From: Kari Argillander <kari.argillan...@gmail.com>
Sent: Monday, January 4, 2021 12:58 AM
> To: Konstantin Komarov <almaz.alexandrov...@paragon-software.com>
> Cc: linux-fsde...@vger.kernel.org; v...@zeniv.linux.org.uk;
> linux-kernel@vger.kernel.org; p...@kernel.org; dste...@suse.cz;
> aap...@suse.com; wi...@infradead.org; rdun...@infradead.org;
> j...@perches.com; m...@harmstone.com; nbori...@suse.com;
> linux-ntfs-...@lists.sourceforge.net; an...@tuxera.com;
> dan.carpen...@oracle.com; h...@lst.de; ebigg...@kernel.org;
> andy.l...@gmail.com
> Subject: Re: [PATCH v17 04/10] fs/ntfs3: Add file operations and
> implementation
>
> On Thu, Dec 31, 2020 at 06:23:55PM +0300, Konstantin Komarov wrote:
> > This adds file operations and implementation
> >
> > Signed-off-by: Konstantin Komarov <almaz.alexandrov...@paragon-software.com>
> > ---
> > fs/ntfs3/dir.c | 570 ++++++++
> > fs/ntfs3/file.c | 1140 ++++++++++++++++
> > fs/ntfs3/frecord.c | 3088 ++++++++++++++++++++++++++++++++++++++++++++
> > fs/ntfs3/namei.c | 590 +++++++++
> > fs/ntfs3/record.c | 614 +++++++++
> > fs/ntfs3/run.c | 1254 ++++++++++++++++++
> > 6 files changed, 7256 insertions(+)
> > create mode 100644 fs/ntfs3/dir.c
> > create mode 100644 fs/ntfs3/file.c
> > create mode 100644 fs/ntfs3/frecord.c
> > create mode 100644 fs/ntfs3/namei.c
> > create mode 100644 fs/ntfs3/record.c
> > create mode 100644 fs/ntfs3/run.c
> >
> > diff --git a/fs/ntfs3/file.c b/fs/ntfs3/file.c
>
> > +int ntfs_file_fsync(struct file *filp, loff_t start, loff_t end, int
> > datasync)
> > +{
> > + return generic_file_fsync(filp, start, end, datasync);
> > +}
>
> Do we have a reson why we implement this if we just use generic. Isn't
> it more clear if we use generic fsync straight away?
>
Hi! Indeed. Migration to the generic fsync will be introduced in v18.
> > +static long ntfs_fallocate(struct file *file, int mode, loff_t vbo, loff_t
> > len)
> > +{
>
> > + /* Return error if mode is not supported */
> > + if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
> > + FALLOC_FL_COLLAPSE_RANGE))
> > + return -EOPNOTSUPP;
>
> > +
> > + if (mode & FALLOC_FL_PUNCH_HOLE) {
>
> > + } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
>
> > + } else {
> > + /*
> > + * normal file: allocate clusters, do not change 'valid' size
> > + */
> > + err = ntfs_set_size(inode, max(end, i_size));
> > + if (err)
> > + goto out;
> > +
> > + if (is_sparsed(ni) || is_compressed(ni)) {
> > + CLST vcn_v = ni->i_valid >> sbi->cluster_bits;
> > + CLST vcn = vbo >> sbi->cluster_bits;
> > + CLST cend = bytes_to_cluster(sbi, end);
> > + CLST lcn, clen;
> > + bool new;
> > +
> > + /*
> > + * allocate but not zero new clusters (see below
> > comments)
> > + * this breaks security (one can read unused on-disk
> > areas)
> > + * zeroing these clusters may be too long
> > + * may be we should check here for root rights?
> > + */
> > + for (; vcn < cend; vcn += clen) {
> > + err = attr_data_get_block(ni, vcn, cend - vcn,
> > + &lcn, &clen, &new);
> > + if (err)
> > + goto out;
> > + if (!new || vcn >= vcn_v)
> > + continue;
> > +
> > + /*
> > + * Unwritten area
> > + * NTFS is not able to store several unwritten
> > areas
> > + * Activate 'ntfs_sparse_cluster' to zero new
> > allocated clusters
> > + *
> > + * Dangerous in case:
> > + * 1G of sparsed clusters + 1 cluster of data =>
> > + * valid_size == 1G + 1 cluster
> > + * fallocate(1G) will zero 1G and this can be
> > very long
> > + * xfstest 016/086 will fail whithout
> > 'ntfs_sparse_cluster'
> > + */
> > + /*ntfs_sparse_cluster(inode, NULL, vcn,
> > + * min(vcn_v - vcn, clen));
> > + */
> > + }
> > + }
> > +
> > + if (mode & FALLOC_FL_KEEP_SIZE) {
>
> Isn't this hole else already (mode & FALLOC_FL_KEEP_SIZE?
Sorry, can you please clarify your question? Not sure, understood it.
>
> > + ni_lock(ni);
> > + /*true - keep preallocated*/
> > + err = attr_set_size(ni, ATTR_DATA, NULL, 0,
> > + &ni->file.run, i_size, &ni->i_valid,
> > + true, NULL);
> > + ni_unlock(ni);
> > + }
> > + }
> > +
> > + if (!err) {
> > + inode->i_ctime = inode->i_mtime = current_time(inode);
> > + mark_inode_dirty(inode);
> > + }
> > +out:
> > + if (err == -EFBIG)
> > + err = -ENOSPC;
> > +
> > + inode_unlock(inode);
> > + return err;
> > +}
>
> > diff --git a/fs/ntfs3/namei.c b/fs/ntfs3/namei.c
>
> > +int mi_get(struct ntfs_sb_info *sbi, CLST rno, struct mft_inode **mi)
> > +{
> > + int err;
> > + struct mft_inode *m = ntfs_alloc(sizeof(struct mft_inode), 1);
> > +
> > + if (!m)
> > + return -ENOMEM;
> > +
> > + err = mi_init(m, sbi, rno);
>
> If error happend should we just free end exit. Now we call mi_put() to
> clean up.
>
Done, will be fixed in v18.
> > + if (!err)
> > + err = mi_read(m, false);
> > +
> > + if (err) {
> > + mi_put(m);
> > + return err;
> > + }
> > +
> > + *mi = m;
> > + return 0;
> > +}
>
> > +struct ATTRIB *mi_enum_attr(struct mft_inode *mi, struct ATTRIB *attr)
> > +{
> > + const struct MFT_REC *rec = mi->mrec;
> > + u32 used = le32_to_cpu(rec->used);
> > + u32 t32, off, asize;
> > + u16 t16;
> > +
> > + if (!attr) {
> > + u32 total = le32_to_cpu(rec->total);
> > +
> > + off = le16_to_cpu(rec->attr_off);
> > +
> > + if (used > total)
> > + goto out;
> > +
> > + if (off >= used || off < MFTRECORD_FIXUP_OFFSET_1 ||
> > + !IsDwordAligned(off)) {
> > + goto out;
> > + }
> > +
> > + /* Skip non-resident records */
> > + if (!is_rec_inuse(rec))
> > + goto out;
> > +
> > + attr = Add2Ptr(rec, off);
> > + } else {
> > + /* Check if input attr inside record */
> > + off = PtrOffset(rec, attr);
> > + if (off >= used)
> > + goto out;
> > +
> > + asize = le32_to_cpu(attr->size);
> > + if (asize < SIZEOF_RESIDENT)
> > + goto out;
> > +
> > + attr = Add2Ptr(attr, asize);
> > + off += asize;
> > + }
> > +
> > + asize = le32_to_cpu(attr->size);
> > +
> > + /* Can we use the first field (attr->type) */
> > + if (off + 8 > used) {
> > + static_assert(QuadAlign(sizeof(enum ATTR_TYPE)) == 8);
> > + goto out;
> > + }
> > +
> > + if (attr->type == ATTR_END) {
> > + if (used != off + 8)
> > + goto out;
>
> This if is not needed if there is return NULL after. But return
> NULL might also be bug.
>
Thanks, will also be fixed in v18.
> > + return NULL;
> > + }
> > +
> > + t32 = le32_to_cpu(attr->type);
> > + if ((t32 & 0xf) || (t32 > 0x100))
> > + goto out;
> > +
> > + /* Check boundary */
> > + if (off + asize > used)
> > + goto out;
> > +
> > + /* Check size of attribute */
> > + if (!attr->non_res) {
> > + if (asize < SIZEOF_RESIDENT)
> > + goto out;
> > +
> > + t16 = le16_to_cpu(attr->res.data_off);
> > +
> > + if (t16 > asize)
> > + goto out;
> > +
> > + t32 = le32_to_cpu(attr->res.data_size);
> > + if (t16 + t32 > asize)
> > + goto out;
> > +
> > + return attr;
> > + }
> > +
> > + /* Check some nonresident fields */
> > + if (attr->name_len &&
> > + le16_to_cpu(attr->name_off) + sizeof(short) * attr->name_len >
> > + le16_to_cpu(attr->nres.run_off)) {
> > + goto out;
> > + }
> > +
> > + if (attr->nres.svcn || !is_attr_ext(attr)) {
> > + if (asize + 8 < SIZEOF_NONRESIDENT)
> > + goto out;
> > +
> > + if (attr->nres.c_unit)
> > + goto out;
> > + } else if (asize + 8 < SIZEOF_NONRESIDENT_EX)
> > + goto out;
> > +
> > + return attr;
> > +
> > +out:
> > + return NULL;
> > +}
>
> > diff --git a/fs/ntfs3/run.c b/fs/ntfs3/run.c
>
> > +static inline int run_packed_size(const s64 *n)
> > +{
> > +#ifdef __BIG_ENDIAN
>
> These are whole functions with ifdef. It would be maybe more clear
> that there really is whole functions to both endiand.
>
This is controversial question, but fixed as well (in v18).
> > + const u8 *p = (const u8 *)n + sizeof(*n) - 1;
> > +
> > + if (*n >= 0) {
> > + if (p[-7] || p[-6] || p[-5] || p[-4])
> > + p -= 4;
> > + if (p[-3] || p[-2])
> > + p -= 2;
> > + if (p[-1])
> > + p -= 1;
> > + if (p[0] & 0x80)
> > + p -= 1;
> > + } else {
> > + if (p[-7] != 0xff || p[-6] != 0xff || p[-5] != 0xff ||
> > + p[-4] != 0xff)
> > + p -= 4;
> > + if (p[-3] != 0xff || p[-2] != 0xff)
> > + p -= 2;
> > + if (p[-1] != 0xff)
> > + p -= 1;
> > + if (!(p[0] & 0x80))
> > + p -= 1;
> > + }
> > + return (const u8 *)n + sizeof(*n) - p;
>
> }
> #else
> static inline int run_packed_size(const s64 *n)
> {
>
> Something like this.
>
> > + const u8 *p = (const u8 *)n;
> > +
> > + if (*n >= 0) {
> > + if (p[7] || p[6] || p[5] || p[4])
> > + p += 4;
> > + if (p[3] || p[2])
> > + p += 2;
> > + if (p[1])
> > + p += 1;
> > + if (p[0] & 0x80)
> > + p += 1;
> > + } else {
> > + if (p[7] != 0xff || p[6] != 0xff || p[5] != 0xff ||
> > + p[4] != 0xff)
> > + p += 4;
> > + if (p[3] != 0xff || p[2] != 0xff)
> > + p += 2;
> > + if (p[1] != 0xff)
> > + p += 1;
> > + if (!(p[0] & 0x80))
> > + p += 1;
> > + }
> > +
> > + return 1 + p - (const u8 *)n;
> > +#endif
> > +}
> > +
> > +/*
> > + * run_pack
> > + *
> > + * packs runs into buffer
> > + * packed_vcns - how much runs we have packed
> > + * packed_size - how much bytes we have used run_buf
> > + */
> > +int run_pack(const struct runs_tree *run, CLST svcn, CLST len, u8 *run_buf,
> > + u32 run_buf_size, CLST *packed_vcns)
> > +{
> > + CLST next_vcn, vcn, lcn;
> > + CLST prev_lcn = 0;
> > + CLST evcn1 = svcn + len;
> > + int packed_size = 0;
> > + size_t i;
> > + bool ok;
> > + s64 dlcn, len64;
> > + int offset_size, size_size, t;
> > + const u8 *p;
> > +
> > + next_vcn = vcn = svcn;
> > +
> > + *packed_vcns = 0;
> > +
> > + if (!len)
> > + goto out;
> > +
> > + ok = run_lookup_entry(run, vcn, &lcn, &len, &i);
> > +
> > + if (!ok)
> > + goto error;
> > +
> > + if (next_vcn != vcn)
> > + goto error;
> > +
> > + for (;;) {
> > + /* offset of current fragment relatively to previous fragment */
> > + dlcn = 0;
>
> This dlcn
>
> > + next_vcn = vcn + len;
> > +
> > + if (next_vcn > evcn1)
> > + len = evcn1 - vcn;
> > +
> > + /*
> > + * mirror of len, but signed, because run_packed_size()
> > + * works with signed int only
> > + */
> > + len64 = len;
> > +
> > + /* how much bytes is packed len64 */
> > + size_size = run_packed_size(&len64);
>
> Does (s64 *)&len work just fine?
>
In addition to Mattew's response:
NTFS by it's spec supports 64bit clusters, however ntfs.sys driver
uses 32bit clusters only. This is a default for ntfs3 as well, which may
be configured to 64bit as well. I.e. len may be both.
> > +
> > + /* offset_size - how much bytes is packed dlcn */
> > + if (lcn == SPARSE_LCN) {
> > + offset_size = 0;
>
> dlcn might be better to live here?
>
You are right, done.
> > + } else {
> > + /* NOTE: lcn can be less than prev_lcn! */
> > + dlcn = (s64)lcn - prev_lcn;
> > + offset_size = run_packed_size(&dlcn);
> > + prev_lcn = lcn;
> > + }
> > +
> > + t = run_buf_size - packed_size - 2 - offset_size;
> > + if (t <= 0)
> > + goto out;
> > +
> > + /* can we store this entire run */
> > + if (t < size_size)
> > + goto out;
> > +
> > + if (run_buf) {
> > + p = (u8 *)&len64;
> > +
> > + /* pack run header */
> > + run_buf[0] = ((u8)(size_size | (offset_size << 4)));
> > + run_buf += 1;
> > +
> > + /* Pack the length of run */
> > + switch (size_size) {
> > +#ifdef __BIG_ENDIAN
> > + case 8:
> > + run_buf[7] = p[0];
> > + fallthrough;
> > + case 7:
> > + run_buf[6] = p[1];
> > + fallthrough;
> > + case 6:
> > + run_buf[5] = p[2];
> > + fallthrough;
> > + case 5:
> > + run_buf[4] = p[3];
> > + fallthrough;
> > + case 4:
> > + run_buf[3] = p[4];
> > + fallthrough;
> > + case 3:
> > + run_buf[2] = p[5];
> > + fallthrough;
> > + case 2:
> > + run_buf[1] = p[6];
> > + fallthrough;
> > + case 1:
> > + run_buf[0] = p[7];
> > +#else
> > + case 8:
> > + run_buf[7] = p[7];
> > + fallthrough;
> > + case 7:
> > + run_buf[6] = p[6];
> > + fallthrough;
> > + case 6:
> > + run_buf[5] = p[5];
> > + fallthrough;
> > + case 5:
> > + run_buf[4] = p[4];
> > + fallthrough;
> > + case 4:
> > + run_buf[3] = p[3];
> > + fallthrough;
> > + case 3:
> > + run_buf[2] = p[2];
> > + fallthrough;
> > + case 2:
> > + run_buf[1] = p[1];
> > + fallthrough;
> > + case 1:
> > + run_buf[0] = p[0];
> > +#endif
> > + }
>
> Why is this not own function? We use this like 5 places. Also isn't
> little endian just memcopy()
>
You are right, for little endiand this is just memcopy, but if it isn't
inlined, we will have overhead calling the function.
> > +
> > + run_buf += size_size;
> > + p = (u8 *)&dlcn;
>
> I think that when we have function for that switch tmp p is not needed
> anymore.
>
Yes, but this doesn't give more readability or simplification I think.
More of a personal preference.
> > +
> > + /* Pack the offset from previous lcn */
> > + switch (offset_size) {
> > +#ifdef __BIG_ENDIAN
> > + case 8:
> > + run_buf[7] = p[0];
> > + fallthrough;
> > + case 7:
> > + run_buf[6] = p[1];
> > + fallthrough;
> > + case 6:
> > + run_buf[5] = p[2];
> > + fallthrough;
> > + case 5:
> > + run_buf[4] = p[3];
> > + fallthrough;
> > + case 4:
> > + run_buf[3] = p[4];
> > + fallthrough;
> > + case 3:
> > + run_buf[2] = p[5];
> > + fallthrough;
> > + case 2:
> > + run_buf[1] = p[6];
> > + fallthrough;
> > + case 1:
> > + run_buf[0] = p[7];
> > +#else
> > + case 8:
> > + run_buf[7] = p[7];
> > + fallthrough;
> > + case 7:
> > + run_buf[6] = p[6];
> > + fallthrough;
> > + case 6:
> > + run_buf[5] = p[5];
> > + fallthrough;
> > + case 5:
> > + run_buf[4] = p[4];
> > + fallthrough;
> > + case 4:
> > + run_buf[3] = p[3];
> > + fallthrough;
> > + case 3:
> > + run_buf[2] = p[2];
> > + fallthrough;
> > + case 2:
> > + run_buf[1] = p[1];
> > + fallthrough;
> > + case 1:
> > + run_buf[0] = p[0];
> > +#endif
> > + }
>
> > +int run_get_highest_vcn(CLST vcn, const u8 *run_buf, u64 *highest_vcn)
> > +{
>
> > + /* skip size_size */
> > + run += size_size;
> > +
> > + if (!len)
> > + goto error;
> > +
> > + run += offset_size;
>
> Can this be straight
> run += size_size + offset_size;
>
Done.
> > +
> > +#ifdef NTFS3_64BIT_CLUSTER
> > + if ((vcn >> 32) || (len >> 32))
> > + goto error;
> > +#endif
> > + vcn64 += len;
> > + }
> > +
> > + *highest_vcn = vcn64 - 1;
> > + return 0;
> > +}
