@@ -10,8 +10,8 @@ typedef struct V9fsString
typedef struct V9fsQID
{
int8_t type;
- int32_t version;
- int64_t path;
+ uint32_t version;
+ uint64_t path;
} V9fsQID;
typedef struct V9fsStat
@@ -25,6 +25,7 @@
#include "trace.h"
#include "migration/blocker.h"
#include "sysemu/qtest.h"
+#include "qemu/xxhash.h"
int open_fd_hw;
int total_open_fd;
@@ -571,14 +572,135 @@ static void coroutine_fn virtfs_reset(V9fsPDU *pdu)
P9_STAT_MODE_NAMED_PIPE | \
P9_STAT_MODE_SOCKET)
-/* This is the algorithm from ufs in spfs */
-static void stat_to_qid(const struct stat *stbuf, V9fsQID *qidp)
+
+/* creative abuse of qemu_xxhash7, which is based on xxhash */
+static uint32_t qpp_hash(QppEntry e)
{
- size_t size;
+ return qemu_xxhash7(e.ino_prefix, e.dev, 0, 0, 0);
+}
+
+static uint32_t qpf_hash(QpfEntry e)
+{
+ return qemu_xxhash7(e.ino, e.dev, 0, 0, 0);
+}
+
+static bool qpp_cmp_func(const void *obj, const void *userp)
+{
+ const QppEntry *e1 = obj, *e2 = userp;
+ return (e1->dev == e2->dev) && (e1->ino_prefix == e2->ino_prefix);
+}
+
+static bool qpf_cmp_func(const void *obj, const void *userp)
+{
+ const QpfEntry *e1 = obj, *e2 = userp;
+ return (e1->dev == e2->dev) && (e1->ino == e2->ino);
+}
+
+static void qp_table_remove(void *p, uint32_t h, void *up)
+{
+ g_free(p);
+}
+
+static void qp_table_destroy(struct qht *ht)
+{
+ qht_iter(ht, qp_table_remove, NULL);
+ qht_destroy(ht);
+}
+
+static int qid_path_fullmap(V9fsPDU *pdu, const struct stat *stbuf,
+ uint64_t *path)
+{
+ QpfEntry lookup = {
+ .dev = stbuf->st_dev,
+ .ino = stbuf->st_ino
+ }, *val;
+ uint32_t hash = qpf_hash(lookup);
+
+ /* most users won't need the fullmap, so init the table lazily */
+ if (!pdu->s->qpf_table.map) {
+ qht_init(&pdu->s->qpf_table, qpf_cmp_func, 1 << 16, QHT_MODE_AUTO_RESIZE);
+ }
+
+ val = qht_lookup(&pdu->s->qpf_table, &lookup, hash);
+
+ if (!val) {
+ if (pdu->s->qp_fullpath_next == 0) {
+ /* no more files can be mapped :'( */
+ return -ENFILE;
+ }
+
+ val = g_malloc0(sizeof(QppEntry));
+ if (!val) {
+ return -ENOMEM;
+ }
+ *val = lookup;
+
+ /* new unique inode and device combo */
+ val->path = pdu->s->qp_fullpath_next++;
+ pdu->s->qp_fullpath_next &= QPATH_INO_MASK;
+ qht_insert(&pdu->s->qpf_table, val, hash, NULL);
+ }
+
+ *path = val->path;
+ return 0;
+}
+
+/* stat_to_qid needs to map inode number (64 bits) and device id (32 bits)
+ * to a unique QID path (64 bits). To avoid having to map and keep track
+ * of up to 2^64 objects, we map only the 16 highest bits of the inode plus
+ * the device id to the 16 highest bits of the QID path. The 48 lowest bits
+ * of the QID path equal to the lowest bits of the inode number.
+ *
+ * This takes advantage of the fact that inode number are usually not
+ * random but allocated sequentially, so we have fewer items to keep
+ * track of.
+ */
+static int qid_path_prefixmap(V9fsPDU *pdu, const struct stat *stbuf,
+ uint64_t *path)
+{
+ QppEntry lookup = {
+ .dev = stbuf->st_dev,
+ .ino_prefix = (uint16_t) (stbuf->st_ino >> 48)
+ }, *val;
+ uint32_t hash = qpp_hash(lookup);
+
+ val = qht_lookup(&pdu->s->qpp_table, &lookup, hash);
+
+ if (!val) {
+ if (pdu->s->qp_prefix_next == 0) {
+ /* we ran out of prefixes */
+ return -ENFILE;
+ }
+
+ val = g_malloc0(sizeof(QppEntry));
+ if (!val) {
+ return -ENOMEM;
+ }
+ *val = lookup;
+
+ /* new unique inode prefix and device combo */
+ val->qp_prefix = pdu->s->qp_prefix_next++;
+ qht_insert(&pdu->s->qpp_table, val, hash, NULL);
+ }
+
+ *path = ((uint64_t)val->qp_prefix << 48) | (stbuf->st_ino & QPATH_INO_MASK);
+ return 0;
+}
+
+static int stat_to_qid(V9fsPDU *pdu, const struct stat *stbuf, V9fsQID *qidp)
+{
+ int err;
+
+ /* map inode+device to qid path (fast path) */
+ err = qid_path_prefixmap(pdu, stbuf, &qidp->path);
+ if (err == -ENFILE) {
+ /* fast path didn't work, fal back to full map */
+ err = qid_path_fullmap(pdu, stbuf, &qidp->path);
+ }
+ if (err) {
+ return err;
+ }
- memset(&qidp->path, 0, sizeof(qidp->path));
- size = MIN(sizeof(stbuf->st_ino), sizeof(qidp->path));
- memcpy(&qidp->path, &stbuf->st_ino, size);
qidp->version = stbuf->st_mtime ^ (stbuf->st_size << 8);
qidp->type = 0;
if (S_ISDIR(stbuf->st_mode)) {
@@ -587,6 +709,8 @@ static void stat_to_qid(const struct stat *stbuf, V9fsQID *qidp)
if (S_ISLNK(stbuf->st_mode)) {
qidp->type |= P9_QID_TYPE_SYMLINK;
}
+
+ return 0;
}
static int coroutine_fn fid_to_qid(V9fsPDU *pdu, V9fsFidState *fidp,
@@ -599,7 +723,10 @@ static int coroutine_fn fid_to_qid(V9fsPDU *pdu, V9fsFidState *fidp,
if (err < 0) {
return err;
}
- stat_to_qid(&stbuf, qidp);
+ err = stat_to_qid(pdu, &stbuf, qidp);
+ if (err < 0) {
+ return err;
+ }
return 0;
}
@@ -830,7 +957,10 @@ static int coroutine_fn stat_to_v9stat(V9fsPDU *pdu, V9fsPath *path,
memset(v9stat, 0, sizeof(*v9stat));
- stat_to_qid(stbuf, &v9stat->qid);
+ err = stat_to_qid(pdu, stbuf, &v9stat->qid);
+ if (err < 0) {
+ return err;
+ }
v9stat->mode = stat_to_v9mode(stbuf);
v9stat->atime = stbuf->st_atime;
v9stat->mtime = stbuf->st_mtime;
@@ -891,7 +1021,7 @@ static int coroutine_fn stat_to_v9stat(V9fsPDU *pdu, V9fsPath *path,
#define P9_STATS_ALL 0x00003fffULL /* Mask for All fields above */
-static void stat_to_v9stat_dotl(V9fsState *s, const struct stat *stbuf,
+static int stat_to_v9stat_dotl(V9fsPDU *pdu, const struct stat *stbuf,
V9fsStatDotl *v9lstat)
{
memset(v9lstat, 0, sizeof(*v9lstat));
@@ -913,7 +1043,7 @@ static void stat_to_v9stat_dotl(V9fsState *s, const struct stat *stbuf,
/* Currently we only support BASIC fields in stat */
v9lstat->st_result_mask = P9_STATS_BASIC;
- stat_to_qid(stbuf, &v9lstat->qid);
+ return stat_to_qid(pdu, stbuf, &v9lstat->qid);
}
static void print_sg(struct iovec *sg, int cnt)
@@ -1115,7 +1245,6 @@ static void coroutine_fn v9fs_getattr(void *opaque)
uint64_t request_mask;
V9fsStatDotl v9stat_dotl;
V9fsPDU *pdu = opaque;
- V9fsState *s = pdu->s;
retval = pdu_unmarshal(pdu, offset, "dq", &fid, &request_mask);
if (retval < 0) {
@@ -1136,7 +1265,10 @@ static void coroutine_fn v9fs_getattr(void *opaque)
if (retval < 0) {
goto out;
}
- stat_to_v9stat_dotl(s, &stbuf, &v9stat_dotl);
+ retval = stat_to_v9stat_dotl(pdu, &stbuf, &v9stat_dotl);
+ if (retval < 0) {
+ goto out;
+ }
/* fill st_gen if requested and supported by underlying fs */
if (request_mask & P9_STATS_GEN) {
@@ -1381,7 +1513,10 @@ static void coroutine_fn v9fs_walk(void *opaque)
if (err < 0) {
goto out;
}
- stat_to_qid(&stbuf, &qid);
+ err = stat_to_qid(pdu, &stbuf, &qid);
+ if (err < 0) {
+ goto out;
+ }
v9fs_path_copy(&dpath, &path);
}
memcpy(&qids[name_idx], &qid, sizeof(qid));
@@ -1483,7 +1618,10 @@ static void coroutine_fn v9fs_open(void *opaque)
if (err < 0) {
goto out;
}
- stat_to_qid(&stbuf, &qid);
+ err = stat_to_qid(pdu, &stbuf, &qid);
+ if (err < 0) {
+ goto out;
+ }
if (S_ISDIR(stbuf.st_mode)) {
err = v9fs_co_opendir(pdu, fidp);
if (err < 0) {
@@ -1593,7 +1731,10 @@ static void coroutine_fn v9fs_lcreate(void *opaque)
fidp->flags |= FID_NON_RECLAIMABLE;
}
iounit = get_iounit(pdu, &fidp->path);
- stat_to_qid(&stbuf, &qid);
+ err = stat_to_qid(pdu, &stbuf, &qid);
+ if (err < 0) {
+ goto out;
+ }
err = pdu_marshal(pdu, offset, "Qd", &qid, iounit);
if (err < 0) {
goto out;
@@ -2327,7 +2468,10 @@ static void coroutine_fn v9fs_create(void *opaque)
}
}
iounit = get_iounit(pdu, &fidp->path);
- stat_to_qid(&stbuf, &qid);
+ err = stat_to_qid(pdu, &stbuf, &qid);
+ if (err < 0) {
+ goto out;
+ }
err = pdu_marshal(pdu, offset, "Qd", &qid, iounit);
if (err < 0) {
goto out;
@@ -2384,7 +2528,10 @@ static void coroutine_fn v9fs_symlink(void *opaque)
if (err < 0) {
goto out;
}
- stat_to_qid(&stbuf, &qid);
+ err = stat_to_qid(pdu, &stbuf, &qid);
+ if (err < 0) {
+ goto out;
+ }
err = pdu_marshal(pdu, offset, "Q", &qid);
if (err < 0) {
goto out;
@@ -3064,7 +3211,10 @@ static void coroutine_fn v9fs_mknod(void *opaque)
if (err < 0) {
goto out;
}
- stat_to_qid(&stbuf, &qid);
+ err = stat_to_qid(pdu, &stbuf, &qid);
+ if (err < 0) {
+ goto out;
+ }
err = pdu_marshal(pdu, offset, "Q", &qid);
if (err < 0) {
goto out;
@@ -3222,7 +3372,10 @@ static void coroutine_fn v9fs_mkdir(void *opaque)
if (err < 0) {
goto out;
}
- stat_to_qid(&stbuf, &qid);
+ err = stat_to_qid(pdu, &stbuf, &qid);
+ if (err < 0) {
+ goto out;
+ }
err = pdu_marshal(pdu, offset, "Q", &qid);
if (err < 0) {
goto out;
@@ -3633,6 +3786,11 @@ int v9fs_device_realize_common(V9fsState *s, const V9fsTransport *t,
goto out;
}
+ /* QID path hash table. 1 entry ought to be enough for anybody ;) */
+ qht_init(&s->qpp_table, qpp_cmp_func, 1, QHT_MODE_AUTO_RESIZE);
+ s->qp_prefix_next = 1; /* reserve 0 to detect overflow */
+ s->qp_fullpath_next = 1;
+
s->ctx.fst = &fse->fst;
fsdev_throttle_init(s->ctx.fst);
@@ -3646,6 +3804,8 @@ out:
}
g_free(s->tag);
g_free(s->ctx.fs_root);
+ qp_table_destroy(&s->qpp_table);
+ qp_table_destroy(&s->qpf_table);
v9fs_path_free(&path);
}
return rc;
@@ -3658,6 +3818,8 @@ void v9fs_device_unrealize_common(V9fsState *s, Error **errp)
}
fsdev_throttle_cleanup(s->ctx.fst);
g_free(s->tag);
+ qp_table_destroy(&s->qpp_table);
+ qp_table_destroy(&s->qpf_table);
g_free(s->ctx.fs_root);
}
@@ -8,6 +8,7 @@
#include "fsdev/9p-iov-marshal.h"
#include "qemu/thread.h"
#include "qemu/coroutine.h"
+#include "qemu/qht.h"
enum {
P9_TLERROR = 6,
@@ -235,6 +236,22 @@ struct V9fsFidState
V9fsFidState *rclm_lst;
};
+#define QPATH_INO_MASK (((unsigned long)1 << 48) - 1)
+
+/* QID path prefix entry, see stat_to_qid */
+typedef struct {
+ dev_t dev;
+ uint16_t ino_prefix;
+ uint16_t qp_prefix;
+} QppEntry;
+
+/* QID path full entry, as above */
+typedef struct {
+ dev_t dev;
+ ino_t ino;
+ uint64_t path;
+} QpfEntry;
+
struct V9fsState
{
QLIST_HEAD(, V9fsPDU) free_list;
@@ -256,6 +273,10 @@ struct V9fsState
Error *migration_blocker;
V9fsConf fsconf;
V9fsQID root_qid;
+ struct qht qpp_table;
+ struct qht qpf_table;
+ uint16_t qp_prefix_next;
+ uint64_t qp_fullpath_next;
};
/* 9p2000.L open flags */
I am attempting to revive Antonios Motakis' effort to fix the current file ID collisions with 9p. My patch set does not fix all the concerns you had with his original patch set, especially if there is still some (very rare) case of QID path collision the affected files are still visible on guest. However I hope we can bring this overall issue forward, because with the current 9p implementation it is almost inevitable to end up with QID path collisions, which in turn causes very severe misbehaviours like data corruption and data loss on guest. On the other hand the yet unresolved issues that you noted are IMO rather theoretical issues. Because in practice you usually will have around not more than one or two dozens of entries in qpp_table (and not a single entry in qpf_table). So this first patch here is an updated version of Antonios Motakis' original 4-patch set, merged to one patch: https://lists.gnu.org/archive/html/qemu-devel/2018-02/msg02283.html * Updated to latest git master, specifically to new qht interface. * Merged the original 4 patches to this single patch. Signed-off-by: Christian Schoenebeck <qemu_oss@crudebyte.com> --- fsdev/9p-marshal.h | 4 +- hw/9pfs/9p.c | 200 ++++++++++++++++++++++++++++++++++++++++++++++++----- hw/9pfs/9p.h | 21 ++++++ 3 files changed, 204 insertions(+), 21 deletions(-)