From: Darrick J. Wong <djwong@xxxxxxxxxx>
Augment the media scan when verity files are detected by reading those
files' pagecache to force verity to compare the hash and (optional)
signatures.
Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
---
scrub/phase6.c | 305 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 305 insertions(+)
diff --git a/scrub/phase6.c b/scrub/phase6.c
index de7fcb548fe6..983470b7bece 100644
--- a/scrub/phase6.c
+++ b/scrub/phase6.c
@@ -23,6 +23,8 @@
#include "vfs.h"
#include "common.h"
#include "libfrog/bulkstat.h"
+#include "descr.h"
+#include "progress.h"
/*
* Phase 6: Verify data file integrity.
@@ -34,6 +36,9 @@
* to tell us if metadata are now corrupt. Otherwise, we'll scan the
* whole directory tree looking for files that overlap the bad regions
* and report the paths of the now corrupt files.
+ *
+ * If the filesystem supports verity, read the contents of each verity file to
+ * force it to validate the file contents.
*/
/* Verify disk blocks with GETFSMAP */
@@ -674,6 +679,285 @@ remember_ioerr(
str_liberror(ctx, ret, _("setting bad block bitmap"));
}
+struct verity_ctx {
+ struct scrub_ctx *ctx;
+ struct workqueue wq_ddev;
+ struct workqueue wq_rtdev;
+ bool aborted;
+};
+
+struct verity_file_ctx {
+ struct xfs_handle handle;
+ struct verity_ctx *vc;
+};
+
+static int
+render_ino_from_handle(
+ struct scrub_ctx *ctx,
+ char *buf,
+ size_t buflen,
+ void *data)
+{
+ struct xfs_handle *han = data;
+
+ return scrub_render_ino_descr(ctx, buf, buflen, han->ha_fid.fid_ino,
+ han->ha_fid.fid_gen, NULL);
+}
+
+static inline void
+report_verity_error(
+ struct scrub_ctx *ctx,
+ struct descr *dsc,
+ off_t fail_pos,
+ off_t fail_len)
+{
+ if (fail_pos < 0)
+ return;
+
+ str_unfixable_error(ctx, descr_render(dsc),
+ _("verity error at offsets %llu-%llu"),
+ (unsigned long long)fail_pos,
+ (unsigned long long)(fail_pos + fail_len - 1));
+}
+
+/* Record a verity validation error and maybe log an old error. */
+static inline void
+record_verity_error(
+ struct scrub_ctx *ctx,
+ struct descr *dsc,
+ off_t pos,
+ size_t len,
+ off_t *fail_pos,
+ off_t *fail_len)
+{
+ if (*fail_pos < 0)
+ goto record;
+
+ if (pos == *fail_pos + *fail_len) {
+ *fail_len += len;
+ return;
+ }
+
+ report_verity_error(ctx, dsc, *fail_pos, *fail_len);
+record:
+ *fail_pos = pos;
+ *fail_len = len;
+}
+
+/* Record a verity validation success and maybe log an old error. */
+static inline void
+record_verity_success(
+ struct scrub_ctx *ctx,
+ struct descr *dsc,
+ off_t *fail_pos,
+ off_t *fail_len)
+{
+ if (*fail_pos >= 0)
+ report_verity_error(ctx, dsc, *fail_pos, *fail_len);
+
+ *fail_pos = -1;
+ *fail_len = 0;
+}
+
+/* Scan a verity file's data looking for validation errors. */
+static void
+scan_verity_file(
+ struct workqueue *wq,
+ uint32_t index,
+ void *arg)
+{
+ struct stat sb;
+ struct verity_file_ctx *vf = arg;
+ struct scrub_ctx *ctx = vf->vc->ctx;
+ off_t pos;
+ off_t fail_pos = -1, fail_len = 0;
+ int fd;
+ int ret;
+ DEFINE_DESCR(dsc, ctx, render_ino_from_handle);
+
+ descr_set(&dsc, &vf->handle);
+
+ if (vf->vc->aborted) {
+ ret = ECANCELED;
+ goto out_vf;
+ }
+
+ fd = scrub_open_handle(&vf->handle);
+ if (fd < 0) {
+ /*
+ * Stale file handle means that the verity file is gone.
+ *
+ * Even if there's a replacement file, its contents have been
+ * freshly written and checked. Either way, we can skip
+ * scanning this file.
+ */
+ if (errno == ESTALE) {
+ ret = 0;
+ goto out_vf;
+ }
+
+ /*
+ * If the fsverity metadata is missing, inform the user and
+ * move on to the next file.
+ */
+ if (fsverity_meta_is_missing(errno)) {
+ str_error(ctx, descr_render(&dsc),
+ _("fsverity metadata missing."));
+ ret = 0;
+ goto out_vf;
+ }
+
+ ret = -errno;
+ str_errno(ctx, descr_render(&dsc));
+ goto out_vf;
+ }
+
+ ret = fstat(fd, &sb);
+ if (ret) {
+ str_errno(ctx, descr_render(&dsc));
+ goto out_fd;
+ }
+
+ /* Read a single byte from each block in the file to validate. */
+ for (pos = 0; pos < sb.st_size; pos += sb.st_blksize) {
+ char c;
+ ssize_t bytes_read;
+
+ bytes_read = pread(fd, &c, 1, pos);
+ if (!bytes_read)
+ break;
+ if (bytes_read > 0) {
+ record_verity_success(ctx, &dsc, &fail_pos, &fail_len);
+ progress_add(sb.st_blksize);
+ continue;
+ }
+
+ if (errno == EIO) {
+ size_t length = min(sb.st_size - pos, sb.st_blksize);
+
+ record_verity_error(ctx, &dsc, pos, length, &fail_pos,
+ &fail_len);
+ continue;
+ }
+
+ str_errno(ctx, descr_render(&dsc));
+ break;
+ }
+ report_verity_error(ctx, &dsc, fail_pos, fail_len);
+
+ ret = close(fd);
+ if (ret) {
+ str_errno(ctx, descr_render(&dsc));
+ goto out_vf;
+ }
+ fd = -1;
+
+out_fd:
+ if (fd >= 0)
+ close(fd);
+out_vf:
+ if (ret)
+ vf->vc->aborted = true;
+ free(vf);
+ return;
+}
+
+/* If this is a verity file, queue it for scanning. */
+static int
+schedule_verity_file(
+ struct scrub_ctx *ctx,
+ struct xfs_handle *handle,
+ struct xfs_bulkstat *bs,
+ void *arg)
+{
+ struct verity_ctx *vc = arg;
+ struct verity_file_ctx *vf;
+ int ret;
+
+ if (vc->aborted)
+ return ECANCELED;
+
+ if (!(bs->bs_xflags & FS_XFLAG_VERITY)) {
+ progress_add(bs->bs_size);
+ return 0;
+ }
+
+ vf = malloc(sizeof(struct verity_file_ctx));
+ if (!vf) {
+ str_errno(ctx, _("could not allocate fsverity scan context"));
+ vc->aborted = true;
+ return ENOMEM;
+ }
+
+ /* Queue the validation work. */
+ vf->handle = *handle; /* struct copy */
+ vf->vc = vc;
+
+ if (bs->bs_xflags & FS_XFLAG_REALTIME)
+ ret = -workqueue_add(&vc->wq_rtdev, scan_verity_file, 0, vf);
+ else
+ ret = -workqueue_add(&vc->wq_ddev, scan_verity_file, 0, vf);
+ if (ret) {
+ str_liberror(ctx, ret, _("could not schedule fsverity scan"));
+ vc->aborted = true;
+ return ECANCELED;
+ }
+
+ return 0;
+}
+
+static int
+scan_verity_files(
+ struct scrub_ctx *ctx)
+{
+ struct verity_ctx vc = {
+ .ctx = ctx,
+ };
+ unsigned int verifier_threads;
+ int ret;
+
+ /* Create thread pool for data dev fsverity processing. */
+ verifier_threads = disk_heads(ctx->datadev);
+ if (verifier_threads == 1)
+ verifier_threads = 0;
+ ret = -workqueue_create_bound(&vc.wq_ddev, ctx, verifier_threads, 500);
+ if (ret) {
+ str_liberror(ctx, ret, _("creating data dev fsverity workqueue"));
+ return ret;
+ }
+
+ /* Create thread pool for rtdev fsverity processing. */
+ if (ctx->rtdev) {
+ verifier_threads = disk_heads(ctx->rtdev);
+ if (verifier_threads == 1)
+ verifier_threads = 0;
+ ret = -workqueue_create_bound(&vc.wq_rtdev, ctx,
+ verifier_threads, 500);
+ if (ret) {
+ str_liberror(ctx, ret,
+ _("creating rt dev fsverity workqueue"));
+ goto out_ddev;
+ }
+ }
+
+ /* Find all the verity inodes. */
+ ret = scrub_scan_all_inodes(ctx, schedule_verity_file, 0, &vc);
+ if (ret)
+ goto out_rtdev;
+ if (vc.aborted) {
+ ret = ECANCELED;
+ goto out_rtdev;
+ }
+
+out_rtdev:
+ workqueue_terminate(&vc.wq_rtdev);
+ workqueue_destroy(&vc.wq_rtdev);
+out_ddev:
+ workqueue_terminate(&vc.wq_ddev);
+ workqueue_destroy(&vc.wq_ddev);
+ return ret;
+}
+
/*
* Read verify all the file data blocks in a filesystem. Since XFS doesn't
* do data checksums, we trust that the underlying storage will pass back
@@ -689,6 +973,12 @@ phase6_func(
struct media_verify_state vs = { NULL };
int ret, ret2, ret3;
+ if (ctx->mnt.fsgeom.flags & XFS_FSOP_GEOM_FLAGS_VERITY) {
+ ret = scan_verity_files(ctx);
+ if (ret)
+ return ret;
+ }
+
ret = -bitmap_alloc(&vs.d_bad);
if (ret) {
str_liberror(ctx, ret, _("creating datadev badblock bitmap"));
@@ -816,5 +1106,20 @@ phase6_estimate(
if (ctx->logdev)
*nr_threads += disk_heads(ctx->logdev);
*rshift = 20;
+
+ /*
+ * If fsverity is active, double the amount of progress items because
+ * we will want to validate individual files' data with fsverity.
+ * Bump the thread counts for the separate verity thread pools and the
+ * inode scanner.
+ */
+ if (ctx->mnt.fsgeom.flags & XFS_FSOP_GEOM_FLAGS_VERITY) {
+ *items *= 2;
+ *nr_threads += disk_heads(ctx->datadev);
+ *nr_threads += scrub_nproc_workqueue(ctx);
+ if (ctx->rtdev)
+ *nr_threads += disk_heads(ctx->rtdev);
+ }
+
return 0;
}
