wifi: rtw88: preparations for RTL8723B/RTL8723BS

From: Luka Gejak <[email protected]>

This is the first of two series adding support for the Realtek RTL8723B 802.11n chipset and its RTL8723BS SDIO variant to rtw88. It contains only the changes to the shared rtw88 core that the chip driver depends on. The chip itself, the build glue and the MAINTAINERS entry are a second series.

v1 had 19 patches, v2 had 11, v3 had 5, v4 had 7, v5 had 6, v6 had 6, v7 had 6, v8 had 6, and this has 6.

There is no rtw88-style firmware for this chip and no documentation for the vendor blob it does use, so while bringing it up I deliberately kept the driver as close to the vendor driver's behaviour as I could. With that many unknowns at once, matching the vendor exactly was the only way to tell which difference actually mattered when something did not work, rather than guessing. A number of the v1 patches came from that: they reproduced vendor behaviour that was useful to hold fixed during bring up, not behaviour the chip turns out to require.

Bitterblue Smith's review of v1 prompted me to go back and remove each of those on hardware to find which were actually load bearing; eight were not. Ping-Ke Shih's review of v2 did the same for a further six, and the changelog below gives the measurement for each. In four of those cases I had told the list, or claimed in a commit message, that the chip needed something it does not, on the strength of bring-up notes I had not re-measured. All four corrections are in the changelog.

What is left is there because removing it made something measurably worse, not because the vendor driver does it.

What is left falls into these groups:

- patches 1 and 2: the chip type and helper plus a one-line receive

    fix;
  - patch 3: a longer TX-report purge timeout, needed on slower SDIO
    hosts (see below);

- SDIO (patches 4, 5, 6): software free-page and output-queue accounting, RX aggregation and interrupt setup, and TX back-pressure with a retry on page starvation.

This series no longer touches coex.c, mac80211.c or fw.c at all.

Scope
=====

Everything is gated on rtw_is_8723bs(), which is false for every chip currently supported, so behaviour for existing devices is unchanged. Patch 2 extends an existing RTL8703B zero-length-packet check to this chip. Earlier revisions of this cover letter said that check was gated on the SDIO interface rather than the chip id. That was true of an old version and I did not update the text; the code you have adds rtw_is_8723bs() alongside the existing RTL8703B test, so it is gated on the chip id as well and the RTL8703B behaviour is untouched. Ping-Ke has suggested the check may not need a chip test at all, but I would rather measure that on RTL8703B hardware than assume it, and I do not have that part, so it is unchanged here.

Patch 5 changes how the SDIO interrupt status is acknowledged. In v1 that applied to all SDIO parts; it is scoped to this chip only, and the others keep writing the status word back unchanged.

Everything else is behind rtw_is_8723bs(). Since the point of this revision is to make that checkable rather than asserted, here is the whole of what is not, and what the other parts run as a result.

New code that is not chip gated, all of it structural:

sdio.h struct delayed_work work
rtw_sdio_init_tx() INIT_DELAYED_WORK(), mutex_init(), and
mutex_destroy() on its error path
rtw_sdio_tx_kick_off() mod_delayed_work(..., 0) in place of
queue_work()
rtw_sdio_deinit_tx() cancel_delayed_work_sync(), mutex_destroy()

Only this chip ever arms a delay, so on the other parts the work is still queued immediately, and only this chip ever takes the mutex, so there it is one init and one destroy and nothing else.

Two existing functions were also restructured, and every chip runs through them. rtw_sdio_process_tx_queue() returns a value instead of void, and rtw_sdio_tx_handler() uses it. The behaviour is meant to be unchanged for the other parts: on an empty queue the new "if (ret > 0) break" takes the place of the old skb_queue_empty() break, and on a failed write the frame is still requeued and the loop still retries, because rtw_sdio_8723bs_reschedule_tx() returns false for anything that is not this chip.

Two things worth stating because they can be checked against the base commit rather than taken on trust. rtw_sdio_check_free_txpg() differs by exactly one inserted branch ahead of the existing 8051 test, which becomes an else if, with everything else unchanged. And the generic path through rtw_sdio_write_port(), now rtw_sdio_write_port_generic(), is step for step what it was: the same address from skb->len, the same sdio_align_size(), the same free page check, the same transfer and the same failure message. The only difference visible to another part is that the unaligned SKB warning now prints rtw_sdio_write_to_port as its function name, since that is where the check lives.

rtw88 supports five other SDIO parts that share sdio.c: the RTL8723CS, RTL8723DS, RTL8821CS, RTL8822BS and RTL8822CS. I have none of them, so nothing here has been run on one. If anyone on the list does, a boot and some traffic with this series applied would be worth more than the argument above.

Chip specific code in the common path
=====================================

Ping-Ke asked whether this needs a set of ops rather than a chip test sprinkled through the core. Working through that question is what removed most of this series: once the association and coexistence patches turned out not to be needed, the chip tests they carried went with them.

What remains is one test in rx.c, one in tx.c and eleven in sdio.c. There are none in fw.c, mac80211.c or coex.c. The SDIO ones are in the bus driver rather than the chip driver, and rtw_sdio has no per-chip ops table today; Ping-Ke confirmed an inline chip id check there is fine.

Fixed in the second series
==========================

Two of Bitterblue's review points are addressed in the chip driver rather than here, because that is the right place for both:

- the hardware capability is filled in from a chip-specific read_efuse, instead of special casing rtw_dump_hw_feature() in the core;
- the receive FCS handling is a chip configuration fix. v1 stopped the core advertising RX_INCLUDES_FCS for this chip. The real cause was that the chip cleared BIT_APP_FCS by assigning WLAN_RCR_CFG over hal.rcr wholesale, where the 8723x siblings only write the register and leave hal.rcr alone. Setting that bit in the chip's WLAN_RCR_CFG makes the descriptor carry the FCS like every other rtw88 chip, so the core needs no special case and both hunks are gone from this series.

Neither change is visible here. I planned to send that series once this one is applied to rtw-next, but I can send it now alongside if you would rather review them together, or fold the two into one series if that is easier.

Testing
=======

Tested on RTL8723BS hardware together with the second series: repeated scan, authentication, association, WPA2-PSK/CCMP handshake, DHCP, bidirectional traffic, reconnects and module reloads.

The full driver, this preparation series plus the chip support that follows in a second series, is available for testing at:

https://github.com/MocLG/rtw/tree/8723bs-v9

Throughput is measured against an iperf3 server one hop behind the AP, with the wlan0 byte counters as ground truth because the test machine's wired interface shares the same subnet. On one AP at strong signal, 2.4 GHz HT40, TCP runs at roughly 35 Mbit/s down and 20 Mbit/s up, and UDP reaches 40 Mbit/s down with no loss. The exact figures move by several Mbit/s from one day to the next with the radio environment, so they are a range rather than a benchmark.

Uplink on this band drifts by several Mbit/s from day to day, enough to swamp the differences I was trying to measure, so where the changelog below compares two builds the comparison is a paired test: the two builds alternate within one session, rebuilt and reinstalled each time, rather than being measured on different days.

The locking change in patch 6 was retested that way for v8, against a build carrying the v7 barrier version of the same patch, alternating between the two within one session and reassociating each time. The two came out the same within the usual run to run spread, uplink and downlink alike, and the log was empty after every run.

Suspend and resume are not covered: this machine only offers s2idle and does not reliably come back from it.

A five minute soak: 1495/1495 pings at 0% loss, 15/15 reconnects, 5/5 link up/down cycles, 3/3 module reloads, 5/5 scans under traffic, 24 group rekeys, no deauthentication, no SDIO errors, no warnings in the log. Scanning saw the target AP in 18 of 18 attempts across 6 cold module reloads, with 6/6 associations.

The soak and scan figures were taken on a build that also carried the reserved page patch dropped in v3; the throughput figures above are from the tree posted here.

Every patch builds with W=1 with no warnings on its own, sparse is clean at the tip, and smatch reports nothing in any file this series touches. checkpatch --strict is clean across this series. The other rtw88 bus and chip modules continue to build.

Known limitation
================

After a long idle period the firmware intermittently fails to leave LPS, which rtw88 reports as "firmware failed to leave lps state". That check is the REG_TCR poll in __rtw_fw_leave_lps_check_reg(). On one occurrence here the warning was followed fourteen seconds later by a full reauthentication, so it is not only log noise.

Power save is enabled on my test machine and it reproduces there, roughly once per day of mostly idle uptime, and the occurrences cluster. It was first reported on ARM SDIO boards, but it is not specific to slow hosts.

Instrumenting the poll settles what a failure is, and it is not a budget that is slightly too small. A local debug patch logged REG_TCR on every poll of __rtw_fw_leave_lps_check_reg() and counted how many polls each wake needed. Over 110755 wakes:

poll0=158 poll1=110557 poll2=35 poll3=1 poll4=0 fails=4

pollN is the number of wakes that finished on the Nth read, so poll1 needed one 20 ms sleep and poll0 was already clear on entry; fails is the number still set after all five. Nothing ever finished on the last poll, yet four gave up completely, and in all four the patch kept reading for a further 500 ms with the bit still set. The distribution does not taper into the failures, it is bimodal: the wake is either quick or the firmware never completes it, about once in every 28000 wakes. Raising the poll count would not have caught any of the four.

In v7 I described a failure as the far tail of the ordinary distribution. That was wrong, and the measurement above is what corrects it.

Ping-Ke suggests the firmware may be unable to leave LPS while packets are still sitting in a hardware queue, which fits that shape better than anything I had. I am extending the instrumentation to dump the per-queue lengths and the free page and output queue counters at the failure.

It is not caused by anything in this series, and it needs the chip driver from the second series to be reachable at all, but it is open and I am chasing it.

Changes in v9:
- sdio: rtw_sdio_write_port() is a dispatcher now, with the two paths in rtw_sdio_write_port_8723bs() and rtw_sdio_write_port_generic(), as Ping-Ke asked. The chip specific one takes the mutex with guard() directly, so the extra rtw_sdio_8723bs_write_port() layer, which only existed to give the guard a scope while the generic path shared the outer function, is gone.
- sdio: the unaligned SKB warning sits in rtw_sdio_write_to_port() now, once, next to the transfer whose pointer it checks and after any padding. It was duplicated in both callers in v8 to keep __func__ reading rtw_sdio_write_port on the other parts; Ping-Ke had no preference there, and one copy in the right place seemed better than two in the wrong one.
- the Scope section above is corrected and extended. It described patch 2 as gated on the SDIO interface rather than the chip id, which stopped being true several revisions ago, and it did not mention that rtw_sdio_process_tx_queue() and rtw_sdio_tx_handler() are restructured for every chip.
- patches 1, 2, 3 and 5 are unchanged, and patch 6 is unchanged apart from Ping-Ke's Acked-by.

Changes in v8:
- sdio: patch 4 accounted RTW_TX_QUEUE_VO against the high page pool, but rtw_sdio_get_tx_addr() writes that queue to the normal transmit FIFO, and the FIFO written to is the pool the chip charges. Polling REG_SDIO_FREE_TXPG through a saturating VO flood, the high pool never lost a page while the public pool drained, so the check was reading a pool the traffic does not use and overstating the free space by the whole dedicated high allocation. It is accounted against the normal pool now, which is also what the other two rtw88 SDIO parts do. RTW_TX_QUEUE_MGMT stays on the high pool: it goes to the extra FIFO, which the register has no counter for and this chip gives no pages to, and there is nowhere better to put it.
- sdio: patch 4 claimed an output queue entry before the transfer and did not hand it back when the transfer failed, leaving the cached count low until the next resync from the chip.
- sdio: patch 4 called mutex_init() without a matching mutex_destroy() on teardown or on the init error path, unlike rtw_core_deinit().
- sdio: patch 6 no longer stops a queue and then undoes it. Ping-Ke pointed out that pci.c has no such back and forth, and he is right:
it holds irq_lock across the descriptor check and the stop in rtw_pci_tx_write(), and the wake in rtw_pci_tx_isr() runs with that same lock already held by its caller, which is why there is not one barrier in the whole file. Both SDIO sides now take the TX queue lock across the length check and the flag update, so a queue can only be stopped while it really is above the watermark and there is nothing left to undo. The smp_mb() pair, the READ_ONCE and WRITE_ONCE accessors and the re-check all go with it. It matches pci.c in shape too: the enqueue is its own critical section and the check and the stop are a second one.
- sdio: patch 4 moves the locked region into rtw_sdio_8723bs_write_port() so the mutex can be taken with guard(mutex). guard() expands to a declaration and so cannot be made conditional, and the lock is only wanted for this chip, so the generic path is a short branch that never reaches it. The transfer itself is shared as rtw_sdio_write_to_port(); the unaligned SKB warning stays in both callers so that the message other SDIO chips log is unchanged.
- sdio: lockdep_assert_held() added to the three helpers that must run under tx_credit_lock, and checked on a kernel built with CONFIG_PROVE_LOCKING over bidirectional traffic and three module
reloads: nothing fired, and lockdep reported nothing against the TX queue lock the back-pressure path now takes either.
- sdio: patch 4's padding uses a pad_size local, and its comment is cut to the one point that matters, that __skb_pad() must not free the skb. The comment above the lock is gone, since the one at the declaration of tx_credit_lock already covers it.
- the LPS note under Known limitation is corrected against a much longer instrumentation run; see there.
- patches 1, 2, 3 and 5 are unchanged.

Changes in v7:
- sdio: patch 6 read the mac80211 queue index from the skb after skb_queue_tail() had published it to the TX worker, which can dequeue, transmit and free it first. That is a use after free; the index is now read into a local before the enqueue.
- sdio: patch 4 passed the padded size to rtw_sdio_get_tx_addr(), where upstream passes skb->len. That value is encoded into the CMD53 address as the transfer length, so the encoded length changed for every other SDIO chip whenever sdio_align_size() padded. For the RTL8723BS both expressions are the same value, so the change bought this chip nothing. The padding itself also ran on the generic path, giving the other parts an allocation and an -ENOMEM path they did not have. skb->len is restored and the padding is gated on the chip.
- sdio: the free page check, the output queue wait and the accounting after the transfer are serialised. The TX worker and the H2C path both reach rtw_sdio_write_port(), and two writers could each pass the checks and claim the same pages and output queue entry, after which the chip discards one of the transfers silently. The vendor driver funnels all transmission through one thread and so never had to handle this.
- sdio: the padding called skb_put_zero() after checking only the tailroom, so a cloned skb with room to spare had its shared buffer written. It uses __skb_pad() now, which reallocates in that case and does not move skb->len, so the trim on the way out is gone too.
- sdio: the back-pressure stop could race the drain. The worker could empty the queue after the length check but before the stopped flag became visible, see the flag still clear in its wake check, and never wake the queue, leaving an access category stopped with nothing left to wake it. The stop path now re-checks after setting the flag and undoes the stop if the queue drained.
- patches 1, 2, 3 and 5 are unchanged.

Most of these were found by Sashiko's automated review of v6. Each was verified against the code before acting on it.

Changes in v6:
- Each patch now carries its own changelog after the --- delimiter, as you asked, so the notes sit next to the code they describe.
- sdio: dropped rtw_sdio_reschedule_tx_work(). You did not ask for it in v4 and it was a thin wrapper that hid queue_delayed_work() for no gain. I had read your comment as asking for one, which was my mistake.
- sdio: the reschedule conditions moved into rtw_sdio_8723bs_reschedule_tx(). rtw_sdio_tx_handler() no longer explains any chip specific condition in the common flow, which is what you were actually asking for. No functional change.
- sdio: dropped an unconditional break on a failed transfer that v4 and v5 carried. It changed the other SDIO parts without meaning to:
upstream requeues the frame and the loop retries, and breaking gave up after the first failure. The two errors this chip needs to retry are handled in the helper above, so the rest keep the existing behaviour and the other parts really are untouched again, as the Scope section claims.

Changes in v5:
- Dropped "fw: handle the RTL8723BS management TX reports", which v4 had restored. It cannot do what its commit message claimed. This firmware does not advertise FW_FEATURE_LPS_C2H, so rtw_fw_leave_lps_check() takes the REG_TCR polling path and nothing ever waits on the C2H completion the patch rerouted; and with the driver instrumented, 93 TX reports over three scans all arrive as C2H id 0x03, with no event ever arriving as a top level 0x12 or 0x32. Bitterblue Smith made exactly this point on v1 and was right. A tester on a Rockchip RK3288 board confirmed the timeout patch in this series is enough on its own. This is the fourth correction of something I had claimed on the list; restoring it in v4 was based on a tester reporting that a branch containing it cleared the warnings, which showed the branch helped, not that this patch in it did. With it gone the series no longer touches fw.c.
- Power save is now enabled on the test machine, so LPS is exercised there. The v2 entry below dropped a patch that gated LPS entry on smoothed throughput, partly on the grounds that LPS never engaged on this setup at all, which was never a sound reason to drop it. It is measured now, paired against this series in one session with power save on: 60 idle pings average 5.1 ms without the gating and 6.0 ms with it, and TCP is 30.0 down and 19.1 up against 32.3 and 18.9. It makes no difference, because the gate tests the smoothed throughput and interactive traffic rounds to zero there, so it never fires in the case it was meant to help. It stays dropped, now on a measurement rather than for want of one.
- sdio: fixed a transmit stall in the back-pressure patch, found while auditing a tester report of transmit stopping under bidirectional load with receive unaffected and nothing in the log. The TX work item was only re-armed when a transfer failed with a page shortage. Once the mac80211 queue can be stopped, that is not enough: a stopped queue is handed no further frames, so on any other failure nothing would kick the worker again and the affected access category would stay stopped for good, with the link still up. It now also re-arms on a failed skb expansion, which together with the page shortage covers both of the failures that produce no log message. The remaining errors are each logged where they happen, so unlike those two they are visible rather than an unexplained hang, and they keep the existing behaviour rather than being retried forever.
- rx: the zero length test now evaluates pkt_stat->pkt_len first, so the unlikely case short circuits before the chip test.
- sdio: the commit message now says what OQT is, as far as the vendor driver reveals it: the vendor calls the register the OQT free space and never expands the acronym, and it holds the number of further transfers the SDIO output queue will accept.
- sdio: rtw_sdio_8723bs_sync_free_txpg() now returns whether the chip reported anything and is the only caller of rtw_sdio_8723bs_store_free_txpg(), and rtw_sdio_8723bs_init_free_txpg() returns an error rather than nothing. Working through that suggestion found a real bug: the public pool size was computed as acq_pg_num minus the reserved queues with no check, so a chip that came up with no transmit page allocation at all would underflow a u16 and leave the driver believing it had about 65000 free pages. That calculation is now rtw_sdio_8723bs_pubq_num(), shared with the queue page allocation repair path, and it fails cleanly instead.
- sdio: the output queue wait is now bounded by a jiffies deadline rather than a loop count, so RTW_SDIO_OQT_TIMEOUT_MS really is a timeout in milliseconds. It was 1000 iterations of a 1 to 2 ms sleep before, so the name was only approximately true.
- sdio: rtw_sdio_8723bs_check_rqpn() returns an error instead of silently doing nothing when the transmit page pool cannot cover the reserved queues, and rtw_sdio_start() propagates it. Both early returns are now explained: a non-zero free page count means the allocation latched during power on and must be left alone.
- sdio: fixed the interrupt acknowledgment. It masked the status word with both irq_mask and RTW_SDIO_HISR_CLEAR_MASK, but for this chip irq_mask is REG_SDIO_HIMR_RX_REQUEST alone and that bit is not in the clear mask, so the two had no bits in common and the driver acknowledged nothing at all. It now masks with the defined bits only, which is what the comment described. Thanks to Ping-Ke for catching this. Retested on hardware.
- sdio: the back-pressure and wake conditions moved into rtw_sdio_8723bs_stop_tx_queue() and _wake_tx_queue(), and the two requeue sites into rtw_sdio_reschedule_tx_work(), instead of long conditions inline.
- sdio: rtw_sdio_process_tx_queue() now returns 0 on success and 1 for the empty queue case, rather than the other way round.
- sdio: queue_stopped[] renamed to tx_queue_stopped[].
- RTW_SDIO_TX_RETRY_DELAY is left as msecs_to_jiffies(1). Ping-Ke asked whether it becomes 0 when HZ is below 1000; it does not, since msecs_to_jiffies() rounds up for HZ < 1000 and returns one jiffy.

Changes in v4:
- Restored "tx: extend the TX report purge timeout to RTL8723BS", dropped in v2 because I could not reproduce the failure on my test machine, a fast x86 host, so the path never ran for me. Testers on slower ARM SDIO hosts, Peter Robinson on the RFC and Zhang Ning on a Rockchip RK3288 board, both hit "failed to get tx report from firmware". It extends the exact fix from commit c80788f7c5ae ("wifi:
rtw88: increase TX report timeout to fix race condition"), already applied to the RTL8723DU, to the RTL8723BS, since they run the same firmware and hit the same off-channel-scan race.

Changes in v3:
- Dropped "fw: fix the reserved page upload on RTL8723BS". Its commit message said the BIT_BCN_VALID handshake fails on 8723BS SDIO without it. That does not reproduce. The failure path is an explicit rtw_err("error beacon valid") returning -EBUSY, reached from BSS_CHANGED_ASSOC on every association; over three associations on each of five builds with the patch reverted it never fired, associations were 3/3, and there was no packet loss after 30 s idle in power save, which is the path the reserved page's null data and PS-poll frames serve. As a control, rtw_info messages appear in the same captures and rtw_err outranks them, so a failure would have been logged. AP mode and WoWLAN also download the reserved page and are not tested here, so if it turns out to matter there it can come back with evidence behind it.
- Kept "sdio: add TX back-pressure and retry on page starvation", now with a measurement rather than an assertion, since you asked whether the remaining chip tests are necessary. Counting rtw_sdio_tx_handler invocations under an identical 60 Mbit/s UDP flood, because the retry re-arms that same work item: 1798 and 3095 on two baseline runs against 69 with the patch reverted. Achievable transmit rate falls by about a third without it, 25.5 Mbit/s against 38.3 and 40.2. tx_dropped and tx_errors stay at zero throughout, so the failure is not dropping frames, it is failing to push them.
- Kept "sdio: set up RX aggregation and interrupts". The aggregation half makes no measurable difference to throughput here, and on that evidence alone I would have dropped it. The rest of the patch does not show up in a throughput test at all: the chip keeps raising the interrupt after resume if undefined status bits are written back when acknowledging.

I have not been able to validate the suspend and resume path: the test machine only offers s2idle and does not reliably resume from it, so I cannot exercise the interrupt acknowledgment change. It is kept because the failure it addresses was observed during bring up, not because it has been re-measured. If someone with hardware that suspends cleanly finds it unnecessary, it should go.
- Dropped "run the RTL8723BS association register sequence", 440 lines, which was the vendor join sequence, six sites in rtw_ops_bss_info_changed() and both sites in rtw_ops_set_key().

Ping-Ke asked for this to run from rtw_chip_prepare_tx() rather than a chip test in mgd_prepare_tx(), and I had done that with a new chip_ops::prepare_tx. Since the whole patch is gone, so is the callback; the core is left as it was rather than gaining an op with no user.

I had told the list that association fails without these registers. That was wrong, and it was bring-up-era data I repeated without re-measuring. With the whole patch reverted: 18/18 scans saw the AP, 6/6 associations, and the five minute soak above, whose 15 reconnects and 24 rekeys exercise the exact path the sequence ran on, since it ran from mgd_prepare_tx() on every authentication. Uplink measured paired against the full series, 19.0 against 19.2 Mbit/s over ten runs each.

The default key search handling in that patch was also simply wrong:
rtw_sec_enable_sec_engine() sets sec->default_key_search itself and programs all four USE_DK bits on that basis, so the enable path was a no-op and the disable path was clearing bits the core had deliberately set.
- Dropped "coex: add the RTL8723BS scan antenna workaround" and "coex:
reassert the antenna path when associating", and with them "fw: add the GNT_BT firmware command", whose only caller they were.

The commit message on the first claimed the site survey does not hear the AP reliably without it. That is also wrong: 18/18 scans without it. Both were gated on the chip and on coex bt_disabled, so that gate is their entire scope rather than a subset of it, and the measurements above cover it.
- Dropped "record beacons from the target BSSID before authenticating". Ping-Ke doubted the beacon wait affected the connection and he was right: with the wait removed, association succeeded 12/12 from a cold module reload and 12/12 from a warm reconnect, against 10/10 and 10/10 with it, at the same latency. The pre-auth deauth already sleeps 100 ms with the recording window open, so on a 100 ms beacon interval the wait was almost always already satisfied when it was reached. struct rtw_auth_sync and the RX-path hook go with it.
- Dropped the BT_MP version queries and the BT_INFO query at scan start, and the BT_INFO query plus two of three repeated PS-TDMA H2Cs in the pre-auth replay. Measured over cold module reloads with three scans each, 8 reloads with the queries and 10 without: 29/30 scans saw the target AP and 10/10 associations succeeded without them, against 24/24 and 8/8 with them.
- The cached SDIO page counters are clamped at zero. A lost update between the free page check and the accounting could previously drive the public counter negative, and the sum was assigned to an unsigned, which made the check pass unconditionally from then on and suppressed the resync that was supposed to recover it.
- REG_SDIO_FREE_TXPG now has field masks and is decoded with u32_get_bits(); the RX DMA burst count uses BIT_DMA_BURST_CNT; open-coded BIT(0)|BIT(1) on REG_SYS_FUNC_EN uses the existing names.
- The RTL8723BS-specific SDIO helpers are named _8723bs_ and the chip test moved to their callers.
- rtw_sdio_process_tx_queue() returns a value instead of taking a "processed" out-parameter, which makes the retry and back-pressure logic in the TX handler readable.
- mod_delayed_work() in rtw_sdio_tx_kick_off() carries a comment saying why it is not queue_delayed_work(): a page-starvation retry may already be armed with a delay, and the newly queued frame should not wait for it.
- An argument that had a single value at every call site is gone, along with its dead branch.
- Block comments use the general kernel style, per commit 82b8000c28b5 ("net: drop special comment style").

Changes in v2:
- Dropped "tx: extend the TX report purge timeout". Instrumenting the report path shows payload[6] & 0xfc equals the enqueued sequence number, 32 out of 32 times, with rtw88's existing decode, and no report is missed. The longer timeout was covering for the next patch, not for the hardware.
- Dropped "fw: handle the RTL8723BS management TX reports". 0x12 and 0x32 are the first payload byte of C2H_CCX_TX_RPT, not C2H IDs; every C2H event this chip sends carries a known id, so the existing handler already covers them.
- Dropped "fw: send rate adaptation and RSSI info in the vendor layout". The vendor byte layout is equivalent to what the existing macros produce; uplink and the negotiated rate are unchanged without it (16.4/16.5/17.0 against 15.4/16.7/16.6 Mbit/s, MCS7 both ways).
- Dropped "fw: send the media status report in the vendor layout". The role field the vendor sets makes no difference here, as with the other chips.
- Dropped "sdio: handle the RTL8723BS management TX path". Sequence numbers do work for management frames on this chip, so reporting completion at DMA completion is unnecessary.
- Dropped "calibrate and tune the PHY" entirely. All three parts were vendor-matching scaffolding and none survived measurement: the scan-time initial gain override is worse than rtw_phy_dig_set_max_coverage() (five scans found 41 BSSes without it against 34 with it); the IQ calibration works from the normal phy_calibration path; and the hardcoded per-rate TX AGC table overrode the efuse calibration by up to 22 index units and bypassed the regulatory limit the by-rate path applies, which is not something a driver should do.
- Dropped "match the RTL8723BS firmware connect and power save behaviour". Deferring the connect report changes nothing, and the LPS gating cannot be justified from measurement: with RTW_DBG_PS enabled, LPS never engages on this setup at all, so the gating never takes effect here.
- Dropped "advertise the correct receive capabilities". Both halves move into the chip driver, as described above.
- "fw: add the vendor firmware commands" now adds only GNT_BT. rtw88 already implements MACID_CFG and the WL channel info report, and the coexistence antenna select reserve became unused once the PHY patch went.
- The zero-length packet check is gated on SDIO rather than the chip

    id.

- The SDIO interrupt acknowledgment change is scoped to this chip.
- The association sequence no longer re-applies the BSS capability at association time; BSS_CHANGED_ERP_PREAMBLE and BSS_CHANGED_ERP_SLOT are handled later in the same callback.
- Register accesses that had raw addresses now use the existing REG_GNT_BT and REG_BT_COEX_ENH_INTR_CTRL, plus a named define for the BB antenna select register.

The implementation is based on the initial RTL8723B work by Michael Straube <[email protected]>:

https://github.com/mistraube/rtw88/tree/rtl8723bs

Based on the rtw-next branch from pkshih/rtw.

Luka Gejak (6):
wifi: rtw88: add the RTL8723B chip type and SDIO helper
wifi: rtw88: rx: mark zero length packets on RTL8723BS
wifi: rtw88: tx: extend the TX report purge timeout to RTL8723BS
wifi: rtw88: sdio: track free TX pages and OQT credits for RTL8723BS
wifi: rtw88: sdio: set up RX aggregation and interrupts for RTL8723BS
wifi: rtw88: sdio: add TX back-pressure and retry on page starvation

 drivers/net/wireless/realtek/rtw88/main.h |   7 +
 drivers/net/wireless/realtek/rtw88/rx.c   |   8 +-
 drivers/net/wireless/realtek/rtw88/sdio.c | 541 ++++++++++++++++++++--
 drivers/net/wireless/realtek/rtw88/sdio.h |  24 +-
 drivers/net/wireless/realtek/rtw88/tx.c   |   5 +-
 5 files changed, 552 insertions(+), 33 deletions(-)

2.53.0

From: Luka Gejak <[email protected]>

The RTL8723B is a 802.11n combo chip whose SDIO variant, RTL8723BS, runs a Realtek vendor firmware that differs from the firmware used by the other rtw88 8723 family devices. Supporting it needs a number of small adjustments spread across the shared core, all of which have to be restricted to this one chip and bus combination.

Add the chip type and a helper that tests for it, so the changes that follow can be gated without repeating the chip and bus comparison.

---

Notes:
Changes in v9: none.
Changes in v8: none.
Changes in v7: none.
Changes in v6: none.
Changes in v5: none.

 drivers/net/wireless/realtek/rtw88/main.h | 7 +++++++
 1 file changed, 7 insertions(+)

index c6e981ba7986..8f86f7c12de5 100644
@@ -194,6 +194,7 @@ enum rtw_chip_type {
 	RTW_CHIP_TYPE_8723D,
 	RTW_CHIP_TYPE_8821C,
 	RTW_CHIP_TYPE_8703B,
+	RTW_CHIP_TYPE_8723B,
 	RTW_CHIP_TYPE_8821A,
 	RTW_CHIP_TYPE_8812A,
 	RTW_CHIP_TYPE_8814A,

@@ -2194,6 +2195,12 @@ static inline bool rtw_chip_has_tx_stbc(struct rtw_dev *rtwdev)
 	return rtwdev->chip->tx_stbc;
 }
 
+static inline bool rtw_is_8723bs(struct rtw_dev *rtwdev)
+{
+	return rtwdev->chip->id == RTW_CHIP_TYPE_8723B &&
+	       rtwdev->hci.type == RTW_HCI_TYPE_SDIO;
+}
+
 static inline u8 rtw_acquire_macid(struct rtw_dev *rtwdev)
 {
 	unsigned long mac_id;

2.53.0

From: Luka Gejak <[email protected]>

Like the RTL8703B, the RTL8723BS reports receive descriptors with a zero packet length, which the vendor driver drops outright. rtw88 already flags these as having no PSDU for the RTL8703B, so extend the same handling rather than passing an empty frame up.

This is gated on the SDIO interface rather than the chip id, because it has only been observed there; the USB variant is not known to do it.

---

Notes:
Changes in v9: none.
Changes in v8: none.
Changes in v7: none.
Changes in v6: none.
Changes in v5: the zero length test now evaluates pkt_stat->pkt_len first, so the unlikely case short circuits before the chip test.

 drivers/net/wireless/realtek/rtw88/rx.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

index 01fd299abb7f..fde21c840ac0 100644
@@ -253,10 +253,12 @@ static void rtw_rx_fill_rx_status(struct rtw_dev *rtwdev,
 
 	rtw_rx_addr_match(rtwdev, pkt_stat, hdr);
 
-	/* Rtl8723cs driver checks for size < 14 or size > 8192 and
-	 * simply drops the packet.
+	/*
+	 * Rtl8723cs and rtl8723bs drivers check for size < 14 or size > 8192
+	 * and simply drop the packet.
 	 */
-	if (rtwdev->chip->id == RTW_CHIP_TYPE_8703B && pkt_stat->pkt_len == 0) {
+	if (pkt_stat->pkt_len == 0 &&
+	    (rtwdev->chip->id == RTW_CHIP_TYPE_8703B || rtw_is_8723bs(rtwdev))) {
 		rx_status->flag |= RX_FLAG_NO_PSDU;
 		rtw_dbg(rtwdev, RTW_DBG_RX, "zero length packet");
 	}

2.53.0

From: Luka Gejak <[email protected]>

Commit c80788f7c5ae ("wifi: rtw88: increase TX report timeout to fix race condition") raised the purge timeout to 2500 ms for the RTL8723DU, because the firmware can stay off channel during background scans for longer than the 500 ms default, which delays the TX reports and lets the purge timer drop the tracking skbs. The host stack then reads the missing status as loss and collapses TCP throughput.

The RTL8723BS runs the same vendor firmware over a slower SDIO host and hits the same race. Testers on ARM SDIO boards see "failed to get tx report from firmware" under load, with the same throughput collapse. Extend the 2500 ms timeout to the RTL8723BS.

---

Notes:
Changes in v9: none.
Changes in v8: none.
Changes in v7: none.
Changes in v6: none.
Changes in v5: none. Restored in v4 after being dropped in v2, because testers on slower ARM SDIO hosts hit the warning that commit c80788f7c5ae already fixes for the RTL8723DU.

 drivers/net/wireless/realtek/rtw88/tx.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

index 9d747a060b98..797c1e0402f2 100644
@@ -208,8 +208,9 @@ void rtw_tx_report_enqueue(struct rtw_dev *rtwdev, struct sk_buff *skb, u8 sn)
 	__skb_queue_tail(&tx_report->queue, skb);
 	spin_unlock_irqrestore(&tx_report->q_lock, flags);
 
-	if (rtwdev->chip->id == RTW_CHIP_TYPE_8723D &&
-	    rtwdev->hci.type == RTW_HCI_TYPE_USB)
+	if ((rtwdev->chip->id == RTW_CHIP_TYPE_8723D &&
+	     rtwdev->hci.type == RTW_HCI_TYPE_USB) ||
+	    rtw_is_8723bs(rtwdev))
 		timeout = msecs_to_jiffies(2500);
 
 	mod_timer(&tx_report->purge_timer, jiffies + timeout);

2.53.0

From: Luka Gejak <[email protected]>

The RTL8723BS reports free TX page counts that the generic 8051 path reads back from the chip on every transfer, which is both slow over SDIO and unreliable on this part: the register frequently reads back zero while pages are in fact available. It also gates transmission on a free count in the SDIO output queue, REG_SDIO_OQT_FREE_PG, which rtw88 does not track at all. The vendor driver calls this the OQT free space and never expands the acronym; the register holds the number of further transfers the SDIO output queue can accept, and the chip discards writes that arrive when it has run out.

Mirror the vendor driver and keep the per-queue and public page counts in software, seeded at start and resynchronised from the chip only when the cached counts say there is not enough room. Wait for a free output queue entry before writing, and account for the pages consumed after a successful transfer.

Transfers also have to be padded up to the SDIO block size for this chip rather than using the generic alignment, so size the write separately from the frame and zero the padding with __skb_pad(), which also reallocates a cloned skb instead of writing into a buffer a clone still shares.

The check, the output queue wait and the accounting are serialised by a mutex. The TX worker and the H2C path reach this function concurrently, and two writers that both pass the checks can otherwise claim the same pages and output queue entry, after which the chip silently discards whichever transfer arrives second. The vendor driver avoids the same race by funnelling all transmission through one thread.

Measured on RTL8723BS hardware against an iperf3 server one hop behind the AP, with the wlan0 byte counters as ground truth. On the generic path the association completes but no data passes at all: TCP and UDP both measure 0 bit/s in either direction. With this patch TCP is 25.3 Mbit/s up and 37.3 Mbit/s down, and UDP is 25.0 Mbit/s up at 0% loss.

---

Notes:
Changes in v9:
- rtw_sdio_write_port() is now only a dispatcher, with the two paths in rtw_sdio_write_port_8723bs() and rtw_sdio_write_port_generic(), as you asked. The chip specific one takes the mutex with guard() directly, so the extra rtw_sdio_8723bs_write_port() layer, which only existed to give the guard a scope while the generic path shared the outer function, is gone.
- the unaligned SKB warning is in rtw_sdio_write_to_port() now, once, next to the transfer whose pointer it checks and after any padding. __func__ therefore reads rtw_sdio_write_to_port on every SDIO part.
Changes in v8:
- RTW_TX_QUEUE_VO is accounted against the normal page pool rather than the high one. rtw_sdio_get_tx_addr() writes it to the normal transmit FIFO, and that is the pool the chip charges: over a saturating VO flood the high pool never lost a single page while the public pool drained. The vendor maps VO to the high queue, but it also writes VO to the high FIFO, which rtw88 does not.
- the output queue entry claimed before the transfer is handed back when the transfer fails, rather than staying claimed until the cache next resyncs from the chip.
- mutex_destroy() for tx_credit_lock on teardown and on the init error path, matching rtw_core_deinit().
- the comment on the cached page counters no longer explains the atomic_t by the several contexts that write them. Since v7 the transmit paths all update them under tx_credit_lock; what is left outside the lock is rtw_sdio_start() seeding them, and that is the reason they stay atomic_t.
- the free page check, the output queue wait, the transfer and the accounting moved into rtw_sdio_8723bs_write_port(), which takes the lock with guard(mutex) as you suggested. The generic path is a short branch in rtw_sdio_write_port() that never touches the lock, and the transfer itself is shared as rtw_sdio_write_to_port().
- lockdep_assert_held() added to rtw_sdio_8723bs_write_port(), rtw_sdio_8723bs_wait_tx_oqt() and rtw_sdio_8723bs_consume_txpg(), and checked on a kernel built with CONFIG_PROVE_LOCKING over bidirectional traffic and three module reloads: nothing fired.
- the padding uses a pad_size local and the comment is down to the one point that matters, that __skb_pad() must not free the skb.
- the comment above the lock is gone; the one at the declaration of tx_credit_lock covers it.
Changes in v7:
- the free page check, the output queue wait and the accounting after the transfer are serialised by a mutex. The TX worker and the H2C path run concurrently, and two writers that both passed the checks could claim the same pages and output queue entry; the OQT refill could also read the register while the other writer was between its claim and its transfer. The chip silently discards the overcommitted transfer, which for H2C means a lost firmware command.
- the CMD53 address is computed from skb->len again, as upstream does. Passing the aligned size changed the encoded transfer length for every other SDIO chip whenever sdio_align_size() padded; for the RTL8723BS the two encodings are the same value.
- the padding is applied with __skb_pad() and only on the RTL8723BS path. The open coded version wrote into a cloned skb's shared buffer when the tailroom happened to be large enough, and it also ran on the generic path, giving other chips a new allocation and failure mode. __skb_pad() does not move skb->len, so the trim on the way out is gone too. It must not free the skb on failure, since one caller requeues it and the other frees it.

Changes in v6: none.
Changes in v5:
- the commit message now says what OQT is, as far as the vendor driver reveals it.
- rtw_sdio_8723bs_sync_free_txpg() returns whether the chip reported anything and is the only caller of _store_free_txpg(), and _init_free_txpg() returns an error rather than nothing. That found a real bug: the public pool size was acq_pg_num minus the reserved queues with no check, so a chip coming up with no transmit page allocation would underflow a u16 and leave about 65000 free pages. It is now rtw_sdio_8723bs_pubq_num(), shared with the queue page allocation repair path, and it fails cleanly.
- the output queue wait is bounded by a jiffies deadline rather than a loop count, so RTW_SDIO_OQT_TIMEOUT_MS is really milliseconds.
- rtw_sdio_8723bs_check_rqpn() returns an error instead of silently doing nothing when the pool cannot cover the reserved queues, and rtw_sdio_start() propagates it. Both early returns are explained.

 drivers/net/wireless/realtek/rtw88/sdio.c | 360 +++++++++++++++++++++-
 drivers/net/wireless/realtek/rtw88/sdio.h |  12 +
 2 files changed, 356 insertions(+), 16 deletions(-)

index 5b40d74b16ee..2991acc21b03 100644
@@ -20,6 +20,7 @@
 #include "tx.h"
 
 #define RTW_SDIO_INDIRECT_RW_RETRIES			50
+#define RTW_SDIO_OQT_TIMEOUT_MS				1000
 
 static bool rtw_sdio_is_bus_addr(u32 addr)
 {

@@ -548,12 +549,164 @@ static int rtw_sdio_read_port(struct rtw_dev *rtwdev, u8 *buf, size_t count)
 	return ret;
 }
 
+/*
+ * The cached free page counters are a fast path hint only. The transmit
+ * paths read and update them under tx_credit_lock; rtw_sdio_start() seeds
+ * them outside it, so they are atomic_t. Whenever they claim there is not
+ * enough room they are resynchronised from the chip before the caller gives
+ * up, which also absorbs any drift.
+ */
+static void rtw_sdio_8723bs_store_free_txpg(struct rtw_dev *rtwdev,
+					    u32 free_txpg)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+
+	atomic_set(&rtwsdio->free_pg_high,
+		   u32_get_bits(free_txpg, BIT_FREE_TXPG_HIGH));
+	atomic_set(&rtwsdio->free_pg_normal,
+		   u32_get_bits(free_txpg, BIT_FREE_TXPG_NORMAL));
+	atomic_set(&rtwsdio->free_pg_low,
+		   u32_get_bits(free_txpg, BIT_FREE_TXPG_LOW));
+	atomic_set(&rtwsdio->free_pg_pub,
+		   u32_get_bits(free_txpg, BIT_FREE_TXPG_PUB));
+}
+
+/*
+ * Refresh the cached counters from the chip. Returns false when the chip
+ * reports no free pages at all, which means the counts cannot be trusted
+ * and the caller has to decide what to do instead.
+ */
+static bool rtw_sdio_8723bs_sync_free_txpg(struct rtw_dev *rtwdev)
+{
+	u32 free_txpg = rtw_read32(rtwdev, REG_SDIO_FREE_TXPG);
+
+	if (!free_txpg)
+		return false;
+
+	rtw_sdio_8723bs_store_free_txpg(rtwdev, free_txpg);
+
+	return true;
+}
+
+/*
+ * Size of the public page pool: whatever the transmit page allocation has
+ * left once the per queue pools are taken out. Fails if the allocation
+ * cannot cover the reserved queues, since the remainder would underflow and
+ * there would be no sensible pool to hand out.
+ */
+static int rtw_sdio_8723bs_pubq_num(struct rtw_dev *rtwdev, u16 *pubq_num)
+{
+	const struct rtw_page_table *pg_tbl = &rtwdev->chip->page_table[0];
+	u16 acq_pg_num = rtwdev->fifo.acq_pg_num;
+	u16 reserved_num;
+
+	reserved_num = pg_tbl->hq_num + pg_tbl->lq_num + pg_tbl->nq_num +
+		       pg_tbl->exq_num + pg_tbl->gapq_num;
+	if (acq_pg_num <= reserved_num) {
+		rtw_err(rtwdev,
+			"no transmit pages left for the public queue: %u of %u reserved\n",
+			reserved_num, acq_pg_num);
+		return -EINVAL;
+	}
+
+	*pubq_num = acq_pg_num - reserved_num;
+
+	return 0;
+}
+
+static int rtw_sdio_8723bs_init_free_txpg(struct rtw_dev *rtwdev)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+	const struct rtw_page_table *pg_tbl;
+	u16 pubq_num;
+	int ret;
+
+	/* Seed from the page table when the chip has nothing to report yet. */
+	if (!rtw_sdio_8723bs_sync_free_txpg(rtwdev)) {
+		ret = rtw_sdio_8723bs_pubq_num(rtwdev, &pubq_num);
+		if (ret)
+			return ret;
+
+		pg_tbl = &rtwdev->chip->page_table[0];
+		atomic_set(&rtwsdio->free_pg_high, pg_tbl->hq_num);
+		atomic_set(&rtwsdio->free_pg_normal, pg_tbl->nq_num);
+		atomic_set(&rtwsdio->free_pg_low, pg_tbl->lq_num);
+		atomic_set(&rtwsdio->free_pg_pub, pubq_num);
+	}
+
+	atomic_set(&rtwsdio->tx_oqt_free,
+		   rtw_read8(rtwdev, REG_SDIO_OQT_FREE_PG));
+
+	return 0;
+}
+
+/*
+ * Sum of the queue's dedicated counter and the public pool, clamped at zero:
+ * a lost update between the check below and rtw_sdio_8723bs_consume_txpg()
+ * can briefly drive a counter negative, and letting that wrap would hide the
+ * shortage instead of triggering a resync from the chip.
+ */
+static unsigned int rtw_sdio_8723bs_pages_free(struct rtw_dev *rtwdev,
+					       atomic_t *dedicated)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+	int free;
+
+	free = atomic_read(dedicated) + atomic_read(&rtwsdio->free_pg_pub);
+
+	return free > 0 ? free : 0;
+}
+
+/*
+ * The pool a queue draws from follows the transmit FIFO that
+ * rtw_sdio_get_tx_addr() writes into, since that is the one the chip charges.
+ * RTW_TX_QUEUE_MGMT is the exception: it goes to the extra FIFO, which
+ * REG_SDIO_FREE_TXPG has no counter for and this chip allocates no pages to,
+ * so it is accounted against the high pool and served from the public one.
+ */
+static atomic_t *rtw_sdio_8723bs_free_txpg(struct rtw_dev *rtwdev, u8 queue)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+
+	switch (queue) {
+	case RTW_TX_QUEUE_VI:
+	case RTW_TX_QUEUE_VO:
+		return &rtwsdio->free_pg_normal;
+	case RTW_TX_QUEUE_BE:
+	case RTW_TX_QUEUE_BK:
+		return &rtwsdio->free_pg_low;
+	case RTW_TX_QUEUE_BCN:
+	case RTW_TX_QUEUE_H2C:
+	case RTW_TX_QUEUE_HI0:
+	case RTW_TX_QUEUE_MGMT:
+		return &rtwsdio->free_pg_high;
+	default:
+		return NULL;
+	}
+}
+
 static int rtw_sdio_check_free_txpg(struct rtw_dev *rtwdev, u8 queue,
 				    size_t count)
 {
 	unsigned int pages_free, pages_needed;
 
-	if (rtw_chip_wcpu_8051(rtwdev)) {
+	if (rtw_is_8723bs(rtwdev)) {
+		atomic_t *dedicated;
+
+		dedicated = rtw_sdio_8723bs_free_txpg(rtwdev, queue);
+		if (!dedicated) {
+			rtw_warn(rtwdev, "Unknown mapping for queue %u\n", queue);
+			return -EINVAL;
+		}
+
+		pages_free = rtw_sdio_8723bs_pages_free(rtwdev, dedicated);
+		pages_needed = DIV_ROUND_UP(count, rtwdev->chip->page_size);
+		if (pages_needed <= pages_free)
+			return 0;
+
+		rtw_sdio_8723bs_sync_free_txpg(rtwdev);
+		pages_free = rtw_sdio_8723bs_pages_free(rtwdev, dedicated);
+	} else if (rtw_chip_wcpu_8051(rtwdev)) {
 		u32 free_txpg;
 
 		free_txpg = rtw_sdio_read32(rtwdev, REG_SDIO_FREE_TXPG);

@@ -632,24 +785,60 @@ static int rtw_sdio_check_free_txpg(struct rtw_dev *rtwdev, u8 queue,
 	return 0;
 }
 
-static int rtw_sdio_write_port(struct rtw_dev *rtwdev, struct sk_buff *skb,
-			       enum rtw_tx_queue_type queue)
+static int rtw_sdio_8723bs_wait_tx_oqt(struct rtw_dev *rtwdev)
 {
 	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
-	bool bus_claim;
-	size_t txsize;
-	u32 txaddr;
-	int ret;
+	unsigned long deadline;
+	u8 free;
 
-	txaddr = rtw_sdio_get_tx_addr(rtwdev, skb->len, queue);
-	if (!txaddr)
-		return -EINVAL;
+	lockdep_assert_held(&rtwsdio->tx_credit_lock);
 
-	txsize = sdio_align_size(rtwsdio->sdio_func, skb->len);
+	if (atomic_add_unless(&rtwsdio->tx_oqt_free, -1, 0))
+		return 0;
 
-	ret = rtw_sdio_check_free_txpg(rtwdev, queue, txsize);
-	if (ret)
-		return ret;
+	deadline = jiffies + msecs_to_jiffies(RTW_SDIO_OQT_TIMEOUT_MS);
+	do {
+		free = rtw_read8(rtwdev, REG_SDIO_OQT_FREE_PG);
+		if (free) {
+			atomic_set(&rtwsdio->tx_oqt_free, free - 1);
+			return 0;
+		}
+		usleep_range(1000, 2000);
+	} while (time_before(jiffies, deadline));
+
+	return -EBUSY;
+}
+
+static void rtw_sdio_8723bs_consume_txpg(struct rtw_dev *rtwdev, u8 queue,
+					 unsigned int pages)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+	atomic_t *dedicated;
+	unsigned int taken;
+	int free;
+
+	lockdep_assert_held(&rtwsdio->tx_credit_lock);
+
+	dedicated = rtw_sdio_8723bs_free_txpg(rtwdev, queue);
+	if (!dedicated)
+		return;
+
+	free = atomic_read(dedicated);
+	taken = min_t(unsigned int, pages, free > 0 ? free : 0);
+	atomic_sub(taken, dedicated);
+
+	pages -= taken;
+	if (pages && atomic_sub_return(pages, &rtwsdio->free_pg_pub) < 0)
+		atomic_set(&rtwsdio->free_pg_pub, 0);
+}
+
+static int rtw_sdio_write_to_port(struct rtw_dev *rtwdev, struct sk_buff *skb,
+				  enum rtw_tx_queue_type queue, u32 txaddr,
+				  size_t write_size)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+	bool bus_claim;
+	int ret;
 
 	if (!IS_ALIGNED((unsigned long)skb->data, RTW_SDIO_DATA_PTR_ALIGN))
 		rtw_warn(rtwdev, "Got unaligned SKB in %s() for queue %u\n",

@@ -660,7 +849,8 @@ static int rtw_sdio_write_port(struct rtw_dev *rtwdev, struct sk_buff *skb,
 	if (bus_claim)
 		sdio_claim_host(rtwsdio->sdio_func);
 
-	ret = sdio_memcpy_toio(rtwsdio->sdio_func, txaddr, skb->data, txsize);
+	ret = sdio_memcpy_toio(rtwsdio->sdio_func, txaddr, skb->data,
+			       write_size);
 
 	if (bus_claim)
 		sdio_release_host(rtwsdio->sdio_func);

@@ -668,11 +858,101 @@ static int rtw_sdio_write_port(struct rtw_dev *rtwdev, struct sk_buff *skb,
 	if (ret)
 		rtw_warn(rtwdev,
 			 "Failed to write %zu byte(s) to SDIO port 0x%08x",
-			 txsize, txaddr);
+			 write_size, txaddr);
 
 	return ret;
 }
 
+/*
+ * The free page check, the output queue wait and the accounting after the
+ * transfer have to be one unit, or two writers can both pass the checks and
+ * claim the same pages and output queue entry.
+ */
+static int rtw_sdio_write_port_8723bs(struct rtw_dev *rtwdev,
+				      struct sk_buff *skb,
+				      enum rtw_tx_queue_type queue)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+	unsigned int pages;
+	size_t write_size;
+	size_t txsize;
+	u32 txaddr;
+	int ret;
+
+	txaddr = rtw_sdio_get_tx_addr(rtwdev, skb->len, queue);
+	if (!txaddr)
+		return -EINVAL;
+
+	txsize = round_up(skb->len, 4);
+	write_size = txsize > RTW_SDIO_BLOCK_SIZE ?
+		     round_up(txsize, RTW_SDIO_BLOCK_SIZE) : txsize;
+
+	if (write_size > skb->len) {
+		size_t pad_size = write_size - skb->len;
+
+		/*
+		 * __skb_pad() must not free the skb on failure: both callers
+		 * still own it, one requeues it and the other frees it.
+		 */
+		ret = __skb_pad(skb, pad_size, false);
+		if (ret)
+			return ret;
+	}
+
+	guard(mutex)(&rtwsdio->tx_credit_lock);
+
+	ret = rtw_sdio_check_free_txpg(rtwdev, queue, txsize);
+	if (ret)
+		return ret;
+
+	ret = rtw_sdio_8723bs_wait_tx_oqt(rtwdev);
+	if (ret)
+		return ret;
+
+	ret = rtw_sdio_write_to_port(rtwdev, skb, queue, txaddr, write_size);
+	if (ret) {
+		/* nothing was queued, so hand the output queue entry back */
+		atomic_inc(&rtwsdio->tx_oqt_free);
+		return ret;
+	}
+
+	pages = DIV_ROUND_UP(txsize, rtwdev->chip->page_size);
+	rtw_sdio_8723bs_consume_txpg(rtwdev, queue, pages);
+
+	return 0;
+}
+
+static int rtw_sdio_write_port_generic(struct rtw_dev *rtwdev,
+				       struct sk_buff *skb,
+				       enum rtw_tx_queue_type queue)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+	size_t txsize;
+	u32 txaddr;
+	int ret;
+
+	txaddr = rtw_sdio_get_tx_addr(rtwdev, skb->len, queue);
+	if (!txaddr)
+		return -EINVAL;
+
+	txsize = sdio_align_size(rtwsdio->sdio_func, skb->len);
+
+	ret = rtw_sdio_check_free_txpg(rtwdev, queue, txsize);
+	if (ret)
+		return ret;
+
+	return rtw_sdio_write_to_port(rtwdev, skb, queue, txaddr, txsize);
+}
+
+static int rtw_sdio_write_port(struct rtw_dev *rtwdev, struct sk_buff *skb,
+			       enum rtw_tx_queue_type queue)
+{
+	if (rtw_is_8723bs(rtwdev))
+		return rtw_sdio_write_port_8723bs(rtwdev, skb, queue);
+
+	return rtw_sdio_write_port_generic(rtwdev, skb, queue);
+}
+
 static void rtw_sdio_init(struct rtw_dev *rtwdev)
 {
 	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;

@@ -749,8 +1029,51 @@ static int rtw_sdio_setup(struct rtw_dev *rtwdev)
 	return 0;
 }
 
+/*
+ * Reprogram the queue page allocation if the chip came up with none. This is
+ * a repair path, not part of the normal start sequence: a non-zero free page
+ * count means the allocation latched during power on and must be left alone,
+ * and without a transmit page pool there is nothing to divide up either.
+ */
+static int rtw_sdio_8723bs_check_rqpn(struct rtw_dev *rtwdev)
+{
+	const struct rtw_chip_info *chip = rtwdev->chip;
+	struct rtw_fifo_conf *fifo = &rtwdev->fifo;
+	const struct rtw_page_table *pg_tbl;
+	u32 free_txpg;
+	u16 pubq_num;
+	int ret;
+
+	free_txpg = rtw_read32(rtwdev, REG_SDIO_FREE_TXPG);
+	if (free_txpg || !fifo->acq_pg_num)
+		return 0;
+
+	ret = rtw_sdio_8723bs_pubq_num(rtwdev, &pubq_num);
+	if (ret)
+		return ret;
+
+	pg_tbl = &chip->page_table[0];
+	rtw_write32(rtwdev, REG_RQPN_NPQ,
+		    BIT_RQPN_NE(pg_tbl->nq_num, pg_tbl->exq_num));
+	rtw_write32(rtwdev, REG_RQPN,
+		    BIT_RQPN_HLP(pg_tbl->hq_num, pg_tbl->lq_num, pubq_num));
+
+	return 0;
+}
+
 static int rtw_sdio_start(struct rtw_dev *rtwdev)
 {
+	if (rtw_is_8723bs(rtwdev)) {
+		int ret = rtw_sdio_8723bs_check_rqpn(rtwdev);
+
+		if (ret)
+			return ret;
+
+		ret = rtw_sdio_8723bs_init_free_txpg(rtwdev);
+		if (ret)
+			return ret;
+	}
+
 	rtw_sdio_enable_rx_aggregation(rtwdev);
 	rtw_sdio_enable_interrupt(rtwdev);
 

@@ -1294,6 +1617,8 @@ static int rtw_sdio_init_tx(struct rtw_dev *rtwdev)
 		return -ENOMEM;
 	}
 
+	mutex_init(&rtwsdio->tx_credit_lock);
+
 	for (i = 0; i < RTK_MAX_TX_QUEUE_NUM; i++)
 		skb_queue_head_init(&rtwsdio->tx_queue[i]);
 	rtwsdio->tx_handler_data = kmalloc_obj(*rtwsdio->tx_handler_data);

@@ -1306,6 +1631,7 @@ static int rtw_sdio_init_tx(struct rtw_dev *rtwdev)
 	return 0;
 
 err_destroy_wq:
+	mutex_destroy(&rtwsdio->tx_credit_lock);
 	destroy_workqueue(rtwsdio->txwq);
 	return -ENOMEM;
 }

@@ -1320,6 +1646,8 @@ static void rtw_sdio_deinit_tx(struct rtw_dev *rtwdev)
 
 	for (i = 0; i < RTK_MAX_TX_QUEUE_NUM; i++)
 		ieee80211_purge_tx_queue(rtwdev->hw, &rtwsdio->tx_queue[i]);
+
+	mutex_destroy(&rtwsdio->tx_credit_lock);
 }
 
 int rtw_sdio_probe(struct sdio_func *sdio_func,

index 457e8b02380e..f43f1c6309b7 100644
@@ -86,6 +86,10 @@
 #define REG_SDIO_OQT_FREE_PG			(SDIO_LOCAL_OFFSET + 0x001E)
 /* Free Tx Buffer Page */
 #define REG_SDIO_FREE_TXPG			(SDIO_LOCAL_OFFSET + 0x0020)
+#define BIT_FREE_TXPG_HIGH			GENMASK(7, 0)
+#define BIT_FREE_TXPG_NORMAL			GENMASK(15, 8)
+#define BIT_FREE_TXPG_LOW			GENMASK(23, 16)
+#define BIT_FREE_TXPG_PUB			GENMASK(31, 24)
 /* HCI Current Power Mode 1 */
 #define REG_SDIO_HCPWM1				(SDIO_LOCAL_OFFSET + 0x0024)
 /* HCI Current Power Mode 2 */

@@ -159,6 +163,14 @@ struct rtw_sdio {
 	struct workqueue_struct *txwq;
 	struct rtw_sdio_work_data *tx_handler_data;
 	struct sk_buff_head tx_queue[RTK_MAX_TX_QUEUE_NUM];
+
+	atomic_t free_pg_high;
+	atomic_t free_pg_normal;
+	atomic_t free_pg_low;
+	atomic_t free_pg_pub;
+	atomic_t tx_oqt_free;
+	/* one writer at a time between the credit check and the accounting */
+	struct mutex tx_credit_lock;
 };
 
 extern const struct dev_pm_ops rtw_sdio_pm_ops;

2.53.0

If the __skb_padd() is necessary for original (generic) part, please add another patch to fix it. (see below comment)

[...]

In this patchset, you add many comments. Please review them across whole patchset to see if they actually need. Like this one, do you think the code can't explain itself?

    write_size = ...
    pad_size = write_size - skb->len;

    if (pad_size > 0) {
         ...
         __skb_pad( ... pad_size ...);
    }

On Sun Sep 6, 2026 at 4:22 AM CEST, Ping-Ke Shih wrote:
[...]

[...]

There is. rtw_sdio_write_port() rounds the transfer up with sdio_align_size() and passes that length to sdio_memcpy_toio() while the skb still holds only skb->len bytes, so between one and 511 bytes from past the end of the frame get transmitted. Whether that stays inside the skb's allocation depends on how much tailroom it happens to have.

It gets its own patch, no chip condition:

wifi: rtw88: sdio: zero the padding added to a TX transfer

Which also answers the question the other way round: the RTL8723BS does not need the padding for its own sake. I tested that rather than assume it, with a local patch that transfers unpadded exactly as the generic path does. Uplink is 20.9 to 21.1 Mbit/s unpadded against 19.5 to 19.8 padded, in an interleaved A/B in one session, and scans, reconnection and a UDP flood are clean either way. So the padding is a correctness fix that belongs in the shared path, and patch 4 no longer carries it.

There is an ordering point in that, which I would rather raise with you. The separate fix patches rtw_sdio_write_port(), which after this series is rtw_sdio_write_port_generic(), so it applies cleanly on top but only covers the generic path. The RTL8723BS would then be the one part still transmitting the padding unzeroed, which is not where I want to leave it.

The shared place is rtw_sdio_write_to_port(), which both paths call, but that only exists once this series is in. So I plan to send it once this series is applied, against that function, where a single patch covers every SDIO part including this one and nothing here needs respinning.

There is one more thing. With the padding gone from patch 4, nothing on that path returns -ENOMEM any more, so retrying it in patch 6 was dead code and I removed it. If the padding fix is applied then -ENOMEM becomes reachable again, for every chip, and that retry has to come back with it:
without it a stopped access category has nothing left to kick it.

Yes, and v10 does. I checked it is really the same rather than only looking the same: on this card sdio_align_size() matches the open coded rounding at the block boundaries, and over a full session it agreed on every frame, 80000 of them with no difference.

It is equivalent because the card reports multi block support and a 512 byte block size, which puts the byte mode limit at 512. Without those two properties sdio_align_size() falls into chunked byte transfers that are not block multiples, so the equivalence is a property of this card and host rather than something unconditional. It is still the right call:
it is the core helper and it is what the generic path already uses.

[...]

You are right that it can never be negative. The whole block is gone in v10 with the padding, so the question does not arise any more, but it would have been the better shape.

I went through the series and removed redundant comments: v10 adds 62 comment lines where v9 added 103.

What I kept is the things the code cannot show: which page pool a queue draws from and how that was measured on hardware, why the counters are atomic_t, why only the defined interrupt status bits are acknowledged, and the two places where an skb has already been handed on and must not be touched. Two more are there because they were asked for in earlier
review: the comment in patch 2 and the mod_delayed_work() one in patch 6. Say the word if any of those should go too.

Best regards,
Luka Gejak

Will/should you add this in v10 or v11 before your patch 4/6?

As your v10 didn't get ran by NIPA because of [1], you can send v10 with RESEND or just send v11 (including the additional one).

(To answer the ordering point question below.) I prefer to have the patch in advance, for two reasons. One is to backport easier, and another is to clear to know it is existing problem.

[1] https://lore.kernel.org/linux-wireless/[email protected]/T/#u

In v10, the beginning of rtw_sdio_write_port_8723bs() and rtw_sdio_write_port_generic become the similar. And I think ' txsize = round_up(skb->len, 4);' is not necessary to rtw_sdio_write_port_8723bs() because it is 512-byte aligned.

Therefore, you can do the same thing in rtw_sdio_write_port() and then call

    rtw_sdio_write_port_8723bs()

or
rtw_sdio_write_port_generic()
by condition.

Will this be needed by the additional patch

"wifi: rtw88: sdio: zero the padding added to a TX transfer"?

I will check this by v11, because I'd like you can have v11 with

1. "wifi: rtw88: sdio: zero the padding added to a TX transfer"
2. move the same beginning of rtw_sdio_write_port_8723bs() and rtw_sdio_write_port_generic() to rtw_sdio_write_port()

Is it okay to you?

Ping-Ke

[...]

[...]

I have decided to send v11, with the fix as patch 4, ahead of the accounting patch. Your two reasons are better than mine, so I have dropped the plan to send it afterwards.

It is not gated on anything, so it does change behaviour for the other five SDIO parts: the padding they send is zeroed now, and a transmit can fail with -ENOMEM if the skb has to be reallocated. That is called out in the Scope section rather than buried in the changelog.

One consequence worth flagging. The back-pressure patch stopped retrying -ENOMEM in v10, because with no padding in the series nothing could return it. It is reachable again now, so the retry is back. Without it a stopped access category has nothing left to kick it. That patch is therefore code identical to what you acked in v9; only its comments differ, so I have carried your Acked-by.

[...]

Done. rtw_sdio_write_port() works out the transfer address and the aligned size and passes both down, and the padding stays there too, so it runs once for both paths and outside the credit mutex.

I kept the separate length in the RTL8723BS path, though, and I think it is needed. It is not the transfer size, it feeds the page accounting:
rtw_sdio_check_free_txpg() and DIV_ROUND_UP(txsize, page_size), with a 128 byte page. The chip charges pages by the frame length rather than by the padded transfer, which is what the vendor driver does, and the two differ just above a block boundary:

frame pages by length pages by padded size
513 5 8
1025 9 12

So using the aligned size there would charge up to three pages per frame that the chip never took. There is a comment on the line now saying why.

[...]

Yes, with pad_size declared at the top, computed unconditionally, and tested with if (pad_size > 0).

[...]

Both are in v11, so yes.

Best regards,
Luka Gejak

From: Luka Gejak <[email protected]>

Enable the existing RX aggregation setup for this chip and select the larger DMA burst count it needs. The RTL8723BS does not raise CPWM1, so leave that source out of its interrupt mask, and set the SDIO TX control bit the vendor driver uses to have transfers always recognised.

The chip was also seen to keep raising the interrupt after resume when undefined status bits were written back on acknowledgment, so acknowledge only the bits this driver defines. That observation dates from bring up and I cannot re-measure it: the test machine only offers s2idle and does not reliably return from it, so the resume path is not exercised here. The change is scoped to this chip; the other SDIO parts keep writing the status word back unchanged.

---

Notes:
Changes in v9: none.
Changes in v8: none.
Changes in v7: none.
Changes in v6: none.
Changes in v5: fixed the interrupt acknowledgment. It masked the status word with both irq_mask and RTW_SDIO_HISR_CLEAR_MASK, but for this chip irq_mask is REG_SDIO_HIMR_RX_REQUEST alone and that bit is not in the clear mask, so the two had no bits in common and the driver acknowledged nothing at all. It now masks with the defined bits only, which is what the comment described. The commit message also no longer presents the resume behaviour as re-measured, because it is not.

 drivers/net/wireless/realtek/rtw88/sdio.c | 28 ++++++++++++++++++++++-
 drivers/net/wireless/realtek/rtw88/sdio.h |  9 ++++++++
 2 files changed, 36 insertions(+), 1 deletion(-)

index 2991acc21b03..4e6864b4f8ea 100644
@@ -957,7 +957,11 @@ static void rtw_sdio_init(struct rtw_dev *rtwdev)
 {
 	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
 
-	rtwsdio->irq_mask = REG_SDIO_HIMR_RX_REQUEST | REG_SDIO_HIMR_CPWM1;
+	if (rtw_is_8723bs(rtwdev))
+		rtwsdio->irq_mask = REG_SDIO_HIMR_RX_REQUEST;
+	else
+		rtwsdio->irq_mask = REG_SDIO_HIMR_RX_REQUEST |
+				    REG_SDIO_HIMR_CPWM1;
 }
 
 static void rtw_sdio_enable_rx_aggregation(struct rtw_dev *rtwdev)

@@ -965,6 +969,7 @@ static void rtw_sdio_enable_rx_aggregation(struct rtw_dev *rtwdev)
 	u8 size, timeout;
 
 	switch (rtwdev->chip->id) {
+	case RTW_CHIP_TYPE_8723B:
 	case RTW_CHIP_TYPE_8703B:
 	case RTW_CHIP_TYPE_8821A:
 	case RTW_CHIP_TYPE_8812A:

@@ -992,6 +997,8 @@ static void rtw_sdio_enable_rx_aggregation(struct rtw_dev *rtwdev)
 		    FIELD_PREP(BIT_DMA_AGG_TO_V1, timeout));
 
 	rtw_write8_set(rtwdev, REG_RXDMA_MODE, BIT_DMA_MODE);
+	if (rtw_is_8723bs(rtwdev))
+		rtw_write8_mask(rtwdev, REG_RXDMA_MODE, BIT_DMA_BURST_CNT, 0x3);
 }
 
 static void rtw_sdio_enable_interrupt(struct rtw_dev *rtwdev)

@@ -1063,6 +1070,8 @@ static int rtw_sdio_8723bs_check_rqpn(struct rtw_dev *rtwdev)
 
 static int rtw_sdio_start(struct rtw_dev *rtwdev)
 {
+	u32 clear;
+
 	if (rtw_is_8723bs(rtwdev)) {
 		int ret = rtw_sdio_8723bs_check_rqpn(rtwdev);
 

@@ -1075,6 +1084,13 @@ static int rtw_sdio_start(struct rtw_dev *rtwdev)
 	}
 
 	rtw_sdio_enable_rx_aggregation(rtwdev);
+
+	if (rtw_is_8723bs(rtwdev)) {
+		clear = rtw_read32(rtwdev, REG_SDIO_HISR) & RTW_SDIO_HISR_CLEAR_MASK;
+		if (clear)
+			rtw_write32(rtwdev, REG_SDIO_HISR, clear);
+	}
+
 	rtw_sdio_enable_interrupt(rtwdev);
 
 	return 0;

@@ -1154,6 +1170,8 @@ static void rtw_sdio_interface_cfg(struct rtw_dev *rtwdev)
 
 	val = rtw_read32(rtwdev, REG_SDIO_TX_CTRL);
 	val &= 0xfff8;
+	if (rtw_is_8723bs(rtwdev))
+		val |= BIT_SDIO_TX_CTRL_ALWAYS_RECOGNIZE;
 	rtw_write32(rtwdev, REG_SDIO_TX_CTRL, val);
 }
 

@@ -1418,6 +1436,14 @@ static void rtw_sdio_handle_interrupt(struct sdio_func *sdio_func)
 		rtw_sdio_rx_isr(rtwdev);
 	}
 
+	/*
+	 * RTL8723BS keeps raising the interrupt after resume if undefined
+	 * status bits are written back, so acknowledge only the bits this
+	 * driver defines. Other chips keep the existing behaviour.
+	 */
+	if (rtw_is_8723bs(rtwdev))
+		hisr &= RTW_SDIO_HISR_CLEAR_MASK;
+
 	rtw_write32(rtwdev, REG_SDIO_HISR, hisr);
 
 	rtwsdio->irq_thread = NULL;

index f43f1c6309b7..634c0b1339bb 100644
@@ -22,6 +22,7 @@
 
 /* SDIO Tx Control */
 #define REG_SDIO_TX_CTRL			(SDIO_LOCAL_OFFSET + 0x0000)
+#define BIT_SDIO_TX_CTRL_ALWAYS_RECOGNIZE	BIT(4)
 
 /*SDIO status timeout*/
 #define REG_SDIO_TIMEOUT			(SDIO_LOCAL_OFFSET + 0x0002)

@@ -77,6 +78,14 @@
 /* the following two are RTL8188 SDIO Specific */
 #define REG_SDIO_HISR_MCU_ERR			BIT(28)
 #define REG_SDIO_HISR_TSF_BIT32_TOGGLE		BIT(29)
+#define RTW_SDIO_HISR_CLEAR_MASK		\
+	(REG_SDIO_HISR_TXERR | REG_SDIO_HISR_RXERR | \
+	 REG_SDIO_HISR_TXFOVW | REG_SDIO_HISR_RXFOVW | \
+	 REG_SDIO_HISR_TXBCNOK | REG_SDIO_HISR_TXBCNERR | \
+	 REG_SDIO_HISR_C2HCMD | REG_SDIO_HISR_CPWM1 | \
+	 REG_SDIO_HISR_CPWM2 | REG_SDIO_HISR_HSISR_IND | \
+	 REG_SDIO_HISR_GTINT3_IND | REG_SDIO_HISR_GTINT4_IND | \
+	 REG_SDIO_HISR_PSTIMEOUT | REG_SDIO_HISR_OCPINT)
 
 /* HCI Current Power Mode */
 #define REG_SDIO_HCPWM				(SDIO_LOCAL_OFFSET + 0x0019)

2.53.0

From: Luka Gejak <[email protected]>

Two problems show up on RTL8723BS uplink. The per-AC software FIFO is unbounded, so mac80211 keeps handing frames down until latency collapses under load. And when a transfer cannot be completed the queue is simply abandoned for that pass, which stalls the AC until something else kicks the worker.

Stop the mac80211 queue once a data AC fills past a high watermark and wake it from the drain path when it falls back to a low one. Both sides take the TX queue lock across the length check and the flag update, the way rtw_pci_tx_write() and rtw_pci_tx_isr() use irq_lock. Without it the producer could stop a queue on a length the drain path had already emptied, having seen the flag still clear and so skipped the wake, and the AC would have stayed stopped with nothing left to wake it.

Convert the TX work item to a delayed work and re-arm it when a transfer fails for a reason that can clear on its own, so it is retried rather than the AC abandoned, and cancel the work on teardown.

Retrying matters once the queue can be stopped. A stopped queue is handed no further frames, so nothing else would kick the worker, and the AC would stay stopped for good with the link still up and receive unaffected. The two retried cases, a transmit page or output queue shortage and a failed skb expansion, are also the two that fail silently; the rest are logged where they happen, so they are visible rather than an unexplained hang, and they keep the existing behaviour rather than being retried indefinitely.

Measured on RTL8723BS hardware, uplink goes from 11.9 Mbit/s with 204 TCP retransmits to 20.1 Mbit/s with 2.

---

Notes:
Changes in v9: none, beyond recording your Acked-by.
Changes in v8:
- the stop and the wake take the TX queue lock with guard(spinlock_irqsave) across the length check and the flag update, following pci.c, which holds irq_lock on both sides and needs no barrier because of it. The smp_mb() pair, the READ_ONCE and WRITE_ONCE accessors and the re-check that undid a stop all went with it.
- q_map is read later, just before the call that uses it. It cannot move past rtw_sdio_indicate_tx_status(), which consumes the skb, so that is as far down as it goes.
Changes in v7:
- the mac80211 queue index is read before skb_queue_tail() publishes the skb to the TX worker, which may process and free it immediately; reading it afterwards was a use after free.
- the stop path re-checks the FIFO after setting the stopped flag and undoes the stop if the worker drained it meanwhile. In that window the drain side still saw the flag clear, so neither side would have woken the queue and the AC stayed stopped for good. The flag is accessed with READ_ONCE/WRITE_ONCE and the re-check is ordered against the drain path with a barrier pair.

Changes in v6:
- dropped rtw_sdio_reschedule_tx_work(). It was a thin wrapper around queue_delayed_work() and hid the kernel API for no gain.
- the reschedule conditions moved into rtw_sdio_8723bs_reschedule_tx(), so rtw_sdio_tx_handler() no longer explains any chip specific condition in the common flow.
- dropped the unconditional break on a failed transfer. v4 and v5 had it, and it quietly changed the other SDIO parts: upstream requeues the frame and the loop retries, and breaking gave up after the first failure. The two errors this chip needs to retry are handled in the helper above, so the rest can keep the existing behaviour and the other parts are untouched again.
Changes in v5:
- the back-pressure and wake conditions moved into rtw_sdio_8723bs_stop_tx_queue() and _wake_tx_queue().
- rtw_sdio_process_tx_queue() returns 0 on success and 1 for the empty queue case, rather than the other way round.
- queue_stopped[] renamed to tx_queue_stopped[].
- fixed a transmit stall: the work item was only re-armed on a page shortage, so on any other failure a stopped access category would stay stopped for good. It now also re-arms on a failed skb expansion, which with the page shortage covers both failures that produce no log message.
- RTW_SDIO_TX_RETRY_DELAY is left as msecs_to_jiffies(1): it does not become 0 for HZ < 1000, since msecs_to_jiffies() rounds up.

 drivers/net/wireless/realtek/rtw88/sdio.c | 153 ++++++++++++++++++++--
 drivers/net/wireless/realtek/rtw88/sdio.h |   3 +-
 2 files changed, 145 insertions(+), 11 deletions(-)

index 4e6864b4f8ea..b281e111502c 100644
@@ -22,6 +22,16 @@
 #define RTW_SDIO_INDIRECT_RW_RETRIES			50
 #define RTW_SDIO_OQT_TIMEOUT_MS				1000
 
+/*
+ * 8723BS SDIO TX FIFO back-pressure watermarks: stop the mac80211 queue once
+ * the per-AC software FIFO fills past the high watermark, and wake it from the
+ * TX drain path once it falls back to the low one. Bounds the queueing latency
+ * that otherwise causes uplink bufferbloat / congestion collapse.
+ */
+#define RTW_SDIO_TX_FIFO_HIWATER			16
+#define RTW_SDIO_TX_FIFO_LOWATER			8
+#define RTW_SDIO_TX_RETRY_DELAY			msecs_to_jiffies(1)
+
 static bool rtw_sdio_is_bus_addr(u32 addr)
 {
 	return !!(addr & RTW_SDIO_BUS_MSK);

@@ -1154,7 +1164,11 @@ static void rtw_sdio_tx_kick_off(struct rtw_dev *rtwdev)
 {
 	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
 
-	queue_work(rtwsdio->txwq, &rtwsdio->tx_handler_data->work);
+	/*
+	 * A retry may already be pending with a delay; re-arm it so a newly
+	 * queued frame is not held back by it.
+	 */
+	mod_delayed_work(rtwsdio->txwq, &rtwsdio->tx_handler_data->work, 0);
 }
 
 static void rtw_sdio_link_ps(struct rtw_dev *rtwdev, bool enter)

@@ -1263,12 +1277,65 @@ static int rtw_sdio_write_data_h2c(struct rtw_dev *rtwdev, u8 *buf, u32 size)
 	return rtw_sdio_write_data(rtwdev, &pkt_info, skb, RTW_TX_QUEUE_H2C);
 }
 
+/*
+ * Back-pressure on the data ACs (BK/BE/VI/VO): once the software FIFO fills
+ * past the high watermark, stop the corresponding mac80211 queue so it stops
+ * handing frames down, which bounds the queueing latency. The queue is woken
+ * again from the TX drain path once the FIFO falls back to the low watermark.
+ *
+ * Both sides hold the TX queue lock across the length check and the flag
+ * update, so a queue is only ever stopped while it really is above the
+ * watermark, and the drain path always sees the flag the producer set.
+ */
+static void rtw_sdio_8723bs_stop_tx_queue(struct rtw_dev *rtwdev,
+					  enum rtw_tx_queue_type queue,
+					  u16 q_map)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+
+	if (!rtw_is_8723bs(rtwdev) || queue >= RTW_TX_QUEUE_BCN)
+		return;
+
+	guard(spinlock_irqsave)(&rtwsdio->tx_queue[queue].lock);
+
+	if (rtwsdio->tx_queue_stopped[queue])
+		return;
+
+	if (skb_queue_len(&rtwsdio->tx_queue[queue]) < RTW_SDIO_TX_FIFO_HIWATER)
+		return;
+
+	rtwsdio->tx_queue_stopped[queue] = true;
+	ieee80211_stop_queue(rtwdev->hw, q_map);
+}
+
+static void rtw_sdio_8723bs_wake_tx_queue(struct rtw_dev *rtwdev,
+					  enum rtw_tx_queue_type queue,
+					  u16 q_map)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+
+	if (!rtw_is_8723bs(rtwdev) || queue >= RTW_TX_QUEUE_BCN)
+		return;
+
+	guard(spinlock_irqsave)(&rtwsdio->tx_queue[queue].lock);
+
+	if (!rtwsdio->tx_queue_stopped[queue])
+		return;
+
+	if (skb_queue_len(&rtwsdio->tx_queue[queue]) > RTW_SDIO_TX_FIFO_LOWATER)
+		return;
+
+	rtwsdio->tx_queue_stopped[queue] = false;
+	ieee80211_wake_queue(rtwdev->hw, q_map);
+}
+
 static int rtw_sdio_tx_write(struct rtw_dev *rtwdev,
 			     struct rtw_tx_pkt_info *pkt_info,
 			     struct sk_buff *skb)
 {
 	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
 	enum rtw_tx_queue_type queue = rtw_tx_queue_mapping(skb);
+	u16 q_map = skb_get_queue_mapping(skb);
 	struct rtw_sdio_tx_data *tx_data;
 
 	rtw_sdio_tx_skb_prepare(rtwdev, pkt_info, skb, queue);

@@ -1276,8 +1343,15 @@ static int rtw_sdio_tx_write(struct rtw_dev *rtwdev,
 	tx_data = rtw_sdio_get_tx_data(skb);
 	tx_data->sn = pkt_info->sn;
 
+	/*
+	 * skb_queue_tail() publishes the skb to the TX worker, which may
+	 * process and free it immediately, so nothing may touch the skb
+	 * past this point.
+	 */
 	skb_queue_tail(&rtwsdio->tx_queue[queue], skb);
 
+	rtw_sdio_8723bs_stop_tx_queue(rtwdev, queue, q_map);
+
 	return 0;
 }
 

@@ -1580,33 +1654,83 @@ static void rtw_sdio_indicate_tx_status(struct rtw_dev *rtwdev,
 	ieee80211_tx_status_irqsafe(hw, skb);
 }
 
-static void rtw_sdio_process_tx_queue(struct rtw_dev *rtwdev,
-				      enum rtw_tx_queue_type queue)
+/*
+ * Send one frame from @queue. Returns 0 when a frame was written, 1 when the
+ * queue was empty and a negative errno when the write failed, in which case
+ * the frame is put back at the head of the queue.
+ */
+static int rtw_sdio_process_tx_queue(struct rtw_dev *rtwdev,
+				     enum rtw_tx_queue_type queue)
 {
 	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
 	struct sk_buff *skb;
+	u16 q_map;
 	int ret;
 
 	skb = skb_dequeue(&rtwsdio->tx_queue[queue]);
 	if (!skb)
-		return;
+		return 1;
 
 	ret = rtw_sdio_write_port(rtwdev, skb, queue);
 	if (ret) {
 		skb_queue_head(&rtwsdio->tx_queue[queue], skb);
-		return;
+		return ret;
 	}
 
+	/* rtw_sdio_indicate_tx_status() consumes the skb, so read this first */
+	q_map = skb_get_queue_mapping(skb);
+
 	rtw_sdio_indicate_tx_status(rtwdev, skb);
+
+	rtw_sdio_8723bs_wake_tx_queue(rtwdev, queue, q_map);
+
+	return 0;
+}
+
+/*
+ * Decide whether the RTL8723BS wants the TX work to run again, and if so
+ * arrange it and tell the caller to stop draining. Two cases need it.
+ *
+ * A transmit page or output queue shortage and a failed skb expansion are
+ * transient and leave the frame queued, so come back for it shortly. That
+ * matters once the mac80211 queue can be stopped: a stopped queue is handed
+ * no further frames, so nothing else would kick this work item and the access
+ * category would stay stopped for good. The remaining errors are logged where
+ * they happen and are not retried.
+ *
+ * After a management frame, restart from the highest priority queue so the
+ * join sequence is not held up behind a data backlog.
+ */
+static bool rtw_sdio_8723bs_reschedule_tx(struct rtw_dev *rtwdev,
+					  struct rtw_sdio_work_data *work_data,
+					  enum rtw_tx_queue_type queue, int ret)
+{
+	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
+	unsigned long delay;
+
+	if (!rtw_is_8723bs(rtwdev))
+		return false;
+
+	if (ret == -EBUSY || ret == -ENOMEM)
+		delay = RTW_SDIO_TX_RETRY_DELAY;
+	else if (ret == 0 && queue == RTW_TX_QUEUE_MGMT)
+		delay = 0;
+	else
+		return false;
+
+	queue_delayed_work(rtwsdio->txwq, &work_data->work, delay);
+
+	return true;
 }
 
 static void rtw_sdio_tx_handler(struct work_struct *work)
 {
 	struct rtw_sdio_work_data *work_data =
-		container_of(work, struct rtw_sdio_work_data, work);
+		container_of(to_delayed_work(work), struct rtw_sdio_work_data,
+			     work);
 	struct rtw_sdio *rtwsdio;
 	struct rtw_dev *rtwdev;
-	int limit, queue;
+	int limit, queue, ret;
 
 	rtwdev = work_data->rtwdev;
 	rtwsdio = (struct rtw_sdio *)rtwdev->priv;

@@ -1616,7 +1740,13 @@ static void rtw_sdio_tx_handler(struct work_struct *work)
 
 	for (queue = RTK_MAX_TX_QUEUE_NUM - 1; queue >= 0; queue--) {
 		for (limit = 0; limit < 1000; limit++) {
-			rtw_sdio_process_tx_queue(rtwdev, queue);
+			ret = rtw_sdio_process_tx_queue(rtwdev, queue);
+			if (ret > 0)
+				break;
+
+			if (rtw_sdio_8723bs_reschedule_tx(rtwdev, work_data,
+							  queue, ret))
+				return;
 
 			if (skb_queue_empty(&rtwsdio->tx_queue[queue]))
 				break;

@@ -1645,14 +1775,16 @@ static int rtw_sdio_init_tx(struct rtw_dev *rtwdev)
 
 	mutex_init(&rtwsdio->tx_credit_lock);
 
-	for (i = 0; i < RTK_MAX_TX_QUEUE_NUM; i++)
+	for (i = 0; i < RTK_MAX_TX_QUEUE_NUM; i++) {
 		skb_queue_head_init(&rtwsdio->tx_queue[i]);
+		rtwsdio->tx_queue_stopped[i] = false;
+	}
 	rtwsdio->tx_handler_data = kmalloc_obj(*rtwsdio->tx_handler_data);
 	if (!rtwsdio->tx_handler_data)
 		goto err_destroy_wq;
 
 	rtwsdio->tx_handler_data->rtwdev = rtwdev;
-	INIT_WORK(&rtwsdio->tx_handler_data->work, rtw_sdio_tx_handler);
+	INIT_DELAYED_WORK(&rtwsdio->tx_handler_data->work, rtw_sdio_tx_handler);
 
 	return 0;
 

@@ -1667,6 +1799,7 @@ static void rtw_sdio_deinit_tx(struct rtw_dev *rtwdev)
 	struct rtw_sdio *rtwsdio = (struct rtw_sdio *)rtwdev->priv;
 	int i;
 
+	cancel_delayed_work_sync(&rtwsdio->tx_handler_data->work);
 	destroy_workqueue(rtwsdio->txwq);
 	kfree(rtwsdio->tx_handler_data);
 

index 634c0b1339bb..6e7e6009744b 100644
@@ -156,7 +156,7 @@ struct rtw_sdio_tx_data {
 };
 
 struct rtw_sdio_work_data {
-	struct work_struct work;
+	struct delayed_work work;
 	struct rtw_dev *rtwdev;
 };
 

@@ -172,6 +172,7 @@ struct rtw_sdio {
 	struct workqueue_struct *txwq;
 	struct rtw_sdio_work_data *tx_handler_data;
 	struct sk_buff_head tx_queue[RTK_MAX_TX_QUEUE_NUM];
+	bool tx_queue_stopped[RTK_MAX_TX_QUEUE_NUM];
 
 	atomic_t free_pg_high;
 	atomic_t free_pg_normal;

2.53.0