fd = fd;
-
- result = cl_io_loop(env, io);
- if (result == 0) {
- LASSERT(fio->ft_page != NULL);
- page = cl_page_vmpage(env, fio->ft_page);
- } else if (result == -EFAULT) {
- page = NOPAGE_SIGBUS;
- } else if (result == -ENOMEM) {
- page = NOPAGE_OOM;
+ result = io->ci_result;
+ if (result < 0)
+ goto out_err;
+
+ vio = vvp_env_io(env);
+ vio->u.fault.ft_vma = vma;
+ vio->u.fault.ft_vmpage = NULL;
+ vio->u.fault.fault.ft_vmf = vmf;
+
+ result = cl_io_loop(env, io);
+
+ vmpage = vio->u.fault.ft_vmpage;
+ if (result != 0 && vmpage != NULL) {
+ page_cache_release(vmpage);
+ vmf->page = NULL;
+ }
+
+ fault_ret = vio->u.fault.fault.ft_flags;
+
+out_err:
+ if (result != 0 && fault_ret == 0)
+ fault_ret = VM_FAULT_ERROR;
+
+ vma->vm_flags |= ra_flags;
+
+ cl_io_fini(env, io);
+ cl_env_nested_put(&nest, env);
+
+ CDEBUG(D_MMAP, "%s fault %d/%d\n",
+ cfs_current()->comm, fault_ret, result);
+ RETURN(fault_ret);
+}
+
+static int ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
+{
+ int count = 0;
+ bool printed = false;
+ int result;
+
+restart:
+ result = ll_fault0(vma, vmf);
+ LASSERT(!(result & VM_FAULT_LOCKED));
+ if (result == 0) {
+ struct page *vmpage = vmf->page;
+
+ /* check if this page has been truncated */
+ lock_page(vmpage);
+ if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
+ unlock_page(vmpage);
+ page_cache_release(vmpage);
+ vmf->page = NULL;
+
+ if (!printed && ++count > 16) {
+ CWARN("the page is under heavy contention,"
+ "maybe your app(%s) needs revising :-)\n",
+ current->comm);
+ printed = true;
}
- } else
- result = io->ci_result;
- vma->vm_flags &= ~VM_RAND_READ;
- vma->vm_flags |= ra_flags;
+ goto restart;
+ }
- cl_io_fini(env, io);
- cl_env_nested_put(&nest, env);
+ result |= VM_FAULT_LOCKED;
}
- RETURN(page);
+ return result;
+}
+#endif
+
+#ifndef HAVE_PGMKWRITE_USE_VMFAULT
+static int ll_page_mkwrite(struct vm_area_struct *vma, struct page *vmpage)
+{
+ int count = 0;
+ bool printed = false;
+ bool retry;
+ int result;
+
+ do {
+ retry = false;
+ result = ll_page_mkwrite0(vma, vmpage, &retry);
+
+ if (!printed && ++count > 16) {
+ CWARN("app(%s): the page %lu of file %lu is under heavy"
+ " contention.\n",
+ current->comm, page_index(vmpage),
+ vma->vm_file->f_dentry->d_inode->i_ino);
+ printed = true;
+ }
+ } while (retry);
+
+ if (result == 0)
+ unlock_page(vmpage);
+ else if (result == -ENODATA)
+ result = 0; /* kernel will know truncate has happened and
+ * retry */
+
+ return result;
+}
+#else
+static int ll_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
+{
+ int count = 0;
+ bool printed = false;
+ bool retry;
+ int result;
+
+ do {
+ retry = false;
+ result = ll_page_mkwrite0(vma, vmf->page, &retry);
+
+ if (!printed && ++count > 16) {
+ CWARN("app(%s): the page %lu of file %lu is under heavy"
+ " contention.\n",
+ current->comm, vmf->pgoff,
+ vma->vm_file->f_dentry->d_inode->i_ino);
+ printed = true;
+ }
+ } while (retry);
+
+ switch(result) {
+ case 0:
+ LASSERT(PageLocked(vmf->page));
+ result = VM_FAULT_LOCKED;
+ break;
+ case -ENODATA:
+ case -EFAULT:
+ result = VM_FAULT_NOPAGE;
+ break;
+ case -ENOMEM:
+ result = VM_FAULT_OOM;
+ break;
+ case -EAGAIN:
+ result = VM_FAULT_RETRY;
+ break;
+ default:
+ result = VM_FAULT_SIGBUS;
+ break;
+ }
+
+ return result;
}
+#endif
/**
* To avoid cancel the locks covering mmapped region for lock cache pressure,
ENTRY;
LASSERT(vma->vm_file);
- LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0);
- atomic_inc(&vob->cob_mmap_cnt);
+ LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
+ cfs_atomic_inc(&vob->cob_mmap_cnt);
EXIT;
}
ENTRY;
LASSERT(vma->vm_file);
- atomic_dec(&vob->cob_mmap_cnt);
- LASSERT(atomic_read(&vob->cob_mmap_cnt) >= 0);
+ cfs_atomic_dec(&vob->cob_mmap_cnt);
+ LASSERT(cfs_atomic_read(&vob->cob_mmap_cnt) >= 0);
EXIT;
}
+#ifndef HAVE_VM_OP_FAULT
#ifndef HAVE_FILEMAP_POPULATE
static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
#endif
rc = filemap_populate(area, address, len, prot, pgoff, 1);
RETURN(rc);
}
+#endif
/* return the user space pointer that maps to a file offset via a vma */
static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
}
static struct vm_operations_struct ll_file_vm_ops = {
+#ifndef HAVE_VM_OP_FAULT
.nopage = ll_nopage,
+ .populate = ll_populate,
+
+#else
+ .fault = ll_fault,
+#endif
+ .page_mkwrite = ll_page_mkwrite,
.open = ll_vm_open,
.close = ll_vm_close,
- .populate = ll_populate,
};
int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MAP, 1);
rc = generic_file_mmap(file, vma);
if (rc == 0) {
-#if !defined(HAVE_FILEMAP_POPULATE)
+#if !defined(HAVE_FILEMAP_POPULATE) && !defined(HAVE_VM_OP_FAULT)
if (!filemap_populate)
filemap_populate = vma->vm_ops->populate;
#endif
vma->vm_ops = &ll_file_vm_ops;
vma->vm_ops->open(vma);
/* update the inode's size and mtime */
- rc = cl_glimpse_size(inode);
+ rc = ll_glimpse_size(inode);
}
RETURN(rc);