"value":"In the Linux kernel, the following vulnerability has been resolved:\n\nfix bitmap corruption on close_range() with CLOSE_RANGE_UNSHARE\n\ncopy_fd_bitmaps(new, old, count) is expected to copy the first\ncount/BITS_PER_LONG bits from old->full_fds_bits[] and fill\nthe rest with zeroes. What it does is copying enough words\n(BITS_TO_LONGS(count/BITS_PER_LONG)), then memsets the rest.\nThat works fine, *if* all bits past the cutoff point are\nclear. Otherwise we are risking garbage from the last word\nwe'd copied.\n\nFor most of the callers that is true - expand_fdtable() has\ncount equal to old->max_fds, so there's no open descriptors\npast count, let alone fully occupied words in ->open_fds[],\nwhich is what bits in ->full_fds_bits[] correspond to.\n\nThe other caller (dup_fd()) passes sane_fdtable_size(old_fdt, max_fds),\nwhich is the smallest multiple of BITS_PER_LONG that covers all\nopened descriptors below max_fds. In the common case (copying on\nfork()) max_fds is ~0U, so all opened descriptors will be below\nit and we are fine, by the same reasons why the call in expand_fdtable()\nis safe.\n\nUnfortunately, there is a case where max_fds is less than that\nand where we might, indeed, end up with junk in ->full_fds_bits[] -\nclose_range(from, to, CLOSE_RANGE_UNSHARE) with\n\t* descriptor table being currently shared\n\t* 'to' being above the current capacity of descriptor table\n\t* 'from' being just under some chunk of opened descriptors.\nIn that case we end up with observably wrong behaviour - e.g. spawn\na child with CLONE_FILES, get all descriptors in range 0..127 open,\nthen close_range(64, ~0U, CLOSE_RANGE_UNSHARE) and watch dup(0) ending\nup with descriptor #128, despite #64 being observably not open.\n\nThe minimally invasive fix would be to deal with that in dup_fd().\nIf this proves to add measurable overhead, we can go that way, but\nlet's try to fix copy_fd_bitmaps() first.\n\n* new helper: bitmap_copy_and_expand(to, from, bits_to_copy, size).\n* make copy_fd_bitmaps() take the bitmap size in words, rather than\nbits; it's 'count' argument is always a multiple of BITS_PER_LONG,\nso we are not losing any information, and that way we can use the\nsame helper for all three bitmaps - compiler will see that count\nis a multiple of BITS_PER_LONG for the large ones, so it'll generate\nplain memcpy()+memset().\n\nReproducer added to tools/testing/selftests/core/close_range_test.c"
