On arm64 systems with 64K pages, the selftest task_local_data has the following
failures:
...
test_task_local_data_basic:PASS:tld_create_key 0 nsec
test_task_local_data_basic:FAIL:tld_create_key unexpected tld_create_key: actual 0 != expected -28
...
test_task_local_data_basic_thread:PASS:run task_main 0 nsec
test_task_local_data_basic_thread:FAIL:task_main retval unexpected error: 2 (errno 0)
test_task_local_data_basic_thread:FAIL:tld_get_data value0 unexpected tld_get_data value0: actual 0 != expected 6268
...
#447/1 task_local_data/task_local_data_basic:FAIL
...
#447/2 task_local_data/task_local_data_race:FAIL
#447 task_local_data:FAIL
When TLD_DYN_DATA_SIZE is 64K page size, for
struct tld_meta_u {
_Atomic __u8 cnt;
__u16 size;
struct tld_metadata metadata[];
};
field 'cnt' would overflow. For example, for 4K page, 'cnt' will
be 4096/64 = 64. But for 64K page, 'cnt' will be 65536/64 = 1024
and 'cnt' is not enough for 1024. To accommodate 64K page,
'_Atomic __u8 cnt' becomes '_Atomic __u16 cnt'. A few other places
are adjusted accordingly.
In test_task_local_data.c, the value for TLD_DYN_DATA_SIZE is changed
from 4096 to (getpagesize() - 8) since the maximum buffer size for
TLD_DYN_DATA_SIZE is (getpagesize() - 8).
Reviewed-by: Alan Maguire <alan.maguire@oracle.com>
Tested-by: Alan Maguire <alan.maguire@oracle.com>
Cc: Amery Hung <ameryhung@gmail.com>
Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
Acked-by: Amery Hung <ameryhung@gmail.com>
Link: https://lore.kernel.org/r/20260123055122.494352-1-yonghong.song@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Task local data defines an abstract storage type for storing task-
specific data (TLD). This patch provides user space and bpf
implementation as header-only libraries for accessing task local data.
Task local data is a bpf task local storage map with two UPTRs:
- tld_meta_u, shared by all tasks of a process, consists of the total
count and size of TLDs and an array of metadata of TLDs. A TLD
metadata contains the size and name. The name is used to identify a
specific TLD in bpf programs.
- u_tld_data points to a task-specific memory. It stores TLD data and
the starting offset of data in a page.
Task local design decouple user space and bpf programs. Since bpf
program does not know the size of TLDs in compile time, u_tld_data
is declared as a page to accommodate TLDs up to a page. As a result,
while user space will likely allocate memory smaller than a page for
actual TLDs, it needs to pin a page to kernel. It will pin the page
that contains enough memory if the allocated memory spans across the
page boundary.
The library also creates another task local storage map, tld_key_map,
to cache keys for bpf programs to speed up the access.
Below are the core task local data API:
User space BPF
Define TLD TLD_DEFINE_KEY(), tld_create_key() -
Init TLD object - tld_object_init()
Get TLD data tld_get_data() tld_get_data()
- TLD_DEFINE_KEY(), tld_create_key()
A TLD is first defined by the user space with TLD_DEFINE_KEY() or
tld_create_key(). TLD_DEFINE_KEY() defines a TLD statically and
allocates just enough memory during initialization. tld_create_key()
allows creating TLDs on the fly, but has a fix memory budget,
TLD_DYN_DATA_SIZE.
Internally, they all call __tld_create_key(), which iterates
tld_meta_u->metadata to check if a TLD can be added. The total TLD
size needs to fit into a page (limit of UPTR), and no two TLDs can
have the same name. If a TLD can be added, u_tld_meta->cnt is
increased using cmpxchg as there may be other concurrent
__tld_create_key(). After a successful cmpxchg, the last available
tld_meta_u->metadata now belongs to the calling thread. To prevent
other threads from reading incomplete metadata while it is being
updated, tld_meta_u->metadata->size is used to signal the completion.
Finally, the offset, derived from adding up prior TLD sizes is then
encapsulated as an opaque object key to prevent user misuse. The
offset is guaranteed to be 8-byte aligned to prevent load/store
tearing and allow atomic operations on it.
- tld_get_data()
User space programs can pass the key to tld_get_data() to get a
pointer to the associated TLD. The pointer will remain valid for the
lifetime of the thread.
tld_data_u is lazily allocated on the first call to tld_get_data().
Trying to read task local data from bpf will result in -ENODATA
during tld_object_init(). The task-specific memory need to be freed
manually by calling tld_free() on thread exit to prevent memory leak
or use TLD_FREE_DATA_ON_THREAD_EXIT.
- tld_object_init() (BPF)
BPF programs need to call tld_object_init() before calling
tld_get_data(). This is to avoid redundant map lookup in
tld_get_data() by storing pointers to the map values on stack.
The pointers are encapsulated as tld_object.
tld_key_map is also created on the first time tld_object_init()
is called to cache TLD keys successfully fetched by tld_get_data().
bpf_task_storage_get(.., F_CREATE) needs to be retried since it may
fail when another thread has already taken the percpu counter lock
for the task local storage.
- tld_get_data() (BPF)
BPF programs can also get a pointer to a TLD with tld_get_data().
It uses the cached key in tld_key_map to locate the data in
tld_data_u->data. If the cached key is not set yet (<= 0),
__tld_fetch_key() will be called to iterate tld_meta_u->metadata
and find the TLD by name. To prevent redundant string comparison
in the future when the search fail, the tld_meta_u->cnt is stored
in the non-positive range of the key. Next time, __tld_fetch_key()
will be called only if there are new TLDs and the search will start
from the newly added tld_meta_u->metadata using the old
tld_meta_u-cnt.
Signed-off-by: Amery Hung <ameryhung@gmail.com>
Reviewed-by: Emil Tsalapatis <emil@etsalapatis.com>
Link: https://lore.kernel.org/r/20250730185903.3574598-3-ameryhung@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
